1 /*
   2  * Copyright (c) 1999, 2015, 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 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.hpp"
  27 #include "classfile/systemDictionary.hpp"
  28 #include "classfile/vmSymbols.hpp"
  29 #include "compiler/compileBroker.hpp"
  30 #include "compiler/compileLog.hpp"
  31 #include "oops/objArrayKlass.hpp"
  32 #include "opto/addnode.hpp"
  33 #include "opto/arraycopynode.hpp"
  34 #include "opto/callGenerator.hpp"
  35 #include "opto/castnode.hpp"
  36 #include "opto/cfgnode.hpp"
  37 #include "opto/convertnode.hpp"
  38 #include "opto/countbitsnode.hpp"
  39 #include "opto/intrinsicnode.hpp"
  40 #include "opto/idealKit.hpp"
  41 #include "opto/mathexactnode.hpp"
  42 #include "opto/movenode.hpp"
  43 #include "opto/mulnode.hpp"
  44 #include "opto/narrowptrnode.hpp"
  45 #include "opto/opaquenode.hpp"
  46 #include "opto/parse.hpp"
  47 #include "opto/runtime.hpp"
  48 #include "opto/subnode.hpp"
  49 #include "prims/nativeLookup.hpp"
  50 #include "runtime/sharedRuntime.hpp"
  51 #include "trace/traceMacros.hpp"
  52 
  53 class LibraryIntrinsic : public InlineCallGenerator {
  54   // Extend the set of intrinsics known to the runtime:
  55  public:
  56  private:
  57   bool             _is_virtual;
  58   bool             _does_virtual_dispatch;
  59   int8_t           _predicates_count;  // Intrinsic is predicated by several conditions
  60   int8_t           _last_predicate; // Last generated predicate
  61   vmIntrinsics::ID _intrinsic_id;
  62 
  63  public:
  64   LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
  65     : InlineCallGenerator(m),
  66       _is_virtual(is_virtual),
  67       _does_virtual_dispatch(does_virtual_dispatch),
  68       _predicates_count((int8_t)predicates_count),
  69       _last_predicate((int8_t)-1),
  70       _intrinsic_id(id)
  71   {
  72   }
  73   virtual bool is_intrinsic() const { return true; }
  74   virtual bool is_virtual()   const { return _is_virtual; }
  75   virtual bool is_predicated() const { return _predicates_count > 0; }
  76   virtual int  predicates_count() const { return _predicates_count; }
  77   virtual bool does_virtual_dispatch()   const { return _does_virtual_dispatch; }
  78   virtual JVMState* generate(JVMState* jvms);
  79   virtual Node* generate_predicate(JVMState* jvms, int predicate);
  80   vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
  81 };
  82 
  83 
  84 // Local helper class for LibraryIntrinsic:
  85 class LibraryCallKit : public GraphKit {
  86  private:
  87   LibraryIntrinsic* _intrinsic;     // the library intrinsic being called
  88   Node*             _result;        // the result node, if any
  89   int               _reexecute_sp;  // the stack pointer when bytecode needs to be reexecuted
  90 
  91   const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
  92 
  93  public:
  94   LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
  95     : GraphKit(jvms),
  96       _intrinsic(intrinsic),
  97       _result(NULL)
  98   {
  99     // Check if this is a root compile.  In that case we don't have a caller.
 100     if (!jvms->has_method()) {
 101       _reexecute_sp = sp();
 102     } else {
 103       // Find out how many arguments the interpreter needs when deoptimizing
 104       // and save the stack pointer value so it can used by uncommon_trap.
 105       // We find the argument count by looking at the declared signature.
 106       bool ignored_will_link;
 107       ciSignature* declared_signature = NULL;
 108       ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
 109       const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
 110       _reexecute_sp = sp() + nargs;  // "push" arguments back on stack
 111     }
 112   }
 113 
 114   virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
 115 
 116   ciMethod*         caller()    const    { return jvms()->method(); }
 117   int               bci()       const    { return jvms()->bci(); }
 118   LibraryIntrinsic* intrinsic() const    { return _intrinsic; }
 119   vmIntrinsics::ID  intrinsic_id() const { return _intrinsic->intrinsic_id(); }
 120   ciMethod*         callee()    const    { return _intrinsic->method(); }
 121 
 122   bool  try_to_inline(int predicate);
 123   Node* try_to_predicate(int predicate);
 124 
 125   void push_result() {
 126     // Push the result onto the stack.
 127     if (!stopped() && result() != NULL) {
 128       BasicType bt = result()->bottom_type()->basic_type();
 129       push_node(bt, result());
 130     }
 131   }
 132 
 133  private:
 134   void fatal_unexpected_iid(vmIntrinsics::ID iid) {
 135     fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
 136   }
 137 
 138   void  set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
 139   void  set_result(RegionNode* region, PhiNode* value);
 140   Node*     result() { return _result; }
 141 
 142   virtual int reexecute_sp() { return _reexecute_sp; }
 143 
 144   // Helper functions to inline natives
 145   Node* generate_guard(Node* test, RegionNode* region, float true_prob);
 146   Node* generate_slow_guard(Node* test, RegionNode* region);
 147   Node* generate_fair_guard(Node* test, RegionNode* region);
 148   Node* generate_negative_guard(Node* index, RegionNode* region,
 149                                 // resulting CastII of index:
 150                                 Node* *pos_index = NULL);
 151   Node* generate_limit_guard(Node* offset, Node* subseq_length,
 152                              Node* array_length,
 153                              RegionNode* region);
 154   Node* generate_current_thread(Node* &tls_output);
 155   Node* load_mirror_from_klass(Node* klass);
 156   Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
 157                                       RegionNode* region, int null_path,
 158                                       int offset);
 159   Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
 160                                RegionNode* region, int null_path) {
 161     int offset = java_lang_Class::klass_offset_in_bytes();
 162     return load_klass_from_mirror_common(mirror, never_see_null,
 163                                          region, null_path,
 164                                          offset);
 165   }
 166   Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
 167                                      RegionNode* region, int null_path) {
 168     int offset = java_lang_Class::array_klass_offset_in_bytes();
 169     return load_klass_from_mirror_common(mirror, never_see_null,
 170                                          region, null_path,
 171                                          offset);
 172   }
 173   Node* generate_access_flags_guard(Node* kls,
 174                                     int modifier_mask, int modifier_bits,
 175                                     RegionNode* region);
 176   Node* generate_interface_guard(Node* kls, RegionNode* region);
 177   Node* generate_array_guard(Node* kls, RegionNode* region) {
 178     return generate_array_guard_common(kls, region, false, false);
 179   }
 180   Node* generate_non_array_guard(Node* kls, RegionNode* region) {
 181     return generate_array_guard_common(kls, region, false, true);
 182   }
 183   Node* generate_objArray_guard(Node* kls, RegionNode* region) {
 184     return generate_array_guard_common(kls, region, true, false);
 185   }
 186   Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
 187     return generate_array_guard_common(kls, region, true, true);
 188   }
 189   Node* generate_array_guard_common(Node* kls, RegionNode* region,
 190                                     bool obj_array, bool not_array);
 191   Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
 192   CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
 193                                      bool is_virtual = false, bool is_static = false);
 194   CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
 195     return generate_method_call(method_id, false, true);
 196   }
 197   CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
 198     return generate_method_call(method_id, true, false);
 199   }
 200   Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
 201 
 202   Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
 203   Node* make_string_method_node(int opcode, Node* str1, Node* str2);
 204   bool inline_string_compareTo();
 205   bool inline_string_indexOf();
 206   Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
 207   bool inline_string_equals();
 208   Node* round_double_node(Node* n);
 209   bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
 210   bool inline_math_native(vmIntrinsics::ID id);
 211   bool inline_trig(vmIntrinsics::ID id);
 212   bool inline_math(vmIntrinsics::ID id);
 213   template <typename OverflowOp>
 214   bool inline_math_overflow(Node* arg1, Node* arg2);
 215   void inline_math_mathExact(Node* math, Node* test);
 216   bool inline_math_addExactI(bool is_increment);
 217   bool inline_math_addExactL(bool is_increment);
 218   bool inline_math_multiplyExactI();
 219   bool inline_math_multiplyExactL();
 220   bool inline_math_negateExactI();
 221   bool inline_math_negateExactL();
 222   bool inline_math_subtractExactI(bool is_decrement);
 223   bool inline_math_subtractExactL(bool is_decrement);
 224   bool inline_exp();
 225   bool inline_pow();
 226   Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
 227   bool inline_min_max(vmIntrinsics::ID id);
 228   Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
 229   // This returns Type::AnyPtr, RawPtr, or OopPtr.
 230   int classify_unsafe_addr(Node* &base, Node* &offset);
 231   Node* make_unsafe_address(Node* base, Node* offset);
 232   // Helper for inline_unsafe_access.
 233   // Generates the guards that check whether the result of
 234   // Unsafe.getObject should be recorded in an SATB log buffer.
 235   void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
 236   bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
 237   static bool klass_needs_init_guard(Node* kls);
 238   bool inline_unsafe_allocate();
 239   bool inline_unsafe_copyMemory();
 240   bool inline_native_currentThread();
 241 #ifdef TRACE_HAVE_INTRINSICS
 242   bool inline_native_classID();
 243   bool inline_native_threadID();
 244 #endif
 245   bool inline_native_time_funcs(address method, const char* funcName);
 246   bool inline_native_isInterrupted();
 247   bool inline_native_Class_query(vmIntrinsics::ID id);
 248   bool inline_native_subtype_check();
 249 
 250   bool inline_native_newArray();
 251   bool inline_native_getLength();
 252   bool inline_array_copyOf(bool is_copyOfRange);
 253   bool inline_array_equals();
 254   void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
 255   bool inline_native_clone(bool is_virtual);
 256   bool inline_native_Reflection_getCallerClass();
 257   // Helper function for inlining native object hash method
 258   bool inline_native_hashcode(bool is_virtual, bool is_static);
 259   bool inline_native_getClass();
 260 
 261   // Helper functions for inlining arraycopy
 262   bool inline_arraycopy();
 263   AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
 264                                                 RegionNode* slow_region);
 265   JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
 266   void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
 267 
 268   typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
 269   bool inline_unsafe_load_store(BasicType type,  LoadStoreKind kind);
 270   bool inline_unsafe_ordered_store(BasicType type);
 271   bool inline_unsafe_fence(vmIntrinsics::ID id);
 272   bool inline_fp_conversions(vmIntrinsics::ID id);
 273   bool inline_number_methods(vmIntrinsics::ID id);
 274   bool inline_reference_get();
 275   bool inline_Class_cast();
 276   bool inline_aescrypt_Block(vmIntrinsics::ID id);
 277   bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
 278   Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
 279   Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
 280   Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
 281   bool inline_ghash_processBlocks();
 282   bool inline_sha_implCompress(vmIntrinsics::ID id);
 283   bool inline_digestBase_implCompressMB(int predicate);
 284   bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
 285                                  bool long_state, address stubAddr, const char *stubName,
 286                                  Node* src_start, Node* ofs, Node* limit);
 287   Node* get_state_from_sha_object(Node *sha_object);
 288   Node* get_state_from_sha5_object(Node *sha_object);
 289   Node* inline_digestBase_implCompressMB_predicate(int predicate);
 290   bool inline_encodeISOArray();
 291   bool inline_updateCRC32();
 292   bool inline_updateBytesCRC32();
 293   bool inline_updateByteBufferCRC32();
 294   Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
 295   bool inline_updateBytesCRC32C();
 296   bool inline_updateDirectByteBufferCRC32C();
 297   bool inline_multiplyToLen();
 298   bool inline_squareToLen();
 299   bool inline_mulAdd();
 300   bool inline_montgomeryMultiply();
 301   bool inline_montgomerySquare();
 302 
 303   bool inline_profileBoolean();
 304   bool inline_isCompileConstant();
 305 };
 306 
 307 
 308 //---------------------------make_vm_intrinsic----------------------------
 309 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
 310   vmIntrinsics::ID id = m->intrinsic_id();
 311   assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
 312 
 313   ccstr disable_intr = NULL;
 314 
 315   if ((DisableIntrinsic[0] != '\0'
 316        && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
 317       (method_has_option_value("DisableIntrinsic", disable_intr)
 318        && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
 319     // disabled by a user request on the command line:
 320     // example: -XX:DisableIntrinsic=_hashCode,_getClass
 321     return NULL;
 322   }
 323 
 324   if (!m->is_loaded()) {
 325     // do not attempt to inline unloaded methods
 326     return NULL;
 327   }
 328 
 329   // Only a few intrinsics implement a virtual dispatch.
 330   // They are expensive calls which are also frequently overridden.
 331   if (is_virtual) {
 332     switch (id) {
 333     case vmIntrinsics::_hashCode:
 334     case vmIntrinsics::_clone:
 335       // OK, Object.hashCode and Object.clone intrinsics come in both flavors
 336       break;
 337     default:
 338       return NULL;
 339     }
 340   }
 341 
 342   // -XX:-InlineNatives disables nearly all intrinsics:
 343   if (!InlineNatives) {
 344     switch (id) {
 345     case vmIntrinsics::_indexOf:
 346     case vmIntrinsics::_compareTo:
 347     case vmIntrinsics::_equals:
 348     case vmIntrinsics::_equalsC:
 349     case vmIntrinsics::_getAndAddInt:
 350     case vmIntrinsics::_getAndAddLong:
 351     case vmIntrinsics::_getAndSetInt:
 352     case vmIntrinsics::_getAndSetLong:
 353     case vmIntrinsics::_getAndSetObject:
 354     case vmIntrinsics::_loadFence:
 355     case vmIntrinsics::_storeFence:
 356     case vmIntrinsics::_fullFence:
 357       break;  // InlineNatives does not control String.compareTo
 358     case vmIntrinsics::_Reference_get:
 359       break;  // InlineNatives does not control Reference.get
 360     default:
 361       return NULL;
 362     }
 363   }
 364 
 365   int predicates = 0;
 366   bool does_virtual_dispatch = false;
 367 
 368   switch (id) {
 369   case vmIntrinsics::_compareTo:
 370     if (!SpecialStringCompareTo)  return NULL;
 371     if (!Matcher::match_rule_supported(Op_StrComp))  return NULL;
 372     break;
 373   case vmIntrinsics::_indexOf:
 374     if (!SpecialStringIndexOf)  return NULL;
 375     break;
 376   case vmIntrinsics::_equals:
 377     if (!SpecialStringEquals)  return NULL;
 378     if (!Matcher::match_rule_supported(Op_StrEquals))  return NULL;
 379     break;
 380   case vmIntrinsics::_equalsC:
 381     if (!SpecialArraysEquals)  return NULL;
 382     if (!Matcher::match_rule_supported(Op_AryEq))  return NULL;
 383     break;
 384   case vmIntrinsics::_arraycopy:
 385     if (!InlineArrayCopy)  return NULL;
 386     break;
 387   case vmIntrinsics::_copyMemory:
 388     if (StubRoutines::unsafe_arraycopy() == NULL)  return NULL;
 389     if (!InlineArrayCopy)  return NULL;
 390     break;
 391   case vmIntrinsics::_hashCode:
 392     if (!InlineObjectHash)  return NULL;
 393     does_virtual_dispatch = true;
 394     break;
 395   case vmIntrinsics::_clone:
 396     does_virtual_dispatch = true;
 397   case vmIntrinsics::_copyOf:
 398   case vmIntrinsics::_copyOfRange:
 399     if (!InlineObjectCopy)  return NULL;
 400     // These also use the arraycopy intrinsic mechanism:
 401     if (!InlineArrayCopy)  return NULL;
 402     break;
 403   case vmIntrinsics::_encodeISOArray:
 404     if (!SpecialEncodeISOArray)  return NULL;
 405     if (!Matcher::match_rule_supported(Op_EncodeISOArray))  return NULL;
 406     break;
 407   case vmIntrinsics::_checkIndex:
 408     // We do not intrinsify this.  The optimizer does fine with it.
 409     return NULL;
 410 
 411   case vmIntrinsics::_getCallerClass:
 412     if (!InlineReflectionGetCallerClass)  return NULL;
 413     if (SystemDictionary::reflect_CallerSensitive_klass() == NULL)  return NULL;
 414     break;
 415 
 416   case vmIntrinsics::_bitCount_i:
 417     if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
 418     break;
 419 
 420   case vmIntrinsics::_bitCount_l:
 421     if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
 422     break;
 423 
 424   case vmIntrinsics::_numberOfLeadingZeros_i:
 425     if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
 426     break;
 427 
 428   case vmIntrinsics::_numberOfLeadingZeros_l:
 429     if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
 430     break;
 431 
 432   case vmIntrinsics::_numberOfTrailingZeros_i:
 433     if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
 434     break;
 435 
 436   case vmIntrinsics::_numberOfTrailingZeros_l:
 437     if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
 438     break;
 439 
 440   case vmIntrinsics::_reverseBytes_c:
 441     if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
 442     break;
 443   case vmIntrinsics::_reverseBytes_s:
 444     if (!Matcher::match_rule_supported(Op_ReverseBytesS))  return NULL;
 445     break;
 446   case vmIntrinsics::_reverseBytes_i:
 447     if (!Matcher::match_rule_supported(Op_ReverseBytesI))  return NULL;
 448     break;
 449   case vmIntrinsics::_reverseBytes_l:
 450     if (!Matcher::match_rule_supported(Op_ReverseBytesL))  return NULL;
 451     break;
 452 
 453   case vmIntrinsics::_Reference_get:
 454     // Use the intrinsic version of Reference.get() so that the value in
 455     // the referent field can be registered by the G1 pre-barrier code.
 456     // Also add memory barrier to prevent commoning reads from this field
 457     // across safepoint since GC can change it value.
 458     break;
 459 
 460   case vmIntrinsics::_compareAndSwapObject:
 461 #ifdef _LP64
 462     if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
 463 #endif
 464     break;
 465 
 466   case vmIntrinsics::_compareAndSwapLong:
 467     if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
 468     break;
 469 
 470   case vmIntrinsics::_getAndAddInt:
 471     if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
 472     break;
 473 
 474   case vmIntrinsics::_getAndAddLong:
 475     if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
 476     break;
 477 
 478   case vmIntrinsics::_getAndSetInt:
 479     if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
 480     break;
 481 
 482   case vmIntrinsics::_getAndSetLong:
 483     if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
 484     break;
 485 
 486   case vmIntrinsics::_getAndSetObject:
 487 #ifdef _LP64
 488     if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
 489     if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
 490     break;
 491 #else
 492     if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
 493     break;
 494 #endif
 495 
 496   case vmIntrinsics::_aescrypt_encryptBlock:
 497   case vmIntrinsics::_aescrypt_decryptBlock:
 498     if (!UseAESIntrinsics) return NULL;
 499     break;
 500 
 501   case vmIntrinsics::_multiplyToLen:
 502     if (!UseMultiplyToLenIntrinsic) return NULL;
 503     break;
 504 
 505   case vmIntrinsics::_squareToLen:
 506     if (!UseSquareToLenIntrinsic) return NULL;
 507     break;
 508 
 509   case vmIntrinsics::_mulAdd:
 510     if (!UseMulAddIntrinsic) return NULL;
 511     break;
 512 
 513   case vmIntrinsics::_montgomeryMultiply:
 514      if (!UseMontgomeryMultiplyIntrinsic) return NULL;
 515     break;
 516   case vmIntrinsics::_montgomerySquare:
 517      if (!UseMontgomerySquareIntrinsic) return NULL;
 518     break;
 519 
 520   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 521   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 522     if (!UseAESIntrinsics) return NULL;
 523     // these two require the predicated logic
 524     predicates = 1;
 525     break;
 526 
 527   case vmIntrinsics::_sha_implCompress:
 528     if (!UseSHA1Intrinsics) return NULL;
 529     break;
 530 
 531   case vmIntrinsics::_sha2_implCompress:
 532     if (!UseSHA256Intrinsics) return NULL;
 533     break;
 534 
 535   case vmIntrinsics::_sha5_implCompress:
 536     if (!UseSHA512Intrinsics) return NULL;
 537     break;
 538 
 539   case vmIntrinsics::_digestBase_implCompressMB:
 540     if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
 541     predicates = 3;
 542     break;
 543 
 544   case vmIntrinsics::_ghash_processBlocks:
 545     if (!UseGHASHIntrinsics) return NULL;
 546     break;
 547 
 548   case vmIntrinsics::_updateCRC32:
 549   case vmIntrinsics::_updateBytesCRC32:
 550   case vmIntrinsics::_updateByteBufferCRC32:
 551     if (!UseCRC32Intrinsics) return NULL;
 552     break;
 553 
 554   case vmIntrinsics::_updateBytesCRC32C:
 555   case vmIntrinsics::_updateDirectByteBufferCRC32C:
 556     if (!UseCRC32CIntrinsics) return NULL;
 557     break;
 558 
 559   case vmIntrinsics::_incrementExactI:
 560   case vmIntrinsics::_addExactI:
 561     if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
 562     break;
 563   case vmIntrinsics::_incrementExactL:
 564   case vmIntrinsics::_addExactL:
 565     if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
 566     break;
 567   case vmIntrinsics::_decrementExactI:
 568   case vmIntrinsics::_subtractExactI:
 569     if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
 570     break;
 571   case vmIntrinsics::_decrementExactL:
 572   case vmIntrinsics::_subtractExactL:
 573     if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
 574     break;
 575   case vmIntrinsics::_negateExactI:
 576     if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
 577     break;
 578   case vmIntrinsics::_negateExactL:
 579     if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
 580     break;
 581   case vmIntrinsics::_multiplyExactI:
 582     if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
 583     break;
 584   case vmIntrinsics::_multiplyExactL:
 585     if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
 586     break;
 587 
 588   case vmIntrinsics::_getShortUnaligned:
 589   case vmIntrinsics::_getCharUnaligned:
 590   case vmIntrinsics::_getIntUnaligned:
 591   case vmIntrinsics::_getLongUnaligned:
 592   case vmIntrinsics::_putShortUnaligned:
 593   case vmIntrinsics::_putCharUnaligned:
 594   case vmIntrinsics::_putIntUnaligned:
 595   case vmIntrinsics::_putLongUnaligned:
 596     if (!UseUnalignedAccesses) return NULL;
 597     break;
 598 
 599  default:
 600     assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
 601     assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
 602     break;
 603   }
 604 
 605   // -XX:-InlineClassNatives disables natives from the Class class.
 606   // The flag applies to all reflective calls, notably Array.newArray
 607   // (visible to Java programmers as Array.newInstance).
 608   if (m->holder()->name() == ciSymbol::java_lang_Class() ||
 609       m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
 610     if (!InlineClassNatives)  return NULL;
 611   }
 612 
 613   // -XX:-InlineThreadNatives disables natives from the Thread class.
 614   if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
 615     if (!InlineThreadNatives)  return NULL;
 616   }
 617 
 618   // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
 619   if (m->holder()->name() == ciSymbol::java_lang_Math() ||
 620       m->holder()->name() == ciSymbol::java_lang_Float() ||
 621       m->holder()->name() == ciSymbol::java_lang_Double()) {
 622     if (!InlineMathNatives)  return NULL;
 623   }
 624 
 625   // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
 626   if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
 627     if (!InlineUnsafeOps)  return NULL;
 628   }
 629 
 630   return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
 631 }
 632 
 633 //----------------------register_library_intrinsics-----------------------
 634 // Initialize this file's data structures, for each Compile instance.
 635 void Compile::register_library_intrinsics() {
 636   // Nothing to do here.
 637 }
 638 
 639 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
 640   LibraryCallKit kit(jvms, this);
 641   Compile* C = kit.C;
 642   int nodes = C->unique();
 643 #ifndef PRODUCT
 644   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
 645     char buf[1000];
 646     const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
 647     tty->print_cr("Intrinsic %s", str);
 648   }
 649 #endif
 650   ciMethod* callee = kit.callee();
 651   const int bci    = kit.bci();
 652 
 653   // Try to inline the intrinsic.
 654   if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
 655       kit.try_to_inline(_last_predicate)) {
 656     if (C->print_intrinsics() || C->print_inlining()) {
 657       C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
 658     }
 659     C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
 660     if (C->log()) {
 661       C->log()->elem("intrinsic id='%s'%s nodes='%d'",
 662                      vmIntrinsics::name_at(intrinsic_id()),
 663                      (is_virtual() ? " virtual='1'" : ""),
 664                      C->unique() - nodes);
 665     }
 666     // Push the result from the inlined method onto the stack.
 667     kit.push_result();
 668     C->print_inlining_update(this);
 669     return kit.transfer_exceptions_into_jvms();
 670   }
 671 
 672   // The intrinsic bailed out
 673   if (C->print_intrinsics() || C->print_inlining()) {
 674     if (jvms->has_method()) {
 675       // Not a root compile.
 676       const char* msg;
 677       if (callee->intrinsic_candidate()) {
 678         msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
 679       } else {
 680         msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
 681                            : "failed to inline (intrinsic), method not annotated";
 682       }
 683       C->print_inlining(callee, jvms->depth() - 1, bci, msg);
 684     } else {
 685       // Root compile
 686       tty->print("Did not generate intrinsic %s%s at bci:%d in",
 687                vmIntrinsics::name_at(intrinsic_id()),
 688                (is_virtual() ? " (virtual)" : ""), bci);
 689     }
 690   }
 691   C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
 692   C->print_inlining_update(this);
 693   return NULL;
 694 }
 695 
 696 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
 697   LibraryCallKit kit(jvms, this);
 698   Compile* C = kit.C;
 699   int nodes = C->unique();
 700   _last_predicate = predicate;
 701 #ifndef PRODUCT
 702   assert(is_predicated() && predicate < predicates_count(), "sanity");
 703   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
 704     char buf[1000];
 705     const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
 706     tty->print_cr("Predicate for intrinsic %s", str);
 707   }
 708 #endif
 709   ciMethod* callee = kit.callee();
 710   const int bci    = kit.bci();
 711 
 712   Node* slow_ctl = kit.try_to_predicate(predicate);
 713   if (!kit.failing()) {
 714     if (C->print_intrinsics() || C->print_inlining()) {
 715       C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
 716     }
 717     C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
 718     if (C->log()) {
 719       C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
 720                      vmIntrinsics::name_at(intrinsic_id()),
 721                      (is_virtual() ? " virtual='1'" : ""),
 722                      C->unique() - nodes);
 723     }
 724     return slow_ctl; // Could be NULL if the check folds.
 725   }
 726 
 727   // The intrinsic bailed out
 728   if (C->print_intrinsics() || C->print_inlining()) {
 729     if (jvms->has_method()) {
 730       // Not a root compile.
 731       const char* msg = "failed to generate predicate for intrinsic";
 732       C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
 733     } else {
 734       // Root compile
 735       C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
 736                                         vmIntrinsics::name_at(intrinsic_id()),
 737                                         (is_virtual() ? " (virtual)" : ""), bci);
 738     }
 739   }
 740   C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
 741   return NULL;
 742 }
 743 
 744 bool LibraryCallKit::try_to_inline(int predicate) {
 745   // Handle symbolic names for otherwise undistinguished boolean switches:
 746   const bool is_store       = true;
 747   const bool is_native_ptr  = true;
 748   const bool is_static      = true;
 749   const bool is_volatile    = true;
 750 
 751   if (!jvms()->has_method()) {
 752     // Root JVMState has a null method.
 753     assert(map()->memory()->Opcode() == Op_Parm, "");
 754     // Insert the memory aliasing node
 755     set_all_memory(reset_memory());
 756   }
 757   assert(merged_memory(), "");
 758 
 759 
 760   switch (intrinsic_id()) {
 761   case vmIntrinsics::_hashCode:                 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
 762   case vmIntrinsics::_identityHashCode:         return inline_native_hashcode(/*!virtual*/ false,         is_static);
 763   case vmIntrinsics::_getClass:                 return inline_native_getClass();
 764 
 765   case vmIntrinsics::_dsin:
 766   case vmIntrinsics::_dcos:
 767   case vmIntrinsics::_dtan:
 768   case vmIntrinsics::_dabs:
 769   case vmIntrinsics::_datan2:
 770   case vmIntrinsics::_dsqrt:
 771   case vmIntrinsics::_dexp:
 772   case vmIntrinsics::_dlog:
 773   case vmIntrinsics::_dlog10:
 774   case vmIntrinsics::_dpow:                     return inline_math_native(intrinsic_id());
 775 
 776   case vmIntrinsics::_min:
 777   case vmIntrinsics::_max:                      return inline_min_max(intrinsic_id());
 778 
 779   case vmIntrinsics::_addExactI:                return inline_math_addExactI(false /* add */);
 780   case vmIntrinsics::_addExactL:                return inline_math_addExactL(false /* add */);
 781   case vmIntrinsics::_decrementExactI:          return inline_math_subtractExactI(true /* decrement */);
 782   case vmIntrinsics::_decrementExactL:          return inline_math_subtractExactL(true /* decrement */);
 783   case vmIntrinsics::_incrementExactI:          return inline_math_addExactI(true /* increment */);
 784   case vmIntrinsics::_incrementExactL:          return inline_math_addExactL(true /* increment */);
 785   case vmIntrinsics::_multiplyExactI:           return inline_math_multiplyExactI();
 786   case vmIntrinsics::_multiplyExactL:           return inline_math_multiplyExactL();
 787   case vmIntrinsics::_negateExactI:             return inline_math_negateExactI();
 788   case vmIntrinsics::_negateExactL:             return inline_math_negateExactL();
 789   case vmIntrinsics::_subtractExactI:           return inline_math_subtractExactI(false /* subtract */);
 790   case vmIntrinsics::_subtractExactL:           return inline_math_subtractExactL(false /* subtract */);
 791 
 792   case vmIntrinsics::_arraycopy:                return inline_arraycopy();
 793 
 794   case vmIntrinsics::_compareTo:                return inline_string_compareTo();
 795   case vmIntrinsics::_indexOf:                  return inline_string_indexOf();
 796   case vmIntrinsics::_equals:                   return inline_string_equals();
 797 
 798   case vmIntrinsics::_getObject:                return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT,  !is_volatile);
 799   case vmIntrinsics::_getBoolean:               return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
 800   case vmIntrinsics::_getByte:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE,    !is_volatile);
 801   case vmIntrinsics::_getShort:                 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,   !is_volatile);
 802   case vmIntrinsics::_getChar:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,    !is_volatile);
 803   case vmIntrinsics::_getInt:                   return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,     !is_volatile);
 804   case vmIntrinsics::_getLong:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,    !is_volatile);
 805   case vmIntrinsics::_getFloat:                 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT,   !is_volatile);
 806   case vmIntrinsics::_getDouble:                return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE,  !is_volatile);
 807 
 808   case vmIntrinsics::_putObject:                return inline_unsafe_access(!is_native_ptr,  is_store, T_OBJECT,  !is_volatile);
 809   case vmIntrinsics::_putBoolean:               return inline_unsafe_access(!is_native_ptr,  is_store, T_BOOLEAN, !is_volatile);
 810   case vmIntrinsics::_putByte:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_BYTE,    !is_volatile);
 811   case vmIntrinsics::_putShort:                 return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,   !is_volatile);
 812   case vmIntrinsics::_putChar:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,    !is_volatile);
 813   case vmIntrinsics::_putInt:                   return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,     !is_volatile);
 814   case vmIntrinsics::_putLong:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,    !is_volatile);
 815   case vmIntrinsics::_putFloat:                 return inline_unsafe_access(!is_native_ptr,  is_store, T_FLOAT,   !is_volatile);
 816   case vmIntrinsics::_putDouble:                return inline_unsafe_access(!is_native_ptr,  is_store, T_DOUBLE,  !is_volatile);
 817 
 818   case vmIntrinsics::_getByte_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE,    !is_volatile);
 819   case vmIntrinsics::_getShort_raw:             return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT,   !is_volatile);
 820   case vmIntrinsics::_getChar_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR,    !is_volatile);
 821   case vmIntrinsics::_getInt_raw:               return inline_unsafe_access( is_native_ptr, !is_store, T_INT,     !is_volatile);
 822   case vmIntrinsics::_getLong_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_LONG,    !is_volatile);
 823   case vmIntrinsics::_getFloat_raw:             return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT,   !is_volatile);
 824   case vmIntrinsics::_getDouble_raw:            return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE,  !is_volatile);
 825   case vmIntrinsics::_getAddress_raw:           return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
 826 
 827   case vmIntrinsics::_putByte_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_BYTE,    !is_volatile);
 828   case vmIntrinsics::_putShort_raw:             return inline_unsafe_access( is_native_ptr,  is_store, T_SHORT,   !is_volatile);
 829   case vmIntrinsics::_putChar_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_CHAR,    !is_volatile);
 830   case vmIntrinsics::_putInt_raw:               return inline_unsafe_access( is_native_ptr,  is_store, T_INT,     !is_volatile);
 831   case vmIntrinsics::_putLong_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_LONG,    !is_volatile);
 832   case vmIntrinsics::_putFloat_raw:             return inline_unsafe_access( is_native_ptr,  is_store, T_FLOAT,   !is_volatile);
 833   case vmIntrinsics::_putDouble_raw:            return inline_unsafe_access( is_native_ptr,  is_store, T_DOUBLE,  !is_volatile);
 834   case vmIntrinsics::_putAddress_raw:           return inline_unsafe_access( is_native_ptr,  is_store, T_ADDRESS, !is_volatile);
 835 
 836   case vmIntrinsics::_getObjectVolatile:        return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT,   is_volatile);
 837   case vmIntrinsics::_getBooleanVolatile:       return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN,  is_volatile);
 838   case vmIntrinsics::_getByteVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE,     is_volatile);
 839   case vmIntrinsics::_getShortVolatile:         return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,    is_volatile);
 840   case vmIntrinsics::_getCharVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,     is_volatile);
 841   case vmIntrinsics::_getIntVolatile:           return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,      is_volatile);
 842   case vmIntrinsics::_getLongVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,     is_volatile);
 843   case vmIntrinsics::_getFloatVolatile:         return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT,    is_volatile);
 844   case vmIntrinsics::_getDoubleVolatile:        return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE,   is_volatile);
 845 
 846   case vmIntrinsics::_putObjectVolatile:        return inline_unsafe_access(!is_native_ptr,  is_store, T_OBJECT,   is_volatile);
 847   case vmIntrinsics::_putBooleanVolatile:       return inline_unsafe_access(!is_native_ptr,  is_store, T_BOOLEAN,  is_volatile);
 848   case vmIntrinsics::_putByteVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_BYTE,     is_volatile);
 849   case vmIntrinsics::_putShortVolatile:         return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,    is_volatile);
 850   case vmIntrinsics::_putCharVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,     is_volatile);
 851   case vmIntrinsics::_putIntVolatile:           return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,      is_volatile);
 852   case vmIntrinsics::_putLongVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,     is_volatile);
 853   case vmIntrinsics::_putFloatVolatile:         return inline_unsafe_access(!is_native_ptr,  is_store, T_FLOAT,    is_volatile);
 854   case vmIntrinsics::_putDoubleVolatile:        return inline_unsafe_access(!is_native_ptr,  is_store, T_DOUBLE,   is_volatile);
 855 
 856   case vmIntrinsics::_getShortUnaligned:        return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,   !is_volatile);
 857   case vmIntrinsics::_getCharUnaligned:         return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,    !is_volatile);
 858   case vmIntrinsics::_getIntUnaligned:          return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,     !is_volatile);
 859   case vmIntrinsics::_getLongUnaligned:         return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,    !is_volatile);
 860 
 861   case vmIntrinsics::_putShortUnaligned:        return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,   !is_volatile);
 862   case vmIntrinsics::_putCharUnaligned:         return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,    !is_volatile);
 863   case vmIntrinsics::_putIntUnaligned:          return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,     !is_volatile);
 864   case vmIntrinsics::_putLongUnaligned:         return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,    !is_volatile);
 865 
 866   case vmIntrinsics::_compareAndSwapObject:     return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
 867   case vmIntrinsics::_compareAndSwapInt:        return inline_unsafe_load_store(T_INT,    LS_cmpxchg);
 868   case vmIntrinsics::_compareAndSwapLong:       return inline_unsafe_load_store(T_LONG,   LS_cmpxchg);
 869 
 870   case vmIntrinsics::_putOrderedObject:         return inline_unsafe_ordered_store(T_OBJECT);
 871   case vmIntrinsics::_putOrderedInt:            return inline_unsafe_ordered_store(T_INT);
 872   case vmIntrinsics::_putOrderedLong:           return inline_unsafe_ordered_store(T_LONG);
 873 
 874   case vmIntrinsics::_getAndAddInt:             return inline_unsafe_load_store(T_INT,    LS_xadd);
 875   case vmIntrinsics::_getAndAddLong:            return inline_unsafe_load_store(T_LONG,   LS_xadd);
 876   case vmIntrinsics::_getAndSetInt:             return inline_unsafe_load_store(T_INT,    LS_xchg);
 877   case vmIntrinsics::_getAndSetLong:            return inline_unsafe_load_store(T_LONG,   LS_xchg);
 878   case vmIntrinsics::_getAndSetObject:          return inline_unsafe_load_store(T_OBJECT, LS_xchg);
 879 
 880   case vmIntrinsics::_loadFence:
 881   case vmIntrinsics::_storeFence:
 882   case vmIntrinsics::_fullFence:                return inline_unsafe_fence(intrinsic_id());
 883 
 884   case vmIntrinsics::_currentThread:            return inline_native_currentThread();
 885   case vmIntrinsics::_isInterrupted:            return inline_native_isInterrupted();
 886 
 887 #ifdef TRACE_HAVE_INTRINSICS
 888   case vmIntrinsics::_classID:                  return inline_native_classID();
 889   case vmIntrinsics::_threadID:                 return inline_native_threadID();
 890   case vmIntrinsics::_counterTime:              return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
 891 #endif
 892   case vmIntrinsics::_currentTimeMillis:        return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
 893   case vmIntrinsics::_nanoTime:                 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
 894   case vmIntrinsics::_allocateInstance:         return inline_unsafe_allocate();
 895   case vmIntrinsics::_copyMemory:               return inline_unsafe_copyMemory();
 896   case vmIntrinsics::_newArray:                 return inline_native_newArray();
 897   case vmIntrinsics::_getLength:                return inline_native_getLength();
 898   case vmIntrinsics::_copyOf:                   return inline_array_copyOf(false);
 899   case vmIntrinsics::_copyOfRange:              return inline_array_copyOf(true);
 900   case vmIntrinsics::_equalsC:                  return inline_array_equals();
 901   case vmIntrinsics::_clone:                    return inline_native_clone(intrinsic()->is_virtual());
 902 
 903   case vmIntrinsics::_isAssignableFrom:         return inline_native_subtype_check();
 904 
 905   case vmIntrinsics::_isInstance:
 906   case vmIntrinsics::_getModifiers:
 907   case vmIntrinsics::_isInterface:
 908   case vmIntrinsics::_isArray:
 909   case vmIntrinsics::_isPrimitive:
 910   case vmIntrinsics::_getSuperclass:
 911   case vmIntrinsics::_getClassAccessFlags:      return inline_native_Class_query(intrinsic_id());
 912 
 913   case vmIntrinsics::_floatToRawIntBits:
 914   case vmIntrinsics::_floatToIntBits:
 915   case vmIntrinsics::_intBitsToFloat:
 916   case vmIntrinsics::_doubleToRawLongBits:
 917   case vmIntrinsics::_doubleToLongBits:
 918   case vmIntrinsics::_longBitsToDouble:         return inline_fp_conversions(intrinsic_id());
 919 
 920   case vmIntrinsics::_numberOfLeadingZeros_i:
 921   case vmIntrinsics::_numberOfLeadingZeros_l:
 922   case vmIntrinsics::_numberOfTrailingZeros_i:
 923   case vmIntrinsics::_numberOfTrailingZeros_l:
 924   case vmIntrinsics::_bitCount_i:
 925   case vmIntrinsics::_bitCount_l:
 926   case vmIntrinsics::_reverseBytes_i:
 927   case vmIntrinsics::_reverseBytes_l:
 928   case vmIntrinsics::_reverseBytes_s:
 929   case vmIntrinsics::_reverseBytes_c:           return inline_number_methods(intrinsic_id());
 930 
 931   case vmIntrinsics::_getCallerClass:           return inline_native_Reflection_getCallerClass();
 932 
 933   case vmIntrinsics::_Reference_get:            return inline_reference_get();
 934 
 935   case vmIntrinsics::_Class_cast:               return inline_Class_cast();
 936 
 937   case vmIntrinsics::_aescrypt_encryptBlock:
 938   case vmIntrinsics::_aescrypt_decryptBlock:    return inline_aescrypt_Block(intrinsic_id());
 939 
 940   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 941   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 942     return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
 943 
 944   case vmIntrinsics::_sha_implCompress:
 945   case vmIntrinsics::_sha2_implCompress:
 946   case vmIntrinsics::_sha5_implCompress:
 947     return inline_sha_implCompress(intrinsic_id());
 948 
 949   case vmIntrinsics::_digestBase_implCompressMB:
 950     return inline_digestBase_implCompressMB(predicate);
 951 
 952   case vmIntrinsics::_multiplyToLen:
 953     return inline_multiplyToLen();
 954 
 955   case vmIntrinsics::_squareToLen:
 956     return inline_squareToLen();
 957 
 958   case vmIntrinsics::_mulAdd:
 959     return inline_mulAdd();
 960 
 961   case vmIntrinsics::_montgomeryMultiply:
 962     return inline_montgomeryMultiply();
 963   case vmIntrinsics::_montgomerySquare:
 964     return inline_montgomerySquare();
 965 
 966   case vmIntrinsics::_ghash_processBlocks:
 967     return inline_ghash_processBlocks();
 968 
 969   case vmIntrinsics::_encodeISOArray:
 970     return inline_encodeISOArray();
 971 
 972   case vmIntrinsics::_updateCRC32:
 973     return inline_updateCRC32();
 974   case vmIntrinsics::_updateBytesCRC32:
 975     return inline_updateBytesCRC32();
 976   case vmIntrinsics::_updateByteBufferCRC32:
 977     return inline_updateByteBufferCRC32();
 978 
 979   case vmIntrinsics::_updateBytesCRC32C:
 980     return inline_updateBytesCRC32C();
 981   case vmIntrinsics::_updateDirectByteBufferCRC32C:
 982     return inline_updateDirectByteBufferCRC32C();
 983 
 984   case vmIntrinsics::_profileBoolean:
 985     return inline_profileBoolean();
 986   case vmIntrinsics::_isCompileConstant:
 987     return inline_isCompileConstant();
 988 
 989   default:
 990     // If you get here, it may be that someone has added a new intrinsic
 991     // to the list in vmSymbols.hpp without implementing it here.
 992 #ifndef PRODUCT
 993     if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
 994       tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
 995                     vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
 996     }
 997 #endif
 998     return false;
 999   }
1000 }
1001 
1002 Node* LibraryCallKit::try_to_predicate(int predicate) {
1003   if (!jvms()->has_method()) {
1004     // Root JVMState has a null method.
1005     assert(map()->memory()->Opcode() == Op_Parm, "");
1006     // Insert the memory aliasing node
1007     set_all_memory(reset_memory());
1008   }
1009   assert(merged_memory(), "");
1010 
1011   switch (intrinsic_id()) {
1012   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
1013     return inline_cipherBlockChaining_AESCrypt_predicate(false);
1014   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
1015     return inline_cipherBlockChaining_AESCrypt_predicate(true);
1016   case vmIntrinsics::_digestBase_implCompressMB:
1017     return inline_digestBase_implCompressMB_predicate(predicate);
1018 
1019   default:
1020     // If you get here, it may be that someone has added a new intrinsic
1021     // to the list in vmSymbols.hpp without implementing it here.
1022 #ifndef PRODUCT
1023     if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
1024       tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
1025                     vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
1026     }
1027 #endif
1028     Node* slow_ctl = control();
1029     set_control(top()); // No fast path instrinsic
1030     return slow_ctl;
1031   }
1032 }
1033 
1034 //------------------------------set_result-------------------------------
1035 // Helper function for finishing intrinsics.
1036 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
1037   record_for_igvn(region);
1038   set_control(_gvn.transform(region));
1039   set_result( _gvn.transform(value));
1040   assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1041 }
1042 
1043 //------------------------------generate_guard---------------------------
1044 // Helper function for generating guarded fast-slow graph structures.
1045 // The given 'test', if true, guards a slow path.  If the test fails
1046 // then a fast path can be taken.  (We generally hope it fails.)
1047 // In all cases, GraphKit::control() is updated to the fast path.
1048 // The returned value represents the control for the slow path.
1049 // The return value is never 'top'; it is either a valid control
1050 // or NULL if it is obvious that the slow path can never be taken.
1051 // Also, if region and the slow control are not NULL, the slow edge
1052 // is appended to the region.
1053 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1054   if (stopped()) {
1055     // Already short circuited.
1056     return NULL;
1057   }
1058 
1059   // Build an if node and its projections.
1060   // If test is true we take the slow path, which we assume is uncommon.
1061   if (_gvn.type(test) == TypeInt::ZERO) {
1062     // The slow branch is never taken.  No need to build this guard.
1063     return NULL;
1064   }
1065 
1066   IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1067 
1068   Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1069   if (if_slow == top()) {
1070     // The slow branch is never taken.  No need to build this guard.
1071     return NULL;
1072   }
1073 
1074   if (region != NULL)
1075     region->add_req(if_slow);
1076 
1077   Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1078   set_control(if_fast);
1079 
1080   return if_slow;
1081 }
1082 
1083 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1084   return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1085 }
1086 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1087   return generate_guard(test, region, PROB_FAIR);
1088 }
1089 
1090 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1091                                                      Node* *pos_index) {
1092   if (stopped())
1093     return NULL;                // already stopped
1094   if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1095     return NULL;                // index is already adequately typed
1096   Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1097   Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1098   Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1099   if (is_neg != NULL && pos_index != NULL) {
1100     // Emulate effect of Parse::adjust_map_after_if.
1101     Node* ccast = new CastIINode(index, TypeInt::POS);
1102     ccast->set_req(0, control());
1103     (*pos_index) = _gvn.transform(ccast);
1104   }
1105   return is_neg;
1106 }
1107 
1108 // Make sure that 'position' is a valid limit index, in [0..length].
1109 // There are two equivalent plans for checking this:
1110 //   A. (offset + copyLength)  unsigned<=  arrayLength
1111 //   B. offset  <=  (arrayLength - copyLength)
1112 // We require that all of the values above, except for the sum and
1113 // difference, are already known to be non-negative.
1114 // Plan A is robust in the face of overflow, if offset and copyLength
1115 // are both hugely positive.
1116 //
1117 // Plan B is less direct and intuitive, but it does not overflow at
1118 // all, since the difference of two non-negatives is always
1119 // representable.  Whenever Java methods must perform the equivalent
1120 // check they generally use Plan B instead of Plan A.
1121 // For the moment we use Plan A.
1122 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1123                                                   Node* subseq_length,
1124                                                   Node* array_length,
1125                                                   RegionNode* region) {
1126   if (stopped())
1127     return NULL;                // already stopped
1128   bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1129   if (zero_offset && subseq_length->eqv_uncast(array_length))
1130     return NULL;                // common case of whole-array copy
1131   Node* last = subseq_length;
1132   if (!zero_offset)             // last += offset
1133     last = _gvn.transform(new AddINode(last, offset));
1134   Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1135   Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1136   Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1137   return is_over;
1138 }
1139 
1140 
1141 //--------------------------generate_current_thread--------------------
1142 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1143   ciKlass*    thread_klass = env()->Thread_klass();
1144   const Type* thread_type  = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1145   Node* thread = _gvn.transform(new ThreadLocalNode());
1146   Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1147   Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1148   tls_output = thread;
1149   return threadObj;
1150 }
1151 
1152 
1153 //------------------------------make_string_method_node------------------------
1154 // Helper method for String intrinsic functions. This version is called
1155 // with str1 and str2 pointing to String object nodes.
1156 //
1157 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1158   Node* no_ctrl = NULL;
1159 
1160   // Get start addr of string
1161   Node* str1_value   = load_String_value(no_ctrl, str1);
1162   Node* str1_offset  = load_String_offset(no_ctrl, str1);
1163   Node* str1_start   = array_element_address(str1_value, str1_offset, T_CHAR);
1164 
1165   // Get length of string 1
1166   Node* str1_len  = load_String_length(no_ctrl, str1);
1167 
1168   Node* str2_value   = load_String_value(no_ctrl, str2);
1169   Node* str2_offset  = load_String_offset(no_ctrl, str2);
1170   Node* str2_start   = array_element_address(str2_value, str2_offset, T_CHAR);
1171 
1172   Node* str2_len = NULL;
1173   Node* result = NULL;
1174 
1175   switch (opcode) {
1176   case Op_StrIndexOf:
1177     // Get length of string 2
1178     str2_len = load_String_length(no_ctrl, str2);
1179 
1180     result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1181                                 str1_start, str1_len, str2_start, str2_len);
1182     break;
1183   case Op_StrComp:
1184     // Get length of string 2
1185     str2_len = load_String_length(no_ctrl, str2);
1186 
1187     result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1188                              str1_start, str1_len, str2_start, str2_len);
1189     break;
1190   case Op_StrEquals:
1191     result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1192                                str1_start, str2_start, str1_len);
1193     break;
1194   default:
1195     ShouldNotReachHere();
1196     return NULL;
1197   }
1198 
1199   // All these intrinsics have checks.
1200   C->set_has_split_ifs(true); // Has chance for split-if optimization
1201 
1202   return _gvn.transform(result);
1203 }
1204 
1205 // Helper method for String intrinsic functions. This version is called
1206 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1207 // to Int nodes containing the lenghts of str1 and str2.
1208 //
1209 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1210   Node* result = NULL;
1211   switch (opcode) {
1212   case Op_StrIndexOf:
1213     result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1214                                 str1_start, cnt1, str2_start, cnt2);
1215     break;
1216   case Op_StrComp:
1217     result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1218                              str1_start, cnt1, str2_start, cnt2);
1219     break;
1220   case Op_StrEquals:
1221     result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1222                                str1_start, str2_start, cnt1);
1223     break;
1224   default:
1225     ShouldNotReachHere();
1226     return NULL;
1227   }
1228 
1229   // All these intrinsics have checks.
1230   C->set_has_split_ifs(true); // Has chance for split-if optimization
1231 
1232   return _gvn.transform(result);
1233 }
1234 
1235 //------------------------------inline_string_compareTo------------------------
1236 // public int java.lang.String.compareTo(String anotherString);
1237 bool LibraryCallKit::inline_string_compareTo() {
1238   Node* receiver = null_check(argument(0));
1239   Node* arg      = null_check(argument(1));
1240   if (stopped()) {
1241     return true;
1242   }
1243   set_result(make_string_method_node(Op_StrComp, receiver, arg));
1244   return true;
1245 }
1246 
1247 //------------------------------inline_string_equals------------------------
1248 bool LibraryCallKit::inline_string_equals() {
1249   Node* receiver = null_check_receiver();
1250   // NOTE: Do not null check argument for String.equals() because spec
1251   // allows to specify NULL as argument.
1252   Node* argument = this->argument(1);
1253   if (stopped()) {
1254     return true;
1255   }
1256 
1257   // paths (plus control) merge
1258   RegionNode* region = new RegionNode(5);
1259   Node* phi = new PhiNode(region, TypeInt::BOOL);
1260 
1261   // does source == target string?
1262   Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1263   Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1264 
1265   Node* if_eq = generate_slow_guard(bol, NULL);
1266   if (if_eq != NULL) {
1267     // receiver == argument
1268     phi->init_req(2, intcon(1));
1269     region->init_req(2, if_eq);
1270   }
1271 
1272   // get String klass for instanceOf
1273   ciInstanceKlass* klass = env()->String_klass();
1274 
1275   if (!stopped()) {
1276     Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1277     Node* cmp  = _gvn.transform(new CmpINode(inst, intcon(1)));
1278     Node* bol  = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1279 
1280     Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1281     //instanceOf == true, fallthrough
1282 
1283     if (inst_false != NULL) {
1284       phi->init_req(3, intcon(0));
1285       region->init_req(3, inst_false);
1286     }
1287   }
1288 
1289   if (!stopped()) {
1290     const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1291 
1292     // Properly cast the argument to String
1293     argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1294     // This path is taken only when argument's type is String:NotNull.
1295     argument = cast_not_null(argument, false);
1296 
1297     Node* no_ctrl = NULL;
1298 
1299     // Get start addr of receiver
1300     Node* receiver_val    = load_String_value(no_ctrl, receiver);
1301     Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1302     Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1303 
1304     // Get length of receiver
1305     Node* receiver_cnt  = load_String_length(no_ctrl, receiver);
1306 
1307     // Get start addr of argument
1308     Node* argument_val    = load_String_value(no_ctrl, argument);
1309     Node* argument_offset = load_String_offset(no_ctrl, argument);
1310     Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1311 
1312     // Get length of argument
1313     Node* argument_cnt  = load_String_length(no_ctrl, argument);
1314 
1315     // Check for receiver count != argument count
1316     Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1317     Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1318     Node* if_ne = generate_slow_guard(bol, NULL);
1319     if (if_ne != NULL) {
1320       phi->init_req(4, intcon(0));
1321       region->init_req(4, if_ne);
1322     }
1323 
1324     // Check for count == 0 is done by assembler code for StrEquals.
1325 
1326     if (!stopped()) {
1327       Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1328       phi->init_req(1, equals);
1329       region->init_req(1, control());
1330     }
1331   }
1332 
1333   // post merge
1334   set_control(_gvn.transform(region));
1335   record_for_igvn(region);
1336 
1337   set_result(_gvn.transform(phi));
1338   return true;
1339 }
1340 
1341 //------------------------------inline_array_equals----------------------------
1342 bool LibraryCallKit::inline_array_equals() {
1343   Node* arg1 = argument(0);
1344   Node* arg2 = argument(1);
1345   set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1346   return true;
1347 }
1348 
1349 // Java version of String.indexOf(constant string)
1350 // class StringDecl {
1351 //   StringDecl(char[] ca) {
1352 //     offset = 0;
1353 //     count = ca.length;
1354 //     value = ca;
1355 //   }
1356 //   int offset;
1357 //   int count;
1358 //   char[] value;
1359 // }
1360 //
1361 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1362 //                             int targetOffset, int cache_i, int md2) {
1363 //   int cache = cache_i;
1364 //   int sourceOffset = string_object.offset;
1365 //   int sourceCount = string_object.count;
1366 //   int targetCount = target_object.length;
1367 //
1368 //   int targetCountLess1 = targetCount - 1;
1369 //   int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1370 //
1371 //   char[] source = string_object.value;
1372 //   char[] target = target_object;
1373 //   int lastChar = target[targetCountLess1];
1374 //
1375 //  outer_loop:
1376 //   for (int i = sourceOffset; i < sourceEnd; ) {
1377 //     int src = source[i + targetCountLess1];
1378 //     if (src == lastChar) {
1379 //       // With random strings and a 4-character alphabet,
1380 //       // reverse matching at this point sets up 0.8% fewer
1381 //       // frames, but (paradoxically) makes 0.3% more probes.
1382 //       // Since those probes are nearer the lastChar probe,
1383 //       // there is may be a net D$ win with reverse matching.
1384 //       // But, reversing loop inhibits unroll of inner loop
1385 //       // for unknown reason.  So, does running outer loop from
1386 //       // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1387 //       for (int j = 0; j < targetCountLess1; j++) {
1388 //         if (target[targetOffset + j] != source[i+j]) {
1389 //           if ((cache & (1 << source[i+j])) == 0) {
1390 //             if (md2 < j+1) {
1391 //               i += j+1;
1392 //               continue outer_loop;
1393 //             }
1394 //           }
1395 //           i += md2;
1396 //           continue outer_loop;
1397 //         }
1398 //       }
1399 //       return i - sourceOffset;
1400 //     }
1401 //     if ((cache & (1 << src)) == 0) {
1402 //       i += targetCountLess1;
1403 //     } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1404 //     i++;
1405 //   }
1406 //   return -1;
1407 // }
1408 
1409 //------------------------------string_indexOf------------------------
1410 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1411                                      jint cache_i, jint md2_i) {
1412 
1413   Node* no_ctrl  = NULL;
1414   float likely   = PROB_LIKELY(0.9);
1415   float unlikely = PROB_UNLIKELY(0.9);
1416 
1417   const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1418 
1419   Node* source        = load_String_value(no_ctrl, string_object);
1420   Node* sourceOffset  = load_String_offset(no_ctrl, string_object);
1421   Node* sourceCount   = load_String_length(no_ctrl, string_object);
1422 
1423   Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1424   jint target_length = target_array->length();
1425   const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1426   const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1427 
1428   // String.value field is known to be @Stable.
1429   if (UseImplicitStableValues) {
1430     target = cast_array_to_stable(target, target_type);
1431   }
1432 
1433   IdealKit kit(this, false, true);
1434 #define __ kit.
1435   Node* zero             = __ ConI(0);
1436   Node* one              = __ ConI(1);
1437   Node* cache            = __ ConI(cache_i);
1438   Node* md2              = __ ConI(md2_i);
1439   Node* lastChar         = __ ConI(target_array->char_at(target_length - 1));
1440   Node* targetCountLess1 = __ ConI(target_length - 1);
1441   Node* targetOffset     = __ ConI(targetOffset_i);
1442   Node* sourceEnd        = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1443 
1444   IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1445   Node* outer_loop = __ make_label(2 /* goto */);
1446   Node* return_    = __ make_label(1);
1447 
1448   __ set(rtn,__ ConI(-1));
1449   __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1450        Node* i2  = __ AddI(__ value(i), targetCountLess1);
1451        // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1452        Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1453        __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1454          __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1455               Node* tpj = __ AddI(targetOffset, __ value(j));
1456               Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1457               Node* ipj  = __ AddI(__ value(i), __ value(j));
1458               Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1459               __ if_then(targ, BoolTest::ne, src2); {
1460                 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1461                   __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1462                     __ increment(i, __ AddI(__ value(j), one));
1463                     __ goto_(outer_loop);
1464                   } __ end_if(); __ dead(j);
1465                 }__ end_if(); __ dead(j);
1466                 __ increment(i, md2);
1467                 __ goto_(outer_loop);
1468               }__ end_if();
1469               __ increment(j, one);
1470          }__ end_loop(); __ dead(j);
1471          __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1472          __ goto_(return_);
1473        }__ end_if();
1474        __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1475          __ increment(i, targetCountLess1);
1476        }__ end_if();
1477        __ increment(i, one);
1478        __ bind(outer_loop);
1479   }__ end_loop(); __ dead(i);
1480   __ bind(return_);
1481 
1482   // Final sync IdealKit and GraphKit.
1483   final_sync(kit);
1484   Node* result = __ value(rtn);
1485 #undef __
1486   C->set_has_loops(true);
1487   return result;
1488 }
1489 
1490 //------------------------------inline_string_indexOf------------------------
1491 bool LibraryCallKit::inline_string_indexOf() {
1492   Node* receiver = argument(0);
1493   Node* arg      = argument(1);
1494 
1495   Node* result;
1496   if (Matcher::has_match_rule(Op_StrIndexOf) &&
1497       UseSSE42Intrinsics) {
1498     // Generate SSE4.2 version of indexOf
1499     // We currently only have match rules that use SSE4.2
1500 
1501     receiver = null_check(receiver);
1502     arg      = null_check(arg);
1503     if (stopped()) {
1504       return true;
1505     }
1506 
1507     // Make the merge point
1508     RegionNode* result_rgn = new RegionNode(4);
1509     Node*       result_phi = new PhiNode(result_rgn, TypeInt::INT);
1510     Node* no_ctrl  = NULL;
1511 
1512     // Get start addr of source string
1513     Node* source = load_String_value(no_ctrl, receiver);
1514     Node* source_offset = load_String_offset(no_ctrl, receiver);
1515     Node* source_start = array_element_address(source, source_offset, T_CHAR);
1516 
1517     // Get length of source string
1518     Node* source_cnt  = load_String_length(no_ctrl, receiver);
1519 
1520     // Get start addr of substring
1521     Node* substr = load_String_value(no_ctrl, arg);
1522     Node* substr_offset = load_String_offset(no_ctrl, arg);
1523     Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1524 
1525     // Get length of source string
1526     Node* substr_cnt  = load_String_length(no_ctrl, arg);
1527 
1528     // Check for substr count > string count
1529     Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1530     Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1531     Node* if_gt = generate_slow_guard(bol, NULL);
1532     if (if_gt != NULL) {
1533       result_phi->init_req(2, intcon(-1));
1534       result_rgn->init_req(2, if_gt);
1535     }
1536 
1537     if (!stopped()) {
1538       // Check for substr count == 0
1539       cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1540       bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1541       Node* if_zero = generate_slow_guard(bol, NULL);
1542       if (if_zero != NULL) {
1543         result_phi->init_req(3, intcon(0));
1544         result_rgn->init_req(3, if_zero);
1545       }
1546     }
1547 
1548     if (!stopped()) {
1549       result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1550       result_phi->init_req(1, result);
1551       result_rgn->init_req(1, control());
1552     }
1553     set_control(_gvn.transform(result_rgn));
1554     record_for_igvn(result_rgn);
1555     result = _gvn.transform(result_phi);
1556 
1557   } else { // Use LibraryCallKit::string_indexOf
1558     // don't intrinsify if argument isn't a constant string.
1559     if (!arg->is_Con()) {
1560      return false;
1561     }
1562     const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1563     if (str_type == NULL) {
1564       return false;
1565     }
1566     ciInstanceKlass* klass = env()->String_klass();
1567     ciObject* str_const = str_type->const_oop();
1568     if (str_const == NULL || str_const->klass() != klass) {
1569       return false;
1570     }
1571     ciInstance* str = str_const->as_instance();
1572     assert(str != NULL, "must be instance");
1573 
1574     ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1575     ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1576 
1577     int o;
1578     int c;
1579     if (java_lang_String::has_offset_field()) {
1580       o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1581       c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1582     } else {
1583       o = 0;
1584       c = pat->length();
1585     }
1586 
1587     // constant strings have no offset and count == length which
1588     // simplifies the resulting code somewhat so lets optimize for that.
1589     if (o != 0 || c != pat->length()) {
1590      return false;
1591     }
1592 
1593     receiver = null_check(receiver, T_OBJECT);
1594     // NOTE: No null check on the argument is needed since it's a constant String oop.
1595     if (stopped()) {
1596       return true;
1597     }
1598 
1599     // The null string as a pattern always returns 0 (match at beginning of string)
1600     if (c == 0) {
1601       set_result(intcon(0));
1602       return true;
1603     }
1604 
1605     // Generate default indexOf
1606     jchar lastChar = pat->char_at(o + (c - 1));
1607     int cache = 0;
1608     int i;
1609     for (i = 0; i < c - 1; i++) {
1610       assert(i < pat->length(), "out of range");
1611       cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1612     }
1613 
1614     int md2 = c;
1615     for (i = 0; i < c - 1; i++) {
1616       assert(i < pat->length(), "out of range");
1617       if (pat->char_at(o + i) == lastChar) {
1618         md2 = (c - 1) - i;
1619       }
1620     }
1621 
1622     result = string_indexOf(receiver, pat, o, cache, md2);
1623   }
1624   set_result(result);
1625   return true;
1626 }
1627 
1628 //--------------------------round_double_node--------------------------------
1629 // Round a double node if necessary.
1630 Node* LibraryCallKit::round_double_node(Node* n) {
1631   if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1632     n = _gvn.transform(new RoundDoubleNode(0, n));
1633   return n;
1634 }
1635 
1636 //------------------------------inline_math-----------------------------------
1637 // public static double Math.abs(double)
1638 // public static double Math.sqrt(double)
1639 // public static double Math.log(double)
1640 // public static double Math.log10(double)
1641 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1642   Node* arg = round_double_node(argument(0));
1643   Node* n;
1644   switch (id) {
1645   case vmIntrinsics::_dabs:   n = new AbsDNode(                arg);  break;
1646   case vmIntrinsics::_dsqrt:  n = new SqrtDNode(C, control(),  arg);  break;
1647   case vmIntrinsics::_dlog:   n = new LogDNode(C, control(),   arg);  break;
1648   case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg);  break;
1649   default:  fatal_unexpected_iid(id);  break;
1650   }
1651   set_result(_gvn.transform(n));
1652   return true;
1653 }
1654 
1655 //------------------------------inline_trig----------------------------------
1656 // Inline sin/cos/tan instructions, if possible.  If rounding is required, do
1657 // argument reduction which will turn into a fast/slow diamond.
1658 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1659   Node* arg = round_double_node(argument(0));
1660   Node* n = NULL;
1661 
1662   switch (id) {
1663   case vmIntrinsics::_dsin:  n = new SinDNode(C, control(), arg);  break;
1664   case vmIntrinsics::_dcos:  n = new CosDNode(C, control(), arg);  break;
1665   case vmIntrinsics::_dtan:  n = new TanDNode(C, control(), arg);  break;
1666   default:  fatal_unexpected_iid(id);  break;
1667   }
1668   n = _gvn.transform(n);
1669 
1670   // Rounding required?  Check for argument reduction!
1671   if (Matcher::strict_fp_requires_explicit_rounding) {
1672     static const double     pi_4 =  0.7853981633974483;
1673     static const double neg_pi_4 = -0.7853981633974483;
1674     // pi/2 in 80-bit extended precision
1675     // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1676     // -pi/2 in 80-bit extended precision
1677     // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00};
1678     // Cutoff value for using this argument reduction technique
1679     //static const double    pi_2_minus_epsilon =  1.564660403643354;
1680     //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1681 
1682     // Pseudocode for sin:
1683     // if (x <= Math.PI / 4.0) {
1684     //   if (x >= -Math.PI / 4.0) return  fsin(x);
1685     //   if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1686     // } else {
1687     //   if (x <=  Math.PI / 2.0) return  fcos(x - Math.PI / 2.0);
1688     // }
1689     // return StrictMath.sin(x);
1690 
1691     // Pseudocode for cos:
1692     // if (x <= Math.PI / 4.0) {
1693     //   if (x >= -Math.PI / 4.0) return  fcos(x);
1694     //   if (x >= -Math.PI / 2.0) return  fsin(x + Math.PI / 2.0);
1695     // } else {
1696     //   if (x <=  Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1697     // }
1698     // return StrictMath.cos(x);
1699 
1700     // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1701     // requires a special machine instruction to load it.  Instead we'll try
1702     // the 'easy' case.  If we really need the extra range +/- PI/2 we'll
1703     // probably do the math inside the SIN encoding.
1704 
1705     // Make the merge point
1706     RegionNode* r = new RegionNode(3);
1707     Node* phi = new PhiNode(r, Type::DOUBLE);
1708 
1709     // Flatten arg so we need only 1 test
1710     Node *abs = _gvn.transform(new AbsDNode(arg));
1711     // Node for PI/4 constant
1712     Node *pi4 = makecon(TypeD::make(pi_4));
1713     // Check PI/4 : abs(arg)
1714     Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1715     // Check: If PI/4 < abs(arg) then go slow
1716     Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1717     // Branch either way
1718     IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1719     set_control(opt_iff(r,iff));
1720 
1721     // Set fast path result
1722     phi->init_req(2, n);
1723 
1724     // Slow path - non-blocking leaf call
1725     Node* call = NULL;
1726     switch (id) {
1727     case vmIntrinsics::_dsin:
1728       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1729                                CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1730                                "Sin", NULL, arg, top());
1731       break;
1732     case vmIntrinsics::_dcos:
1733       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1734                                CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1735                                "Cos", NULL, arg, top());
1736       break;
1737     case vmIntrinsics::_dtan:
1738       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1739                                CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1740                                "Tan", NULL, arg, top());
1741       break;
1742     }
1743     assert(control()->in(0) == call, "");
1744     Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1745     r->init_req(1, control());
1746     phi->init_req(1, slow_result);
1747 
1748     // Post-merge
1749     set_control(_gvn.transform(r));
1750     record_for_igvn(r);
1751     n = _gvn.transform(phi);
1752 
1753     C->set_has_split_ifs(true); // Has chance for split-if optimization
1754   }
1755   set_result(n);
1756   return true;
1757 }
1758 
1759 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1760   //-------------------
1761   //result=(result.isNaN())? funcAddr():result;
1762   // Check: If isNaN() by checking result!=result? then either trap
1763   // or go to runtime
1764   Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1765   // Build the boolean node
1766   Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1767 
1768   if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1769     { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1770       // The pow or exp intrinsic returned a NaN, which requires a call
1771       // to the runtime.  Recompile with the runtime call.
1772       uncommon_trap(Deoptimization::Reason_intrinsic,
1773                     Deoptimization::Action_make_not_entrant);
1774     }
1775     return result;
1776   } else {
1777     // If this inlining ever returned NaN in the past, we compile a call
1778     // to the runtime to properly handle corner cases
1779 
1780     IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1781     Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1782     Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1783 
1784     if (!if_slow->is_top()) {
1785       RegionNode* result_region = new RegionNode(3);
1786       PhiNode*    result_val = new PhiNode(result_region, Type::DOUBLE);
1787 
1788       result_region->init_req(1, if_fast);
1789       result_val->init_req(1, result);
1790 
1791       set_control(if_slow);
1792 
1793       const TypePtr* no_memory_effects = NULL;
1794       Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1795                                    no_memory_effects,
1796                                    x, top(), y, y ? top() : NULL);
1797       Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1798 #ifdef ASSERT
1799       Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1800       assert(value_top == top(), "second value must be top");
1801 #endif
1802 
1803       result_region->init_req(2, control());
1804       result_val->init_req(2, value);
1805       set_control(_gvn.transform(result_region));
1806       return _gvn.transform(result_val);
1807     } else {
1808       return result;
1809     }
1810   }
1811 }
1812 
1813 //------------------------------inline_exp-------------------------------------
1814 // Inline exp instructions, if possible.  The Intel hardware only misses
1815 // really odd corner cases (+/- Infinity).  Just uncommon-trap them.
1816 bool LibraryCallKit::inline_exp() {
1817   Node* arg = round_double_node(argument(0));
1818   Node* n   = _gvn.transform(new ExpDNode(C, control(), arg));
1819 
1820   n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1821   set_result(n);
1822 
1823   C->set_has_split_ifs(true); // Has chance for split-if optimization
1824   return true;
1825 }
1826 
1827 //------------------------------inline_pow-------------------------------------
1828 // Inline power instructions, if possible.
1829 bool LibraryCallKit::inline_pow() {
1830   // Pseudocode for pow
1831   // if (y == 2) {
1832   //   return x * x;
1833   // } else {
1834   //   if (x <= 0.0) {
1835   //     long longy = (long)y;
1836   //     if ((double)longy == y) { // if y is long
1837   //       if (y + 1 == y) longy = 0; // huge number: even
1838   //       result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1839   //     } else {
1840   //       result = NaN;
1841   //     }
1842   //   } else {
1843   //     result = DPow(x,y);
1844   //   }
1845   //   if (result != result)?  {
1846   //     result = uncommon_trap() or runtime_call();
1847   //   }
1848   //   return result;
1849   // }
1850 
1851   Node* x = round_double_node(argument(0));
1852   Node* y = round_double_node(argument(2));
1853 
1854   Node* result = NULL;
1855 
1856   Node*   const_two_node = makecon(TypeD::make(2.0));
1857   Node*   cmp_node       = _gvn.transform(new CmpDNode(y, const_two_node));
1858   Node*   bool_node      = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1859   IfNode* if_node        = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1860   Node*   if_true        = _gvn.transform(new IfTrueNode(if_node));
1861   Node*   if_false       = _gvn.transform(new IfFalseNode(if_node));
1862 
1863   RegionNode* region_node = new RegionNode(3);
1864   region_node->init_req(1, if_true);
1865 
1866   Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1867   // special case for x^y where y == 2, we can convert it to x * x
1868   phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1869 
1870   // set control to if_false since we will now process the false branch
1871   set_control(if_false);
1872 
1873   if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1874     // Short form: skip the fancy tests and just check for NaN result.
1875     result = _gvn.transform(new PowDNode(C, control(), x, y));
1876   } else {
1877     // If this inlining ever returned NaN in the past, include all
1878     // checks + call to the runtime.
1879 
1880     // Set the merge point for If node with condition of (x <= 0.0)
1881     // There are four possible paths to region node and phi node
1882     RegionNode *r = new RegionNode(4);
1883     Node *phi = new PhiNode(r, Type::DOUBLE);
1884 
1885     // Build the first if node: if (x <= 0.0)
1886     // Node for 0 constant
1887     Node *zeronode = makecon(TypeD::ZERO);
1888     // Check x:0
1889     Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1890     // Check: If (x<=0) then go complex path
1891     Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1892     // Branch either way
1893     IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1894     // Fast path taken; set region slot 3
1895     Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1896     r->init_req(3,fast_taken); // Capture fast-control
1897 
1898     // Fast path not-taken, i.e. slow path
1899     Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1900 
1901     // Set fast path result
1902     Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1903     phi->init_req(3, fast_result);
1904 
1905     // Complex path
1906     // Build the second if node (if y is long)
1907     // Node for (long)y
1908     Node *longy = _gvn.transform(new ConvD2LNode(y));
1909     // Node for (double)((long) y)
1910     Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1911     // Check (double)((long) y) : y
1912     Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1913     // Check if (y isn't long) then go to slow path
1914 
1915     Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1916     // Branch either way
1917     IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1918     Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1919 
1920     Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1921 
1922     // Calculate DPow(abs(x), y)*(1 & (long)y)
1923     // Node for constant 1
1924     Node *conone = longcon(1);
1925     // 1& (long)y
1926     Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1927 
1928     // A huge number is always even. Detect a huge number by checking
1929     // if y + 1 == y and set integer to be tested for parity to 0.
1930     // Required for corner case:
1931     // (long)9.223372036854776E18 = max_jlong
1932     // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1933     // max_jlong is odd but 9.223372036854776E18 is even
1934     Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1935     Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1936     Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1937     Node* correctedsign = NULL;
1938     if (ConditionalMoveLimit != 0) {
1939       correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1940     } else {
1941       IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1942       RegionNode *r = new RegionNode(3);
1943       Node *phi = new PhiNode(r, TypeLong::LONG);
1944       r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1945       r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1946       phi->init_req(1, signnode);
1947       phi->init_req(2, longcon(0));
1948       correctedsign = _gvn.transform(phi);
1949       ylong_path = _gvn.transform(r);
1950       record_for_igvn(r);
1951     }
1952 
1953     // zero node
1954     Node *conzero = longcon(0);
1955     // Check (1&(long)y)==0?
1956     Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1957     // Check if (1&(long)y)!=0?, if so the result is negative
1958     Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1959     // abs(x)
1960     Node *absx=_gvn.transform(new AbsDNode(x));
1961     // abs(x)^y
1962     Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1963     // -abs(x)^y
1964     Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1965     // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1966     Node *signresult = NULL;
1967     if (ConditionalMoveLimit != 0) {
1968       signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1969     } else {
1970       IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1971       RegionNode *r = new RegionNode(3);
1972       Node *phi = new PhiNode(r, Type::DOUBLE);
1973       r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1974       r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1975       phi->init_req(1, absxpowy);
1976       phi->init_req(2, negabsxpowy);
1977       signresult = _gvn.transform(phi);
1978       ylong_path = _gvn.transform(r);
1979       record_for_igvn(r);
1980     }
1981     // Set complex path fast result
1982     r->init_req(2, ylong_path);
1983     phi->init_req(2, signresult);
1984 
1985     static const jlong nan_bits = CONST64(0x7ff8000000000000);
1986     Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1987     r->init_req(1,slow_path);
1988     phi->init_req(1,slow_result);
1989 
1990     // Post merge
1991     set_control(_gvn.transform(r));
1992     record_for_igvn(r);
1993     result = _gvn.transform(phi);
1994   }
1995 
1996   result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1997 
1998   // control from finish_pow_exp is now input to the region node
1999   region_node->set_req(2, control());
2000   // the result from finish_pow_exp is now input to the phi node
2001   phi_node->init_req(2, result);
2002   set_control(_gvn.transform(region_node));
2003   record_for_igvn(region_node);
2004   set_result(_gvn.transform(phi_node));
2005 
2006   C->set_has_split_ifs(true); // Has chance for split-if optimization
2007   return true;
2008 }
2009 
2010 //------------------------------runtime_math-----------------------------
2011 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
2012   assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
2013          "must be (DD)D or (D)D type");
2014 
2015   // Inputs
2016   Node* a = round_double_node(argument(0));
2017   Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
2018 
2019   const TypePtr* no_memory_effects = NULL;
2020   Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
2021                                  no_memory_effects,
2022                                  a, top(), b, b ? top() : NULL);
2023   Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
2024 #ifdef ASSERT
2025   Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
2026   assert(value_top == top(), "second value must be top");
2027 #endif
2028 
2029   set_result(value);
2030   return true;
2031 }
2032 
2033 //------------------------------inline_math_native-----------------------------
2034 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
2035 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
2036   switch (id) {
2037     // These intrinsics are not properly supported on all hardware
2038   case vmIntrinsics::_dcos:   return Matcher::has_match_rule(Op_CosD)   ? inline_trig(id) :
2039     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos),   "COS");
2040   case vmIntrinsics::_dsin:   return Matcher::has_match_rule(Op_SinD)   ? inline_trig(id) :
2041     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin),   "SIN");
2042   case vmIntrinsics::_dtan:   return Matcher::has_match_rule(Op_TanD)   ? inline_trig(id) :
2043     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan),   "TAN");
2044 
2045   case vmIntrinsics::_dlog:   return Matcher::has_match_rule(Op_LogD)   ? inline_math(id) :
2046     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog),   "LOG");
2047   case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2048     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2049 
2050     // These intrinsics are supported on all hardware
2051   case vmIntrinsics::_dsqrt:  return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2052   case vmIntrinsics::_dabs:   return Matcher::has_match_rule(Op_AbsD)   ? inline_math(id) : false;
2053 
2054   case vmIntrinsics::_dexp:   return Matcher::has_match_rule(Op_ExpD)   ? inline_exp()    :
2055     runtime_math(OptoRuntime::Math_D_D_Type(),  FN_PTR(SharedRuntime::dexp),  "EXP");
2056   case vmIntrinsics::_dpow:   return Matcher::has_match_rule(Op_PowD)   ? inline_pow()    :
2057     runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow),  "POW");
2058 #undef FN_PTR
2059 
2060    // These intrinsics are not yet correctly implemented
2061   case vmIntrinsics::_datan2:
2062     return false;
2063 
2064   default:
2065     fatal_unexpected_iid(id);
2066     return false;
2067   }
2068 }
2069 
2070 static bool is_simple_name(Node* n) {
2071   return (n->req() == 1         // constant
2072           || (n->is_Type() && n->as_Type()->type()->singleton())
2073           || n->is_Proj()       // parameter or return value
2074           || n->is_Phi()        // local of some sort
2075           );
2076 }
2077 
2078 //----------------------------inline_min_max-----------------------------------
2079 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2080   set_result(generate_min_max(id, argument(0), argument(1)));
2081   return true;
2082 }
2083 
2084 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2085   Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2086   IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2087   Node* fast_path = _gvn.transform( new IfFalseNode(check));
2088   Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2089 
2090   {
2091     PreserveJVMState pjvms(this);
2092     PreserveReexecuteState preexecs(this);
2093     jvms()->set_should_reexecute(true);
2094 
2095     set_control(slow_path);
2096     set_i_o(i_o());
2097 
2098     uncommon_trap(Deoptimization::Reason_intrinsic,
2099                   Deoptimization::Action_none);
2100   }
2101 
2102   set_control(fast_path);
2103   set_result(math);
2104 }
2105 
2106 template <typename OverflowOp>
2107 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2108   typedef typename OverflowOp::MathOp MathOp;
2109 
2110   MathOp* mathOp = new MathOp(arg1, arg2);
2111   Node* operation = _gvn.transform( mathOp );
2112   Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2113   inline_math_mathExact(operation, ofcheck);
2114   return true;
2115 }
2116 
2117 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2118   return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2119 }
2120 
2121 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2122   return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2123 }
2124 
2125 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2126   return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2127 }
2128 
2129 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2130   return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2131 }
2132 
2133 bool LibraryCallKit::inline_math_negateExactI() {
2134   return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2135 }
2136 
2137 bool LibraryCallKit::inline_math_negateExactL() {
2138   return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2139 }
2140 
2141 bool LibraryCallKit::inline_math_multiplyExactI() {
2142   return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2143 }
2144 
2145 bool LibraryCallKit::inline_math_multiplyExactL() {
2146   return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2147 }
2148 
2149 Node*
2150 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2151   // These are the candidate return value:
2152   Node* xvalue = x0;
2153   Node* yvalue = y0;
2154 
2155   if (xvalue == yvalue) {
2156     return xvalue;
2157   }
2158 
2159   bool want_max = (id == vmIntrinsics::_max);
2160 
2161   const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2162   const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2163   if (txvalue == NULL || tyvalue == NULL)  return top();
2164   // This is not really necessary, but it is consistent with a
2165   // hypothetical MaxINode::Value method:
2166   int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2167 
2168   // %%% This folding logic should (ideally) be in a different place.
2169   // Some should be inside IfNode, and there to be a more reliable
2170   // transformation of ?: style patterns into cmoves.  We also want
2171   // more powerful optimizations around cmove and min/max.
2172 
2173   // Try to find a dominating comparison of these guys.
2174   // It can simplify the index computation for Arrays.copyOf
2175   // and similar uses of System.arraycopy.
2176   // First, compute the normalized version of CmpI(x, y).
2177   int   cmp_op = Op_CmpI;
2178   Node* xkey = xvalue;
2179   Node* ykey = yvalue;
2180   Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2181   if (ideal_cmpxy->is_Cmp()) {
2182     // E.g., if we have CmpI(length - offset, count),
2183     // it might idealize to CmpI(length, count + offset)
2184     cmp_op = ideal_cmpxy->Opcode();
2185     xkey = ideal_cmpxy->in(1);
2186     ykey = ideal_cmpxy->in(2);
2187   }
2188 
2189   // Start by locating any relevant comparisons.
2190   Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2191   Node* cmpxy = NULL;
2192   Node* cmpyx = NULL;
2193   for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2194     Node* cmp = start_from->fast_out(k);
2195     if (cmp->outcnt() > 0 &&            // must have prior uses
2196         cmp->in(0) == NULL &&           // must be context-independent
2197         cmp->Opcode() == cmp_op) {      // right kind of compare
2198       if (cmp->in(1) == xkey && cmp->in(2) == ykey)  cmpxy = cmp;
2199       if (cmp->in(1) == ykey && cmp->in(2) == xkey)  cmpyx = cmp;
2200     }
2201   }
2202 
2203   const int NCMPS = 2;
2204   Node* cmps[NCMPS] = { cmpxy, cmpyx };
2205   int cmpn;
2206   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2207     if (cmps[cmpn] != NULL)  break;     // find a result
2208   }
2209   if (cmpn < NCMPS) {
2210     // Look for a dominating test that tells us the min and max.
2211     int depth = 0;                // Limit search depth for speed
2212     Node* dom = control();
2213     for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2214       if (++depth >= 100)  break;
2215       Node* ifproj = dom;
2216       if (!ifproj->is_Proj())  continue;
2217       Node* iff = ifproj->in(0);
2218       if (!iff->is_If())  continue;
2219       Node* bol = iff->in(1);
2220       if (!bol->is_Bool())  continue;
2221       Node* cmp = bol->in(1);
2222       if (cmp == NULL)  continue;
2223       for (cmpn = 0; cmpn < NCMPS; cmpn++)
2224         if (cmps[cmpn] == cmp)  break;
2225       if (cmpn == NCMPS)  continue;
2226       BoolTest::mask btest = bol->as_Bool()->_test._test;
2227       if (ifproj->is_IfFalse())  btest = BoolTest(btest).negate();
2228       if (cmp->in(1) == ykey)    btest = BoolTest(btest).commute();
2229       // At this point, we know that 'x btest y' is true.
2230       switch (btest) {
2231       case BoolTest::eq:
2232         // They are proven equal, so we can collapse the min/max.
2233         // Either value is the answer.  Choose the simpler.
2234         if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2235           return yvalue;
2236         return xvalue;
2237       case BoolTest::lt:          // x < y
2238       case BoolTest::le:          // x <= y
2239         return (want_max ? yvalue : xvalue);
2240       case BoolTest::gt:          // x > y
2241       case BoolTest::ge:          // x >= y
2242         return (want_max ? xvalue : yvalue);
2243       }
2244     }
2245   }
2246 
2247   // We failed to find a dominating test.
2248   // Let's pick a test that might GVN with prior tests.
2249   Node*          best_bol   = NULL;
2250   BoolTest::mask best_btest = BoolTest::illegal;
2251   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2252     Node* cmp = cmps[cmpn];
2253     if (cmp == NULL)  continue;
2254     for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2255       Node* bol = cmp->fast_out(j);
2256       if (!bol->is_Bool())  continue;
2257       BoolTest::mask btest = bol->as_Bool()->_test._test;
2258       if (btest == BoolTest::eq || btest == BoolTest::ne)  continue;
2259       if (cmp->in(1) == ykey)   btest = BoolTest(btest).commute();
2260       if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2261         best_bol   = bol->as_Bool();
2262         best_btest = btest;
2263       }
2264     }
2265   }
2266 
2267   Node* answer_if_true  = NULL;
2268   Node* answer_if_false = NULL;
2269   switch (best_btest) {
2270   default:
2271     if (cmpxy == NULL)
2272       cmpxy = ideal_cmpxy;
2273     best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2274     // and fall through:
2275   case BoolTest::lt:          // x < y
2276   case BoolTest::le:          // x <= y
2277     answer_if_true  = (want_max ? yvalue : xvalue);
2278     answer_if_false = (want_max ? xvalue : yvalue);
2279     break;
2280   case BoolTest::gt:          // x > y
2281   case BoolTest::ge:          // x >= y
2282     answer_if_true  = (want_max ? xvalue : yvalue);
2283     answer_if_false = (want_max ? yvalue : xvalue);
2284     break;
2285   }
2286 
2287   jint hi, lo;
2288   if (want_max) {
2289     // We can sharpen the minimum.
2290     hi = MAX2(txvalue->_hi, tyvalue->_hi);
2291     lo = MAX2(txvalue->_lo, tyvalue->_lo);
2292   } else {
2293     // We can sharpen the maximum.
2294     hi = MIN2(txvalue->_hi, tyvalue->_hi);
2295     lo = MIN2(txvalue->_lo, tyvalue->_lo);
2296   }
2297 
2298   // Use a flow-free graph structure, to avoid creating excess control edges
2299   // which could hinder other optimizations.
2300   // Since Math.min/max is often used with arraycopy, we want
2301   // tightly_coupled_allocation to be able to see beyond min/max expressions.
2302   Node* cmov = CMoveNode::make(NULL, best_bol,
2303                                answer_if_false, answer_if_true,
2304                                TypeInt::make(lo, hi, widen));
2305 
2306   return _gvn.transform(cmov);
2307 
2308   /*
2309   // This is not as desirable as it may seem, since Min and Max
2310   // nodes do not have a full set of optimizations.
2311   // And they would interfere, anyway, with 'if' optimizations
2312   // and with CMoveI canonical forms.
2313   switch (id) {
2314   case vmIntrinsics::_min:
2315     result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2316   case vmIntrinsics::_max:
2317     result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2318   default:
2319     ShouldNotReachHere();
2320   }
2321   */
2322 }
2323 
2324 inline int
2325 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2326   const TypePtr* base_type = TypePtr::NULL_PTR;
2327   if (base != NULL)  base_type = _gvn.type(base)->isa_ptr();
2328   if (base_type == NULL) {
2329     // Unknown type.
2330     return Type::AnyPtr;
2331   } else if (base_type == TypePtr::NULL_PTR) {
2332     // Since this is a NULL+long form, we have to switch to a rawptr.
2333     base   = _gvn.transform(new CastX2PNode(offset));
2334     offset = MakeConX(0);
2335     return Type::RawPtr;
2336   } else if (base_type->base() == Type::RawPtr) {
2337     return Type::RawPtr;
2338   } else if (base_type->isa_oopptr()) {
2339     // Base is never null => always a heap address.
2340     if (base_type->ptr() == TypePtr::NotNull) {
2341       return Type::OopPtr;
2342     }
2343     // Offset is small => always a heap address.
2344     const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2345     if (offset_type != NULL &&
2346         base_type->offset() == 0 &&     // (should always be?)
2347         offset_type->_lo >= 0 &&
2348         !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2349       return Type::OopPtr;
2350     }
2351     // Otherwise, it might either be oop+off or NULL+addr.
2352     return Type::AnyPtr;
2353   } else {
2354     // No information:
2355     return Type::AnyPtr;
2356   }
2357 }
2358 
2359 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2360   int kind = classify_unsafe_addr(base, offset);
2361   if (kind == Type::RawPtr) {
2362     return basic_plus_adr(top(), base, offset);
2363   } else {
2364     return basic_plus_adr(base, offset);
2365   }
2366 }
2367 
2368 //--------------------------inline_number_methods-----------------------------
2369 // inline int     Integer.numberOfLeadingZeros(int)
2370 // inline int        Long.numberOfLeadingZeros(long)
2371 //
2372 // inline int     Integer.numberOfTrailingZeros(int)
2373 // inline int        Long.numberOfTrailingZeros(long)
2374 //
2375 // inline int     Integer.bitCount(int)
2376 // inline int        Long.bitCount(long)
2377 //
2378 // inline char  Character.reverseBytes(char)
2379 // inline short     Short.reverseBytes(short)
2380 // inline int     Integer.reverseBytes(int)
2381 // inline long       Long.reverseBytes(long)
2382 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2383   Node* arg = argument(0);
2384   Node* n;
2385   switch (id) {
2386   case vmIntrinsics::_numberOfLeadingZeros_i:   n = new CountLeadingZerosINode( arg);  break;
2387   case vmIntrinsics::_numberOfLeadingZeros_l:   n = new CountLeadingZerosLNode( arg);  break;
2388   case vmIntrinsics::_numberOfTrailingZeros_i:  n = new CountTrailingZerosINode(arg);  break;
2389   case vmIntrinsics::_numberOfTrailingZeros_l:  n = new CountTrailingZerosLNode(arg);  break;
2390   case vmIntrinsics::_bitCount_i:               n = new PopCountINode(          arg);  break;
2391   case vmIntrinsics::_bitCount_l:               n = new PopCountLNode(          arg);  break;
2392   case vmIntrinsics::_reverseBytes_c:           n = new ReverseBytesUSNode(0,   arg);  break;
2393   case vmIntrinsics::_reverseBytes_s:           n = new ReverseBytesSNode( 0,   arg);  break;
2394   case vmIntrinsics::_reverseBytes_i:           n = new ReverseBytesINode( 0,   arg);  break;
2395   case vmIntrinsics::_reverseBytes_l:           n = new ReverseBytesLNode( 0,   arg);  break;
2396   default:  fatal_unexpected_iid(id);  break;
2397   }
2398   set_result(_gvn.transform(n));
2399   return true;
2400 }
2401 
2402 //----------------------------inline_unsafe_access----------------------------
2403 
2404 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2405 
2406 // Helper that guards and inserts a pre-barrier.
2407 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2408                                         Node* pre_val, bool need_mem_bar) {
2409   // We could be accessing the referent field of a reference object. If so, when G1
2410   // is enabled, we need to log the value in the referent field in an SATB buffer.
2411   // This routine performs some compile time filters and generates suitable
2412   // runtime filters that guard the pre-barrier code.
2413   // Also add memory barrier for non volatile load from the referent field
2414   // to prevent commoning of loads across safepoint.
2415   if (!UseG1GC && !need_mem_bar)
2416     return;
2417 
2418   // Some compile time checks.
2419 
2420   // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2421   const TypeX* otype = offset->find_intptr_t_type();
2422   if (otype != NULL && otype->is_con() &&
2423       otype->get_con() != java_lang_ref_Reference::referent_offset) {
2424     // Constant offset but not the reference_offset so just return
2425     return;
2426   }
2427 
2428   // We only need to generate the runtime guards for instances.
2429   const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2430   if (btype != NULL) {
2431     if (btype->isa_aryptr()) {
2432       // Array type so nothing to do
2433       return;
2434     }
2435 
2436     const TypeInstPtr* itype = btype->isa_instptr();
2437     if (itype != NULL) {
2438       // Can the klass of base_oop be statically determined to be
2439       // _not_ a sub-class of Reference and _not_ Object?
2440       ciKlass* klass = itype->klass();
2441       if ( klass->is_loaded() &&
2442           !klass->is_subtype_of(env()->Reference_klass()) &&
2443           !env()->Object_klass()->is_subtype_of(klass)) {
2444         return;
2445       }
2446     }
2447   }
2448 
2449   // The compile time filters did not reject base_oop/offset so
2450   // we need to generate the following runtime filters
2451   //
2452   // if (offset == java_lang_ref_Reference::_reference_offset) {
2453   //   if (instance_of(base, java.lang.ref.Reference)) {
2454   //     pre_barrier(_, pre_val, ...);
2455   //   }
2456   // }
2457 
2458   float likely   = PROB_LIKELY(  0.999);
2459   float unlikely = PROB_UNLIKELY(0.999);
2460 
2461   IdealKit ideal(this);
2462 #define __ ideal.
2463 
2464   Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2465 
2466   __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2467       // Update graphKit memory and control from IdealKit.
2468       sync_kit(ideal);
2469 
2470       Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2471       Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2472 
2473       // Update IdealKit memory and control from graphKit.
2474       __ sync_kit(this);
2475 
2476       Node* one = __ ConI(1);
2477       // is_instof == 0 if base_oop == NULL
2478       __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2479 
2480         // Update graphKit from IdeakKit.
2481         sync_kit(ideal);
2482 
2483         // Use the pre-barrier to record the value in the referent field
2484         pre_barrier(false /* do_load */,
2485                     __ ctrl(),
2486                     NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2487                     pre_val /* pre_val */,
2488                     T_OBJECT);
2489         if (need_mem_bar) {
2490           // Add memory barrier to prevent commoning reads from this field
2491           // across safepoint since GC can change its value.
2492           insert_mem_bar(Op_MemBarCPUOrder);
2493         }
2494         // Update IdealKit from graphKit.
2495         __ sync_kit(this);
2496 
2497       } __ end_if(); // _ref_type != ref_none
2498   } __ end_if(); // offset == referent_offset
2499 
2500   // Final sync IdealKit and GraphKit.
2501   final_sync(ideal);
2502 #undef __
2503 }
2504 
2505 
2506 // Interpret Unsafe.fieldOffset cookies correctly:
2507 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2508 
2509 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2510   // Attempt to infer a sharper value type from the offset and base type.
2511   ciKlass* sharpened_klass = NULL;
2512 
2513   // See if it is an instance field, with an object type.
2514   if (alias_type->field() != NULL) {
2515     assert(!is_native_ptr, "native pointer op cannot use a java address");
2516     if (alias_type->field()->type()->is_klass()) {
2517       sharpened_klass = alias_type->field()->type()->as_klass();
2518     }
2519   }
2520 
2521   // See if it is a narrow oop array.
2522   if (adr_type->isa_aryptr()) {
2523     if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2524       const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2525       if (elem_type != NULL) {
2526         sharpened_klass = elem_type->klass();
2527       }
2528     }
2529   }
2530 
2531   // The sharpened class might be unloaded if there is no class loader
2532   // contraint in place.
2533   if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2534     const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2535 
2536 #ifndef PRODUCT
2537     if (C->print_intrinsics() || C->print_inlining()) {
2538       tty->print("  from base type: ");  adr_type->dump();
2539       tty->print("  sharpened value: ");  tjp->dump();
2540     }
2541 #endif
2542     // Sharpen the value type.
2543     return tjp;
2544   }
2545   return NULL;
2546 }
2547 
2548 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2549   if (callee()->is_static())  return false;  // caller must have the capability!
2550 
2551 #ifndef PRODUCT
2552   {
2553     ResourceMark rm;
2554     // Check the signatures.
2555     ciSignature* sig = callee()->signature();
2556 #ifdef ASSERT
2557     if (!is_store) {
2558       // Object getObject(Object base, int/long offset), etc.
2559       BasicType rtype = sig->return_type()->basic_type();
2560       if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2561           rtype = T_ADDRESS;  // it is really a C void*
2562       assert(rtype == type, "getter must return the expected value");
2563       if (!is_native_ptr) {
2564         assert(sig->count() == 2, "oop getter has 2 arguments");
2565         assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2566         assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2567       } else {
2568         assert(sig->count() == 1, "native getter has 1 argument");
2569         assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2570       }
2571     } else {
2572       // void putObject(Object base, int/long offset, Object x), etc.
2573       assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2574       if (!is_native_ptr) {
2575         assert(sig->count() == 3, "oop putter has 3 arguments");
2576         assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2577         assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2578       } else {
2579         assert(sig->count() == 2, "native putter has 2 arguments");
2580         assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2581       }
2582       BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2583       if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2584         vtype = T_ADDRESS;  // it is really a C void*
2585       assert(vtype == type, "putter must accept the expected value");
2586     }
2587 #endif // ASSERT
2588  }
2589 #endif //PRODUCT
2590 
2591   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2592 
2593   Node* receiver = argument(0);  // type: oop
2594 
2595   // Build address expression.
2596   Node* adr;
2597   Node* heap_base_oop = top();
2598   Node* offset = top();
2599   Node* val;
2600 
2601   if (!is_native_ptr) {
2602     // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2603     Node* base = argument(1);  // type: oop
2604     // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2605     offset = argument(2);  // type: long
2606     // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2607     // to be plain byte offsets, which are also the same as those accepted
2608     // by oopDesc::field_base.
2609     assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2610            "fieldOffset must be byte-scaled");
2611     // 32-bit machines ignore the high half!
2612     offset = ConvL2X(offset);
2613     adr = make_unsafe_address(base, offset);
2614     heap_base_oop = base;
2615     val = is_store ? argument(4) : NULL;
2616   } else {
2617     Node* ptr = argument(1);  // type: long
2618     ptr = ConvL2X(ptr);  // adjust Java long to machine word
2619     adr = make_unsafe_address(NULL, ptr);
2620     val = is_store ? argument(3) : NULL;
2621   }
2622 
2623   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2624 
2625   // First guess at the value type.
2626   const Type *value_type = Type::get_const_basic_type(type);
2627 
2628   // Try to categorize the address.  If it comes up as TypeJavaPtr::BOTTOM,
2629   // there was not enough information to nail it down.
2630   Compile::AliasType* alias_type = C->alias_type(adr_type);
2631   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2632 
2633   // We will need memory barriers unless we can determine a unique
2634   // alias category for this reference.  (Note:  If for some reason
2635   // the barriers get omitted and the unsafe reference begins to "pollute"
2636   // the alias analysis of the rest of the graph, either Compile::can_alias
2637   // or Compile::must_alias will throw a diagnostic assert.)
2638   bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2639 
2640   // If we are reading the value of the referent field of a Reference
2641   // object (either by using Unsafe directly or through reflection)
2642   // then, if G1 is enabled, we need to record the referent in an
2643   // SATB log buffer using the pre-barrier mechanism.
2644   // Also we need to add memory barrier to prevent commoning reads
2645   // from this field across safepoint since GC can change its value.
2646   bool need_read_barrier = !is_native_ptr && !is_store &&
2647                            offset != top() && heap_base_oop != top();
2648 
2649   if (!is_store && type == T_OBJECT) {
2650     const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2651     if (tjp != NULL) {
2652       value_type = tjp;
2653     }
2654   }
2655 
2656   receiver = null_check(receiver);
2657   if (stopped()) {
2658     return true;
2659   }
2660   // Heap pointers get a null-check from the interpreter,
2661   // as a courtesy.  However, this is not guaranteed by Unsafe,
2662   // and it is not possible to fully distinguish unintended nulls
2663   // from intended ones in this API.
2664 
2665   if (is_volatile) {
2666     // We need to emit leading and trailing CPU membars (see below) in
2667     // addition to memory membars when is_volatile. This is a little
2668     // too strong, but avoids the need to insert per-alias-type
2669     // volatile membars (for stores; compare Parse::do_put_xxx), which
2670     // we cannot do effectively here because we probably only have a
2671     // rough approximation of type.
2672     need_mem_bar = true;
2673     // For Stores, place a memory ordering barrier now.
2674     if (is_store) {
2675       insert_mem_bar(Op_MemBarRelease);
2676     } else {
2677       if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2678         insert_mem_bar(Op_MemBarVolatile);
2679       }
2680     }
2681   }
2682 
2683   // Memory barrier to prevent normal and 'unsafe' accesses from
2684   // bypassing each other.  Happens after null checks, so the
2685   // exception paths do not take memory state from the memory barrier,
2686   // so there's no problems making a strong assert about mixing users
2687   // of safe & unsafe memory.
2688   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2689 
2690   if (!is_store) {
2691     MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2692     // To be valid, unsafe loads may depend on other conditions than
2693     // the one that guards them: pin the Load node
2694     Node* p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2695     // load value
2696     switch (type) {
2697     case T_BOOLEAN:
2698     case T_CHAR:
2699     case T_BYTE:
2700     case T_SHORT:
2701     case T_INT:
2702     case T_LONG:
2703     case T_FLOAT:
2704     case T_DOUBLE:
2705       break;
2706     case T_OBJECT:
2707       if (need_read_barrier) {
2708         insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2709       }
2710       break;
2711     case T_ADDRESS:
2712       // Cast to an int type.
2713       p = _gvn.transform(new CastP2XNode(NULL, p));
2714       p = ConvX2UL(p);
2715       break;
2716     default:
2717       fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2718       break;
2719     }
2720     // The load node has the control of the preceding MemBarCPUOrder.  All
2721     // following nodes will have the control of the MemBarCPUOrder inserted at
2722     // the end of this method.  So, pushing the load onto the stack at a later
2723     // point is fine.
2724     set_result(p);
2725   } else {
2726     // place effect of store into memory
2727     switch (type) {
2728     case T_DOUBLE:
2729       val = dstore_rounding(val);
2730       break;
2731     case T_ADDRESS:
2732       // Repackage the long as a pointer.
2733       val = ConvL2X(val);
2734       val = _gvn.transform(new CastX2PNode(val));
2735       break;
2736     }
2737 
2738     MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2739     if (type != T_OBJECT ) {
2740       (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2741     } else {
2742       // Possibly an oop being stored to Java heap or native memory
2743       if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2744         // oop to Java heap.
2745         (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2746       } else {
2747         // We can't tell at compile time if we are storing in the Java heap or outside
2748         // of it. So we need to emit code to conditionally do the proper type of
2749         // store.
2750 
2751         IdealKit ideal(this);
2752 #define __ ideal.
2753         // QQQ who knows what probability is here??
2754         __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2755           // Sync IdealKit and graphKit.
2756           sync_kit(ideal);
2757           Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2758           // Update IdealKit memory.
2759           __ sync_kit(this);
2760         } __ else_(); {
2761           __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2762         } __ end_if();
2763         // Final sync IdealKit and GraphKit.
2764         final_sync(ideal);
2765 #undef __
2766       }
2767     }
2768   }
2769 
2770   if (is_volatile) {
2771     if (!is_store) {
2772       insert_mem_bar(Op_MemBarAcquire);
2773     } else {
2774       if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2775         insert_mem_bar(Op_MemBarVolatile);
2776       }
2777     }
2778   }
2779 
2780   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2781 
2782   return true;
2783 }
2784 
2785 //----------------------------inline_unsafe_load_store----------------------------
2786 // This method serves a couple of different customers (depending on LoadStoreKind):
2787 //
2788 // LS_cmpxchg:
2789 //   public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2790 //   public final native boolean compareAndSwapInt(   Object o, long offset, int    expected, int    x);
2791 //   public final native boolean compareAndSwapLong(  Object o, long offset, long   expected, long   x);
2792 //
2793 // LS_xadd:
2794 //   public int  getAndAddInt( Object o, long offset, int  delta)
2795 //   public long getAndAddLong(Object o, long offset, long delta)
2796 //
2797 // LS_xchg:
2798 //   int    getAndSet(Object o, long offset, int    newValue)
2799 //   long   getAndSet(Object o, long offset, long   newValue)
2800 //   Object getAndSet(Object o, long offset, Object newValue)
2801 //
2802 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2803   // This basic scheme here is the same as inline_unsafe_access, but
2804   // differs in enough details that combining them would make the code
2805   // overly confusing.  (This is a true fact! I originally combined
2806   // them, but even I was confused by it!) As much code/comments as
2807   // possible are retained from inline_unsafe_access though to make
2808   // the correspondences clearer. - dl
2809 
2810   if (callee()->is_static())  return false;  // caller must have the capability!
2811 
2812 #ifndef PRODUCT
2813   BasicType rtype;
2814   {
2815     ResourceMark rm;
2816     // Check the signatures.
2817     ciSignature* sig = callee()->signature();
2818     rtype = sig->return_type()->basic_type();
2819     if (kind == LS_xadd || kind == LS_xchg) {
2820       // Check the signatures.
2821 #ifdef ASSERT
2822       assert(rtype == type, "get and set must return the expected type");
2823       assert(sig->count() == 3, "get and set has 3 arguments");
2824       assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2825       assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2826       assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2827 #endif // ASSERT
2828     } else if (kind == LS_cmpxchg) {
2829       // Check the signatures.
2830 #ifdef ASSERT
2831       assert(rtype == T_BOOLEAN, "CAS must return boolean");
2832       assert(sig->count() == 4, "CAS has 4 arguments");
2833       assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2834       assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2835 #endif // ASSERT
2836     } else {
2837       ShouldNotReachHere();
2838     }
2839   }
2840 #endif //PRODUCT
2841 
2842   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2843 
2844   // Get arguments:
2845   Node* receiver = NULL;
2846   Node* base     = NULL;
2847   Node* offset   = NULL;
2848   Node* oldval   = NULL;
2849   Node* newval   = NULL;
2850   if (kind == LS_cmpxchg) {
2851     const bool two_slot_type = type2size[type] == 2;
2852     receiver = argument(0);  // type: oop
2853     base     = argument(1);  // type: oop
2854     offset   = argument(2);  // type: long
2855     oldval   = argument(4);  // type: oop, int, or long
2856     newval   = argument(two_slot_type ? 6 : 5);  // type: oop, int, or long
2857   } else if (kind == LS_xadd || kind == LS_xchg){
2858     receiver = argument(0);  // type: oop
2859     base     = argument(1);  // type: oop
2860     offset   = argument(2);  // type: long
2861     oldval   = NULL;
2862     newval   = argument(4);  // type: oop, int, or long
2863   }
2864 
2865   // Null check receiver.
2866   receiver = null_check(receiver);
2867   if (stopped()) {
2868     return true;
2869   }
2870 
2871   // Build field offset expression.
2872   // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2873   // to be plain byte offsets, which are also the same as those accepted
2874   // by oopDesc::field_base.
2875   assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2876   // 32-bit machines ignore the high half of long offsets
2877   offset = ConvL2X(offset);
2878   Node* adr = make_unsafe_address(base, offset);
2879   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2880 
2881   // For CAS, unlike inline_unsafe_access, there seems no point in
2882   // trying to refine types. Just use the coarse types here.
2883   const Type *value_type = Type::get_const_basic_type(type);
2884   Compile::AliasType* alias_type = C->alias_type(adr_type);
2885   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2886 
2887   if (kind == LS_xchg && type == T_OBJECT) {
2888     const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2889     if (tjp != NULL) {
2890       value_type = tjp;
2891     }
2892   }
2893 
2894   int alias_idx = C->get_alias_index(adr_type);
2895 
2896   // Memory-model-wise, a LoadStore acts like a little synchronized
2897   // block, so needs barriers on each side.  These don't translate
2898   // into actual barriers on most machines, but we still need rest of
2899   // compiler to respect ordering.
2900 
2901   insert_mem_bar(Op_MemBarRelease);
2902   insert_mem_bar(Op_MemBarCPUOrder);
2903 
2904   // 4984716: MemBars must be inserted before this
2905   //          memory node in order to avoid a false
2906   //          dependency which will confuse the scheduler.
2907   Node *mem = memory(alias_idx);
2908 
2909   // For now, we handle only those cases that actually exist: ints,
2910   // longs, and Object. Adding others should be straightforward.
2911   Node* load_store;
2912   switch(type) {
2913   case T_INT:
2914     if (kind == LS_xadd) {
2915       load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2916     } else if (kind == LS_xchg) {
2917       load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2918     } else if (kind == LS_cmpxchg) {
2919       load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2920     } else {
2921       ShouldNotReachHere();
2922     }
2923     break;
2924   case T_LONG:
2925     if (kind == LS_xadd) {
2926       load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2927     } else if (kind == LS_xchg) {
2928       load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2929     } else if (kind == LS_cmpxchg) {
2930       load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2931     } else {
2932       ShouldNotReachHere();
2933     }
2934     break;
2935   case T_OBJECT:
2936     // Transformation of a value which could be NULL pointer (CastPP #NULL)
2937     // could be delayed during Parse (for example, in adjust_map_after_if()).
2938     // Execute transformation here to avoid barrier generation in such case.
2939     if (_gvn.type(newval) == TypePtr::NULL_PTR)
2940       newval = _gvn.makecon(TypePtr::NULL_PTR);
2941 
2942     // Reference stores need a store barrier.
2943     if (kind == LS_xchg) {
2944       // If pre-barrier must execute before the oop store, old value will require do_load here.
2945       if (!can_move_pre_barrier()) {
2946         pre_barrier(true /* do_load*/,
2947                     control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2948                     NULL /* pre_val*/,
2949                     T_OBJECT);
2950       } // Else move pre_barrier to use load_store value, see below.
2951     } else if (kind == LS_cmpxchg) {
2952       // Same as for newval above:
2953       if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2954         oldval = _gvn.makecon(TypePtr::NULL_PTR);
2955       }
2956       // The only known value which might get overwritten is oldval.
2957       pre_barrier(false /* do_load */,
2958                   control(), NULL, NULL, max_juint, NULL, NULL,
2959                   oldval /* pre_val */,
2960                   T_OBJECT);
2961     } else {
2962       ShouldNotReachHere();
2963     }
2964 
2965 #ifdef _LP64
2966     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2967       Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2968       if (kind == LS_xchg) {
2969         load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2970                                                        newval_enc, adr_type, value_type->make_narrowoop()));
2971       } else {
2972         assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2973         Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2974         load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2975                                                                 newval_enc, oldval_enc));
2976       }
2977     } else
2978 #endif
2979     {
2980       if (kind == LS_xchg) {
2981         load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2982       } else {
2983         assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2984         load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2985       }
2986     }
2987     post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2988     break;
2989   default:
2990     fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2991     break;
2992   }
2993 
2994   // SCMemProjNodes represent the memory state of a LoadStore. Their
2995   // main role is to prevent LoadStore nodes from being optimized away
2996   // when their results aren't used.
2997   Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2998   set_memory(proj, alias_idx);
2999 
3000   if (type == T_OBJECT && kind == LS_xchg) {
3001 #ifdef _LP64
3002     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3003       load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3004     }
3005 #endif
3006     if (can_move_pre_barrier()) {
3007       // Don't need to load pre_val. The old value is returned by load_store.
3008       // The pre_barrier can execute after the xchg as long as no safepoint
3009       // gets inserted between them.
3010       pre_barrier(false /* do_load */,
3011                   control(), NULL, NULL, max_juint, NULL, NULL,
3012                   load_store /* pre_val */,
3013                   T_OBJECT);
3014     }
3015   }
3016 
3017   // Add the trailing membar surrounding the access
3018   insert_mem_bar(Op_MemBarCPUOrder);
3019   insert_mem_bar(Op_MemBarAcquire);
3020 
3021   assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3022   set_result(load_store);
3023   return true;
3024 }
3025 
3026 //----------------------------inline_unsafe_ordered_store----------------------
3027 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3028 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3029 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3030 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3031   // This is another variant of inline_unsafe_access, differing in
3032   // that it always issues store-store ("release") barrier and ensures
3033   // store-atomicity (which only matters for "long").
3034 
3035   if (callee()->is_static())  return false;  // caller must have the capability!
3036 
3037 #ifndef PRODUCT
3038   {
3039     ResourceMark rm;
3040     // Check the signatures.
3041     ciSignature* sig = callee()->signature();
3042 #ifdef ASSERT
3043     BasicType rtype = sig->return_type()->basic_type();
3044     assert(rtype == T_VOID, "must return void");
3045     assert(sig->count() == 3, "has 3 arguments");
3046     assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3047     assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3048 #endif // ASSERT
3049   }
3050 #endif //PRODUCT
3051 
3052   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
3053 
3054   // Get arguments:
3055   Node* receiver = argument(0);  // type: oop
3056   Node* base     = argument(1);  // type: oop
3057   Node* offset   = argument(2);  // type: long
3058   Node* val      = argument(4);  // type: oop, int, or long
3059 
3060   // Null check receiver.
3061   receiver = null_check(receiver);
3062   if (stopped()) {
3063     return true;
3064   }
3065 
3066   // Build field offset expression.
3067   assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3068   // 32-bit machines ignore the high half of long offsets
3069   offset = ConvL2X(offset);
3070   Node* adr = make_unsafe_address(base, offset);
3071   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3072   const Type *value_type = Type::get_const_basic_type(type);
3073   Compile::AliasType* alias_type = C->alias_type(adr_type);
3074 
3075   insert_mem_bar(Op_MemBarRelease);
3076   insert_mem_bar(Op_MemBarCPUOrder);
3077   // Ensure that the store is atomic for longs:
3078   const bool require_atomic_access = true;
3079   Node* store;
3080   if (type == T_OBJECT) // reference stores need a store barrier.
3081     store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3082   else {
3083     store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3084   }
3085   insert_mem_bar(Op_MemBarCPUOrder);
3086   return true;
3087 }
3088 
3089 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3090   // Regardless of form, don't allow previous ld/st to move down,
3091   // then issue acquire, release, or volatile mem_bar.
3092   insert_mem_bar(Op_MemBarCPUOrder);
3093   switch(id) {
3094     case vmIntrinsics::_loadFence:
3095       insert_mem_bar(Op_LoadFence);
3096       return true;
3097     case vmIntrinsics::_storeFence:
3098       insert_mem_bar(Op_StoreFence);
3099       return true;
3100     case vmIntrinsics::_fullFence:
3101       insert_mem_bar(Op_MemBarVolatile);
3102       return true;
3103     default:
3104       fatal_unexpected_iid(id);
3105       return false;
3106   }
3107 }
3108 
3109 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3110   if (!kls->is_Con()) {
3111     return true;
3112   }
3113   const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3114   if (klsptr == NULL) {
3115     return true;
3116   }
3117   ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3118   // don't need a guard for a klass that is already initialized
3119   return !ik->is_initialized();
3120 }
3121 
3122 //----------------------------inline_unsafe_allocate---------------------------
3123 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3124 bool LibraryCallKit::inline_unsafe_allocate() {
3125   if (callee()->is_static())  return false;  // caller must have the capability!
3126 
3127   null_check_receiver();  // null-check, then ignore
3128   Node* cls = null_check(argument(1));
3129   if (stopped())  return true;
3130 
3131   Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3132   kls = null_check(kls);
3133   if (stopped())  return true;  // argument was like int.class
3134 
3135   Node* test = NULL;
3136   if (LibraryCallKit::klass_needs_init_guard(kls)) {
3137     // Note:  The argument might still be an illegal value like
3138     // Serializable.class or Object[].class.   The runtime will handle it.
3139     // But we must make an explicit check for initialization.
3140     Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3141     // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3142     // can generate code to load it as unsigned byte.
3143     Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3144     Node* bits = intcon(InstanceKlass::fully_initialized);
3145     test = _gvn.transform(new SubINode(inst, bits));
3146     // The 'test' is non-zero if we need to take a slow path.
3147   }
3148 
3149   Node* obj = new_instance(kls, test);
3150   set_result(obj);
3151   return true;
3152 }
3153 
3154 #ifdef TRACE_HAVE_INTRINSICS
3155 /*
3156  * oop -> myklass
3157  * myklass->trace_id |= USED
3158  * return myklass->trace_id & ~0x3
3159  */
3160 bool LibraryCallKit::inline_native_classID() {
3161   null_check_receiver();  // null-check, then ignore
3162   Node* cls = null_check(argument(1), T_OBJECT);
3163   Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3164   kls = null_check(kls, T_OBJECT);
3165   ByteSize offset = TRACE_ID_OFFSET;
3166   Node* insp = basic_plus_adr(kls, in_bytes(offset));
3167   Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3168   Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3169   Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3170   Node* clsused = longcon(0x01l); // set the class bit
3171   Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3172 
3173   const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3174   store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3175   set_result(andl);
3176   return true;
3177 }
3178 
3179 bool LibraryCallKit::inline_native_threadID() {
3180   Node* tls_ptr = NULL;
3181   Node* cur_thr = generate_current_thread(tls_ptr);
3182   Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3183   Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3184   p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3185 
3186   Node* threadid = NULL;
3187   size_t thread_id_size = OSThread::thread_id_size();
3188   if (thread_id_size == (size_t) BytesPerLong) {
3189     threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3190   } else if (thread_id_size == (size_t) BytesPerInt) {
3191     threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3192   } else {
3193     ShouldNotReachHere();
3194   }
3195   set_result(threadid);
3196   return true;
3197 }
3198 #endif
3199 
3200 //------------------------inline_native_time_funcs--------------
3201 // inline code for System.currentTimeMillis() and System.nanoTime()
3202 // these have the same type and signature
3203 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3204   const TypeFunc* tf = OptoRuntime::void_long_Type();
3205   const TypePtr* no_memory_effects = NULL;
3206   Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3207   Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3208 #ifdef ASSERT
3209   Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3210   assert(value_top == top(), "second value must be top");
3211 #endif
3212   set_result(value);
3213   return true;
3214 }
3215 
3216 //------------------------inline_native_currentThread------------------
3217 bool LibraryCallKit::inline_native_currentThread() {
3218   Node* junk = NULL;
3219   set_result(generate_current_thread(junk));
3220   return true;
3221 }
3222 
3223 //------------------------inline_native_isInterrupted------------------
3224 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3225 bool LibraryCallKit::inline_native_isInterrupted() {
3226   // Add a fast path to t.isInterrupted(clear_int):
3227   //   (t == Thread.current() &&
3228   //    (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3229   //   ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3230   // So, in the common case that the interrupt bit is false,
3231   // we avoid making a call into the VM.  Even if the interrupt bit
3232   // is true, if the clear_int argument is false, we avoid the VM call.
3233   // However, if the receiver is not currentThread, we must call the VM,
3234   // because there must be some locking done around the operation.
3235 
3236   // We only go to the fast case code if we pass two guards.
3237   // Paths which do not pass are accumulated in the slow_region.
3238 
3239   enum {
3240     no_int_result_path   = 1, // t == Thread.current() && !TLS._osthread._interrupted
3241     no_clear_result_path = 2, // t == Thread.current() &&  TLS._osthread._interrupted && !clear_int
3242     slow_result_path     = 3, // slow path: t.isInterrupted(clear_int)
3243     PATH_LIMIT
3244   };
3245 
3246   // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3247   // out of the function.
3248   insert_mem_bar(Op_MemBarCPUOrder);
3249 
3250   RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3251   PhiNode*    result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3252 
3253   RegionNode* slow_region = new RegionNode(1);
3254   record_for_igvn(slow_region);
3255 
3256   // (a) Receiving thread must be the current thread.
3257   Node* rec_thr = argument(0);
3258   Node* tls_ptr = NULL;
3259   Node* cur_thr = generate_current_thread(tls_ptr);
3260   Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3261   Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3262 
3263   generate_slow_guard(bol_thr, slow_region);
3264 
3265   // (b) Interrupt bit on TLS must be false.
3266   Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3267   Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3268   p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3269 
3270   // Set the control input on the field _interrupted read to prevent it floating up.
3271   Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3272   Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3273   Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3274 
3275   IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3276 
3277   // First fast path:  if (!TLS._interrupted) return false;
3278   Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3279   result_rgn->init_req(no_int_result_path, false_bit);
3280   result_val->init_req(no_int_result_path, intcon(0));
3281 
3282   // drop through to next case
3283   set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3284 
3285 #ifndef TARGET_OS_FAMILY_windows
3286   // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3287   Node* clr_arg = argument(1);
3288   Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3289   Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3290   IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3291 
3292   // Second fast path:  ... else if (!clear_int) return true;
3293   Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3294   result_rgn->init_req(no_clear_result_path, false_arg);
3295   result_val->init_req(no_clear_result_path, intcon(1));
3296 
3297   // drop through to next case
3298   set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3299 #else
3300   // To return true on Windows you must read the _interrupted field
3301   // and check the the event state i.e. take the slow path.
3302 #endif // TARGET_OS_FAMILY_windows
3303 
3304   // (d) Otherwise, go to the slow path.
3305   slow_region->add_req(control());
3306   set_control( _gvn.transform(slow_region));
3307 
3308   if (stopped()) {
3309     // There is no slow path.
3310     result_rgn->init_req(slow_result_path, top());
3311     result_val->init_req(slow_result_path, top());
3312   } else {
3313     // non-virtual because it is a private non-static
3314     CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3315 
3316     Node* slow_val = set_results_for_java_call(slow_call);
3317     // this->control() comes from set_results_for_java_call
3318 
3319     Node* fast_io  = slow_call->in(TypeFunc::I_O);
3320     Node* fast_mem = slow_call->in(TypeFunc::Memory);
3321 
3322     // These two phis are pre-filled with copies of of the fast IO and Memory
3323     PhiNode* result_mem  = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3324     PhiNode* result_io   = PhiNode::make(result_rgn, fast_io,  Type::ABIO);
3325 
3326     result_rgn->init_req(slow_result_path, control());
3327     result_io ->init_req(slow_result_path, i_o());
3328     result_mem->init_req(slow_result_path, reset_memory());
3329     result_val->init_req(slow_result_path, slow_val);
3330 
3331     set_all_memory(_gvn.transform(result_mem));
3332     set_i_o(       _gvn.transform(result_io));
3333   }
3334 
3335   C->set_has_split_ifs(true); // Has chance for split-if optimization
3336   set_result(result_rgn, result_val);
3337   return true;
3338 }
3339 
3340 //---------------------------load_mirror_from_klass----------------------------
3341 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3342 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3343   Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3344   return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3345 }
3346 
3347 //-----------------------load_klass_from_mirror_common-------------------------
3348 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3349 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3350 // and branch to the given path on the region.
3351 // If never_see_null, take an uncommon trap on null, so we can optimistically
3352 // compile for the non-null case.
3353 // If the region is NULL, force never_see_null = true.
3354 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3355                                                     bool never_see_null,
3356                                                     RegionNode* region,
3357                                                     int null_path,
3358                                                     int offset) {
3359   if (region == NULL)  never_see_null = true;
3360   Node* p = basic_plus_adr(mirror, offset);
3361   const TypeKlassPtr*  kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3362   Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3363   Node* null_ctl = top();
3364   kls = null_check_oop(kls, &null_ctl, never_see_null);
3365   if (region != NULL) {
3366     // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3367     region->init_req(null_path, null_ctl);
3368   } else {
3369     assert(null_ctl == top(), "no loose ends");
3370   }
3371   return kls;
3372 }
3373 
3374 //--------------------(inline_native_Class_query helpers)---------------------
3375 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3376 // Fall through if (mods & mask) == bits, take the guard otherwise.
3377 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3378   // Branch around if the given klass has the given modifier bit set.
3379   // Like generate_guard, adds a new path onto the region.
3380   Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3381   Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3382   Node* mask = intcon(modifier_mask);
3383   Node* bits = intcon(modifier_bits);
3384   Node* mbit = _gvn.transform(new AndINode(mods, mask));
3385   Node* cmp  = _gvn.transform(new CmpINode(mbit, bits));
3386   Node* bol  = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3387   return generate_fair_guard(bol, region);
3388 }
3389 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3390   return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3391 }
3392 
3393 //-------------------------inline_native_Class_query-------------------
3394 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3395   const Type* return_type = TypeInt::BOOL;
3396   Node* prim_return_value = top();  // what happens if it's a primitive class?
3397   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3398   bool expect_prim = false;     // most of these guys expect to work on refs
3399 
3400   enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3401 
3402   Node* mirror = argument(0);
3403   Node* obj    = top();
3404 
3405   switch (id) {
3406   case vmIntrinsics::_isInstance:
3407     // nothing is an instance of a primitive type
3408     prim_return_value = intcon(0);
3409     obj = argument(1);
3410     break;
3411   case vmIntrinsics::_getModifiers:
3412     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3413     assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3414     return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3415     break;
3416   case vmIntrinsics::_isInterface:
3417     prim_return_value = intcon(0);
3418     break;
3419   case vmIntrinsics::_isArray:
3420     prim_return_value = intcon(0);
3421     expect_prim = true;  // cf. ObjectStreamClass.getClassSignature
3422     break;
3423   case vmIntrinsics::_isPrimitive:
3424     prim_return_value = intcon(1);
3425     expect_prim = true;  // obviously
3426     break;
3427   case vmIntrinsics::_getSuperclass:
3428     prim_return_value = null();
3429     return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3430     break;
3431   case vmIntrinsics::_getClassAccessFlags:
3432     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3433     return_type = TypeInt::INT;  // not bool!  6297094
3434     break;
3435   default:
3436     fatal_unexpected_iid(id);
3437     break;
3438   }
3439 
3440   const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3441   if (mirror_con == NULL)  return false;  // cannot happen?
3442 
3443 #ifndef PRODUCT
3444   if (C->print_intrinsics() || C->print_inlining()) {
3445     ciType* k = mirror_con->java_mirror_type();
3446     if (k) {
3447       tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3448       k->print_name();
3449       tty->cr();
3450     }
3451   }
3452 #endif
3453 
3454   // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3455   RegionNode* region = new RegionNode(PATH_LIMIT);
3456   record_for_igvn(region);
3457   PhiNode* phi = new PhiNode(region, return_type);
3458 
3459   // The mirror will never be null of Reflection.getClassAccessFlags, however
3460   // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3461   // if it is. See bug 4774291.
3462 
3463   // For Reflection.getClassAccessFlags(), the null check occurs in
3464   // the wrong place; see inline_unsafe_access(), above, for a similar
3465   // situation.
3466   mirror = null_check(mirror);
3467   // If mirror or obj is dead, only null-path is taken.
3468   if (stopped())  return true;
3469 
3470   if (expect_prim)  never_see_null = false;  // expect nulls (meaning prims)
3471 
3472   // Now load the mirror's klass metaobject, and null-check it.
3473   // Side-effects region with the control path if the klass is null.
3474   Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3475   // If kls is null, we have a primitive mirror.
3476   phi->init_req(_prim_path, prim_return_value);
3477   if (stopped()) { set_result(region, phi); return true; }
3478   bool safe_for_replace = (region->in(_prim_path) == top());
3479 
3480   Node* p;  // handy temp
3481   Node* null_ctl;
3482 
3483   // Now that we have the non-null klass, we can perform the real query.
3484   // For constant classes, the query will constant-fold in LoadNode::Value.
3485   Node* query_value = top();
3486   switch (id) {
3487   case vmIntrinsics::_isInstance:
3488     // nothing is an instance of a primitive type
3489     query_value = gen_instanceof(obj, kls, safe_for_replace);
3490     break;
3491 
3492   case vmIntrinsics::_getModifiers:
3493     p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3494     query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3495     break;
3496 
3497   case vmIntrinsics::_isInterface:
3498     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3499     if (generate_interface_guard(kls, region) != NULL)
3500       // A guard was added.  If the guard is taken, it was an interface.
3501       phi->add_req(intcon(1));
3502     // If we fall through, it's a plain class.
3503     query_value = intcon(0);
3504     break;
3505 
3506   case vmIntrinsics::_isArray:
3507     // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3508     if (generate_array_guard(kls, region) != NULL)
3509       // A guard was added.  If the guard is taken, it was an array.
3510       phi->add_req(intcon(1));
3511     // If we fall through, it's a plain class.
3512     query_value = intcon(0);
3513     break;
3514 
3515   case vmIntrinsics::_isPrimitive:
3516     query_value = intcon(0); // "normal" path produces false
3517     break;
3518 
3519   case vmIntrinsics::_getSuperclass:
3520     // The rules here are somewhat unfortunate, but we can still do better
3521     // with random logic than with a JNI call.
3522     // Interfaces store null or Object as _super, but must report null.
3523     // Arrays store an intermediate super as _super, but must report Object.
3524     // Other types can report the actual _super.
3525     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3526     if (generate_interface_guard(kls, region) != NULL)
3527       // A guard was added.  If the guard is taken, it was an interface.
3528       phi->add_req(null());
3529     if (generate_array_guard(kls, region) != NULL)
3530       // A guard was added.  If the guard is taken, it was an array.
3531       phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3532     // If we fall through, it's a plain class.  Get its _super.
3533     p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3534     kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3535     null_ctl = top();
3536     kls = null_check_oop(kls, &null_ctl);
3537     if (null_ctl != top()) {
3538       // If the guard is taken, Object.superClass is null (both klass and mirror).
3539       region->add_req(null_ctl);
3540       phi   ->add_req(null());
3541     }
3542     if (!stopped()) {
3543       query_value = load_mirror_from_klass(kls);
3544     }
3545     break;
3546 
3547   case vmIntrinsics::_getClassAccessFlags:
3548     p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3549     query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3550     break;
3551 
3552   default:
3553     fatal_unexpected_iid(id);
3554     break;
3555   }
3556 
3557   // Fall-through is the normal case of a query to a real class.
3558   phi->init_req(1, query_value);
3559   region->init_req(1, control());
3560 
3561   C->set_has_split_ifs(true); // Has chance for split-if optimization
3562   set_result(region, phi);
3563   return true;
3564 }
3565 
3566 //-------------------------inline_Class_cast-------------------
3567 bool LibraryCallKit::inline_Class_cast() {
3568   Node* mirror = argument(0); // Class
3569   Node* obj    = argument(1);
3570   const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3571   if (mirror_con == NULL) {
3572     return false;  // dead path (mirror->is_top()).
3573   }
3574   if (obj == NULL || obj->is_top()) {
3575     return false;  // dead path
3576   }
3577   const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3578 
3579   // First, see if Class.cast() can be folded statically.
3580   // java_mirror_type() returns non-null for compile-time Class constants.
3581   ciType* tm = mirror_con->java_mirror_type();
3582   if (tm != NULL && tm->is_klass() &&
3583       tp != NULL && tp->klass() != NULL) {
3584     if (!tp->klass()->is_loaded()) {
3585       // Don't use intrinsic when class is not loaded.
3586       return false;
3587     } else {
3588       int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3589       if (static_res == Compile::SSC_always_true) {
3590         // isInstance() is true - fold the code.
3591         set_result(obj);
3592         return true;
3593       } else if (static_res == Compile::SSC_always_false) {
3594         // Don't use intrinsic, have to throw ClassCastException.
3595         // If the reference is null, the non-intrinsic bytecode will
3596         // be optimized appropriately.
3597         return false;
3598       }
3599     }
3600   }
3601 
3602   // Bailout intrinsic and do normal inlining if exception path is frequent.
3603   if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3604     return false;
3605   }
3606 
3607   // Generate dynamic checks.
3608   // Class.cast() is java implementation of _checkcast bytecode.
3609   // Do checkcast (Parse::do_checkcast()) optimizations here.
3610 
3611   mirror = null_check(mirror);
3612   // If mirror is dead, only null-path is taken.
3613   if (stopped()) {
3614     return true;
3615   }
3616 
3617   // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3618   enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3619   RegionNode* region = new RegionNode(PATH_LIMIT);
3620   record_for_igvn(region);
3621 
3622   // Now load the mirror's klass metaobject, and null-check it.
3623   // If kls is null, we have a primitive mirror and
3624   // nothing is an instance of a primitive type.
3625   Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3626 
3627   Node* res = top();
3628   if (!stopped()) {
3629     Node* bad_type_ctrl = top();
3630     // Do checkcast optimizations.
3631     res = gen_checkcast(obj, kls, &bad_type_ctrl);
3632     region->init_req(_bad_type_path, bad_type_ctrl);
3633   }
3634   if (region->in(_prim_path) != top() ||
3635       region->in(_bad_type_path) != top()) {
3636     // Let Interpreter throw ClassCastException.
3637     PreserveJVMState pjvms(this);
3638     set_control(_gvn.transform(region));
3639     uncommon_trap(Deoptimization::Reason_intrinsic,
3640                   Deoptimization::Action_maybe_recompile);
3641   }
3642   if (!stopped()) {
3643     set_result(res);
3644   }
3645   return true;
3646 }
3647 
3648 
3649 //--------------------------inline_native_subtype_check------------------------
3650 // This intrinsic takes the JNI calls out of the heart of
3651 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3652 bool LibraryCallKit::inline_native_subtype_check() {
3653   // Pull both arguments off the stack.
3654   Node* args[2];                // two java.lang.Class mirrors: superc, subc
3655   args[0] = argument(0);
3656   args[1] = argument(1);
3657   Node* klasses[2];             // corresponding Klasses: superk, subk
3658   klasses[0] = klasses[1] = top();
3659 
3660   enum {
3661     // A full decision tree on {superc is prim, subc is prim}:
3662     _prim_0_path = 1,           // {P,N} => false
3663                                 // {P,P} & superc!=subc => false
3664     _prim_same_path,            // {P,P} & superc==subc => true
3665     _prim_1_path,               // {N,P} => false
3666     _ref_subtype_path,          // {N,N} & subtype check wins => true
3667     _both_ref_path,             // {N,N} & subtype check loses => false
3668     PATH_LIMIT
3669   };
3670 
3671   RegionNode* region = new RegionNode(PATH_LIMIT);
3672   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
3673   record_for_igvn(region);
3674 
3675   const TypePtr* adr_type = TypeRawPtr::BOTTOM;   // memory type of loads
3676   const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3677   int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3678 
3679   // First null-check both mirrors and load each mirror's klass metaobject.
3680   int which_arg;
3681   for (which_arg = 0; which_arg <= 1; which_arg++) {
3682     Node* arg = args[which_arg];
3683     arg = null_check(arg);
3684     if (stopped())  break;
3685     args[which_arg] = arg;
3686 
3687     Node* p = basic_plus_adr(arg, class_klass_offset);
3688     Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3689     klasses[which_arg] = _gvn.transform(kls);
3690   }
3691 
3692   // Having loaded both klasses, test each for null.
3693   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3694   for (which_arg = 0; which_arg <= 1; which_arg++) {
3695     Node* kls = klasses[which_arg];
3696     Node* null_ctl = top();
3697     kls = null_check_oop(kls, &null_ctl, never_see_null);
3698     int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3699     region->init_req(prim_path, null_ctl);
3700     if (stopped())  break;
3701     klasses[which_arg] = kls;
3702   }
3703 
3704   if (!stopped()) {
3705     // now we have two reference types, in klasses[0..1]
3706     Node* subk   = klasses[1];  // the argument to isAssignableFrom
3707     Node* superk = klasses[0];  // the receiver
3708     region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3709     // now we have a successful reference subtype check
3710     region->set_req(_ref_subtype_path, control());
3711   }
3712 
3713   // If both operands are primitive (both klasses null), then
3714   // we must return true when they are identical primitives.
3715   // It is convenient to test this after the first null klass check.
3716   set_control(region->in(_prim_0_path)); // go back to first null check
3717   if (!stopped()) {
3718     // Since superc is primitive, make a guard for the superc==subc case.
3719     Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3720     Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3721     generate_guard(bol_eq, region, PROB_FAIR);
3722     if (region->req() == PATH_LIMIT+1) {
3723       // A guard was added.  If the added guard is taken, superc==subc.
3724       region->swap_edges(PATH_LIMIT, _prim_same_path);
3725       region->del_req(PATH_LIMIT);
3726     }
3727     region->set_req(_prim_0_path, control()); // Not equal after all.
3728   }
3729 
3730   // these are the only paths that produce 'true':
3731   phi->set_req(_prim_same_path,   intcon(1));
3732   phi->set_req(_ref_subtype_path, intcon(1));
3733 
3734   // pull together the cases:
3735   assert(region->req() == PATH_LIMIT, "sane region");
3736   for (uint i = 1; i < region->req(); i++) {
3737     Node* ctl = region->in(i);
3738     if (ctl == NULL || ctl == top()) {
3739       region->set_req(i, top());
3740       phi   ->set_req(i, top());
3741     } else if (phi->in(i) == NULL) {
3742       phi->set_req(i, intcon(0)); // all other paths produce 'false'
3743     }
3744   }
3745 
3746   set_control(_gvn.transform(region));
3747   set_result(_gvn.transform(phi));
3748   return true;
3749 }
3750 
3751 //---------------------generate_array_guard_common------------------------
3752 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3753                                                   bool obj_array, bool not_array) {
3754 
3755   if (stopped()) {
3756     return NULL;
3757   }
3758 
3759   // If obj_array/non_array==false/false:
3760   // Branch around if the given klass is in fact an array (either obj or prim).
3761   // If obj_array/non_array==false/true:
3762   // Branch around if the given klass is not an array klass of any kind.
3763   // If obj_array/non_array==true/true:
3764   // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3765   // If obj_array/non_array==true/false:
3766   // Branch around if the kls is an oop array (Object[] or subtype)
3767   //
3768   // Like generate_guard, adds a new path onto the region.
3769   jint  layout_con = 0;
3770   Node* layout_val = get_layout_helper(kls, layout_con);
3771   if (layout_val == NULL) {
3772     bool query = (obj_array
3773                   ? Klass::layout_helper_is_objArray(layout_con)
3774                   : Klass::layout_helper_is_array(layout_con));
3775     if (query == not_array) {
3776       return NULL;                       // never a branch
3777     } else {                             // always a branch
3778       Node* always_branch = control();
3779       if (region != NULL)
3780         region->add_req(always_branch);
3781       set_control(top());
3782       return always_branch;
3783     }
3784   }
3785   // Now test the correct condition.
3786   jint  nval = (obj_array
3787                 ? ((jint)Klass::_lh_array_tag_type_value
3788                    <<    Klass::_lh_array_tag_shift)
3789                 : Klass::_lh_neutral_value);
3790   Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3791   BoolTest::mask btest = BoolTest::lt;  // correct for testing is_[obj]array
3792   // invert the test if we are looking for a non-array
3793   if (not_array)  btest = BoolTest(btest).negate();
3794   Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3795   return generate_fair_guard(bol, region);
3796 }
3797 
3798 
3799 //-----------------------inline_native_newArray--------------------------
3800 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3801 bool LibraryCallKit::inline_native_newArray() {
3802   Node* mirror    = argument(0);
3803   Node* count_val = argument(1);
3804 
3805   mirror = null_check(mirror);
3806   // If mirror or obj is dead, only null-path is taken.
3807   if (stopped())  return true;
3808 
3809   enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3810   RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3811   PhiNode*    result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3812   PhiNode*    result_io  = new PhiNode(result_reg, Type::ABIO);
3813   PhiNode*    result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3814 
3815   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3816   Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3817                                                   result_reg, _slow_path);
3818   Node* normal_ctl   = control();
3819   Node* no_array_ctl = result_reg->in(_slow_path);
3820 
3821   // Generate code for the slow case.  We make a call to newArray().
3822   set_control(no_array_ctl);
3823   if (!stopped()) {
3824     // Either the input type is void.class, or else the
3825     // array klass has not yet been cached.  Either the
3826     // ensuing call will throw an exception, or else it
3827     // will cache the array klass for next time.
3828     PreserveJVMState pjvms(this);
3829     CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3830     Node* slow_result = set_results_for_java_call(slow_call);
3831     // this->control() comes from set_results_for_java_call
3832     result_reg->set_req(_slow_path, control());
3833     result_val->set_req(_slow_path, slow_result);
3834     result_io ->set_req(_slow_path, i_o());
3835     result_mem->set_req(_slow_path, reset_memory());
3836   }
3837 
3838   set_control(normal_ctl);
3839   if (!stopped()) {
3840     // Normal case:  The array type has been cached in the java.lang.Class.
3841     // The following call works fine even if the array type is polymorphic.
3842     // It could be a dynamic mix of int[], boolean[], Object[], etc.
3843     Node* obj = new_array(klass_node, count_val, 0);  // no arguments to push
3844     result_reg->init_req(_normal_path, control());
3845     result_val->init_req(_normal_path, obj);
3846     result_io ->init_req(_normal_path, i_o());
3847     result_mem->init_req(_normal_path, reset_memory());
3848   }
3849 
3850   // Return the combined state.
3851   set_i_o(        _gvn.transform(result_io)  );
3852   set_all_memory( _gvn.transform(result_mem));
3853 
3854   C->set_has_split_ifs(true); // Has chance for split-if optimization
3855   set_result(result_reg, result_val);
3856   return true;
3857 }
3858 
3859 //----------------------inline_native_getLength--------------------------
3860 // public static native int java.lang.reflect.Array.getLength(Object array);
3861 bool LibraryCallKit::inline_native_getLength() {
3862   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3863 
3864   Node* array = null_check(argument(0));
3865   // If array is dead, only null-path is taken.
3866   if (stopped())  return true;
3867 
3868   // Deoptimize if it is a non-array.
3869   Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3870 
3871   if (non_array != NULL) {
3872     PreserveJVMState pjvms(this);
3873     set_control(non_array);
3874     uncommon_trap(Deoptimization::Reason_intrinsic,
3875                   Deoptimization::Action_maybe_recompile);
3876   }
3877 
3878   // If control is dead, only non-array-path is taken.
3879   if (stopped())  return true;
3880 
3881   // The works fine even if the array type is polymorphic.
3882   // It could be a dynamic mix of int[], boolean[], Object[], etc.
3883   Node* result = load_array_length(array);
3884 
3885   C->set_has_split_ifs(true);  // Has chance for split-if optimization
3886   set_result(result);
3887   return true;
3888 }
3889 
3890 //------------------------inline_array_copyOf----------------------------
3891 // public static <T,U> T[] java.util.Arrays.copyOf(     U[] original, int newLength,         Class<? extends T[]> newType);
3892 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from,      int to, Class<? extends T[]> newType);
3893 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3894   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3895 
3896   // Get the arguments.
3897   Node* original          = argument(0);
3898   Node* start             = is_copyOfRange? argument(1): intcon(0);
3899   Node* end               = is_copyOfRange? argument(2): argument(1);
3900   Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3901 
3902   Node* newcopy;
3903 
3904   // Set the original stack and the reexecute bit for the interpreter to reexecute
3905   // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3906   { PreserveReexecuteState preexecs(this);
3907     jvms()->set_should_reexecute(true);
3908 
3909     array_type_mirror = null_check(array_type_mirror);
3910     original          = null_check(original);
3911 
3912     // Check if a null path was taken unconditionally.
3913     if (stopped())  return true;
3914 
3915     Node* orig_length = load_array_length(original);
3916 
3917     Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3918     klass_node = null_check(klass_node);
3919 
3920     RegionNode* bailout = new RegionNode(1);
3921     record_for_igvn(bailout);
3922 
3923     // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3924     // Bail out if that is so.
3925     Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3926     if (not_objArray != NULL) {
3927       // Improve the klass node's type from the new optimistic assumption:
3928       ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3929       const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3930       Node* cast = new CastPPNode(klass_node, akls);
3931       cast->init_req(0, control());
3932       klass_node = _gvn.transform(cast);
3933     }
3934 
3935     // Bail out if either start or end is negative.
3936     generate_negative_guard(start, bailout, &start);
3937     generate_negative_guard(end,   bailout, &end);
3938 
3939     Node* length = end;
3940     if (_gvn.type(start) != TypeInt::ZERO) {
3941       length = _gvn.transform(new SubINode(end, start));
3942     }
3943 
3944     // Bail out if length is negative.
3945     // Without this the new_array would throw
3946     // NegativeArraySizeException but IllegalArgumentException is what
3947     // should be thrown
3948     generate_negative_guard(length, bailout, &length);
3949 
3950     if (bailout->req() > 1) {
3951       PreserveJVMState pjvms(this);
3952       set_control(_gvn.transform(bailout));
3953       uncommon_trap(Deoptimization::Reason_intrinsic,
3954                     Deoptimization::Action_maybe_recompile);
3955     }
3956 
3957     if (!stopped()) {
3958       // How many elements will we copy from the original?
3959       // The answer is MinI(orig_length - start, length).
3960       Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3961       Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3962 
3963       // Generate a direct call to the right arraycopy function(s).
3964       // We know the copy is disjoint but we might not know if the
3965       // oop stores need checking.
3966       // Extreme case:  Arrays.copyOf((Integer[])x, 10, String[].class).
3967       // This will fail a store-check if x contains any non-nulls.
3968 
3969       // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3970       // loads/stores but it is legal only if we're sure the
3971       // Arrays.copyOf would succeed. So we need all input arguments
3972       // to the copyOf to be validated, including that the copy to the
3973       // new array won't trigger an ArrayStoreException. That subtype
3974       // check can be optimized if we know something on the type of
3975       // the input array from type speculation.
3976       if (_gvn.type(klass_node)->singleton()) {
3977         ciKlass* subk   = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3978         ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3979 
3980         int test = C->static_subtype_check(superk, subk);
3981         if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3982           const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3983           if (t_original->speculative_type() != NULL) {
3984             original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3985           }
3986         }
3987       }
3988 
3989       bool validated = false;
3990       // Reason_class_check rather than Reason_intrinsic because we
3991       // want to intrinsify even if this traps.
3992       if (!too_many_traps(Deoptimization::Reason_class_check)) {
3993         Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3994                                                    klass_node);
3995 
3996         if (not_subtype_ctrl != top()) {
3997           PreserveJVMState pjvms(this);
3998           set_control(not_subtype_ctrl);
3999           uncommon_trap(Deoptimization::Reason_class_check,
4000                         Deoptimization::Action_make_not_entrant);
4001           assert(stopped(), "Should be stopped");
4002         }
4003         validated = true;
4004       }
4005 
4006       if (!stopped()) {
4007         newcopy = new_array(klass_node, length, 0);  // no arguments to push
4008 
4009         ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
4010                                                 load_object_klass(original), klass_node);
4011         if (!is_copyOfRange) {
4012           ac->set_copyof(validated);
4013         } else {
4014           ac->set_copyofrange(validated);
4015         }
4016         Node* n = _gvn.transform(ac);
4017         if (n == ac) {
4018           ac->connect_outputs(this);
4019         } else {
4020           assert(validated, "shouldn't transform if all arguments not validated");
4021           set_all_memory(n);
4022         }
4023       }
4024     }
4025   } // original reexecute is set back here
4026 
4027   C->set_has_split_ifs(true); // Has chance for split-if optimization
4028   if (!stopped()) {
4029     set_result(newcopy);
4030   }
4031   return true;
4032 }
4033 
4034 
4035 //----------------------generate_virtual_guard---------------------------
4036 // Helper for hashCode and clone.  Peeks inside the vtable to avoid a call.
4037 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4038                                              RegionNode* slow_region) {
4039   ciMethod* method = callee();
4040   int vtable_index = method->vtable_index();
4041   assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4042          err_msg_res("bad index %d", vtable_index));
4043   // Get the Method* out of the appropriate vtable entry.
4044   int entry_offset  = (InstanceKlass::vtable_start_offset() +
4045                      vtable_index*vtableEntry::size()) * wordSize +
4046                      vtableEntry::method_offset_in_bytes();
4047   Node* entry_addr  = basic_plus_adr(obj_klass, entry_offset);
4048   Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4049 
4050   // Compare the target method with the expected method (e.g., Object.hashCode).
4051   const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4052 
4053   Node* native_call = makecon(native_call_addr);
4054   Node* chk_native  = _gvn.transform(new CmpPNode(target_call, native_call));
4055   Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4056 
4057   return generate_slow_guard(test_native, slow_region);
4058 }
4059 
4060 //-----------------------generate_method_call----------------------------
4061 // Use generate_method_call to make a slow-call to the real
4062 // method if the fast path fails.  An alternative would be to
4063 // use a stub like OptoRuntime::slow_arraycopy_Java.
4064 // This only works for expanding the current library call,
4065 // not another intrinsic.  (E.g., don't use this for making an
4066 // arraycopy call inside of the copyOf intrinsic.)
4067 CallJavaNode*
4068 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4069   // When compiling the intrinsic method itself, do not use this technique.
4070   guarantee(callee() != C->method(), "cannot make slow-call to self");
4071 
4072   ciMethod* method = callee();
4073   // ensure the JVMS we have will be correct for this call
4074   guarantee(method_id == method->intrinsic_id(), "must match");
4075 
4076   const TypeFunc* tf = TypeFunc::make(method);
4077   CallJavaNode* slow_call;
4078   if (is_static) {
4079     assert(!is_virtual, "");
4080     slow_call = new CallStaticJavaNode(C, tf,
4081                            SharedRuntime::get_resolve_static_call_stub(),
4082                            method, bci());
4083   } else if (is_virtual) {
4084     null_check_receiver();
4085     int vtable_index = Method::invalid_vtable_index;
4086     if (UseInlineCaches) {
4087       // Suppress the vtable call
4088     } else {
4089       // hashCode and clone are not a miranda methods,
4090       // so the vtable index is fixed.
4091       // No need to use the linkResolver to get it.
4092        vtable_index = method->vtable_index();
4093        assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4094               err_msg_res("bad index %d", vtable_index));
4095     }
4096     slow_call = new CallDynamicJavaNode(tf,
4097                           SharedRuntime::get_resolve_virtual_call_stub(),
4098                           method, vtable_index, bci());
4099   } else {  // neither virtual nor static:  opt_virtual
4100     null_check_receiver();
4101     slow_call = new CallStaticJavaNode(C, tf,
4102                                 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4103                                 method, bci());
4104     slow_call->set_optimized_virtual(true);
4105   }
4106   set_arguments_for_java_call(slow_call);
4107   set_edges_for_java_call(slow_call);
4108   return slow_call;
4109 }
4110 
4111 
4112 /**
4113  * Build special case code for calls to hashCode on an object. This call may
4114  * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4115  * slightly different code.
4116  */
4117 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4118   assert(is_static == callee()->is_static(), "correct intrinsic selection");
4119   assert(!(is_virtual && is_static), "either virtual, special, or static");
4120 
4121   enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4122 
4123   RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4124   PhiNode*    result_val = new PhiNode(result_reg, TypeInt::INT);
4125   PhiNode*    result_io  = new PhiNode(result_reg, Type::ABIO);
4126   PhiNode*    result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4127   Node* obj = NULL;
4128   if (!is_static) {
4129     // Check for hashing null object
4130     obj = null_check_receiver();
4131     if (stopped())  return true;        // unconditionally null
4132     result_reg->init_req(_null_path, top());
4133     result_val->init_req(_null_path, top());
4134   } else {
4135     // Do a null check, and return zero if null.
4136     // System.identityHashCode(null) == 0
4137     obj = argument(0);
4138     Node* null_ctl = top();
4139     obj = null_check_oop(obj, &null_ctl);
4140     result_reg->init_req(_null_path, null_ctl);
4141     result_val->init_req(_null_path, _gvn.intcon(0));
4142   }
4143 
4144   // Unconditionally null?  Then return right away.
4145   if (stopped()) {
4146     set_control( result_reg->in(_null_path));
4147     if (!stopped())
4148       set_result(result_val->in(_null_path));
4149     return true;
4150   }
4151 
4152   // We only go to the fast case code if we pass a number of guards.  The
4153   // paths which do not pass are accumulated in the slow_region.
4154   RegionNode* slow_region = new RegionNode(1);
4155   record_for_igvn(slow_region);
4156 
4157   // If this is a virtual call, we generate a funny guard.  We pull out
4158   // the vtable entry corresponding to hashCode() from the target object.
4159   // If the target method which we are calling happens to be the native
4160   // Object hashCode() method, we pass the guard.  We do not need this
4161   // guard for non-virtual calls -- the caller is known to be the native
4162   // Object hashCode().
4163   if (is_virtual) {
4164     // After null check, get the object's klass.
4165     Node* obj_klass = load_object_klass(obj);
4166     generate_virtual_guard(obj_klass, slow_region);
4167   }
4168 
4169   // Get the header out of the object, use LoadMarkNode when available
4170   Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4171   // The control of the load must be NULL. Otherwise, the load can move before
4172   // the null check after castPP removal.
4173   Node* no_ctrl = NULL;
4174   Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4175 
4176   // Test the header to see if it is unlocked.
4177   Node *lock_mask      = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4178   Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4179   Node *unlocked_val   = _gvn.MakeConX(markOopDesc::unlocked_value);
4180   Node *chk_unlocked   = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4181   Node *test_unlocked  = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4182 
4183   generate_slow_guard(test_unlocked, slow_region);
4184 
4185   // Get the hash value and check to see that it has been properly assigned.
4186   // We depend on hash_mask being at most 32 bits and avoid the use of
4187   // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4188   // vm: see markOop.hpp.
4189   Node *hash_mask      = _gvn.intcon(markOopDesc::hash_mask);
4190   Node *hash_shift     = _gvn.intcon(markOopDesc::hash_shift);
4191   Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4192   // This hack lets the hash bits live anywhere in the mark object now, as long
4193   // as the shift drops the relevant bits into the low 32 bits.  Note that
4194   // Java spec says that HashCode is an int so there's no point in capturing
4195   // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4196   hshifted_header      = ConvX2I(hshifted_header);
4197   Node *hash_val       = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4198 
4199   Node *no_hash_val    = _gvn.intcon(markOopDesc::no_hash);
4200   Node *chk_assigned   = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4201   Node *test_assigned  = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4202 
4203   generate_slow_guard(test_assigned, slow_region);
4204 
4205   Node* init_mem = reset_memory();
4206   // fill in the rest of the null path:
4207   result_io ->init_req(_null_path, i_o());
4208   result_mem->init_req(_null_path, init_mem);
4209 
4210   result_val->init_req(_fast_path, hash_val);
4211   result_reg->init_req(_fast_path, control());
4212   result_io ->init_req(_fast_path, i_o());
4213   result_mem->init_req(_fast_path, init_mem);
4214 
4215   // Generate code for the slow case.  We make a call to hashCode().
4216   set_control(_gvn.transform(slow_region));
4217   if (!stopped()) {
4218     // No need for PreserveJVMState, because we're using up the present state.
4219     set_all_memory(init_mem);
4220     vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4221     CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4222     Node* slow_result = set_results_for_java_call(slow_call);
4223     // this->control() comes from set_results_for_java_call
4224     result_reg->init_req(_slow_path, control());
4225     result_val->init_req(_slow_path, slow_result);
4226     result_io  ->set_req(_slow_path, i_o());
4227     result_mem ->set_req(_slow_path, reset_memory());
4228   }
4229 
4230   // Return the combined state.
4231   set_i_o(        _gvn.transform(result_io)  );
4232   set_all_memory( _gvn.transform(result_mem));
4233 
4234   set_result(result_reg, result_val);
4235   return true;
4236 }
4237 
4238 //---------------------------inline_native_getClass----------------------------
4239 // public final native Class<?> java.lang.Object.getClass();
4240 //
4241 // Build special case code for calls to getClass on an object.
4242 bool LibraryCallKit::inline_native_getClass() {
4243   Node* obj = null_check_receiver();
4244   if (stopped())  return true;
4245   set_result(load_mirror_from_klass(load_object_klass(obj)));
4246   return true;
4247 }
4248 
4249 //-----------------inline_native_Reflection_getCallerClass---------------------
4250 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4251 //
4252 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4253 //
4254 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4255 // in that it must skip particular security frames and checks for
4256 // caller sensitive methods.
4257 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4258 #ifndef PRODUCT
4259   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4260     tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4261   }
4262 #endif
4263 
4264   if (!jvms()->has_method()) {
4265 #ifndef PRODUCT
4266     if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4267       tty->print_cr("  Bailing out because intrinsic was inlined at top level");
4268     }
4269 #endif
4270     return false;
4271   }
4272 
4273   // Walk back up the JVM state to find the caller at the required
4274   // depth.
4275   JVMState* caller_jvms = jvms();
4276 
4277   // Cf. JVM_GetCallerClass
4278   // NOTE: Start the loop at depth 1 because the current JVM state does
4279   // not include the Reflection.getCallerClass() frame.
4280   for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4281     ciMethod* m = caller_jvms->method();
4282     switch (n) {
4283     case 0:
4284       fatal("current JVM state does not include the Reflection.getCallerClass frame");
4285       break;
4286     case 1:
4287       // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4288       if (!m->caller_sensitive()) {
4289 #ifndef PRODUCT
4290         if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4291           tty->print_cr("  Bailing out: CallerSensitive annotation expected at frame %d", n);
4292         }
4293 #endif
4294         return false;  // bail-out; let JVM_GetCallerClass do the work
4295       }
4296       break;
4297     default:
4298       if (!m->is_ignored_by_security_stack_walk()) {
4299         // We have reached the desired frame; return the holder class.
4300         // Acquire method holder as java.lang.Class and push as constant.
4301         ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4302         ciInstance* caller_mirror = caller_klass->java_mirror();
4303         set_result(makecon(TypeInstPtr::make(caller_mirror)));
4304 
4305 #ifndef PRODUCT
4306         if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4307           tty->print_cr("  Succeeded: caller = %d) %s.%s, JVMS depth = %d", n, caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), jvms()->depth());
4308           tty->print_cr("  JVM state at this point:");
4309           for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4310             ciMethod* m = jvms()->of_depth(i)->method();
4311             tty->print_cr("   %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4312           }
4313         }
4314 #endif
4315         return true;
4316       }
4317       break;
4318     }
4319   }
4320 
4321 #ifndef PRODUCT
4322   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4323     tty->print_cr("  Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4324     tty->print_cr("  JVM state at this point:");
4325     for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4326       ciMethod* m = jvms()->of_depth(i)->method();
4327       tty->print_cr("   %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4328     }
4329   }
4330 #endif
4331 
4332   return false;  // bail-out; let JVM_GetCallerClass do the work
4333 }
4334 
4335 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4336   Node* arg = argument(0);
4337   Node* result;
4338 
4339   switch (id) {
4340   case vmIntrinsics::_floatToRawIntBits:    result = new MoveF2INode(arg);  break;
4341   case vmIntrinsics::_intBitsToFloat:       result = new MoveI2FNode(arg);  break;
4342   case vmIntrinsics::_doubleToRawLongBits:  result = new MoveD2LNode(arg);  break;
4343   case vmIntrinsics::_longBitsToDouble:     result = new MoveL2DNode(arg);  break;
4344 
4345   case vmIntrinsics::_doubleToLongBits: {
4346     // two paths (plus control) merge in a wood
4347     RegionNode *r = new RegionNode(3);
4348     Node *phi = new PhiNode(r, TypeLong::LONG);
4349 
4350     Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4351     // Build the boolean node
4352     Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4353 
4354     // Branch either way.
4355     // NaN case is less traveled, which makes all the difference.
4356     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4357     Node *opt_isnan = _gvn.transform(ifisnan);
4358     assert( opt_isnan->is_If(), "Expect an IfNode");
4359     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4360     Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4361 
4362     set_control(iftrue);
4363 
4364     static const jlong nan_bits = CONST64(0x7ff8000000000000);
4365     Node *slow_result = longcon(nan_bits); // return NaN
4366     phi->init_req(1, _gvn.transform( slow_result ));
4367     r->init_req(1, iftrue);
4368 
4369     // Else fall through
4370     Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4371     set_control(iffalse);
4372 
4373     phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4374     r->init_req(2, iffalse);
4375 
4376     // Post merge
4377     set_control(_gvn.transform(r));
4378     record_for_igvn(r);
4379 
4380     C->set_has_split_ifs(true); // Has chance for split-if optimization
4381     result = phi;
4382     assert(result->bottom_type()->isa_long(), "must be");
4383     break;
4384   }
4385 
4386   case vmIntrinsics::_floatToIntBits: {
4387     // two paths (plus control) merge in a wood
4388     RegionNode *r = new RegionNode(3);
4389     Node *phi = new PhiNode(r, TypeInt::INT);
4390 
4391     Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4392     // Build the boolean node
4393     Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4394 
4395     // Branch either way.
4396     // NaN case is less traveled, which makes all the difference.
4397     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4398     Node *opt_isnan = _gvn.transform(ifisnan);
4399     assert( opt_isnan->is_If(), "Expect an IfNode");
4400     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4401     Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4402 
4403     set_control(iftrue);
4404 
4405     static const jint nan_bits = 0x7fc00000;
4406     Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4407     phi->init_req(1, _gvn.transform( slow_result ));
4408     r->init_req(1, iftrue);
4409 
4410     // Else fall through
4411     Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4412     set_control(iffalse);
4413 
4414     phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4415     r->init_req(2, iffalse);
4416 
4417     // Post merge
4418     set_control(_gvn.transform(r));
4419     record_for_igvn(r);
4420 
4421     C->set_has_split_ifs(true); // Has chance for split-if optimization
4422     result = phi;
4423     assert(result->bottom_type()->isa_int(), "must be");
4424     break;
4425   }
4426 
4427   default:
4428     fatal_unexpected_iid(id);
4429     break;
4430   }
4431   set_result(_gvn.transform(result));
4432   return true;
4433 }
4434 
4435 #ifdef _LP64
4436 #define XTOP ,top() /*additional argument*/
4437 #else  //_LP64
4438 #define XTOP        /*no additional argument*/
4439 #endif //_LP64
4440 
4441 //----------------------inline_unsafe_copyMemory-------------------------
4442 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4443 bool LibraryCallKit::inline_unsafe_copyMemory() {
4444   if (callee()->is_static())  return false;  // caller must have the capability!
4445   null_check_receiver();  // null-check receiver
4446   if (stopped())  return true;
4447 
4448   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
4449 
4450   Node* src_ptr =         argument(1);   // type: oop
4451   Node* src_off = ConvL2X(argument(2));  // type: long
4452   Node* dst_ptr =         argument(4);   // type: oop
4453   Node* dst_off = ConvL2X(argument(5));  // type: long
4454   Node* size    = ConvL2X(argument(7));  // type: long
4455 
4456   assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4457          "fieldOffset must be byte-scaled");
4458 
4459   Node* src = make_unsafe_address(src_ptr, src_off);
4460   Node* dst = make_unsafe_address(dst_ptr, dst_off);
4461 
4462   // Conservatively insert a memory barrier on all memory slices.
4463   // Do not let writes of the copy source or destination float below the copy.
4464   insert_mem_bar(Op_MemBarCPUOrder);
4465 
4466   // Call it.  Note that the length argument is not scaled.
4467   make_runtime_call(RC_LEAF|RC_NO_FP,
4468                     OptoRuntime::fast_arraycopy_Type(),
4469                     StubRoutines::unsafe_arraycopy(),
4470                     "unsafe_arraycopy",
4471                     TypeRawPtr::BOTTOM,
4472                     src, dst, size XTOP);
4473 
4474   // Do not let reads of the copy destination float above the copy.
4475   insert_mem_bar(Op_MemBarCPUOrder);
4476 
4477   return true;
4478 }
4479 
4480 //------------------------clone_coping-----------------------------------
4481 // Helper function for inline_native_clone.
4482 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4483   assert(obj_size != NULL, "");
4484   Node* raw_obj = alloc_obj->in(1);
4485   assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4486 
4487   AllocateNode* alloc = NULL;
4488   if (ReduceBulkZeroing) {
4489     // We will be completely responsible for initializing this object -
4490     // mark Initialize node as complete.
4491     alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4492     // The object was just allocated - there should be no any stores!
4493     guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4494     // Mark as complete_with_arraycopy so that on AllocateNode
4495     // expansion, we know this AllocateNode is initialized by an array
4496     // copy and a StoreStore barrier exists after the array copy.
4497     alloc->initialization()->set_complete_with_arraycopy();
4498   }
4499 
4500   // Copy the fastest available way.
4501   // TODO: generate fields copies for small objects instead.
4502   Node* src  = obj;
4503   Node* dest = alloc_obj;
4504   Node* size = _gvn.transform(obj_size);
4505 
4506   // Exclude the header but include array length to copy by 8 bytes words.
4507   // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4508   int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4509                             instanceOopDesc::base_offset_in_bytes();
4510   // base_off:
4511   // 8  - 32-bit VM
4512   // 12 - 64-bit VM, compressed klass
4513   // 16 - 64-bit VM, normal klass
4514   if (base_off % BytesPerLong != 0) {
4515     assert(UseCompressedClassPointers, "");
4516     if (is_array) {
4517       // Exclude length to copy by 8 bytes words.
4518       base_off += sizeof(int);
4519     } else {
4520       // Include klass to copy by 8 bytes words.
4521       base_off = instanceOopDesc::klass_offset_in_bytes();
4522     }
4523     assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4524   }
4525   src  = basic_plus_adr(src,  base_off);
4526   dest = basic_plus_adr(dest, base_off);
4527 
4528   // Compute the length also, if needed:
4529   Node* countx = size;
4530   countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4531   countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4532 
4533   const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4534 
4535   ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4536   ac->set_clonebasic();
4537   Node* n = _gvn.transform(ac);
4538   if (n == ac) {
4539     set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4540   } else {
4541     set_all_memory(n);
4542   }
4543 
4544   // If necessary, emit some card marks afterwards.  (Non-arrays only.)
4545   if (card_mark) {
4546     assert(!is_array, "");
4547     // Put in store barrier for any and all oops we are sticking
4548     // into this object.  (We could avoid this if we could prove
4549     // that the object type contains no oop fields at all.)
4550     Node* no_particular_value = NULL;
4551     Node* no_particular_field = NULL;
4552     int raw_adr_idx = Compile::AliasIdxRaw;
4553     post_barrier(control(),
4554                  memory(raw_adr_type),
4555                  alloc_obj,
4556                  no_particular_field,
4557                  raw_adr_idx,
4558                  no_particular_value,
4559                  T_OBJECT,
4560                  false);
4561   }
4562 
4563   // Do not let reads from the cloned object float above the arraycopy.
4564   if (alloc != NULL) {
4565     // Do not let stores that initialize this object be reordered with
4566     // a subsequent store that would make this object accessible by
4567     // other threads.
4568     // Record what AllocateNode this StoreStore protects so that
4569     // escape analysis can go from the MemBarStoreStoreNode to the
4570     // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4571     // based on the escape status of the AllocateNode.
4572     insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4573   } else {
4574     insert_mem_bar(Op_MemBarCPUOrder);
4575   }
4576 }
4577 
4578 //------------------------inline_native_clone----------------------------
4579 // protected native Object java.lang.Object.clone();
4580 //
4581 // Here are the simple edge cases:
4582 //  null receiver => normal trap
4583 //  virtual and clone was overridden => slow path to out-of-line clone
4584 //  not cloneable or finalizer => slow path to out-of-line Object.clone
4585 //
4586 // The general case has two steps, allocation and copying.
4587 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4588 //
4589 // Copying also has two cases, oop arrays and everything else.
4590 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4591 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4592 //
4593 // These steps fold up nicely if and when the cloned object's klass
4594 // can be sharply typed as an object array, a type array, or an instance.
4595 //
4596 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4597   PhiNode* result_val;
4598 
4599   // Set the reexecute bit for the interpreter to reexecute
4600   // the bytecode that invokes Object.clone if deoptimization happens.
4601   { PreserveReexecuteState preexecs(this);
4602     jvms()->set_should_reexecute(true);
4603 
4604     Node* obj = null_check_receiver();
4605     if (stopped())  return true;
4606 
4607     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4608 
4609     // If we are going to clone an instance, we need its exact type to
4610     // know the number and types of fields to convert the clone to
4611     // loads/stores. Maybe a speculative type can help us.
4612     if (!obj_type->klass_is_exact() &&
4613         obj_type->speculative_type() != NULL &&
4614         obj_type->speculative_type()->is_instance_klass()) {
4615       ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4616       if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4617           !spec_ik->has_injected_fields()) {
4618         ciKlass* k = obj_type->klass();
4619         if (!k->is_instance_klass() ||
4620             k->as_instance_klass()->is_interface() ||
4621             k->as_instance_klass()->has_subklass()) {
4622           obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4623         }
4624       }
4625     }
4626 
4627     Node* obj_klass = load_object_klass(obj);
4628     const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4629     const TypeOopPtr*   toop   = ((tklass != NULL)
4630                                 ? tklass->as_instance_type()
4631                                 : TypeInstPtr::NOTNULL);
4632 
4633     // Conservatively insert a memory barrier on all memory slices.
4634     // Do not let writes into the original float below the clone.
4635     insert_mem_bar(Op_MemBarCPUOrder);
4636 
4637     // paths into result_reg:
4638     enum {
4639       _slow_path = 1,     // out-of-line call to clone method (virtual or not)
4640       _objArray_path,     // plain array allocation, plus arrayof_oop_arraycopy
4641       _array_path,        // plain array allocation, plus arrayof_long_arraycopy
4642       _instance_path,     // plain instance allocation, plus arrayof_long_arraycopy
4643       PATH_LIMIT
4644     };
4645     RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4646     result_val             = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4647     PhiNode*    result_i_o = new PhiNode(result_reg, Type::ABIO);
4648     PhiNode*    result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4649     record_for_igvn(result_reg);
4650 
4651     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4652     int raw_adr_idx = Compile::AliasIdxRaw;
4653 
4654     Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4655     if (array_ctl != NULL) {
4656       // It's an array.
4657       PreserveJVMState pjvms(this);
4658       set_control(array_ctl);
4659       Node* obj_length = load_array_length(obj);
4660       Node* obj_size  = NULL;
4661       Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);  // no arguments to push
4662 
4663       if (!use_ReduceInitialCardMarks()) {
4664         // If it is an oop array, it requires very special treatment,
4665         // because card marking is required on each card of the array.
4666         Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4667         if (is_obja != NULL) {
4668           PreserveJVMState pjvms2(this);
4669           set_control(is_obja);
4670           // Generate a direct call to the right arraycopy function(s).
4671           Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4672           ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4673           ac->set_cloneoop();
4674           Node* n = _gvn.transform(ac);
4675           assert(n == ac, "cannot disappear");
4676           ac->connect_outputs(this);
4677 
4678           result_reg->init_req(_objArray_path, control());
4679           result_val->init_req(_objArray_path, alloc_obj);
4680           result_i_o ->set_req(_objArray_path, i_o());
4681           result_mem ->set_req(_objArray_path, reset_memory());
4682         }
4683       }
4684       // Otherwise, there are no card marks to worry about.
4685       // (We can dispense with card marks if we know the allocation
4686       //  comes out of eden (TLAB)...  In fact, ReduceInitialCardMarks
4687       //  causes the non-eden paths to take compensating steps to
4688       //  simulate a fresh allocation, so that no further
4689       //  card marks are required in compiled code to initialize
4690       //  the object.)
4691 
4692       if (!stopped()) {
4693         copy_to_clone(obj, alloc_obj, obj_size, true, false);
4694 
4695         // Present the results of the copy.
4696         result_reg->init_req(_array_path, control());
4697         result_val->init_req(_array_path, alloc_obj);
4698         result_i_o ->set_req(_array_path, i_o());
4699         result_mem ->set_req(_array_path, reset_memory());
4700       }
4701     }
4702 
4703     // We only go to the instance fast case code if we pass a number of guards.
4704     // The paths which do not pass are accumulated in the slow_region.
4705     RegionNode* slow_region = new RegionNode(1);
4706     record_for_igvn(slow_region);
4707     if (!stopped()) {
4708       // It's an instance (we did array above).  Make the slow-path tests.
4709       // If this is a virtual call, we generate a funny guard.  We grab
4710       // the vtable entry corresponding to clone() from the target object.
4711       // If the target method which we are calling happens to be the
4712       // Object clone() method, we pass the guard.  We do not need this
4713       // guard for non-virtual calls; the caller is known to be the native
4714       // Object clone().
4715       if (is_virtual) {
4716         generate_virtual_guard(obj_klass, slow_region);
4717       }
4718 
4719       // The object must be cloneable and must not have a finalizer.
4720       // Both of these conditions may be checked in a single test.
4721       // We could optimize the cloneable test further, but we don't care.
4722       generate_access_flags_guard(obj_klass,
4723                                   // Test both conditions:
4724                                   JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4725                                   // Must be cloneable but not finalizer:
4726                                   JVM_ACC_IS_CLONEABLE,
4727                                   slow_region);
4728     }
4729 
4730     if (!stopped()) {
4731       // It's an instance, and it passed the slow-path tests.
4732       PreserveJVMState pjvms(this);
4733       Node* obj_size  = NULL;
4734       // Need to deoptimize on exception from allocation since Object.clone intrinsic
4735       // is reexecuted if deoptimization occurs and there could be problems when merging
4736       // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4737       Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4738 
4739       copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4740 
4741       // Present the results of the slow call.
4742       result_reg->init_req(_instance_path, control());
4743       result_val->init_req(_instance_path, alloc_obj);
4744       result_i_o ->set_req(_instance_path, i_o());
4745       result_mem ->set_req(_instance_path, reset_memory());
4746     }
4747 
4748     // Generate code for the slow case.  We make a call to clone().
4749     set_control(_gvn.transform(slow_region));
4750     if (!stopped()) {
4751       PreserveJVMState pjvms(this);
4752       CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4753       Node* slow_result = set_results_for_java_call(slow_call);
4754       // this->control() comes from set_results_for_java_call
4755       result_reg->init_req(_slow_path, control());
4756       result_val->init_req(_slow_path, slow_result);
4757       result_i_o ->set_req(_slow_path, i_o());
4758       result_mem ->set_req(_slow_path, reset_memory());
4759     }
4760 
4761     // Return the combined state.
4762     set_control(    _gvn.transform(result_reg));
4763     set_i_o(        _gvn.transform(result_i_o));
4764     set_all_memory( _gvn.transform(result_mem));
4765   } // original reexecute is set back here
4766 
4767   set_result(_gvn.transform(result_val));
4768   return true;
4769 }
4770 
4771 // If we have a tighly coupled allocation, the arraycopy may take care
4772 // of the array initialization. If one of the guards we insert between
4773 // the allocation and the arraycopy causes a deoptimization, an
4774 // unitialized array will escape the compiled method. To prevent that
4775 // we set the JVM state for uncommon traps between the allocation and
4776 // the arraycopy to the state before the allocation so, in case of
4777 // deoptimization, we'll reexecute the allocation and the
4778 // initialization.
4779 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4780   if (alloc != NULL) {
4781     ciMethod* trap_method = alloc->jvms()->method();
4782     int trap_bci = alloc->jvms()->bci();
4783 
4784     if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4785           !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4786       // Make sure there's no store between the allocation and the
4787       // arraycopy otherwise visible side effects could be rexecuted
4788       // in case of deoptimization and cause incorrect execution.
4789       bool no_interfering_store = true;
4790       Node* mem = alloc->in(TypeFunc::Memory);
4791       if (mem->is_MergeMem()) {
4792         for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4793           Node* n = mms.memory();
4794           if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4795             assert(n->is_Store(), "what else?");
4796             no_interfering_store = false;
4797             break;
4798           }
4799         }
4800       } else {
4801         for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4802           Node* n = mms.memory();
4803           if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4804             assert(n->is_Store(), "what else?");
4805             no_interfering_store = false;
4806             break;
4807           }
4808         }
4809       }
4810 
4811       if (no_interfering_store) {
4812         JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4813         uint size = alloc->req();
4814         SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4815         old_jvms->set_map(sfpt);
4816         for (uint i = 0; i < size; i++) {
4817           sfpt->init_req(i, alloc->in(i));
4818         }
4819         // re-push array length for deoptimization
4820         sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4821         old_jvms->set_sp(old_jvms->sp()+1);
4822         old_jvms->set_monoff(old_jvms->monoff()+1);
4823         old_jvms->set_scloff(old_jvms->scloff()+1);
4824         old_jvms->set_endoff(old_jvms->endoff()+1);
4825         old_jvms->set_should_reexecute(true);
4826 
4827         sfpt->set_i_o(map()->i_o());
4828         sfpt->set_memory(map()->memory());
4829         sfpt->set_control(map()->control());
4830 
4831         JVMState* saved_jvms = jvms();
4832         saved_reexecute_sp = _reexecute_sp;
4833 
4834         set_jvms(sfpt->jvms());
4835         _reexecute_sp = jvms()->sp();
4836 
4837         return saved_jvms;
4838       }
4839     }
4840   }
4841   return NULL;
4842 }
4843 
4844 // In case of a deoptimization, we restart execution at the
4845 // allocation, allocating a new array. We would leave an uninitialized
4846 // array in the heap that GCs wouldn't expect. Move the allocation
4847 // after the traps so we don't allocate the array if we
4848 // deoptimize. This is possible because tightly_coupled_allocation()
4849 // guarantees there's no observer of the allocated array at this point
4850 // and the control flow is simple enough.
4851 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4852   if (saved_jvms != NULL && !stopped()) {
4853     assert(alloc != NULL, "only with a tightly coupled allocation");
4854     // restore JVM state to the state at the arraycopy
4855     saved_jvms->map()->set_control(map()->control());
4856     assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4857     assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4858     // If we've improved the types of some nodes (null check) while
4859     // emitting the guards, propagate them to the current state
4860     map()->replaced_nodes().apply(saved_jvms->map());
4861     set_jvms(saved_jvms);
4862     _reexecute_sp = saved_reexecute_sp;
4863 
4864     // Remove the allocation from above the guards
4865     CallProjections callprojs;
4866     alloc->extract_projections(&callprojs, true);
4867     InitializeNode* init = alloc->initialization();
4868     Node* alloc_mem = alloc->in(TypeFunc::Memory);
4869     C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4870     C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4871     C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4872 
4873     // move the allocation here (after the guards)
4874     _gvn.hash_delete(alloc);
4875     alloc->set_req(TypeFunc::Control, control());
4876     alloc->set_req(TypeFunc::I_O, i_o());
4877     Node *mem = reset_memory();
4878     set_all_memory(mem);
4879     alloc->set_req(TypeFunc::Memory, mem);
4880     set_control(init->proj_out(TypeFunc::Control));
4881     set_i_o(callprojs.fallthrough_ioproj);
4882 
4883     // Update memory as done in GraphKit::set_output_for_allocation()
4884     const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4885     const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4886     if (ary_type->isa_aryptr() && length_type != NULL) {
4887       ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4888     }
4889     const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4890     int            elemidx  = C->get_alias_index(telemref);
4891     set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4892     set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4893 
4894     Node* allocx = _gvn.transform(alloc);
4895     assert(allocx == alloc, "where has the allocation gone?");
4896     assert(dest->is_CheckCastPP(), "not an allocation result?");
4897 
4898     _gvn.hash_delete(dest);
4899     dest->set_req(0, control());
4900     Node* destx = _gvn.transform(dest);
4901     assert(destx == dest, "where has the allocation result gone?");
4902   }
4903 }
4904 
4905 
4906 //------------------------------inline_arraycopy-----------------------
4907 // public static native void java.lang.System.arraycopy(Object src,  int  srcPos,
4908 //                                                      Object dest, int destPos,
4909 //                                                      int length);
4910 bool LibraryCallKit::inline_arraycopy() {
4911   // Get the arguments.
4912   Node* src         = argument(0);  // type: oop
4913   Node* src_offset  = argument(1);  // type: int
4914   Node* dest        = argument(2);  // type: oop
4915   Node* dest_offset = argument(3);  // type: int
4916   Node* length      = argument(4);  // type: int
4917 
4918 
4919   // Check for allocation before we add nodes that would confuse
4920   // tightly_coupled_allocation()
4921   AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4922 
4923   int saved_reexecute_sp = -1;
4924   JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4925   // See arraycopy_restore_alloc_state() comment
4926   // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4927   // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4928   // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4929   bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4930 
4931   // The following tests must be performed
4932   // (1) src and dest are arrays.
4933   // (2) src and dest arrays must have elements of the same BasicType
4934   // (3) src and dest must not be null.
4935   // (4) src_offset must not be negative.
4936   // (5) dest_offset must not be negative.
4937   // (6) length must not be negative.
4938   // (7) src_offset + length must not exceed length of src.
4939   // (8) dest_offset + length must not exceed length of dest.
4940   // (9) each element of an oop array must be assignable
4941 
4942   // (3) src and dest must not be null.
4943   // always do this here because we need the JVM state for uncommon traps
4944   Node* null_ctl = top();
4945   src  = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src,  T_ARRAY);
4946   assert(null_ctl->is_top(), "no null control here");
4947   dest = null_check(dest, T_ARRAY);
4948 
4949   if (!can_emit_guards) {
4950     // if saved_jvms == NULL and alloc != NULL, we don't emit any
4951     // guards but the arraycopy node could still take advantage of a
4952     // tightly allocated allocation. tightly_coupled_allocation() is
4953     // called again to make sure it takes the null check above into
4954     // account: the null check is mandatory and if it caused an
4955     // uncommon trap to be emitted then the allocation can't be
4956     // considered tightly coupled in this context.
4957     alloc = tightly_coupled_allocation(dest, NULL);
4958   }
4959 
4960   bool validated = false;
4961 
4962   const Type* src_type  = _gvn.type(src);
4963   const Type* dest_type = _gvn.type(dest);
4964   const TypeAryPtr* top_src  = src_type->isa_aryptr();
4965   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4966 
4967   // Do we have the type of src?
4968   bool has_src = (top_src != NULL && top_src->klass() != NULL);
4969   // Do we have the type of dest?
4970   bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4971   // Is the type for src from speculation?
4972   bool src_spec = false;
4973   // Is the type for dest from speculation?
4974   bool dest_spec = false;
4975 
4976   if ((!has_src || !has_dest) && can_emit_guards) {
4977     // We don't have sufficient type information, let's see if
4978     // speculative types can help. We need to have types for both src
4979     // and dest so that it pays off.
4980 
4981     // Do we already have or could we have type information for src
4982     bool could_have_src = has_src;
4983     // Do we already have or could we have type information for dest
4984     bool could_have_dest = has_dest;
4985 
4986     ciKlass* src_k = NULL;
4987     if (!has_src) {
4988       src_k = src_type->speculative_type_not_null();
4989       if (src_k != NULL && src_k->is_array_klass()) {
4990         could_have_src = true;
4991       }
4992     }
4993 
4994     ciKlass* dest_k = NULL;
4995     if (!has_dest) {
4996       dest_k = dest_type->speculative_type_not_null();
4997       if (dest_k != NULL && dest_k->is_array_klass()) {
4998         could_have_dest = true;
4999       }
5000     }
5001 
5002     if (could_have_src && could_have_dest) {
5003       // This is going to pay off so emit the required guards
5004       if (!has_src) {
5005         src = maybe_cast_profiled_obj(src, src_k, true);
5006         src_type  = _gvn.type(src);
5007         top_src  = src_type->isa_aryptr();
5008         has_src = (top_src != NULL && top_src->klass() != NULL);
5009         src_spec = true;
5010       }
5011       if (!has_dest) {
5012         dest = maybe_cast_profiled_obj(dest, dest_k, true);
5013         dest_type  = _gvn.type(dest);
5014         top_dest  = dest_type->isa_aryptr();
5015         has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5016         dest_spec = true;
5017       }
5018     }
5019   }
5020 
5021   if (has_src && has_dest && can_emit_guards) {
5022     BasicType src_elem  = top_src->klass()->as_array_klass()->element_type()->basic_type();
5023     BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5024     if (src_elem  == T_ARRAY)  src_elem  = T_OBJECT;
5025     if (dest_elem == T_ARRAY)  dest_elem = T_OBJECT;
5026 
5027     if (src_elem == dest_elem && src_elem == T_OBJECT) {
5028       // If both arrays are object arrays then having the exact types
5029       // for both will remove the need for a subtype check at runtime
5030       // before the call and may make it possible to pick a faster copy
5031       // routine (without a subtype check on every element)
5032       // Do we have the exact type of src?
5033       bool could_have_src = src_spec;
5034       // Do we have the exact type of dest?
5035       bool could_have_dest = dest_spec;
5036       ciKlass* src_k = top_src->klass();
5037       ciKlass* dest_k = top_dest->klass();
5038       if (!src_spec) {
5039         src_k = src_type->speculative_type_not_null();
5040         if (src_k != NULL && src_k->is_array_klass()) {
5041           could_have_src = true;
5042         }
5043       }
5044       if (!dest_spec) {
5045         dest_k = dest_type->speculative_type_not_null();
5046         if (dest_k != NULL && dest_k->is_array_klass()) {
5047           could_have_dest = true;
5048         }
5049       }
5050       if (could_have_src && could_have_dest) {
5051         // If we can have both exact types, emit the missing guards
5052         if (could_have_src && !src_spec) {
5053           src = maybe_cast_profiled_obj(src, src_k, true);
5054         }
5055         if (could_have_dest && !dest_spec) {
5056           dest = maybe_cast_profiled_obj(dest, dest_k, true);
5057         }
5058       }
5059     }
5060   }
5061 
5062   ciMethod* trap_method = method();
5063   int trap_bci = bci();
5064   if (saved_jvms != NULL) {
5065     trap_method = alloc->jvms()->method();
5066     trap_bci = alloc->jvms()->bci();
5067   }
5068 
5069   if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5070       can_emit_guards &&
5071       !src->is_top() && !dest->is_top()) {
5072     // validate arguments: enables transformation the ArrayCopyNode
5073     validated = true;
5074 
5075     RegionNode* slow_region = new RegionNode(1);
5076     record_for_igvn(slow_region);
5077 
5078     // (1) src and dest are arrays.
5079     generate_non_array_guard(load_object_klass(src), slow_region);
5080     generate_non_array_guard(load_object_klass(dest), slow_region);
5081 
5082     // (2) src and dest arrays must have elements of the same BasicType
5083     // done at macro expansion or at Ideal transformation time
5084 
5085     // (4) src_offset must not be negative.
5086     generate_negative_guard(src_offset, slow_region);
5087 
5088     // (5) dest_offset must not be negative.
5089     generate_negative_guard(dest_offset, slow_region);
5090 
5091     // (7) src_offset + length must not exceed length of src.
5092     generate_limit_guard(src_offset, length,
5093                          load_array_length(src),
5094                          slow_region);
5095 
5096     // (8) dest_offset + length must not exceed length of dest.
5097     generate_limit_guard(dest_offset, length,
5098                          load_array_length(dest),
5099                          slow_region);
5100 
5101     // (9) each element of an oop array must be assignable
5102     Node* src_klass  = load_object_klass(src);
5103     Node* dest_klass = load_object_klass(dest);
5104     Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5105 
5106     if (not_subtype_ctrl != top()) {
5107       PreserveJVMState pjvms(this);
5108       set_control(not_subtype_ctrl);
5109       uncommon_trap(Deoptimization::Reason_intrinsic,
5110                     Deoptimization::Action_make_not_entrant);
5111       assert(stopped(), "Should be stopped");
5112     }
5113     {
5114       PreserveJVMState pjvms(this);
5115       set_control(_gvn.transform(slow_region));
5116       uncommon_trap(Deoptimization::Reason_intrinsic,
5117                     Deoptimization::Action_make_not_entrant);
5118       assert(stopped(), "Should be stopped");
5119     }
5120   }
5121 
5122   arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5123 
5124   if (stopped()) {
5125     return true;
5126   }
5127 
5128   ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
5129                                           // Create LoadRange and LoadKlass nodes for use during macro expansion here
5130                                           // so the compiler has a chance to eliminate them: during macro expansion,
5131                                           // we have to set their control (CastPP nodes are eliminated).
5132                                           load_object_klass(src), load_object_klass(dest),
5133                                           load_array_length(src), load_array_length(dest));
5134 
5135   ac->set_arraycopy(validated);
5136 
5137   Node* n = _gvn.transform(ac);
5138   if (n == ac) {
5139     ac->connect_outputs(this);
5140   } else {
5141     assert(validated, "shouldn't transform if all arguments not validated");
5142     set_all_memory(n);
5143   }
5144 
5145   return true;
5146 }
5147 
5148 
5149 // Helper function which determines if an arraycopy immediately follows
5150 // an allocation, with no intervening tests or other escapes for the object.
5151 AllocateArrayNode*
5152 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5153                                            RegionNode* slow_region) {
5154   if (stopped())             return NULL;  // no fast path
5155   if (C->AliasLevel() == 0)  return NULL;  // no MergeMems around
5156 
5157   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5158   if (alloc == NULL)  return NULL;
5159 
5160   Node* rawmem = memory(Compile::AliasIdxRaw);
5161   // Is the allocation's memory state untouched?
5162   if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5163     // Bail out if there have been raw-memory effects since the allocation.
5164     // (Example:  There might have been a call or safepoint.)
5165     return NULL;
5166   }
5167   rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5168   if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5169     return NULL;
5170   }
5171 
5172   // There must be no unexpected observers of this allocation.
5173   for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5174     Node* obs = ptr->fast_out(i);
5175     if (obs != this->map()) {
5176       return NULL;
5177     }
5178   }
5179 
5180   // This arraycopy must unconditionally follow the allocation of the ptr.
5181   Node* alloc_ctl = ptr->in(0);
5182   assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5183 
5184   Node* ctl = control();
5185   while (ctl != alloc_ctl) {
5186     // There may be guards which feed into the slow_region.
5187     // Any other control flow means that we might not get a chance
5188     // to finish initializing the allocated object.
5189     if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5190       IfNode* iff = ctl->in(0)->as_If();
5191       Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5192       assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5193       if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5194         ctl = iff->in(0);       // This test feeds the known slow_region.
5195         continue;
5196       }
5197       // One more try:  Various low-level checks bottom out in
5198       // uncommon traps.  If the debug-info of the trap omits
5199       // any reference to the allocation, as we've already
5200       // observed, then there can be no objection to the trap.
5201       bool found_trap = false;
5202       for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5203         Node* obs = not_ctl->fast_out(j);
5204         if (obs->in(0) == not_ctl && obs->is_Call() &&
5205             (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5206           found_trap = true; break;
5207         }
5208       }
5209       if (found_trap) {
5210         ctl = iff->in(0);       // This test feeds a harmless uncommon trap.
5211         continue;
5212       }
5213     }
5214     return NULL;
5215   }
5216 
5217   // If we get this far, we have an allocation which immediately
5218   // precedes the arraycopy, and we can take over zeroing the new object.
5219   // The arraycopy will finish the initialization, and provide
5220   // a new control state to which we will anchor the destination pointer.
5221 
5222   return alloc;
5223 }
5224 
5225 //-------------inline_encodeISOArray-----------------------------------
5226 // encode char[] to byte[] in ISO_8859_1
5227 bool LibraryCallKit::inline_encodeISOArray() {
5228   assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5229   // no receiver since it is static method
5230   Node *src         = argument(0);
5231   Node *src_offset  = argument(1);
5232   Node *dst         = argument(2);
5233   Node *dst_offset  = argument(3);
5234   Node *length      = argument(4);
5235 
5236   const Type* src_type = src->Value(&_gvn);
5237   const Type* dst_type = dst->Value(&_gvn);
5238   const TypeAryPtr* top_src = src_type->isa_aryptr();
5239   const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5240   if (top_src  == NULL || top_src->klass()  == NULL ||
5241       top_dest == NULL || top_dest->klass() == NULL) {
5242     // failed array check
5243     return false;
5244   }
5245 
5246   // Figure out the size and type of the elements we will be copying.
5247   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5248   BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5249   if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5250     return false;
5251   }
5252   Node* src_start = array_element_address(src, src_offset, src_elem);
5253   Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5254   // 'src_start' points to src array + scaled offset
5255   // 'dst_start' points to dst array + scaled offset
5256 
5257   const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5258   Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5259   enc = _gvn.transform(enc);
5260   Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5261   set_memory(res_mem, mtype);
5262   set_result(enc);
5263   return true;
5264 }
5265 
5266 //-------------inline_multiplyToLen-----------------------------------
5267 bool LibraryCallKit::inline_multiplyToLen() {
5268   assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5269 
5270   address stubAddr = StubRoutines::multiplyToLen();
5271   if (stubAddr == NULL) {
5272     return false; // Intrinsic's stub is not implemented on this platform
5273   }
5274   const char* stubName = "multiplyToLen";
5275 
5276   assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5277 
5278   // no receiver because it is a static method
5279   Node* x    = argument(0);
5280   Node* xlen = argument(1);
5281   Node* y    = argument(2);
5282   Node* ylen = argument(3);
5283   Node* z    = argument(4);
5284 
5285   const Type* x_type = x->Value(&_gvn);
5286   const Type* y_type = y->Value(&_gvn);
5287   const TypeAryPtr* top_x = x_type->isa_aryptr();
5288   const TypeAryPtr* top_y = y_type->isa_aryptr();
5289   if (top_x  == NULL || top_x->klass()  == NULL ||
5290       top_y == NULL || top_y->klass() == NULL) {
5291     // failed array check
5292     return false;
5293   }
5294 
5295   BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5296   BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5297   if (x_elem != T_INT || y_elem != T_INT) {
5298     return false;
5299   }
5300 
5301   // Set the original stack and the reexecute bit for the interpreter to reexecute
5302   // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5303   // on the return from z array allocation in runtime.
5304   { PreserveReexecuteState preexecs(this);
5305     jvms()->set_should_reexecute(true);
5306 
5307     Node* x_start = array_element_address(x, intcon(0), x_elem);
5308     Node* y_start = array_element_address(y, intcon(0), y_elem);
5309     // 'x_start' points to x array + scaled xlen
5310     // 'y_start' points to y array + scaled ylen
5311 
5312     // Allocate the result array
5313     Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5314     ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5315     Node* klass_node = makecon(TypeKlassPtr::make(klass));
5316 
5317     IdealKit ideal(this);
5318 
5319 #define __ ideal.
5320      Node* one = __ ConI(1);
5321      Node* zero = __ ConI(0);
5322      IdealVariable need_alloc(ideal), z_alloc(ideal);  __ declarations_done();
5323      __ set(need_alloc, zero);
5324      __ set(z_alloc, z);
5325      __ if_then(z, BoolTest::eq, null()); {
5326        __ increment (need_alloc, one);
5327      } __ else_(); {
5328        // Update graphKit memory and control from IdealKit.
5329        sync_kit(ideal);
5330        Node* zlen_arg = load_array_length(z);
5331        // Update IdealKit memory and control from graphKit.
5332        __ sync_kit(this);
5333        __ if_then(zlen_arg, BoolTest::lt, zlen); {
5334          __ increment (need_alloc, one);
5335        } __ end_if();
5336      } __ end_if();
5337 
5338      __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5339        // Update graphKit memory and control from IdealKit.
5340        sync_kit(ideal);
5341        Node * narr = new_array(klass_node, zlen, 1);
5342        // Update IdealKit memory and control from graphKit.
5343        __ sync_kit(this);
5344        __ set(z_alloc, narr);
5345      } __ end_if();
5346 
5347      sync_kit(ideal);
5348      z = __ value(z_alloc);
5349      // Can't use TypeAryPtr::INTS which uses Bottom offset.
5350      _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5351      // Final sync IdealKit and GraphKit.
5352      final_sync(ideal);
5353 #undef __
5354 
5355     Node* z_start = array_element_address(z, intcon(0), T_INT);
5356 
5357     Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5358                                    OptoRuntime::multiplyToLen_Type(),
5359                                    stubAddr, stubName, TypePtr::BOTTOM,
5360                                    x_start, xlen, y_start, ylen, z_start, zlen);
5361   } // original reexecute is set back here
5362 
5363   C->set_has_split_ifs(true); // Has chance for split-if optimization
5364   set_result(z);
5365   return true;
5366 }
5367 
5368 //-------------inline_squareToLen------------------------------------
5369 bool LibraryCallKit::inline_squareToLen() {
5370   assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5371 
5372   address stubAddr = StubRoutines::squareToLen();
5373   if (stubAddr == NULL) {
5374     return false; // Intrinsic's stub is not implemented on this platform
5375   }
5376   const char* stubName = "squareToLen";
5377 
5378   assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5379 
5380   Node* x    = argument(0);
5381   Node* len  = argument(1);
5382   Node* z    = argument(2);
5383   Node* zlen = argument(3);
5384 
5385   const Type* x_type = x->Value(&_gvn);
5386   const Type* z_type = z->Value(&_gvn);
5387   const TypeAryPtr* top_x = x_type->isa_aryptr();
5388   const TypeAryPtr* top_z = z_type->isa_aryptr();
5389   if (top_x  == NULL || top_x->klass()  == NULL ||
5390       top_z  == NULL || top_z->klass()  == NULL) {
5391     // failed array check
5392     return false;
5393   }
5394 
5395   BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5396   BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5397   if (x_elem != T_INT || z_elem != T_INT) {
5398     return false;
5399   }
5400 
5401 
5402   Node* x_start = array_element_address(x, intcon(0), x_elem);
5403   Node* z_start = array_element_address(z, intcon(0), z_elem);
5404 
5405   Node*  call = make_runtime_call(RC_LEAF|RC_NO_FP,
5406                                   OptoRuntime::squareToLen_Type(),
5407                                   stubAddr, stubName, TypePtr::BOTTOM,
5408                                   x_start, len, z_start, zlen);
5409 
5410   set_result(z);
5411   return true;
5412 }
5413 
5414 //-------------inline_mulAdd------------------------------------------
5415 bool LibraryCallKit::inline_mulAdd() {
5416   assert(UseMulAddIntrinsic, "not implementated on this platform");
5417 
5418   address stubAddr = StubRoutines::mulAdd();
5419   if (stubAddr == NULL) {
5420     return false; // Intrinsic's stub is not implemented on this platform
5421   }
5422   const char* stubName = "mulAdd";
5423 
5424   assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5425 
5426   Node* out      = argument(0);
5427   Node* in       = argument(1);
5428   Node* offset   = argument(2);
5429   Node* len      = argument(3);
5430   Node* k        = argument(4);
5431 
5432   const Type* out_type = out->Value(&_gvn);
5433   const Type* in_type = in->Value(&_gvn);
5434   const TypeAryPtr* top_out = out_type->isa_aryptr();
5435   const TypeAryPtr* top_in = in_type->isa_aryptr();
5436   if (top_out  == NULL || top_out->klass()  == NULL ||
5437       top_in == NULL || top_in->klass() == NULL) {
5438     // failed array check
5439     return false;
5440   }
5441 
5442   BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5443   BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5444   if (out_elem != T_INT || in_elem != T_INT) {
5445     return false;
5446   }
5447 
5448   Node* outlen = load_array_length(out);
5449   Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5450   Node* out_start = array_element_address(out, intcon(0), out_elem);
5451   Node* in_start = array_element_address(in, intcon(0), in_elem);
5452 
5453   Node*  call = make_runtime_call(RC_LEAF|RC_NO_FP,
5454                                   OptoRuntime::mulAdd_Type(),
5455                                   stubAddr, stubName, TypePtr::BOTTOM,
5456                                   out_start,in_start, new_offset, len, k);
5457   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5458   set_result(result);
5459   return true;
5460 }
5461 
5462 //-------------inline_montgomeryMultiply-----------------------------------
5463 bool LibraryCallKit::inline_montgomeryMultiply() {
5464   address stubAddr = StubRoutines::montgomeryMultiply();
5465   if (stubAddr == NULL) {
5466     return false; // Intrinsic's stub is not implemented on this platform
5467   }
5468 
5469   assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5470   const char* stubName = "montgomery_square";
5471 
5472   assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5473 
5474   Node* a    = argument(0);
5475   Node* b    = argument(1);
5476   Node* n    = argument(2);
5477   Node* len  = argument(3);
5478   Node* inv  = argument(4);
5479   Node* m    = argument(6);
5480 
5481   const Type* a_type = a->Value(&_gvn);
5482   const TypeAryPtr* top_a = a_type->isa_aryptr();
5483   const Type* b_type = b->Value(&_gvn);
5484   const TypeAryPtr* top_b = b_type->isa_aryptr();
5485   const Type* n_type = a->Value(&_gvn);
5486   const TypeAryPtr* top_n = n_type->isa_aryptr();
5487   const Type* m_type = a->Value(&_gvn);
5488   const TypeAryPtr* top_m = m_type->isa_aryptr();
5489   if (top_a  == NULL || top_a->klass()  == NULL ||
5490       top_b == NULL || top_b->klass()  == NULL ||
5491       top_n == NULL || top_n->klass()  == NULL ||
5492       top_m == NULL || top_m->klass()  == NULL) {
5493     // failed array check
5494     return false;
5495   }
5496 
5497   BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5498   BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5499   BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5500   BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5501   if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5502     return false;
5503   }
5504 
5505   // Make the call
5506   {
5507     Node* a_start = array_element_address(a, intcon(0), a_elem);
5508     Node* b_start = array_element_address(b, intcon(0), b_elem);
5509     Node* n_start = array_element_address(n, intcon(0), n_elem);
5510     Node* m_start = array_element_address(m, intcon(0), m_elem);
5511 
5512     Node* call = make_runtime_call(RC_LEAF,
5513                                    OptoRuntime::montgomeryMultiply_Type(),
5514                                    stubAddr, stubName, TypePtr::BOTTOM,
5515                                    a_start, b_start, n_start, len, inv, top(),
5516                                    m_start);
5517     set_result(m);
5518   }
5519 
5520   return true;
5521 }
5522 
5523 bool LibraryCallKit::inline_montgomerySquare() {
5524   address stubAddr = StubRoutines::montgomerySquare();
5525   if (stubAddr == NULL) {
5526     return false; // Intrinsic's stub is not implemented on this platform
5527   }
5528 
5529   assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5530   const char* stubName = "montgomery_square";
5531 
5532   assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5533 
5534   Node* a    = argument(0);
5535   Node* n    = argument(1);
5536   Node* len  = argument(2);
5537   Node* inv  = argument(3);
5538   Node* m    = argument(5);
5539 
5540   const Type* a_type = a->Value(&_gvn);
5541   const TypeAryPtr* top_a = a_type->isa_aryptr();
5542   const Type* n_type = a->Value(&_gvn);
5543   const TypeAryPtr* top_n = n_type->isa_aryptr();
5544   const Type* m_type = a->Value(&_gvn);
5545   const TypeAryPtr* top_m = m_type->isa_aryptr();
5546   if (top_a  == NULL || top_a->klass()  == NULL ||
5547       top_n == NULL || top_n->klass()  == NULL ||
5548       top_m == NULL || top_m->klass()  == NULL) {
5549     // failed array check
5550     return false;
5551   }
5552 
5553   BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5554   BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5555   BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5556   if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5557     return false;
5558   }
5559 
5560   // Make the call
5561   {
5562     Node* a_start = array_element_address(a, intcon(0), a_elem);
5563     Node* n_start = array_element_address(n, intcon(0), n_elem);
5564     Node* m_start = array_element_address(m, intcon(0), m_elem);
5565 
5566     Node* call = make_runtime_call(RC_LEAF,
5567                                    OptoRuntime::montgomerySquare_Type(),
5568                                    stubAddr, stubName, TypePtr::BOTTOM,
5569                                    a_start, n_start, len, inv, top(),
5570                                    m_start);
5571     set_result(m);
5572   }
5573 
5574   return true;
5575 }
5576 
5577 
5578 /**
5579  * Calculate CRC32 for byte.
5580  * int java.util.zip.CRC32.update(int crc, int b)
5581  */
5582 bool LibraryCallKit::inline_updateCRC32() {
5583   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5584   assert(callee()->signature()->size() == 2, "update has 2 parameters");
5585   // no receiver since it is static method
5586   Node* crc  = argument(0); // type: int
5587   Node* b    = argument(1); // type: int
5588 
5589   /*
5590    *    int c = ~ crc;
5591    *    b = timesXtoThe32[(b ^ c) & 0xFF];
5592    *    b = b ^ (c >>> 8);
5593    *    crc = ~b;
5594    */
5595 
5596   Node* M1 = intcon(-1);
5597   crc = _gvn.transform(new XorINode(crc, M1));
5598   Node* result = _gvn.transform(new XorINode(crc, b));
5599   result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5600 
5601   Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5602   Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5603   Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5604   result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5605 
5606   crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5607   result = _gvn.transform(new XorINode(crc, result));
5608   result = _gvn.transform(new XorINode(result, M1));
5609   set_result(result);
5610   return true;
5611 }
5612 
5613 /**
5614  * Calculate CRC32 for byte[] array.
5615  * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5616  */
5617 bool LibraryCallKit::inline_updateBytesCRC32() {
5618   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5619   assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5620   // no receiver since it is static method
5621   Node* crc     = argument(0); // type: int
5622   Node* src     = argument(1); // type: oop
5623   Node* offset  = argument(2); // type: int
5624   Node* length  = argument(3); // type: int
5625 
5626   const Type* src_type = src->Value(&_gvn);
5627   const TypeAryPtr* top_src = src_type->isa_aryptr();
5628   if (top_src  == NULL || top_src->klass()  == NULL) {
5629     // failed array check
5630     return false;
5631   }
5632 
5633   // Figure out the size and type of the elements we will be copying.
5634   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5635   if (src_elem != T_BYTE) {
5636     return false;
5637   }
5638 
5639   // 'src_start' points to src array + scaled offset
5640   Node* src_start = array_element_address(src, offset, src_elem);
5641 
5642   // We assume that range check is done by caller.
5643   // TODO: generate range check (offset+length < src.length) in debug VM.
5644 
5645   // Call the stub.
5646   address stubAddr = StubRoutines::updateBytesCRC32();
5647   const char *stubName = "updateBytesCRC32";
5648 
5649   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5650                                  stubAddr, stubName, TypePtr::BOTTOM,
5651                                  crc, src_start, length);
5652   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5653   set_result(result);
5654   return true;
5655 }
5656 
5657 /**
5658  * Calculate CRC32 for ByteBuffer.
5659  * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5660  */
5661 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5662   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5663   assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5664   // no receiver since it is static method
5665   Node* crc     = argument(0); // type: int
5666   Node* src     = argument(1); // type: long
5667   Node* offset  = argument(3); // type: int
5668   Node* length  = argument(4); // type: int
5669 
5670   src = ConvL2X(src);  // adjust Java long to machine word
5671   Node* base = _gvn.transform(new CastX2PNode(src));
5672   offset = ConvI2X(offset);
5673 
5674   // 'src_start' points to src array + scaled offset
5675   Node* src_start = basic_plus_adr(top(), base, offset);
5676 
5677   // Call the stub.
5678   address stubAddr = StubRoutines::updateBytesCRC32();
5679   const char *stubName = "updateBytesCRC32";
5680 
5681   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5682                                  stubAddr, stubName, TypePtr::BOTTOM,
5683                                  crc, src_start, length);
5684   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5685   set_result(result);
5686   return true;
5687 }
5688 
5689 //------------------------------get_table_from_crc32c_class-----------------------
5690 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5691   Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5692   assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5693 
5694   return table;
5695 }
5696 
5697 //------------------------------inline_updateBytesCRC32C-----------------------
5698 //
5699 // Calculate CRC32C for byte[] array.
5700 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5701 //
5702 bool LibraryCallKit::inline_updateBytesCRC32C() {
5703   assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5704   assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5705   assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5706   // no receiver since it is a static method
5707   Node* crc     = argument(0); // type: int
5708   Node* src     = argument(1); // type: oop
5709   Node* offset  = argument(2); // type: int
5710   Node* end     = argument(3); // type: int
5711 
5712   Node* length = _gvn.transform(new SubINode(end, offset));
5713 
5714   const Type* src_type = src->Value(&_gvn);
5715   const TypeAryPtr* top_src = src_type->isa_aryptr();
5716   if (top_src  == NULL || top_src->klass()  == NULL) {
5717     // failed array check
5718     return false;
5719   }
5720 
5721   // Figure out the size and type of the elements we will be copying.
5722   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5723   if (src_elem != T_BYTE) {
5724     return false;
5725   }
5726 
5727   // 'src_start' points to src array + scaled offset
5728   Node* src_start = array_element_address(src, offset, src_elem);
5729 
5730   // static final int[] byteTable in class CRC32C
5731   Node* table = get_table_from_crc32c_class(callee()->holder());
5732   Node* table_start = array_element_address(table, intcon(0), T_INT);
5733 
5734   // We assume that range check is done by caller.
5735   // TODO: generate range check (offset+length < src.length) in debug VM.
5736 
5737   // Call the stub.
5738   address stubAddr = StubRoutines::updateBytesCRC32C();
5739   const char *stubName = "updateBytesCRC32C";
5740 
5741   Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5742                                  stubAddr, stubName, TypePtr::BOTTOM,
5743                                  crc, src_start, length, table_start);
5744   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5745   set_result(result);
5746   return true;
5747 }
5748 
5749 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5750 //
5751 // Calculate CRC32C for DirectByteBuffer.
5752 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5753 //
5754 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5755   assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5756   assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5757   assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5758   // no receiver since it is a static method
5759   Node* crc     = argument(0); // type: int
5760   Node* src     = argument(1); // type: long
5761   Node* offset  = argument(3); // type: int
5762   Node* end     = argument(4); // type: int
5763 
5764   Node* length = _gvn.transform(new SubINode(end, offset));
5765 
5766   src = ConvL2X(src);  // adjust Java long to machine word
5767   Node* base = _gvn.transform(new CastX2PNode(src));
5768   offset = ConvI2X(offset);
5769 
5770   // 'src_start' points to src array + scaled offset
5771   Node* src_start = basic_plus_adr(top(), base, offset);
5772 
5773   // static final int[] byteTable in class CRC32C
5774   Node* table = get_table_from_crc32c_class(callee()->holder());
5775   Node* table_start = array_element_address(table, intcon(0), T_INT);
5776 
5777   // Call the stub.
5778   address stubAddr = StubRoutines::updateBytesCRC32C();
5779   const char *stubName = "updateBytesCRC32C";
5780 
5781   Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5782                                  stubAddr, stubName, TypePtr::BOTTOM,
5783                                  crc, src_start, length, table_start);
5784   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5785   set_result(result);
5786   return true;
5787 }
5788 
5789 //----------------------------inline_reference_get----------------------------
5790 // public T java.lang.ref.Reference.get();
5791 bool LibraryCallKit::inline_reference_get() {
5792   const int referent_offset = java_lang_ref_Reference::referent_offset;
5793   guarantee(referent_offset > 0, "should have already been set");
5794 
5795   // Get the argument:
5796   Node* reference_obj = null_check_receiver();
5797   if (stopped()) return true;
5798 
5799   Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5800 
5801   ciInstanceKlass* klass = env()->Object_klass();
5802   const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5803 
5804   Node* no_ctrl = NULL;
5805   Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5806 
5807   // Use the pre-barrier to record the value in the referent field
5808   pre_barrier(false /* do_load */,
5809               control(),
5810               NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5811               result /* pre_val */,
5812               T_OBJECT);
5813 
5814   // Add memory barrier to prevent commoning reads from this field
5815   // across safepoint since GC can change its value.
5816   insert_mem_bar(Op_MemBarCPUOrder);
5817 
5818   set_result(result);
5819   return true;
5820 }
5821 
5822 
5823 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5824                                               bool is_exact=true, bool is_static=false,
5825                                               ciInstanceKlass * fromKls=NULL) {
5826   if (fromKls == NULL) {
5827     const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5828     assert(tinst != NULL, "obj is null");
5829     assert(tinst->klass()->is_loaded(), "obj is not loaded");
5830     assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5831     fromKls = tinst->klass()->as_instance_klass();
5832   } else {
5833     assert(is_static, "only for static field access");
5834   }
5835   ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5836                                               ciSymbol::make(fieldTypeString),
5837                                               is_static);
5838 
5839   assert (field != NULL, "undefined field");
5840   if (field == NULL) return (Node *) NULL;
5841 
5842   if (is_static) {
5843     const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5844     fromObj = makecon(tip);
5845   }
5846 
5847   // Next code  copied from Parse::do_get_xxx():
5848 
5849   // Compute address and memory type.
5850   int offset  = field->offset_in_bytes();
5851   bool is_vol = field->is_volatile();
5852   ciType* field_klass = field->type();
5853   assert(field_klass->is_loaded(), "should be loaded");
5854   const TypePtr* adr_type = C->alias_type(field)->adr_type();
5855   Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5856   BasicType bt = field->layout_type();
5857 
5858   // Build the resultant type of the load
5859   const Type *type;
5860   if (bt == T_OBJECT) {
5861     type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5862   } else {
5863     type = Type::get_const_basic_type(bt);
5864   }
5865 
5866   if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5867     insert_mem_bar(Op_MemBarVolatile);   // StoreLoad barrier
5868   }
5869   // Build the load.
5870   MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5871   Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5872   // If reference is volatile, prevent following memory ops from
5873   // floating up past the volatile read.  Also prevents commoning
5874   // another volatile read.
5875   if (is_vol) {
5876     // Memory barrier includes bogus read of value to force load BEFORE membar
5877     insert_mem_bar(Op_MemBarAcquire, loadedField);
5878   }
5879   return loadedField;
5880 }
5881 
5882 
5883 //------------------------------inline_aescrypt_Block-----------------------
5884 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5885   address stubAddr;
5886   const char *stubName;
5887   assert(UseAES, "need AES instruction support");
5888 
5889   switch(id) {
5890   case vmIntrinsics::_aescrypt_encryptBlock:
5891     stubAddr = StubRoutines::aescrypt_encryptBlock();
5892     stubName = "aescrypt_encryptBlock";
5893     break;
5894   case vmIntrinsics::_aescrypt_decryptBlock:
5895     stubAddr = StubRoutines::aescrypt_decryptBlock();
5896     stubName = "aescrypt_decryptBlock";
5897     break;
5898   }
5899   if (stubAddr == NULL) return false;
5900 
5901   Node* aescrypt_object = argument(0);
5902   Node* src             = argument(1);
5903   Node* src_offset      = argument(2);
5904   Node* dest            = argument(3);
5905   Node* dest_offset     = argument(4);
5906 
5907   // (1) src and dest are arrays.
5908   const Type* src_type = src->Value(&_gvn);
5909   const Type* dest_type = dest->Value(&_gvn);
5910   const TypeAryPtr* top_src = src_type->isa_aryptr();
5911   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5912   assert (top_src  != NULL && top_src->klass()  != NULL &&  top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5913 
5914   // for the quick and dirty code we will skip all the checks.
5915   // we are just trying to get the call to be generated.
5916   Node* src_start  = src;
5917   Node* dest_start = dest;
5918   if (src_offset != NULL || dest_offset != NULL) {
5919     assert(src_offset != NULL && dest_offset != NULL, "");
5920     src_start  = array_element_address(src,  src_offset,  T_BYTE);
5921     dest_start = array_element_address(dest, dest_offset, T_BYTE);
5922   }
5923 
5924   // now need to get the start of its expanded key array
5925   // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5926   Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5927   if (k_start == NULL) return false;
5928 
5929   if (Matcher::pass_original_key_for_aes()) {
5930     // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5931     // compatibility issues between Java key expansion and SPARC crypto instructions
5932     Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5933     if (original_k_start == NULL) return false;
5934 
5935     // Call the stub.
5936     make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5937                       stubAddr, stubName, TypePtr::BOTTOM,
5938                       src_start, dest_start, k_start, original_k_start);
5939   } else {
5940     // Call the stub.
5941     make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5942                       stubAddr, stubName, TypePtr::BOTTOM,
5943                       src_start, dest_start, k_start);
5944   }
5945 
5946   return true;
5947 }
5948 
5949 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5950 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5951   address stubAddr;
5952   const char *stubName;
5953 
5954   assert(UseAES, "need AES instruction support");
5955 
5956   switch(id) {
5957   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5958     stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5959     stubName = "cipherBlockChaining_encryptAESCrypt";
5960     break;
5961   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5962     stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5963     stubName = "cipherBlockChaining_decryptAESCrypt";
5964     break;
5965   }
5966   if (stubAddr == NULL) return false;
5967 
5968   Node* cipherBlockChaining_object = argument(0);
5969   Node* src                        = argument(1);
5970   Node* src_offset                 = argument(2);
5971   Node* len                        = argument(3);
5972   Node* dest                       = argument(4);
5973   Node* dest_offset                = argument(5);
5974 
5975   // (1) src and dest are arrays.
5976   const Type* src_type = src->Value(&_gvn);
5977   const Type* dest_type = dest->Value(&_gvn);
5978   const TypeAryPtr* top_src = src_type->isa_aryptr();
5979   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5980   assert (top_src  != NULL && top_src->klass()  != NULL
5981           &&  top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5982 
5983   // checks are the responsibility of the caller
5984   Node* src_start  = src;
5985   Node* dest_start = dest;
5986   if (src_offset != NULL || dest_offset != NULL) {
5987     assert(src_offset != NULL && dest_offset != NULL, "");
5988     src_start  = array_element_address(src,  src_offset,  T_BYTE);
5989     dest_start = array_element_address(dest, dest_offset, T_BYTE);
5990   }
5991 
5992   // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5993   // (because of the predicated logic executed earlier).
5994   // so we cast it here safely.
5995   // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5996 
5997   Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5998   if (embeddedCipherObj == NULL) return false;
5999 
6000   // cast it to what we know it will be at runtime
6001   const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6002   assert(tinst != NULL, "CBC obj is null");
6003   assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6004   ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6005   assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6006 
6007   ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6008   const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6009   const TypeOopPtr* xtype = aklass->as_instance_type();
6010   Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6011   aescrypt_object = _gvn.transform(aescrypt_object);
6012 
6013   // we need to get the start of the aescrypt_object's expanded key array
6014   Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6015   if (k_start == NULL) return false;
6016 
6017   // similarly, get the start address of the r vector
6018   Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6019   if (objRvec == NULL) return false;
6020   Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6021 
6022   Node* cbcCrypt;
6023   if (Matcher::pass_original_key_for_aes()) {
6024     // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6025     // compatibility issues between Java key expansion and SPARC crypto instructions
6026     Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6027     if (original_k_start == NULL) return false;
6028 
6029     // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6030     cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6031                                  OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6032                                  stubAddr, stubName, TypePtr::BOTTOM,
6033                                  src_start, dest_start, k_start, r_start, len, original_k_start);
6034   } else {
6035     // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6036     cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6037                                  OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6038                                  stubAddr, stubName, TypePtr::BOTTOM,
6039                                  src_start, dest_start, k_start, r_start, len);
6040   }
6041 
6042   // return cipher length (int)
6043   Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6044   set_result(retvalue);
6045   return true;
6046 }
6047 
6048 //------------------------------get_key_start_from_aescrypt_object-----------------------
6049 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6050   Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6051   assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6052   if (objAESCryptKey == NULL) return (Node *) NULL;
6053 
6054   // now have the array, need to get the start address of the K array
6055   Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6056   return k_start;
6057 }
6058 
6059 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6060 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6061   Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6062   assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6063   if (objAESCryptKey == NULL) return (Node *) NULL;
6064 
6065   // now have the array, need to get the start address of the lastKey array
6066   Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6067   return original_k_start;
6068 }
6069 
6070 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6071 // Return node representing slow path of predicate check.
6072 // the pseudo code we want to emulate with this predicate is:
6073 // for encryption:
6074 //    if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6075 // for decryption:
6076 //    if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6077 //    note cipher==plain is more conservative than the original java code but that's OK
6078 //
6079 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6080   // The receiver was checked for NULL already.
6081   Node* objCBC = argument(0);
6082 
6083   // Load embeddedCipher field of CipherBlockChaining object.
6084   Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6085 
6086   // get AESCrypt klass for instanceOf check
6087   // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6088   // will have same classloader as CipherBlockChaining object
6089   const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6090   assert(tinst != NULL, "CBCobj is null");
6091   assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6092 
6093   // we want to do an instanceof comparison against the AESCrypt class
6094   ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6095   if (!klass_AESCrypt->is_loaded()) {
6096     // if AESCrypt is not even loaded, we never take the intrinsic fast path
6097     Node* ctrl = control();
6098     set_control(top()); // no regular fast path
6099     return ctrl;
6100   }
6101   ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6102 
6103   Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6104   Node* cmp_instof  = _gvn.transform(new CmpINode(instof, intcon(1)));
6105   Node* bool_instof  = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6106 
6107   Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6108 
6109   // for encryption, we are done
6110   if (!decrypting)
6111     return instof_false;  // even if it is NULL
6112 
6113   // for decryption, we need to add a further check to avoid
6114   // taking the intrinsic path when cipher and plain are the same
6115   // see the original java code for why.
6116   RegionNode* region = new RegionNode(3);
6117   region->init_req(1, instof_false);
6118   Node* src = argument(1);
6119   Node* dest = argument(4);
6120   Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6121   Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6122   Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6123   region->init_req(2, src_dest_conjoint);
6124 
6125   record_for_igvn(region);
6126   return _gvn.transform(region);
6127 }
6128 
6129 //------------------------------inline_ghash_processBlocks
6130 bool LibraryCallKit::inline_ghash_processBlocks() {
6131   address stubAddr;
6132   const char *stubName;
6133   assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6134 
6135   stubAddr = StubRoutines::ghash_processBlocks();
6136   stubName = "ghash_processBlocks";
6137 
6138   Node* data           = argument(0);
6139   Node* offset         = argument(1);
6140   Node* len            = argument(2);
6141   Node* state          = argument(3);
6142   Node* subkeyH        = argument(4);
6143 
6144   Node* state_start  = array_element_address(state, intcon(0), T_LONG);
6145   assert(state_start, "state is NULL");
6146   Node* subkeyH_start  = array_element_address(subkeyH, intcon(0), T_LONG);
6147   assert(subkeyH_start, "subkeyH is NULL");
6148   Node* data_start  = array_element_address(data, offset, T_BYTE);
6149   assert(data_start, "data is NULL");
6150 
6151   Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6152                                   OptoRuntime::ghash_processBlocks_Type(),
6153                                   stubAddr, stubName, TypePtr::BOTTOM,
6154                                   state_start, subkeyH_start, data_start, len);
6155   return true;
6156 }
6157 
6158 //------------------------------inline_sha_implCompress-----------------------
6159 //
6160 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6161 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6162 //
6163 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6164 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6165 //
6166 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6167 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6168 //
6169 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6170   assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6171 
6172   Node* sha_obj = argument(0);
6173   Node* src     = argument(1); // type oop
6174   Node* ofs     = argument(2); // type int
6175 
6176   const Type* src_type = src->Value(&_gvn);
6177   const TypeAryPtr* top_src = src_type->isa_aryptr();
6178   if (top_src  == NULL || top_src->klass()  == NULL) {
6179     // failed array check
6180     return false;
6181   }
6182   // Figure out the size and type of the elements we will be copying.
6183   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6184   if (src_elem != T_BYTE) {
6185     return false;
6186   }
6187   // 'src_start' points to src array + offset
6188   Node* src_start = array_element_address(src, ofs, src_elem);
6189   Node* state = NULL;
6190   address stubAddr;
6191   const char *stubName;
6192 
6193   switch(id) {
6194   case vmIntrinsics::_sha_implCompress:
6195     assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6196     state = get_state_from_sha_object(sha_obj);
6197     stubAddr = StubRoutines::sha1_implCompress();
6198     stubName = "sha1_implCompress";
6199     break;
6200   case vmIntrinsics::_sha2_implCompress:
6201     assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6202     state = get_state_from_sha_object(sha_obj);
6203     stubAddr = StubRoutines::sha256_implCompress();
6204     stubName = "sha256_implCompress";
6205     break;
6206   case vmIntrinsics::_sha5_implCompress:
6207     assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6208     state = get_state_from_sha5_object(sha_obj);
6209     stubAddr = StubRoutines::sha512_implCompress();
6210     stubName = "sha512_implCompress";
6211     break;
6212   default:
6213     fatal_unexpected_iid(id);
6214     return false;
6215   }
6216   if (state == NULL) return false;
6217 
6218   // Call the stub.
6219   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6220                                  stubAddr, stubName, TypePtr::BOTTOM,
6221                                  src_start, state);
6222 
6223   return true;
6224 }
6225 
6226 //------------------------------inline_digestBase_implCompressMB-----------------------
6227 //
6228 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6229 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6230 //
6231 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6232   assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6233          "need SHA1/SHA256/SHA512 instruction support");
6234   assert((uint)predicate < 3, "sanity");
6235   assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6236 
6237   Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6238   Node* src            = argument(1); // byte[] array
6239   Node* ofs            = argument(2); // type int
6240   Node* limit          = argument(3); // type int
6241 
6242   const Type* src_type = src->Value(&_gvn);
6243   const TypeAryPtr* top_src = src_type->isa_aryptr();
6244   if (top_src  == NULL || top_src->klass()  == NULL) {
6245     // failed array check
6246     return false;
6247   }
6248   // Figure out the size and type of the elements we will be copying.
6249   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6250   if (src_elem != T_BYTE) {
6251     return false;
6252   }
6253   // 'src_start' points to src array + offset
6254   Node* src_start = array_element_address(src, ofs, src_elem);
6255 
6256   const char* klass_SHA_name = NULL;
6257   const char* stub_name = NULL;
6258   address     stub_addr = NULL;
6259   bool        long_state = false;
6260 
6261   switch (predicate) {
6262   case 0:
6263     if (UseSHA1Intrinsics) {
6264       klass_SHA_name = "sun/security/provider/SHA";
6265       stub_name = "sha1_implCompressMB";
6266       stub_addr = StubRoutines::sha1_implCompressMB();
6267     }
6268     break;
6269   case 1:
6270     if (UseSHA256Intrinsics) {
6271       klass_SHA_name = "sun/security/provider/SHA2";
6272       stub_name = "sha256_implCompressMB";
6273       stub_addr = StubRoutines::sha256_implCompressMB();
6274     }
6275     break;
6276   case 2:
6277     if (UseSHA512Intrinsics) {
6278       klass_SHA_name = "sun/security/provider/SHA5";
6279       stub_name = "sha512_implCompressMB";
6280       stub_addr = StubRoutines::sha512_implCompressMB();
6281       long_state = true;
6282     }
6283     break;
6284   default:
6285     fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6286   }
6287   if (klass_SHA_name != NULL) {
6288     // get DigestBase klass to lookup for SHA klass
6289     const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6290     assert(tinst != NULL, "digestBase_obj is not instance???");
6291     assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6292 
6293     ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6294     assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6295     ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6296     return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6297   }
6298   return false;
6299 }
6300 //------------------------------inline_sha_implCompressMB-----------------------
6301 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6302                                                bool long_state, address stubAddr, const char *stubName,
6303                                                Node* src_start, Node* ofs, Node* limit) {
6304   const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6305   const TypeOopPtr* xtype = aklass->as_instance_type();
6306   Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6307   sha_obj = _gvn.transform(sha_obj);
6308 
6309   Node* state;
6310   if (long_state) {
6311     state = get_state_from_sha5_object(sha_obj);
6312   } else {
6313     state = get_state_from_sha_object(sha_obj);
6314   }
6315   if (state == NULL) return false;
6316 
6317   // Call the stub.
6318   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6319                                  OptoRuntime::digestBase_implCompressMB_Type(),
6320                                  stubAddr, stubName, TypePtr::BOTTOM,
6321                                  src_start, state, ofs, limit);
6322   // return ofs (int)
6323   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6324   set_result(result);
6325 
6326   return true;
6327 }
6328 
6329 //------------------------------get_state_from_sha_object-----------------------
6330 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6331   Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6332   assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6333   if (sha_state == NULL) return (Node *) NULL;
6334 
6335   // now have the array, need to get the start address of the state array
6336   Node* state = array_element_address(sha_state, intcon(0), T_INT);
6337   return state;
6338 }
6339 
6340 //------------------------------get_state_from_sha5_object-----------------------
6341 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6342   Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6343   assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6344   if (sha_state == NULL) return (Node *) NULL;
6345 
6346   // now have the array, need to get the start address of the state array
6347   Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6348   return state;
6349 }
6350 
6351 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6352 // Return node representing slow path of predicate check.
6353 // the pseudo code we want to emulate with this predicate is:
6354 //    if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6355 //
6356 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6357   assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6358          "need SHA1/SHA256/SHA512 instruction support");
6359   assert((uint)predicate < 3, "sanity");
6360 
6361   // The receiver was checked for NULL already.
6362   Node* digestBaseObj = argument(0);
6363 
6364   // get DigestBase klass for instanceOf check
6365   const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6366   assert(tinst != NULL, "digestBaseObj is null");
6367   assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6368 
6369   const char* klass_SHA_name = NULL;
6370   switch (predicate) {
6371   case 0:
6372     if (UseSHA1Intrinsics) {
6373       // we want to do an instanceof comparison against the SHA class
6374       klass_SHA_name = "sun/security/provider/SHA";
6375     }
6376     break;
6377   case 1:
6378     if (UseSHA256Intrinsics) {
6379       // we want to do an instanceof comparison against the SHA2 class
6380       klass_SHA_name = "sun/security/provider/SHA2";
6381     }
6382     break;
6383   case 2:
6384     if (UseSHA512Intrinsics) {
6385       // we want to do an instanceof comparison against the SHA5 class
6386       klass_SHA_name = "sun/security/provider/SHA5";
6387     }
6388     break;
6389   default:
6390     fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6391   }
6392 
6393   ciKlass* klass_SHA = NULL;
6394   if (klass_SHA_name != NULL) {
6395     klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6396   }
6397   if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6398     // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6399     Node* ctrl = control();
6400     set_control(top()); // no intrinsic path
6401     return ctrl;
6402   }
6403   ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6404 
6405   Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6406   Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6407   Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6408   Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6409 
6410   return instof_false;  // even if it is NULL
6411 }
6412 
6413 bool LibraryCallKit::inline_profileBoolean() {
6414   Node* counts = argument(1);
6415   const TypeAryPtr* ary = NULL;
6416   ciArray* aobj = NULL;
6417   if (counts->is_Con()
6418       && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6419       && (aobj = ary->const_oop()->as_array()) != NULL
6420       && (aobj->length() == 2)) {
6421     // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6422     jint false_cnt = aobj->element_value(0).as_int();
6423     jint  true_cnt = aobj->element_value(1).as_int();
6424 
6425     if (C->log() != NULL) {
6426       C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6427                      false_cnt, true_cnt);
6428     }
6429 
6430     if (false_cnt + true_cnt == 0) {
6431       // According to profile, never executed.
6432       uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6433                           Deoptimization::Action_reinterpret);
6434       return true;
6435     }
6436 
6437     // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6438     // is a number of each value occurrences.
6439     Node* result = argument(0);
6440     if (false_cnt == 0 || true_cnt == 0) {
6441       // According to profile, one value has been never seen.
6442       int expected_val = (false_cnt == 0) ? 1 : 0;
6443 
6444       Node* cmp  = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6445       Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6446 
6447       IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6448       Node* fast_path = _gvn.transform(new IfTrueNode(check));
6449       Node* slow_path = _gvn.transform(new IfFalseNode(check));
6450 
6451       { // Slow path: uncommon trap for never seen value and then reexecute
6452         // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6453         // the value has been seen at least once.
6454         PreserveJVMState pjvms(this);
6455         PreserveReexecuteState preexecs(this);
6456         jvms()->set_should_reexecute(true);
6457 
6458         set_control(slow_path);
6459         set_i_o(i_o());
6460 
6461         uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6462                             Deoptimization::Action_reinterpret);
6463       }
6464       // The guard for never seen value enables sharpening of the result and
6465       // returning a constant. It allows to eliminate branches on the same value
6466       // later on.
6467       set_control(fast_path);
6468       result = intcon(expected_val);
6469     }
6470     // Stop profiling.
6471     // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6472     // By replacing method body with profile data (represented as ProfileBooleanNode
6473     // on IR level) we effectively disable profiling.
6474     // It enables full speed execution once optimized code is generated.
6475     Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6476     C->record_for_igvn(profile);
6477     set_result(profile);
6478     return true;
6479   } else {
6480     // Continue profiling.
6481     // Profile data isn't available at the moment. So, execute method's bytecode version.
6482     // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6483     // is compiled and counters aren't available since corresponding MethodHandle
6484     // isn't a compile-time constant.
6485     return false;
6486   }
6487 }
6488 
6489 bool LibraryCallKit::inline_isCompileConstant() {
6490   Node* n = argument(0);
6491   set_result(n->is_Con() ? intcon(1) : intcon(0));
6492   return true;
6493 }