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