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 "prims/unsafe.hpp"
  51 #include "runtime/sharedRuntime.hpp"
  52 #include "trace/traceMacros.hpp"
  53 
  54 class LibraryIntrinsic : public InlineCallGenerator {
  55   // Extend the set of intrinsics known to the runtime:
  56  public:
  57  private:
  58   bool             _is_virtual;
  59   bool             _does_virtual_dispatch;
  60   int8_t           _predicates_count;  // Intrinsic is predicated by several conditions
  61   int8_t           _last_predicate; // Last generated predicate
  62   vmIntrinsics::ID _intrinsic_id;
  63 
  64  public:
  65   LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
  66     : InlineCallGenerator(m),
  67       _is_virtual(is_virtual),
  68       _does_virtual_dispatch(does_virtual_dispatch),
  69       _predicates_count((int8_t)predicates_count),
  70       _last_predicate((int8_t)-1),
  71       _intrinsic_id(id)
  72   {
  73   }
  74   virtual bool is_intrinsic() const { return true; }
  75   virtual bool is_virtual()   const { return _is_virtual; }
  76   virtual bool is_predicated() const { return _predicates_count > 0; }
  77   virtual int  predicates_count() const { return _predicates_count; }
  78   virtual bool does_virtual_dispatch()   const { return _does_virtual_dispatch; }
  79   virtual JVMState* generate(JVMState* jvms);
  80   virtual Node* generate_predicate(JVMState* jvms, int predicate);
  81   vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
  82 };
  83 
  84 
  85 // Local helper class for LibraryIntrinsic:
  86 class LibraryCallKit : public GraphKit {
  87  private:
  88   LibraryIntrinsic* _intrinsic;     // the library intrinsic being called
  89   Node*             _result;        // the result node, if any
  90   int               _reexecute_sp;  // the stack pointer when bytecode needs to be reexecuted
  91 
  92   const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
  93 
  94  public:
  95   LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
  96     : GraphKit(jvms),
  97       _intrinsic(intrinsic),
  98       _result(NULL)
  99   {
 100     // Check if this is a root compile.  In that case we don't have a caller.
 101     if (!jvms->has_method()) {
 102       _reexecute_sp = sp();
 103     } else {
 104       // Find out how many arguments the interpreter needs when deoptimizing
 105       // and save the stack pointer value so it can used by uncommon_trap.
 106       // We find the argument count by looking at the declared signature.
 107       bool ignored_will_link;
 108       ciSignature* declared_signature = NULL;
 109       ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
 110       const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
 111       _reexecute_sp = sp() + nargs;  // "push" arguments back on stack
 112     }
 113   }
 114 
 115   virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
 116 
 117   ciMethod*         caller()    const    { return jvms()->method(); }
 118   int               bci()       const    { return jvms()->bci(); }
 119   LibraryIntrinsic* intrinsic() const    { return _intrinsic; }
 120   vmIntrinsics::ID  intrinsic_id() const { return _intrinsic->intrinsic_id(); }
 121   ciMethod*         callee()    const    { return _intrinsic->method(); }
 122 
 123   bool  try_to_inline(int predicate);
 124   Node* try_to_predicate(int predicate);
 125 
 126   void push_result() {
 127     // Push the result onto the stack.
 128     if (!stopped() && result() != NULL) {
 129       BasicType bt = result()->bottom_type()->basic_type();
 130       push_node(bt, result());
 131     }
 132   }
 133 
 134  private:
 135   void fatal_unexpected_iid(vmIntrinsics::ID iid) {
 136     fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
 137   }
 138 
 139   void  set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
 140   void  set_result(RegionNode* region, PhiNode* value);
 141   Node*     result() { return _result; }
 142 
 143   virtual int reexecute_sp() { return _reexecute_sp; }
 144 
 145   // Helper functions to inline natives
 146   Node* generate_guard(Node* test, RegionNode* region, float true_prob);
 147   Node* generate_slow_guard(Node* test, RegionNode* region);
 148   Node* generate_fair_guard(Node* test, RegionNode* region);
 149   Node* generate_negative_guard(Node* index, RegionNode* region,
 150                                 // resulting CastII of index:
 151                                 Node* *pos_index = NULL);
 152   Node* generate_limit_guard(Node* offset, Node* subseq_length,
 153                              Node* array_length,
 154                              RegionNode* region);
 155   Node* generate_current_thread(Node* &tls_output);
 156   Node* load_mirror_from_klass(Node* klass);
 157   Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
 158                                       RegionNode* region, int null_path,
 159                                       int offset);
 160   Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
 161                                RegionNode* region, int null_path) {
 162     int offset = java_lang_Class::klass_offset_in_bytes();
 163     return load_klass_from_mirror_common(mirror, never_see_null,
 164                                          region, null_path,
 165                                          offset);
 166   }
 167   Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
 168                                      RegionNode* region, int null_path) {
 169     int offset = java_lang_Class::array_klass_offset_in_bytes();
 170     return load_klass_from_mirror_common(mirror, never_see_null,
 171                                          region, null_path,
 172                                          offset);
 173   }
 174   Node* generate_access_flags_guard(Node* kls,
 175                                     int modifier_mask, int modifier_bits,
 176                                     RegionNode* region);
 177   Node* generate_interface_guard(Node* kls, RegionNode* region);
 178   Node* generate_array_guard(Node* kls, RegionNode* region) {
 179     return generate_array_guard_common(kls, region, false, false);
 180   }
 181   Node* generate_non_array_guard(Node* kls, RegionNode* region) {
 182     return generate_array_guard_common(kls, region, false, true);
 183   }
 184   Node* generate_objArray_guard(Node* kls, RegionNode* region) {
 185     return generate_array_guard_common(kls, region, true, false);
 186   }
 187   Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
 188     return generate_array_guard_common(kls, region, true, true);
 189   }
 190   Node* generate_array_guard_common(Node* kls, RegionNode* region,
 191                                     bool obj_array, bool not_array);
 192   Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
 193   CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
 194                                      bool is_virtual = false, bool is_static = false);
 195   CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
 196     return generate_method_call(method_id, false, true);
 197   }
 198   CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
 199     return generate_method_call(method_id, true, false);
 200   }
 201   Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
 202 
 203   Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
 204   Node* make_string_method_node(int opcode, Node* str1, Node* str2);
 205   bool inline_string_compareTo();
 206   bool inline_string_indexOf();
 207   Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
 208   bool inline_string_equals();
 209   Node* round_double_node(Node* n);
 210   bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
 211   bool inline_math_native(vmIntrinsics::ID id);
 212   bool inline_trig(vmIntrinsics::ID id);
 213   bool inline_math(vmIntrinsics::ID id);
 214   template <typename OverflowOp>
 215   bool inline_math_overflow(Node* arg1, Node* arg2);
 216   void inline_math_mathExact(Node* math, Node* test);
 217   bool inline_math_addExactI(bool is_increment);
 218   bool inline_math_addExactL(bool is_increment);
 219   bool inline_math_multiplyExactI();
 220   bool inline_math_multiplyExactL();
 221   bool inline_math_negateExactI();
 222   bool inline_math_negateExactL();
 223   bool inline_math_subtractExactI(bool is_decrement);
 224   bool inline_math_subtractExactL(bool is_decrement);
 225   bool inline_exp();
 226   bool inline_pow();
 227   Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
 228   bool inline_min_max(vmIntrinsics::ID id);
 229   Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
 230   // This returns Type::AnyPtr, RawPtr, or OopPtr.
 231   int classify_unsafe_addr(Node* &base, Node* &offset, bool decode_offset);
 232   Node* make_unsafe_address(Node* base, Node* offset, bool decode_offset);
 233   // Helper for inline_unsafe_access.
 234   // Generates the guards that check whether the result of
 235   // Unsafe.getObject should be recorded in an SATB log buffer.
 236   void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
 237   bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned = false);
 238   static bool klass_needs_init_guard(Node* kls);
 239   bool inline_unsafe_allocate();
 240   bool inline_unsafe_copyMemory();
 241   bool inline_native_currentThread();
 242 #ifdef TRACE_HAVE_INTRINSICS
 243   bool inline_native_classID();
 244   bool inline_native_threadID();
 245 #endif
 246   bool inline_native_time_funcs(address method, const char* funcName);
 247   bool inline_native_isInterrupted();
 248   bool inline_native_Class_query(vmIntrinsics::ID id);
 249   bool inline_native_subtype_check();
 250 
 251   bool inline_native_newArray();
 252   bool inline_native_getLength();
 253   bool inline_array_copyOf(bool is_copyOfRange);
 254   bool inline_array_equals();
 255   void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
 256   bool inline_native_clone(bool is_virtual);
 257   bool inline_native_Reflection_getCallerClass();
 258   // Helper function for inlining native object hash method
 259   bool inline_native_hashcode(bool is_virtual, bool is_static);
 260   bool inline_native_getClass();
 261 
 262   // Helper functions for inlining arraycopy
 263   bool inline_arraycopy();
 264   AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
 265                                                 RegionNode* slow_region);
 266   JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
 267   void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
 268 
 269   typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
 270   bool inline_unsafe_load_store(BasicType type,  LoadStoreKind kind);
 271   bool inline_unsafe_ordered_store(BasicType type);
 272   bool inline_unsafe_fence(vmIntrinsics::ID id);
 273   bool inline_fp_conversions(vmIntrinsics::ID id);
 274   bool inline_number_methods(vmIntrinsics::ID id);
 275   bool inline_reference_get();
 276   bool inline_Class_cast();
 277   bool inline_aescrypt_Block(vmIntrinsics::ID id);
 278   bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
 279   Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
 280   Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
 281   Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
 282   bool inline_sha_implCompress(vmIntrinsics::ID id);
 283   bool inline_digestBase_implCompressMB(int predicate);
 284   bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
 285                                  bool long_state, address stubAddr, const char *stubName,
 286                                  Node* src_start, Node* ofs, Node* limit);
 287   Node* get_state_from_sha_object(Node *sha_object);
 288   Node* get_state_from_sha5_object(Node *sha_object);
 289   Node* inline_digestBase_implCompressMB_predicate(int predicate);
 290   bool inline_encodeISOArray();
 291   bool inline_updateCRC32();
 292   bool inline_updateBytesCRC32();
 293   bool inline_updateByteBufferCRC32();
 294   bool inline_multiplyToLen();
 295   bool inline_squareToLen();
 296   bool inline_mulAdd();
 297 
 298   bool inline_profileBoolean();
 299   bool inline_isCompileConstant();
 300 };
 301 
 302 
 303 //---------------------------make_vm_intrinsic----------------------------
 304 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
 305   vmIntrinsics::ID id = m->intrinsic_id();
 306   assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
 307 
 308   ccstr disable_intr = NULL;
 309 
 310   if ((DisableIntrinsic[0] != '\0'
 311        && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
 312       (method_has_option_value("DisableIntrinsic", disable_intr)
 313        && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
 314     // disabled by a user request on the command line:
 315     // example: -XX:DisableIntrinsic=_hashCode,_getClass
 316     return NULL;
 317   }
 318 
 319   if (!m->is_loaded()) {
 320     // do not attempt to inline unloaded methods
 321     return NULL;
 322   }
 323 
 324   // Only a few intrinsics implement a virtual dispatch.
 325   // They are expensive calls which are also frequently overridden.
 326   if (is_virtual) {
 327     switch (id) {
 328     case vmIntrinsics::_hashCode:
 329     case vmIntrinsics::_clone:
 330       // OK, Object.hashCode and Object.clone intrinsics come in both flavors
 331       break;
 332     default:
 333       return NULL;
 334     }
 335   }
 336 
 337   // -XX:-InlineNatives disables nearly all intrinsics:
 338   if (!InlineNatives) {
 339     switch (id) {
 340     case vmIntrinsics::_indexOf:
 341     case vmIntrinsics::_compareTo:
 342     case vmIntrinsics::_equals:
 343     case vmIntrinsics::_equalsC:
 344     case vmIntrinsics::_getAndAddInt:
 345     case vmIntrinsics::_getAndAddLong:
 346     case vmIntrinsics::_getAndSetInt:
 347     case vmIntrinsics::_getAndSetLong:
 348     case vmIntrinsics::_getAndSetObject:
 349     case vmIntrinsics::_loadFence:
 350     case vmIntrinsics::_storeFence:
 351     case vmIntrinsics::_fullFence:
 352       break;  // InlineNatives does not control String.compareTo
 353     case vmIntrinsics::_Reference_get:
 354       break;  // InlineNatives does not control Reference.get
 355     default:
 356       return NULL;
 357     }
 358   }
 359 
 360   int predicates = 0;
 361   bool does_virtual_dispatch = false;
 362 
 363   switch (id) {
 364   case vmIntrinsics::_compareTo:
 365     if (!SpecialStringCompareTo)  return NULL;
 366     if (!Matcher::match_rule_supported(Op_StrComp))  return NULL;
 367     break;
 368   case vmIntrinsics::_indexOf:
 369     if (!SpecialStringIndexOf)  return NULL;
 370     break;
 371   case vmIntrinsics::_equals:
 372     if (!SpecialStringEquals)  return NULL;
 373     if (!Matcher::match_rule_supported(Op_StrEquals))  return NULL;
 374     break;
 375   case vmIntrinsics::_equalsC:
 376     if (!SpecialArraysEquals)  return NULL;
 377     if (!Matcher::match_rule_supported(Op_AryEq))  return NULL;
 378     break;
 379   case vmIntrinsics::_arraycopy:
 380     if (!InlineArrayCopy)  return NULL;
 381     break;
 382   case vmIntrinsics::_copyMemory:
 383     if (StubRoutines::unsafe_arraycopy() == NULL)  return NULL;
 384     if (!InlineArrayCopy)  return NULL;
 385     break;
 386   case vmIntrinsics::_hashCode:
 387     if (!InlineObjectHash)  return NULL;
 388     does_virtual_dispatch = true;
 389     break;
 390   case vmIntrinsics::_clone:
 391     does_virtual_dispatch = true;
 392   case vmIntrinsics::_copyOf:
 393   case vmIntrinsics::_copyOfRange:
 394     if (!InlineObjectCopy)  return NULL;
 395     // These also use the arraycopy intrinsic mechanism:
 396     if (!InlineArrayCopy)  return NULL;
 397     break;
 398   case vmIntrinsics::_encodeISOArray:
 399     if (!SpecialEncodeISOArray)  return NULL;
 400     if (!Matcher::match_rule_supported(Op_EncodeISOArray))  return NULL;
 401     break;
 402   case vmIntrinsics::_checkIndex:
 403     // We do not intrinsify this.  The optimizer does fine with it.
 404     return NULL;
 405 
 406   case vmIntrinsics::_getCallerClass:
 407     if (!InlineReflectionGetCallerClass)  return NULL;
 408     if (SystemDictionary::reflect_CallerSensitive_klass() == NULL)  return NULL;
 409     break;
 410 
 411   case vmIntrinsics::_bitCount_i:
 412     if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
 413     break;
 414 
 415   case vmIntrinsics::_bitCount_l:
 416     if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
 417     break;
 418 
 419   case vmIntrinsics::_numberOfLeadingZeros_i:
 420     if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
 421     break;
 422 
 423   case vmIntrinsics::_numberOfLeadingZeros_l:
 424     if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
 425     break;
 426 
 427   case vmIntrinsics::_numberOfTrailingZeros_i:
 428     if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
 429     break;
 430 
 431   case vmIntrinsics::_numberOfTrailingZeros_l:
 432     if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
 433     break;
 434 
 435   case vmIntrinsics::_reverseBytes_c:
 436     if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
 437     break;
 438   case vmIntrinsics::_reverseBytes_s:
 439     if (!Matcher::match_rule_supported(Op_ReverseBytesS))  return NULL;
 440     break;
 441   case vmIntrinsics::_reverseBytes_i:
 442     if (!Matcher::match_rule_supported(Op_ReverseBytesI))  return NULL;
 443     break;
 444   case vmIntrinsics::_reverseBytes_l:
 445     if (!Matcher::match_rule_supported(Op_ReverseBytesL))  return NULL;
 446     break;
 447 
 448   case vmIntrinsics::_Reference_get:
 449     // Use the intrinsic version of Reference.get() so that the value in
 450     // the referent field can be registered by the G1 pre-barrier code.
 451     // Also add memory barrier to prevent commoning reads from this field
 452     // across safepoint since GC can change it value.
 453     break;
 454 
 455   case vmIntrinsics::_compareAndSwapObject:
 456 #ifdef _LP64
 457     if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
 458 #endif
 459     break;
 460 
 461   case vmIntrinsics::_compareAndSwapLong:
 462     if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
 463     break;
 464 
 465   case vmIntrinsics::_getAndAddInt:
 466     if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
 467     break;
 468 
 469   case vmIntrinsics::_getAndAddLong:
 470     if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
 471     break;
 472 
 473   case vmIntrinsics::_getAndSetInt:
 474     if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
 475     break;
 476 
 477   case vmIntrinsics::_getAndSetLong:
 478     if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
 479     break;
 480 
 481   case vmIntrinsics::_getAndSetObject:
 482 #ifdef _LP64
 483     if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
 484     if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
 485     break;
 486 #else
 487     if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
 488     break;
 489 #endif
 490 
 491   case vmIntrinsics::_aescrypt_encryptBlock:
 492   case vmIntrinsics::_aescrypt_decryptBlock:
 493     if (!UseAESIntrinsics) return NULL;
 494     break;
 495 
 496   case vmIntrinsics::_multiplyToLen:
 497     if (!UseMultiplyToLenIntrinsic) return NULL;
 498     break;
 499 
 500   case vmIntrinsics::_squareToLen:
 501     if (!UseSquareToLenIntrinsic) return NULL;
 502     break;
 503 
 504   case vmIntrinsics::_mulAdd:
 505     if (!UseMulAddIntrinsic) return NULL;
 506     break;
 507 
 508   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 509   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 510     if (!UseAESIntrinsics) return NULL;
 511     // these two require the predicated logic
 512     predicates = 1;
 513     break;
 514 
 515   case vmIntrinsics::_sha_implCompress:
 516     if (!UseSHA1Intrinsics) return NULL;
 517     break;
 518 
 519   case vmIntrinsics::_sha2_implCompress:
 520     if (!UseSHA256Intrinsics) return NULL;
 521     break;
 522 
 523   case vmIntrinsics::_sha5_implCompress:
 524     if (!UseSHA512Intrinsics) return NULL;
 525     break;
 526 
 527   case vmIntrinsics::_digestBase_implCompressMB:
 528     if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
 529     predicates = 3;
 530     break;
 531 
 532   case vmIntrinsics::_updateCRC32:
 533   case vmIntrinsics::_updateBytesCRC32:
 534   case vmIntrinsics::_updateByteBufferCRC32:
 535     if (!UseCRC32Intrinsics) return NULL;
 536     break;
 537 
 538   case vmIntrinsics::_incrementExactI:
 539   case vmIntrinsics::_addExactI:
 540     if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
 541     break;
 542   case vmIntrinsics::_incrementExactL:
 543   case vmIntrinsics::_addExactL:
 544     if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
 545     break;
 546   case vmIntrinsics::_decrementExactI:
 547   case vmIntrinsics::_subtractExactI:
 548     if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
 549     break;
 550   case vmIntrinsics::_decrementExactL:
 551   case vmIntrinsics::_subtractExactL:
 552     if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
 553     break;
 554   case vmIntrinsics::_negateExactI:
 555     if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
 556     break;
 557   case vmIntrinsics::_negateExactL:
 558     if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
 559     break;
 560   case vmIntrinsics::_multiplyExactI:
 561     if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
 562     break;
 563   case vmIntrinsics::_multiplyExactL:
 564     if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
 565     break;
 566 
 567   case vmIntrinsics::_getShortUnaligned:
 568   case vmIntrinsics::_getCharUnaligned:
 569   case vmIntrinsics::_getIntUnaligned:
 570   case vmIntrinsics::_getLongUnaligned:
 571   case vmIntrinsics::_putShortUnaligned:
 572   case vmIntrinsics::_putCharUnaligned:
 573   case vmIntrinsics::_putIntUnaligned:
 574   case vmIntrinsics::_putLongUnaligned:
 575     if (!UseUnalignedAccesses) return NULL;
 576     break;
 577 
 578  default:
 579     assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
 580     assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
 581     break;
 582   }
 583 
 584   // -XX:-InlineClassNatives disables natives from the Class class.
 585   // The flag applies to all reflective calls, notably Array.newArray
 586   // (visible to Java programmers as Array.newInstance).
 587   if (m->holder()->name() == ciSymbol::java_lang_Class() ||
 588       m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
 589     if (!InlineClassNatives)  return NULL;
 590   }
 591 
 592   // -XX:-InlineThreadNatives disables natives from the Thread class.
 593   if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
 594     if (!InlineThreadNatives)  return NULL;
 595   }
 596 
 597   // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
 598   if (m->holder()->name() == ciSymbol::java_lang_Math() ||
 599       m->holder()->name() == ciSymbol::java_lang_Float() ||
 600       m->holder()->name() == ciSymbol::java_lang_Double()) {
 601     if (!InlineMathNatives)  return NULL;
 602   }
 603 
 604   // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
 605   if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
 606     if (!InlineUnsafeOps)  return NULL;
 607   }
 608 
 609   return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
 610 }
 611 
 612 //----------------------register_library_intrinsics-----------------------
 613 // Initialize this file's data structures, for each Compile instance.
 614 void Compile::register_library_intrinsics() {
 615   // Nothing to do here.
 616 }
 617 
 618 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
 619   LibraryCallKit kit(jvms, this);
 620   Compile* C = kit.C;
 621   int nodes = C->unique();
 622 #ifndef PRODUCT
 623   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
 624     char buf[1000];
 625     const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
 626     tty->print_cr("Intrinsic %s", str);
 627   }
 628 #endif
 629   ciMethod* callee = kit.callee();
 630   const int bci    = kit.bci();
 631 
 632   // Try to inline the intrinsic.
 633   if (kit.try_to_inline(_last_predicate)) {
 634     if (C->print_intrinsics() || C->print_inlining()) {
 635       C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
 636     }
 637     C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
 638     if (C->log()) {
 639       C->log()->elem("intrinsic id='%s'%s nodes='%d'",
 640                      vmIntrinsics::name_at(intrinsic_id()),
 641                      (is_virtual() ? " virtual='1'" : ""),
 642                      C->unique() - nodes);
 643     }
 644     // Push the result from the inlined method onto the stack.
 645     kit.push_result();
 646     C->print_inlining_update(this);
 647     return kit.transfer_exceptions_into_jvms();
 648   }
 649 
 650   // The intrinsic bailed out
 651   if (C->print_intrinsics() || C->print_inlining()) {
 652     if (jvms->has_method()) {
 653       // Not a root compile.
 654       const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
 655       C->print_inlining(callee, jvms->depth() - 1, bci, msg);
 656     } else {
 657       // Root compile
 658       tty->print("Did not generate intrinsic %s%s at bci:%d in",
 659                vmIntrinsics::name_at(intrinsic_id()),
 660                (is_virtual() ? " (virtual)" : ""), bci);
 661     }
 662   }
 663   C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
 664   C->print_inlining_update(this);
 665   return NULL;
 666 }
 667 
 668 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
 669   LibraryCallKit kit(jvms, this);
 670   Compile* C = kit.C;
 671   int nodes = C->unique();
 672   _last_predicate = predicate;
 673 #ifndef PRODUCT
 674   assert(is_predicated() && predicate < predicates_count(), "sanity");
 675   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
 676     char buf[1000];
 677     const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
 678     tty->print_cr("Predicate for intrinsic %s", str);
 679   }
 680 #endif
 681   ciMethod* callee = kit.callee();
 682   const int bci    = kit.bci();
 683 
 684   Node* slow_ctl = kit.try_to_predicate(predicate);
 685   if (!kit.failing()) {
 686     if (C->print_intrinsics() || C->print_inlining()) {
 687       C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
 688     }
 689     C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
 690     if (C->log()) {
 691       C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
 692                      vmIntrinsics::name_at(intrinsic_id()),
 693                      (is_virtual() ? " virtual='1'" : ""),
 694                      C->unique() - nodes);
 695     }
 696     return slow_ctl; // Could be NULL if the check folds.
 697   }
 698 
 699   // The intrinsic bailed out
 700   if (C->print_intrinsics() || C->print_inlining()) {
 701     if (jvms->has_method()) {
 702       // Not a root compile.
 703       const char* msg = "failed to generate predicate for intrinsic";
 704       C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
 705     } else {
 706       // Root compile
 707       C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
 708                                         vmIntrinsics::name_at(intrinsic_id()),
 709                                         (is_virtual() ? " (virtual)" : ""), bci);
 710     }
 711   }
 712   C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
 713   return NULL;
 714 }
 715 
 716 bool LibraryCallKit::try_to_inline(int predicate) {
 717   // Handle symbolic names for otherwise undistinguished boolean switches:
 718   const bool is_store       = true;
 719   const bool is_native_ptr  = true;
 720   const bool is_static      = true;
 721   const bool is_volatile    = true;
 722 
 723   if (!jvms()->has_method()) {
 724     // Root JVMState has a null method.
 725     assert(map()->memory()->Opcode() == Op_Parm, "");
 726     // Insert the memory aliasing node
 727     set_all_memory(reset_memory());
 728   }
 729   assert(merged_memory(), "");
 730 
 731 
 732   switch (intrinsic_id()) {
 733   case vmIntrinsics::_hashCode:                 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
 734   case vmIntrinsics::_identityHashCode:         return inline_native_hashcode(/*!virtual*/ false,         is_static);
 735   case vmIntrinsics::_getClass:                 return inline_native_getClass();
 736 
 737   case vmIntrinsics::_dsin:
 738   case vmIntrinsics::_dcos:
 739   case vmIntrinsics::_dtan:
 740   case vmIntrinsics::_dabs:
 741   case vmIntrinsics::_datan2:
 742   case vmIntrinsics::_dsqrt:
 743   case vmIntrinsics::_dexp:
 744   case vmIntrinsics::_dlog:
 745   case vmIntrinsics::_dlog10:
 746   case vmIntrinsics::_dpow:                     return inline_math_native(intrinsic_id());
 747 
 748   case vmIntrinsics::_min:
 749   case vmIntrinsics::_max:                      return inline_min_max(intrinsic_id());
 750 
 751   case vmIntrinsics::_addExactI:                return inline_math_addExactI(false /* add */);
 752   case vmIntrinsics::_addExactL:                return inline_math_addExactL(false /* add */);
 753   case vmIntrinsics::_decrementExactI:          return inline_math_subtractExactI(true /* decrement */);
 754   case vmIntrinsics::_decrementExactL:          return inline_math_subtractExactL(true /* decrement */);
 755   case vmIntrinsics::_incrementExactI:          return inline_math_addExactI(true /* increment */);
 756   case vmIntrinsics::_incrementExactL:          return inline_math_addExactL(true /* increment */);
 757   case vmIntrinsics::_multiplyExactI:           return inline_math_multiplyExactI();
 758   case vmIntrinsics::_multiplyExactL:           return inline_math_multiplyExactL();
 759   case vmIntrinsics::_negateExactI:             return inline_math_negateExactI();
 760   case vmIntrinsics::_negateExactL:             return inline_math_negateExactL();
 761   case vmIntrinsics::_subtractExactI:           return inline_math_subtractExactI(false /* subtract */);
 762   case vmIntrinsics::_subtractExactL:           return inline_math_subtractExactL(false /* subtract */);
 763 
 764   case vmIntrinsics::_arraycopy:                return inline_arraycopy();
 765 
 766   case vmIntrinsics::_compareTo:                return inline_string_compareTo();
 767   case vmIntrinsics::_indexOf:                  return inline_string_indexOf();
 768   case vmIntrinsics::_equals:                   return inline_string_equals();
 769 
 770   case vmIntrinsics::_getObject:                return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT,  !is_volatile);
 771   case vmIntrinsics::_getBoolean:               return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
 772   case vmIntrinsics::_getByte:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE,    !is_volatile);
 773   case vmIntrinsics::_getShort:                 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,   !is_volatile);
 774   case vmIntrinsics::_getChar:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,    !is_volatile);
 775   case vmIntrinsics::_getInt:                   return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,     !is_volatile);
 776   case vmIntrinsics::_getLong:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,    !is_volatile);
 777   case vmIntrinsics::_getFloat:                 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT,   !is_volatile);
 778   case vmIntrinsics::_getDouble:                return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE,  !is_volatile);
 779 
 780   case vmIntrinsics::_putObject:                return inline_unsafe_access(!is_native_ptr,  is_store, T_OBJECT,  !is_volatile);
 781   case vmIntrinsics::_putBoolean:               return inline_unsafe_access(!is_native_ptr,  is_store, T_BOOLEAN, !is_volatile);
 782   case vmIntrinsics::_putByte:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_BYTE,    !is_volatile);
 783   case vmIntrinsics::_putShort:                 return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,   !is_volatile);
 784   case vmIntrinsics::_putChar:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,    !is_volatile);
 785   case vmIntrinsics::_putInt:                   return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,     !is_volatile);
 786   case vmIntrinsics::_putLong:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,    !is_volatile);
 787   case vmIntrinsics::_putFloat:                 return inline_unsafe_access(!is_native_ptr,  is_store, T_FLOAT,   !is_volatile);
 788   case vmIntrinsics::_putDouble:                return inline_unsafe_access(!is_native_ptr,  is_store, T_DOUBLE,  !is_volatile);
 789 
 790   case vmIntrinsics::_getByte_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE,    !is_volatile);
 791   case vmIntrinsics::_getShort_raw:             return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT,   !is_volatile);
 792   case vmIntrinsics::_getChar_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR,    !is_volatile);
 793   case vmIntrinsics::_getInt_raw:               return inline_unsafe_access( is_native_ptr, !is_store, T_INT,     !is_volatile);
 794   case vmIntrinsics::_getLong_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_LONG,    !is_volatile);
 795   case vmIntrinsics::_getFloat_raw:             return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT,   !is_volatile);
 796   case vmIntrinsics::_getDouble_raw:            return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE,  !is_volatile);
 797   case vmIntrinsics::_getAddress_raw:           return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
 798 
 799   case vmIntrinsics::_putByte_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_BYTE,    !is_volatile);
 800   case vmIntrinsics::_putShort_raw:             return inline_unsafe_access( is_native_ptr,  is_store, T_SHORT,   !is_volatile);
 801   case vmIntrinsics::_putChar_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_CHAR,    !is_volatile);
 802   case vmIntrinsics::_putInt_raw:               return inline_unsafe_access( is_native_ptr,  is_store, T_INT,     !is_volatile);
 803   case vmIntrinsics::_putLong_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_LONG,    !is_volatile);
 804   case vmIntrinsics::_putFloat_raw:             return inline_unsafe_access( is_native_ptr,  is_store, T_FLOAT,   !is_volatile);
 805   case vmIntrinsics::_putDouble_raw:            return inline_unsafe_access( is_native_ptr,  is_store, T_DOUBLE,  !is_volatile);
 806   case vmIntrinsics::_putAddress_raw:           return inline_unsafe_access( is_native_ptr,  is_store, T_ADDRESS, !is_volatile);
 807 
 808   case vmIntrinsics::_getObjectVolatile:        return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT,   is_volatile);
 809   case vmIntrinsics::_getBooleanVolatile:       return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN,  is_volatile);
 810   case vmIntrinsics::_getByteVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE,     is_volatile);
 811   case vmIntrinsics::_getShortVolatile:         return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,    is_volatile);
 812   case vmIntrinsics::_getCharVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,     is_volatile);
 813   case vmIntrinsics::_getIntVolatile:           return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,      is_volatile);
 814   case vmIntrinsics::_getLongVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,     is_volatile);
 815   case vmIntrinsics::_getFloatVolatile:         return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT,    is_volatile);
 816   case vmIntrinsics::_getDoubleVolatile:        return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE,   is_volatile);
 817 
 818   case vmIntrinsics::_putObjectVolatile:        return inline_unsafe_access(!is_native_ptr,  is_store, T_OBJECT,   is_volatile);
 819   case vmIntrinsics::_putBooleanVolatile:       return inline_unsafe_access(!is_native_ptr,  is_store, T_BOOLEAN,  is_volatile);
 820   case vmIntrinsics::_putByteVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_BYTE,     is_volatile);
 821   case vmIntrinsics::_putShortVolatile:         return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,    is_volatile);
 822   case vmIntrinsics::_putCharVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,     is_volatile);
 823   case vmIntrinsics::_putIntVolatile:           return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,      is_volatile);
 824   case vmIntrinsics::_putLongVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,     is_volatile);
 825   case vmIntrinsics::_putFloatVolatile:         return inline_unsafe_access(!is_native_ptr,  is_store, T_FLOAT,    is_volatile);
 826   case vmIntrinsics::_putDoubleVolatile:        return inline_unsafe_access(!is_native_ptr,  is_store, T_DOUBLE,   is_volatile);
 827 
 828   case vmIntrinsics::_getShortUnaligned:        return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,   !is_volatile, /*unaligned=*/true);
 829   case vmIntrinsics::_getCharUnaligned:         return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,    !is_volatile, /*unaligned=*/true);
 830   case vmIntrinsics::_getIntUnaligned:          return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,     !is_volatile, /*unaligned=*/true);
 831   case vmIntrinsics::_getLongUnaligned:         return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,    !is_volatile, /*unaligned=*/true);
 832 
 833   case vmIntrinsics::_putShortUnaligned:        return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,   !is_volatile, /*unaligned=*/true);
 834   case vmIntrinsics::_putCharUnaligned:         return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,    !is_volatile, /*unaligned=*/true);
 835   case vmIntrinsics::_putIntUnaligned:          return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,     !is_volatile, /*unaligned=*/true);
 836   case vmIntrinsics::_putLongUnaligned:         return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,    !is_volatile, /*unaligned=*/true);
 837 
 838   case vmIntrinsics::_compareAndSwapObject:     return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
 839   case vmIntrinsics::_compareAndSwapInt:        return inline_unsafe_load_store(T_INT,    LS_cmpxchg);
 840   case vmIntrinsics::_compareAndSwapLong:       return inline_unsafe_load_store(T_LONG,   LS_cmpxchg);
 841 
 842   case vmIntrinsics::_putOrderedObject:         return inline_unsafe_ordered_store(T_OBJECT);
 843   case vmIntrinsics::_putOrderedInt:            return inline_unsafe_ordered_store(T_INT);
 844   case vmIntrinsics::_putOrderedLong:           return inline_unsafe_ordered_store(T_LONG);
 845 
 846   case vmIntrinsics::_getAndAddInt:             return inline_unsafe_load_store(T_INT,    LS_xadd);
 847   case vmIntrinsics::_getAndAddLong:            return inline_unsafe_load_store(T_LONG,   LS_xadd);
 848   case vmIntrinsics::_getAndSetInt:             return inline_unsafe_load_store(T_INT,    LS_xchg);
 849   case vmIntrinsics::_getAndSetLong:            return inline_unsafe_load_store(T_LONG,   LS_xchg);
 850   case vmIntrinsics::_getAndSetObject:          return inline_unsafe_load_store(T_OBJECT, LS_xchg);
 851 
 852   case vmIntrinsics::_loadFence:
 853   case vmIntrinsics::_storeFence:
 854   case vmIntrinsics::_fullFence:                return inline_unsafe_fence(intrinsic_id());
 855 
 856   case vmIntrinsics::_currentThread:            return inline_native_currentThread();
 857   case vmIntrinsics::_isInterrupted:            return inline_native_isInterrupted();
 858 
 859 #ifdef TRACE_HAVE_INTRINSICS
 860   case vmIntrinsics::_classID:                  return inline_native_classID();
 861   case vmIntrinsics::_threadID:                 return inline_native_threadID();
 862   case vmIntrinsics::_counterTime:              return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
 863 #endif
 864   case vmIntrinsics::_currentTimeMillis:        return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
 865   case vmIntrinsics::_nanoTime:                 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
 866   case vmIntrinsics::_allocateInstance:         return inline_unsafe_allocate();
 867   case vmIntrinsics::_copyMemory:               return inline_unsafe_copyMemory();
 868   case vmIntrinsics::_newArray:                 return inline_native_newArray();
 869   case vmIntrinsics::_getLength:                return inline_native_getLength();
 870   case vmIntrinsics::_copyOf:                   return inline_array_copyOf(false);
 871   case vmIntrinsics::_copyOfRange:              return inline_array_copyOf(true);
 872   case vmIntrinsics::_equalsC:                  return inline_array_equals();
 873   case vmIntrinsics::_clone:                    return inline_native_clone(intrinsic()->is_virtual());
 874 
 875   case vmIntrinsics::_isAssignableFrom:         return inline_native_subtype_check();
 876 
 877   case vmIntrinsics::_isInstance:
 878   case vmIntrinsics::_getModifiers:
 879   case vmIntrinsics::_isInterface:
 880   case vmIntrinsics::_isArray:
 881   case vmIntrinsics::_isPrimitive:
 882   case vmIntrinsics::_getSuperclass:
 883   case vmIntrinsics::_getClassAccessFlags:      return inline_native_Class_query(intrinsic_id());
 884 
 885   case vmIntrinsics::_floatToRawIntBits:
 886   case vmIntrinsics::_floatToIntBits:
 887   case vmIntrinsics::_intBitsToFloat:
 888   case vmIntrinsics::_doubleToRawLongBits:
 889   case vmIntrinsics::_doubleToLongBits:
 890   case vmIntrinsics::_longBitsToDouble:         return inline_fp_conversions(intrinsic_id());
 891 
 892   case vmIntrinsics::_numberOfLeadingZeros_i:
 893   case vmIntrinsics::_numberOfLeadingZeros_l:
 894   case vmIntrinsics::_numberOfTrailingZeros_i:
 895   case vmIntrinsics::_numberOfTrailingZeros_l:
 896   case vmIntrinsics::_bitCount_i:
 897   case vmIntrinsics::_bitCount_l:
 898   case vmIntrinsics::_reverseBytes_i:
 899   case vmIntrinsics::_reverseBytes_l:
 900   case vmIntrinsics::_reverseBytes_s:
 901   case vmIntrinsics::_reverseBytes_c:           return inline_number_methods(intrinsic_id());
 902 
 903   case vmIntrinsics::_getCallerClass:           return inline_native_Reflection_getCallerClass();
 904 
 905   case vmIntrinsics::_Reference_get:            return inline_reference_get();
 906 
 907   case vmIntrinsics::_Class_cast:               return inline_Class_cast();
 908 
 909   case vmIntrinsics::_aescrypt_encryptBlock:
 910   case vmIntrinsics::_aescrypt_decryptBlock:    return inline_aescrypt_Block(intrinsic_id());
 911 
 912   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 913   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 914     return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
 915 
 916   case vmIntrinsics::_sha_implCompress:
 917   case vmIntrinsics::_sha2_implCompress:
 918   case vmIntrinsics::_sha5_implCompress:
 919     return inline_sha_implCompress(intrinsic_id());
 920 
 921   case vmIntrinsics::_digestBase_implCompressMB:
 922     return inline_digestBase_implCompressMB(predicate);
 923 
 924   case vmIntrinsics::_multiplyToLen:
 925     return inline_multiplyToLen();
 926 
 927   case vmIntrinsics::_squareToLen:
 928     return inline_squareToLen();
 929 
 930   case vmIntrinsics::_mulAdd:
 931     return inline_mulAdd();
 932 
 933   case vmIntrinsics::_encodeISOArray:
 934     return inline_encodeISOArray();
 935 
 936   case vmIntrinsics::_updateCRC32:
 937     return inline_updateCRC32();
 938   case vmIntrinsics::_updateBytesCRC32:
 939     return inline_updateBytesCRC32();
 940   case vmIntrinsics::_updateByteBufferCRC32:
 941     return inline_updateByteBufferCRC32();
 942 
 943   case vmIntrinsics::_profileBoolean:
 944     return inline_profileBoolean();
 945   case vmIntrinsics::_isCompileConstant:
 946     return inline_isCompileConstant();
 947 
 948   default:
 949     // If you get here, it may be that someone has added a new intrinsic
 950     // to the list in vmSymbols.hpp without implementing it here.
 951 #ifndef PRODUCT
 952     if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
 953       tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
 954                     vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
 955     }
 956 #endif
 957     return false;
 958   }
 959 }
 960 
 961 Node* LibraryCallKit::try_to_predicate(int predicate) {
 962   if (!jvms()->has_method()) {
 963     // Root JVMState has a null method.
 964     assert(map()->memory()->Opcode() == Op_Parm, "");
 965     // Insert the memory aliasing node
 966     set_all_memory(reset_memory());
 967   }
 968   assert(merged_memory(), "");
 969 
 970   switch (intrinsic_id()) {
 971   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 972     return inline_cipherBlockChaining_AESCrypt_predicate(false);
 973   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 974     return inline_cipherBlockChaining_AESCrypt_predicate(true);
 975   case vmIntrinsics::_digestBase_implCompressMB:
 976     return inline_digestBase_implCompressMB_predicate(predicate);
 977 
 978   default:
 979     // If you get here, it may be that someone has added a new intrinsic
 980     // to the list in vmSymbols.hpp without implementing it here.
 981 #ifndef PRODUCT
 982     if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
 983       tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
 984                     vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
 985     }
 986 #endif
 987     Node* slow_ctl = control();
 988     set_control(top()); // No fast path instrinsic
 989     return slow_ctl;
 990   }
 991 }
 992 
 993 //------------------------------set_result-------------------------------
 994 // Helper function for finishing intrinsics.
 995 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
 996   record_for_igvn(region);
 997   set_control(_gvn.transform(region));
 998   set_result( _gvn.transform(value));
 999   assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
1000 }
1001 
1002 //------------------------------generate_guard---------------------------
1003 // Helper function for generating guarded fast-slow graph structures.
1004 // The given 'test', if true, guards a slow path.  If the test fails
1005 // then a fast path can be taken.  (We generally hope it fails.)
1006 // In all cases, GraphKit::control() is updated to the fast path.
1007 // The returned value represents the control for the slow path.
1008 // The return value is never 'top'; it is either a valid control
1009 // or NULL if it is obvious that the slow path can never be taken.
1010 // Also, if region and the slow control are not NULL, the slow edge
1011 // is appended to the region.
1012 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
1013   if (stopped()) {
1014     // Already short circuited.
1015     return NULL;
1016   }
1017 
1018   // Build an if node and its projections.
1019   // If test is true we take the slow path, which we assume is uncommon.
1020   if (_gvn.type(test) == TypeInt::ZERO) {
1021     // The slow branch is never taken.  No need to build this guard.
1022     return NULL;
1023   }
1024 
1025   IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
1026 
1027   Node* if_slow = _gvn.transform(new IfTrueNode(iff));
1028   if (if_slow == top()) {
1029     // The slow branch is never taken.  No need to build this guard.
1030     return NULL;
1031   }
1032 
1033   if (region != NULL)
1034     region->add_req(if_slow);
1035 
1036   Node* if_fast = _gvn.transform(new IfFalseNode(iff));
1037   set_control(if_fast);
1038 
1039   return if_slow;
1040 }
1041 
1042 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
1043   return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
1044 }
1045 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1046   return generate_guard(test, region, PROB_FAIR);
1047 }
1048 
1049 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1050                                                      Node* *pos_index) {
1051   if (stopped())
1052     return NULL;                // already stopped
1053   if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1054     return NULL;                // index is already adequately typed
1055   Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1056   Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1057   Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1058   if (is_neg != NULL && pos_index != NULL) {
1059     // Emulate effect of Parse::adjust_map_after_if.
1060     Node* ccast = new CastIINode(index, TypeInt::POS);
1061     ccast->set_req(0, control());
1062     (*pos_index) = _gvn.transform(ccast);
1063   }
1064   return is_neg;
1065 }
1066 
1067 // Make sure that 'position' is a valid limit index, in [0..length].
1068 // There are two equivalent plans for checking this:
1069 //   A. (offset + copyLength)  unsigned<=  arrayLength
1070 //   B. offset  <=  (arrayLength - copyLength)
1071 // We require that all of the values above, except for the sum and
1072 // difference, are already known to be non-negative.
1073 // Plan A is robust in the face of overflow, if offset and copyLength
1074 // are both hugely positive.
1075 //
1076 // Plan B is less direct and intuitive, but it does not overflow at
1077 // all, since the difference of two non-negatives is always
1078 // representable.  Whenever Java methods must perform the equivalent
1079 // check they generally use Plan B instead of Plan A.
1080 // For the moment we use Plan A.
1081 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1082                                                   Node* subseq_length,
1083                                                   Node* array_length,
1084                                                   RegionNode* region) {
1085   if (stopped())
1086     return NULL;                // already stopped
1087   bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1088   if (zero_offset && subseq_length->eqv_uncast(array_length))
1089     return NULL;                // common case of whole-array copy
1090   Node* last = subseq_length;
1091   if (!zero_offset)             // last += offset
1092     last = _gvn.transform(new AddINode(last, offset));
1093   Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1094   Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1095   Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1096   return is_over;
1097 }
1098 
1099 
1100 //--------------------------generate_current_thread--------------------
1101 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1102   ciKlass*    thread_klass = env()->Thread_klass();
1103   const Type* thread_type  = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1104   Node* thread = _gvn.transform(new ThreadLocalNode());
1105   Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1106   Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1107   tls_output = thread;
1108   return threadObj;
1109 }
1110 
1111 
1112 //------------------------------make_string_method_node------------------------
1113 // Helper method for String intrinsic functions. This version is called
1114 // with str1 and str2 pointing to String object nodes.
1115 //
1116 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1117   Node* no_ctrl = NULL;
1118 
1119   // Get start addr of string
1120   Node* str1_value   = load_String_value(no_ctrl, str1);
1121   Node* str1_offset  = load_String_offset(no_ctrl, str1);
1122   Node* str1_start   = array_element_address(str1_value, str1_offset, T_CHAR);
1123 
1124   // Get length of string 1
1125   Node* str1_len  = load_String_length(no_ctrl, str1);
1126 
1127   Node* str2_value   = load_String_value(no_ctrl, str2);
1128   Node* str2_offset  = load_String_offset(no_ctrl, str2);
1129   Node* str2_start   = array_element_address(str2_value, str2_offset, T_CHAR);
1130 
1131   Node* str2_len = NULL;
1132   Node* result = NULL;
1133 
1134   switch (opcode) {
1135   case Op_StrIndexOf:
1136     // Get length of string 2
1137     str2_len = load_String_length(no_ctrl, str2);
1138 
1139     result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1140                                 str1_start, str1_len, str2_start, str2_len);
1141     break;
1142   case Op_StrComp:
1143     // Get length of string 2
1144     str2_len = load_String_length(no_ctrl, str2);
1145 
1146     result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1147                              str1_start, str1_len, str2_start, str2_len);
1148     break;
1149   case Op_StrEquals:
1150     result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1151                                str1_start, str2_start, str1_len);
1152     break;
1153   default:
1154     ShouldNotReachHere();
1155     return NULL;
1156   }
1157 
1158   // All these intrinsics have checks.
1159   C->set_has_split_ifs(true); // Has chance for split-if optimization
1160 
1161   return _gvn.transform(result);
1162 }
1163 
1164 // Helper method for String intrinsic functions. This version is called
1165 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1166 // to Int nodes containing the lenghts of str1 and str2.
1167 //
1168 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1169   Node* result = NULL;
1170   switch (opcode) {
1171   case Op_StrIndexOf:
1172     result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1173                                 str1_start, cnt1, str2_start, cnt2);
1174     break;
1175   case Op_StrComp:
1176     result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1177                              str1_start, cnt1, str2_start, cnt2);
1178     break;
1179   case Op_StrEquals:
1180     result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1181                                str1_start, str2_start, cnt1);
1182     break;
1183   default:
1184     ShouldNotReachHere();
1185     return NULL;
1186   }
1187 
1188   // All these intrinsics have checks.
1189   C->set_has_split_ifs(true); // Has chance for split-if optimization
1190 
1191   return _gvn.transform(result);
1192 }
1193 
1194 //------------------------------inline_string_compareTo------------------------
1195 // public int java.lang.String.compareTo(String anotherString);
1196 bool LibraryCallKit::inline_string_compareTo() {
1197   Node* receiver = null_check(argument(0));
1198   Node* arg      = null_check(argument(1));
1199   if (stopped()) {
1200     return true;
1201   }
1202   set_result(make_string_method_node(Op_StrComp, receiver, arg));
1203   return true;
1204 }
1205 
1206 //------------------------------inline_string_equals------------------------
1207 bool LibraryCallKit::inline_string_equals() {
1208   Node* receiver = null_check_receiver();
1209   // NOTE: Do not null check argument for String.equals() because spec
1210   // allows to specify NULL as argument.
1211   Node* argument = this->argument(1);
1212   if (stopped()) {
1213     return true;
1214   }
1215 
1216   // paths (plus control) merge
1217   RegionNode* region = new RegionNode(5);
1218   Node* phi = new PhiNode(region, TypeInt::BOOL);
1219 
1220   // does source == target string?
1221   Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1222   Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1223 
1224   Node* if_eq = generate_slow_guard(bol, NULL);
1225   if (if_eq != NULL) {
1226     // receiver == argument
1227     phi->init_req(2, intcon(1));
1228     region->init_req(2, if_eq);
1229   }
1230 
1231   // get String klass for instanceOf
1232   ciInstanceKlass* klass = env()->String_klass();
1233 
1234   if (!stopped()) {
1235     Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1236     Node* cmp  = _gvn.transform(new CmpINode(inst, intcon(1)));
1237     Node* bol  = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1238 
1239     Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1240     //instanceOf == true, fallthrough
1241 
1242     if (inst_false != NULL) {
1243       phi->init_req(3, intcon(0));
1244       region->init_req(3, inst_false);
1245     }
1246   }
1247 
1248   if (!stopped()) {
1249     const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1250 
1251     // Properly cast the argument to String
1252     argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1253     // This path is taken only when argument's type is String:NotNull.
1254     argument = cast_not_null(argument, false);
1255 
1256     Node* no_ctrl = NULL;
1257 
1258     // Get start addr of receiver
1259     Node* receiver_val    = load_String_value(no_ctrl, receiver);
1260     Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1261     Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1262 
1263     // Get length of receiver
1264     Node* receiver_cnt  = load_String_length(no_ctrl, receiver);
1265 
1266     // Get start addr of argument
1267     Node* argument_val    = load_String_value(no_ctrl, argument);
1268     Node* argument_offset = load_String_offset(no_ctrl, argument);
1269     Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1270 
1271     // Get length of argument
1272     Node* argument_cnt  = load_String_length(no_ctrl, argument);
1273 
1274     // Check for receiver count != argument count
1275     Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1276     Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1277     Node* if_ne = generate_slow_guard(bol, NULL);
1278     if (if_ne != NULL) {
1279       phi->init_req(4, intcon(0));
1280       region->init_req(4, if_ne);
1281     }
1282 
1283     // Check for count == 0 is done by assembler code for StrEquals.
1284 
1285     if (!stopped()) {
1286       Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1287       phi->init_req(1, equals);
1288       region->init_req(1, control());
1289     }
1290   }
1291 
1292   // post merge
1293   set_control(_gvn.transform(region));
1294   record_for_igvn(region);
1295 
1296   set_result(_gvn.transform(phi));
1297   return true;
1298 }
1299 
1300 //------------------------------inline_array_equals----------------------------
1301 bool LibraryCallKit::inline_array_equals() {
1302   Node* arg1 = argument(0);
1303   Node* arg2 = argument(1);
1304   set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1305   return true;
1306 }
1307 
1308 // Java version of String.indexOf(constant string)
1309 // class StringDecl {
1310 //   StringDecl(char[] ca) {
1311 //     offset = 0;
1312 //     count = ca.length;
1313 //     value = ca;
1314 //   }
1315 //   int offset;
1316 //   int count;
1317 //   char[] value;
1318 // }
1319 //
1320 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1321 //                             int targetOffset, int cache_i, int md2) {
1322 //   int cache = cache_i;
1323 //   int sourceOffset = string_object.offset;
1324 //   int sourceCount = string_object.count;
1325 //   int targetCount = target_object.length;
1326 //
1327 //   int targetCountLess1 = targetCount - 1;
1328 //   int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1329 //
1330 //   char[] source = string_object.value;
1331 //   char[] target = target_object;
1332 //   int lastChar = target[targetCountLess1];
1333 //
1334 //  outer_loop:
1335 //   for (int i = sourceOffset; i < sourceEnd; ) {
1336 //     int src = source[i + targetCountLess1];
1337 //     if (src == lastChar) {
1338 //       // With random strings and a 4-character alphabet,
1339 //       // reverse matching at this point sets up 0.8% fewer
1340 //       // frames, but (paradoxically) makes 0.3% more probes.
1341 //       // Since those probes are nearer the lastChar probe,
1342 //       // there is may be a net D$ win with reverse matching.
1343 //       // But, reversing loop inhibits unroll of inner loop
1344 //       // for unknown reason.  So, does running outer loop from
1345 //       // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1346 //       for (int j = 0; j < targetCountLess1; j++) {
1347 //         if (target[targetOffset + j] != source[i+j]) {
1348 //           if ((cache & (1 << source[i+j])) == 0) {
1349 //             if (md2 < j+1) {
1350 //               i += j+1;
1351 //               continue outer_loop;
1352 //             }
1353 //           }
1354 //           i += md2;
1355 //           continue outer_loop;
1356 //         }
1357 //       }
1358 //       return i - sourceOffset;
1359 //     }
1360 //     if ((cache & (1 << src)) == 0) {
1361 //       i += targetCountLess1;
1362 //     } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1363 //     i++;
1364 //   }
1365 //   return -1;
1366 // }
1367 
1368 //------------------------------string_indexOf------------------------
1369 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1370                                      jint cache_i, jint md2_i) {
1371 
1372   Node* no_ctrl  = NULL;
1373   float likely   = PROB_LIKELY(0.9);
1374   float unlikely = PROB_UNLIKELY(0.9);
1375 
1376   const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1377 
1378   Node* source        = load_String_value(no_ctrl, string_object);
1379   Node* sourceOffset  = load_String_offset(no_ctrl, string_object);
1380   Node* sourceCount   = load_String_length(no_ctrl, string_object);
1381 
1382   Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1383   jint target_length = target_array->length();
1384   const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1385   const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1386 
1387   // String.value field is known to be @Stable.
1388   if (UseImplicitStableValues) {
1389     target = cast_array_to_stable(target, target_type);
1390   }
1391 
1392   IdealKit kit(this, false, true);
1393 #define __ kit.
1394   Node* zero             = __ ConI(0);
1395   Node* one              = __ ConI(1);
1396   Node* cache            = __ ConI(cache_i);
1397   Node* md2              = __ ConI(md2_i);
1398   Node* lastChar         = __ ConI(target_array->char_at(target_length - 1));
1399   Node* targetCountLess1 = __ ConI(target_length - 1);
1400   Node* targetOffset     = __ ConI(targetOffset_i);
1401   Node* sourceEnd        = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1402 
1403   IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1404   Node* outer_loop = __ make_label(2 /* goto */);
1405   Node* return_    = __ make_label(1);
1406 
1407   __ set(rtn,__ ConI(-1));
1408   __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1409        Node* i2  = __ AddI(__ value(i), targetCountLess1);
1410        // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1411        Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1412        __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1413          __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1414               Node* tpj = __ AddI(targetOffset, __ value(j));
1415               Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1416               Node* ipj  = __ AddI(__ value(i), __ value(j));
1417               Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1418               __ if_then(targ, BoolTest::ne, src2); {
1419                 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1420                   __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1421                     __ increment(i, __ AddI(__ value(j), one));
1422                     __ goto_(outer_loop);
1423                   } __ end_if(); __ dead(j);
1424                 }__ end_if(); __ dead(j);
1425                 __ increment(i, md2);
1426                 __ goto_(outer_loop);
1427               }__ end_if();
1428               __ increment(j, one);
1429          }__ end_loop(); __ dead(j);
1430          __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1431          __ goto_(return_);
1432        }__ end_if();
1433        __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1434          __ increment(i, targetCountLess1);
1435        }__ end_if();
1436        __ increment(i, one);
1437        __ bind(outer_loop);
1438   }__ end_loop(); __ dead(i);
1439   __ bind(return_);
1440 
1441   // Final sync IdealKit and GraphKit.
1442   final_sync(kit);
1443   Node* result = __ value(rtn);
1444 #undef __
1445   C->set_has_loops(true);
1446   return result;
1447 }
1448 
1449 //------------------------------inline_string_indexOf------------------------
1450 bool LibraryCallKit::inline_string_indexOf() {
1451   Node* receiver = argument(0);
1452   Node* arg      = argument(1);
1453 
1454   Node* result;
1455   if (Matcher::has_match_rule(Op_StrIndexOf) &&
1456       UseSSE42Intrinsics) {
1457     // Generate SSE4.2 version of indexOf
1458     // We currently only have match rules that use SSE4.2
1459 
1460     receiver = null_check(receiver);
1461     arg      = null_check(arg);
1462     if (stopped()) {
1463       return true;
1464     }
1465 
1466     // Make the merge point
1467     RegionNode* result_rgn = new RegionNode(4);
1468     Node*       result_phi = new PhiNode(result_rgn, TypeInt::INT);
1469     Node* no_ctrl  = NULL;
1470 
1471     // Get start addr of source string
1472     Node* source = load_String_value(no_ctrl, receiver);
1473     Node* source_offset = load_String_offset(no_ctrl, receiver);
1474     Node* source_start = array_element_address(source, source_offset, T_CHAR);
1475 
1476     // Get length of source string
1477     Node* source_cnt  = load_String_length(no_ctrl, receiver);
1478 
1479     // Get start addr of substring
1480     Node* substr = load_String_value(no_ctrl, arg);
1481     Node* substr_offset = load_String_offset(no_ctrl, arg);
1482     Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1483 
1484     // Get length of source string
1485     Node* substr_cnt  = load_String_length(no_ctrl, arg);
1486 
1487     // Check for substr count > string count
1488     Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1489     Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1490     Node* if_gt = generate_slow_guard(bol, NULL);
1491     if (if_gt != NULL) {
1492       result_phi->init_req(2, intcon(-1));
1493       result_rgn->init_req(2, if_gt);
1494     }
1495 
1496     if (!stopped()) {
1497       // Check for substr count == 0
1498       cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1499       bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1500       Node* if_zero = generate_slow_guard(bol, NULL);
1501       if (if_zero != NULL) {
1502         result_phi->init_req(3, intcon(0));
1503         result_rgn->init_req(3, if_zero);
1504       }
1505     }
1506 
1507     if (!stopped()) {
1508       result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1509       result_phi->init_req(1, result);
1510       result_rgn->init_req(1, control());
1511     }
1512     set_control(_gvn.transform(result_rgn));
1513     record_for_igvn(result_rgn);
1514     result = _gvn.transform(result_phi);
1515 
1516   } else { // Use LibraryCallKit::string_indexOf
1517     // don't intrinsify if argument isn't a constant string.
1518     if (!arg->is_Con()) {
1519      return false;
1520     }
1521     const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1522     if (str_type == NULL) {
1523       return false;
1524     }
1525     ciInstanceKlass* klass = env()->String_klass();
1526     ciObject* str_const = str_type->const_oop();
1527     if (str_const == NULL || str_const->klass() != klass) {
1528       return false;
1529     }
1530     ciInstance* str = str_const->as_instance();
1531     assert(str != NULL, "must be instance");
1532 
1533     ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1534     ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1535 
1536     int o;
1537     int c;
1538     if (java_lang_String::has_offset_field()) {
1539       o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1540       c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1541     } else {
1542       o = 0;
1543       c = pat->length();
1544     }
1545 
1546     // constant strings have no offset and count == length which
1547     // simplifies the resulting code somewhat so lets optimize for that.
1548     if (o != 0 || c != pat->length()) {
1549      return false;
1550     }
1551 
1552     receiver = null_check(receiver, T_OBJECT);
1553     // NOTE: No null check on the argument is needed since it's a constant String oop.
1554     if (stopped()) {
1555       return true;
1556     }
1557 
1558     // The null string as a pattern always returns 0 (match at beginning of string)
1559     if (c == 0) {
1560       set_result(intcon(0));
1561       return true;
1562     }
1563 
1564     // Generate default indexOf
1565     jchar lastChar = pat->char_at(o + (c - 1));
1566     int cache = 0;
1567     int i;
1568     for (i = 0; i < c - 1; i++) {
1569       assert(i < pat->length(), "out of range");
1570       cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1571     }
1572 
1573     int md2 = c;
1574     for (i = 0; i < c - 1; i++) {
1575       assert(i < pat->length(), "out of range");
1576       if (pat->char_at(o + i) == lastChar) {
1577         md2 = (c - 1) - i;
1578       }
1579     }
1580 
1581     result = string_indexOf(receiver, pat, o, cache, md2);
1582   }
1583   set_result(result);
1584   return true;
1585 }
1586 
1587 //--------------------------round_double_node--------------------------------
1588 // Round a double node if necessary.
1589 Node* LibraryCallKit::round_double_node(Node* n) {
1590   if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1591     n = _gvn.transform(new RoundDoubleNode(0, n));
1592   return n;
1593 }
1594 
1595 //------------------------------inline_math-----------------------------------
1596 // public static double Math.abs(double)
1597 // public static double Math.sqrt(double)
1598 // public static double Math.log(double)
1599 // public static double Math.log10(double)
1600 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1601   Node* arg = round_double_node(argument(0));
1602   Node* n;
1603   switch (id) {
1604   case vmIntrinsics::_dabs:   n = new AbsDNode(                arg);  break;
1605   case vmIntrinsics::_dsqrt:  n = new SqrtDNode(C, control(),  arg);  break;
1606   case vmIntrinsics::_dlog:   n = new LogDNode(C, control(),   arg);  break;
1607   case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg);  break;
1608   default:  fatal_unexpected_iid(id);  break;
1609   }
1610   set_result(_gvn.transform(n));
1611   return true;
1612 }
1613 
1614 //------------------------------inline_trig----------------------------------
1615 // Inline sin/cos/tan instructions, if possible.  If rounding is required, do
1616 // argument reduction which will turn into a fast/slow diamond.
1617 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1618   Node* arg = round_double_node(argument(0));
1619   Node* n = NULL;
1620 
1621   switch (id) {
1622   case vmIntrinsics::_dsin:  n = new SinDNode(C, control(), arg);  break;
1623   case vmIntrinsics::_dcos:  n = new CosDNode(C, control(), arg);  break;
1624   case vmIntrinsics::_dtan:  n = new TanDNode(C, control(), arg);  break;
1625   default:  fatal_unexpected_iid(id);  break;
1626   }
1627   n = _gvn.transform(n);
1628 
1629   // Rounding required?  Check for argument reduction!
1630   if (Matcher::strict_fp_requires_explicit_rounding) {
1631     static const double     pi_4 =  0.7853981633974483;
1632     static const double neg_pi_4 = -0.7853981633974483;
1633     // pi/2 in 80-bit extended precision
1634     // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1635     // -pi/2 in 80-bit extended precision
1636     // 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};
1637     // Cutoff value for using this argument reduction technique
1638     //static const double    pi_2_minus_epsilon =  1.564660403643354;
1639     //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1640 
1641     // Pseudocode for sin:
1642     // if (x <= Math.PI / 4.0) {
1643     //   if (x >= -Math.PI / 4.0) return  fsin(x);
1644     //   if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1645     // } else {
1646     //   if (x <=  Math.PI / 2.0) return  fcos(x - Math.PI / 2.0);
1647     // }
1648     // return StrictMath.sin(x);
1649 
1650     // Pseudocode for cos:
1651     // if (x <= Math.PI / 4.0) {
1652     //   if (x >= -Math.PI / 4.0) return  fcos(x);
1653     //   if (x >= -Math.PI / 2.0) return  fsin(x + Math.PI / 2.0);
1654     // } else {
1655     //   if (x <=  Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1656     // }
1657     // return StrictMath.cos(x);
1658 
1659     // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1660     // requires a special machine instruction to load it.  Instead we'll try
1661     // the 'easy' case.  If we really need the extra range +/- PI/2 we'll
1662     // probably do the math inside the SIN encoding.
1663 
1664     // Make the merge point
1665     RegionNode* r = new RegionNode(3);
1666     Node* phi = new PhiNode(r, Type::DOUBLE);
1667 
1668     // Flatten arg so we need only 1 test
1669     Node *abs = _gvn.transform(new AbsDNode(arg));
1670     // Node for PI/4 constant
1671     Node *pi4 = makecon(TypeD::make(pi_4));
1672     // Check PI/4 : abs(arg)
1673     Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1674     // Check: If PI/4 < abs(arg) then go slow
1675     Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1676     // Branch either way
1677     IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1678     set_control(opt_iff(r,iff));
1679 
1680     // Set fast path result
1681     phi->init_req(2, n);
1682 
1683     // Slow path - non-blocking leaf call
1684     Node* call = NULL;
1685     switch (id) {
1686     case vmIntrinsics::_dsin:
1687       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1688                                CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1689                                "Sin", NULL, arg, top());
1690       break;
1691     case vmIntrinsics::_dcos:
1692       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1693                                CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1694                                "Cos", NULL, arg, top());
1695       break;
1696     case vmIntrinsics::_dtan:
1697       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1698                                CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1699                                "Tan", NULL, arg, top());
1700       break;
1701     }
1702     assert(control()->in(0) == call, "");
1703     Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1704     r->init_req(1, control());
1705     phi->init_req(1, slow_result);
1706 
1707     // Post-merge
1708     set_control(_gvn.transform(r));
1709     record_for_igvn(r);
1710     n = _gvn.transform(phi);
1711 
1712     C->set_has_split_ifs(true); // Has chance for split-if optimization
1713   }
1714   set_result(n);
1715   return true;
1716 }
1717 
1718 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1719   //-------------------
1720   //result=(result.isNaN())? funcAddr():result;
1721   // Check: If isNaN() by checking result!=result? then either trap
1722   // or go to runtime
1723   Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1724   // Build the boolean node
1725   Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1726 
1727   if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1728     { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1729       // The pow or exp intrinsic returned a NaN, which requires a call
1730       // to the runtime.  Recompile with the runtime call.
1731       uncommon_trap(Deoptimization::Reason_intrinsic,
1732                     Deoptimization::Action_make_not_entrant);
1733     }
1734     return result;
1735   } else {
1736     // If this inlining ever returned NaN in the past, we compile a call
1737     // to the runtime to properly handle corner cases
1738 
1739     IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1740     Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1741     Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1742 
1743     if (!if_slow->is_top()) {
1744       RegionNode* result_region = new RegionNode(3);
1745       PhiNode*    result_val = new PhiNode(result_region, Type::DOUBLE);
1746 
1747       result_region->init_req(1, if_fast);
1748       result_val->init_req(1, result);
1749 
1750       set_control(if_slow);
1751 
1752       const TypePtr* no_memory_effects = NULL;
1753       Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1754                                    no_memory_effects,
1755                                    x, top(), y, y ? top() : NULL);
1756       Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1757 #ifdef ASSERT
1758       Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1759       assert(value_top == top(), "second value must be top");
1760 #endif
1761 
1762       result_region->init_req(2, control());
1763       result_val->init_req(2, value);
1764       set_control(_gvn.transform(result_region));
1765       return _gvn.transform(result_val);
1766     } else {
1767       return result;
1768     }
1769   }
1770 }
1771 
1772 //------------------------------inline_exp-------------------------------------
1773 // Inline exp instructions, if possible.  The Intel hardware only misses
1774 // really odd corner cases (+/- Infinity).  Just uncommon-trap them.
1775 bool LibraryCallKit::inline_exp() {
1776   Node* arg = round_double_node(argument(0));
1777   Node* n   = _gvn.transform(new ExpDNode(C, control(), arg));
1778 
1779   n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1780   set_result(n);
1781 
1782   C->set_has_split_ifs(true); // Has chance for split-if optimization
1783   return true;
1784 }
1785 
1786 //------------------------------inline_pow-------------------------------------
1787 // Inline power instructions, if possible.
1788 bool LibraryCallKit::inline_pow() {
1789   // Pseudocode for pow
1790   // if (y == 2) {
1791   //   return x * x;
1792   // } else {
1793   //   if (x <= 0.0) {
1794   //     long longy = (long)y;
1795   //     if ((double)longy == y) { // if y is long
1796   //       if (y + 1 == y) longy = 0; // huge number: even
1797   //       result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1798   //     } else {
1799   //       result = NaN;
1800   //     }
1801   //   } else {
1802   //     result = DPow(x,y);
1803   //   }
1804   //   if (result != result)?  {
1805   //     result = uncommon_trap() or runtime_call();
1806   //   }
1807   //   return result;
1808   // }
1809 
1810   Node* x = round_double_node(argument(0));
1811   Node* y = round_double_node(argument(2));
1812 
1813   Node* result = NULL;
1814 
1815   Node*   const_two_node = makecon(TypeD::make(2.0));
1816   Node*   cmp_node       = _gvn.transform(new CmpDNode(y, const_two_node));
1817   Node*   bool_node      = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1818   IfNode* if_node        = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1819   Node*   if_true        = _gvn.transform(new IfTrueNode(if_node));
1820   Node*   if_false       = _gvn.transform(new IfFalseNode(if_node));
1821 
1822   RegionNode* region_node = new RegionNode(3);
1823   region_node->init_req(1, if_true);
1824 
1825   Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1826   // special case for x^y where y == 2, we can convert it to x * x
1827   phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1828 
1829   // set control to if_false since we will now process the false branch
1830   set_control(if_false);
1831 
1832   if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1833     // Short form: skip the fancy tests and just check for NaN result.
1834     result = _gvn.transform(new PowDNode(C, control(), x, y));
1835   } else {
1836     // If this inlining ever returned NaN in the past, include all
1837     // checks + call to the runtime.
1838 
1839     // Set the merge point for If node with condition of (x <= 0.0)
1840     // There are four possible paths to region node and phi node
1841     RegionNode *r = new RegionNode(4);
1842     Node *phi = new PhiNode(r, Type::DOUBLE);
1843 
1844     // Build the first if node: if (x <= 0.0)
1845     // Node for 0 constant
1846     Node *zeronode = makecon(TypeD::ZERO);
1847     // Check x:0
1848     Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1849     // Check: If (x<=0) then go complex path
1850     Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1851     // Branch either way
1852     IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1853     // Fast path taken; set region slot 3
1854     Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1855     r->init_req(3,fast_taken); // Capture fast-control
1856 
1857     // Fast path not-taken, i.e. slow path
1858     Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1859 
1860     // Set fast path result
1861     Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1862     phi->init_req(3, fast_result);
1863 
1864     // Complex path
1865     // Build the second if node (if y is long)
1866     // Node for (long)y
1867     Node *longy = _gvn.transform(new ConvD2LNode(y));
1868     // Node for (double)((long) y)
1869     Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1870     // Check (double)((long) y) : y
1871     Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1872     // Check if (y isn't long) then go to slow path
1873 
1874     Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1875     // Branch either way
1876     IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1877     Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1878 
1879     Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1880 
1881     // Calculate DPow(abs(x), y)*(1 & (long)y)
1882     // Node for constant 1
1883     Node *conone = longcon(1);
1884     // 1& (long)y
1885     Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1886 
1887     // A huge number is always even. Detect a huge number by checking
1888     // if y + 1 == y and set integer to be tested for parity to 0.
1889     // Required for corner case:
1890     // (long)9.223372036854776E18 = max_jlong
1891     // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1892     // max_jlong is odd but 9.223372036854776E18 is even
1893     Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1894     Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1895     Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1896     Node* correctedsign = NULL;
1897     if (ConditionalMoveLimit != 0) {
1898       correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1899     } else {
1900       IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1901       RegionNode *r = new RegionNode(3);
1902       Node *phi = new PhiNode(r, TypeLong::LONG);
1903       r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1904       r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1905       phi->init_req(1, signnode);
1906       phi->init_req(2, longcon(0));
1907       correctedsign = _gvn.transform(phi);
1908       ylong_path = _gvn.transform(r);
1909       record_for_igvn(r);
1910     }
1911 
1912     // zero node
1913     Node *conzero = longcon(0);
1914     // Check (1&(long)y)==0?
1915     Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1916     // Check if (1&(long)y)!=0?, if so the result is negative
1917     Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1918     // abs(x)
1919     Node *absx=_gvn.transform(new AbsDNode(x));
1920     // abs(x)^y
1921     Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1922     // -abs(x)^y
1923     Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1924     // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1925     Node *signresult = NULL;
1926     if (ConditionalMoveLimit != 0) {
1927       signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1928     } else {
1929       IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1930       RegionNode *r = new RegionNode(3);
1931       Node *phi = new PhiNode(r, Type::DOUBLE);
1932       r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1933       r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1934       phi->init_req(1, absxpowy);
1935       phi->init_req(2, negabsxpowy);
1936       signresult = _gvn.transform(phi);
1937       ylong_path = _gvn.transform(r);
1938       record_for_igvn(r);
1939     }
1940     // Set complex path fast result
1941     r->init_req(2, ylong_path);
1942     phi->init_req(2, signresult);
1943 
1944     static const jlong nan_bits = CONST64(0x7ff8000000000000);
1945     Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1946     r->init_req(1,slow_path);
1947     phi->init_req(1,slow_result);
1948 
1949     // Post merge
1950     set_control(_gvn.transform(r));
1951     record_for_igvn(r);
1952     result = _gvn.transform(phi);
1953   }
1954 
1955   result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1956 
1957   // control from finish_pow_exp is now input to the region node
1958   region_node->set_req(2, control());
1959   // the result from finish_pow_exp is now input to the phi node
1960   phi_node->init_req(2, result);
1961   set_control(_gvn.transform(region_node));
1962   record_for_igvn(region_node);
1963   set_result(_gvn.transform(phi_node));
1964 
1965   C->set_has_split_ifs(true); // Has chance for split-if optimization
1966   return true;
1967 }
1968 
1969 //------------------------------runtime_math-----------------------------
1970 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1971   assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1972          "must be (DD)D or (D)D type");
1973 
1974   // Inputs
1975   Node* a = round_double_node(argument(0));
1976   Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1977 
1978   const TypePtr* no_memory_effects = NULL;
1979   Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1980                                  no_memory_effects,
1981                                  a, top(), b, b ? top() : NULL);
1982   Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1983 #ifdef ASSERT
1984   Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1985   assert(value_top == top(), "second value must be top");
1986 #endif
1987 
1988   set_result(value);
1989   return true;
1990 }
1991 
1992 //------------------------------inline_math_native-----------------------------
1993 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1994 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1995   switch (id) {
1996     // These intrinsics are not properly supported on all hardware
1997   case vmIntrinsics::_dcos:   return Matcher::has_match_rule(Op_CosD)   ? inline_trig(id) :
1998     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos),   "COS");
1999   case vmIntrinsics::_dsin:   return Matcher::has_match_rule(Op_SinD)   ? inline_trig(id) :
2000     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin),   "SIN");
2001   case vmIntrinsics::_dtan:   return Matcher::has_match_rule(Op_TanD)   ? inline_trig(id) :
2002     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan),   "TAN");
2003 
2004   case vmIntrinsics::_dlog:   return Matcher::has_match_rule(Op_LogD)   ? inline_math(id) :
2005     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog),   "LOG");
2006   case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
2007     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
2008 
2009     // These intrinsics are supported on all hardware
2010   case vmIntrinsics::_dsqrt:  return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
2011   case vmIntrinsics::_dabs:   return Matcher::has_match_rule(Op_AbsD)   ? inline_math(id) : false;
2012 
2013   case vmIntrinsics::_dexp:   return Matcher::has_match_rule(Op_ExpD)   ? inline_exp()    :
2014     runtime_math(OptoRuntime::Math_D_D_Type(),  FN_PTR(SharedRuntime::dexp),  "EXP");
2015   case vmIntrinsics::_dpow:   return Matcher::has_match_rule(Op_PowD)   ? inline_pow()    :
2016     runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow),  "POW");
2017 #undef FN_PTR
2018 
2019    // These intrinsics are not yet correctly implemented
2020   case vmIntrinsics::_datan2:
2021     return false;
2022 
2023   default:
2024     fatal_unexpected_iid(id);
2025     return false;
2026   }
2027 }
2028 
2029 static bool is_simple_name(Node* n) {
2030   return (n->req() == 1         // constant
2031           || (n->is_Type() && n->as_Type()->type()->singleton())
2032           || n->is_Proj()       // parameter or return value
2033           || n->is_Phi()        // local of some sort
2034           );
2035 }
2036 
2037 //----------------------------inline_min_max-----------------------------------
2038 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2039   set_result(generate_min_max(id, argument(0), argument(1)));
2040   return true;
2041 }
2042 
2043 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2044   Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2045   IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2046   Node* fast_path = _gvn.transform( new IfFalseNode(check));
2047   Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2048 
2049   {
2050     PreserveJVMState pjvms(this);
2051     PreserveReexecuteState preexecs(this);
2052     jvms()->set_should_reexecute(true);
2053 
2054     set_control(slow_path);
2055     set_i_o(i_o());
2056 
2057     uncommon_trap(Deoptimization::Reason_intrinsic,
2058                   Deoptimization::Action_none);
2059   }
2060 
2061   set_control(fast_path);
2062   set_result(math);
2063 }
2064 
2065 template <typename OverflowOp>
2066 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2067   typedef typename OverflowOp::MathOp MathOp;
2068 
2069   MathOp* mathOp = new MathOp(arg1, arg2);
2070   Node* operation = _gvn.transform( mathOp );
2071   Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2072   inline_math_mathExact(operation, ofcheck);
2073   return true;
2074 }
2075 
2076 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2077   return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2078 }
2079 
2080 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2081   return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2082 }
2083 
2084 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2085   return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2086 }
2087 
2088 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2089   return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2090 }
2091 
2092 bool LibraryCallKit::inline_math_negateExactI() {
2093   return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2094 }
2095 
2096 bool LibraryCallKit::inline_math_negateExactL() {
2097   return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2098 }
2099 
2100 bool LibraryCallKit::inline_math_multiplyExactI() {
2101   return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2102 }
2103 
2104 bool LibraryCallKit::inline_math_multiplyExactL() {
2105   return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2106 }
2107 
2108 Node*
2109 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2110   // These are the candidate return value:
2111   Node* xvalue = x0;
2112   Node* yvalue = y0;
2113 
2114   if (xvalue == yvalue) {
2115     return xvalue;
2116   }
2117 
2118   bool want_max = (id == vmIntrinsics::_max);
2119 
2120   const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2121   const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2122   if (txvalue == NULL || tyvalue == NULL)  return top();
2123   // This is not really necessary, but it is consistent with a
2124   // hypothetical MaxINode::Value method:
2125   int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2126 
2127   // %%% This folding logic should (ideally) be in a different place.
2128   // Some should be inside IfNode, and there to be a more reliable
2129   // transformation of ?: style patterns into cmoves.  We also want
2130   // more powerful optimizations around cmove and min/max.
2131 
2132   // Try to find a dominating comparison of these guys.
2133   // It can simplify the index computation for Arrays.copyOf
2134   // and similar uses of System.arraycopy.
2135   // First, compute the normalized version of CmpI(x, y).
2136   int   cmp_op = Op_CmpI;
2137   Node* xkey = xvalue;
2138   Node* ykey = yvalue;
2139   Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2140   if (ideal_cmpxy->is_Cmp()) {
2141     // E.g., if we have CmpI(length - offset, count),
2142     // it might idealize to CmpI(length, count + offset)
2143     cmp_op = ideal_cmpxy->Opcode();
2144     xkey = ideal_cmpxy->in(1);
2145     ykey = ideal_cmpxy->in(2);
2146   }
2147 
2148   // Start by locating any relevant comparisons.
2149   Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2150   Node* cmpxy = NULL;
2151   Node* cmpyx = NULL;
2152   for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2153     Node* cmp = start_from->fast_out(k);
2154     if (cmp->outcnt() > 0 &&            // must have prior uses
2155         cmp->in(0) == NULL &&           // must be context-independent
2156         cmp->Opcode() == cmp_op) {      // right kind of compare
2157       if (cmp->in(1) == xkey && cmp->in(2) == ykey)  cmpxy = cmp;
2158       if (cmp->in(1) == ykey && cmp->in(2) == xkey)  cmpyx = cmp;
2159     }
2160   }
2161 
2162   const int NCMPS = 2;
2163   Node* cmps[NCMPS] = { cmpxy, cmpyx };
2164   int cmpn;
2165   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2166     if (cmps[cmpn] != NULL)  break;     // find a result
2167   }
2168   if (cmpn < NCMPS) {
2169     // Look for a dominating test that tells us the min and max.
2170     int depth = 0;                // Limit search depth for speed
2171     Node* dom = control();
2172     for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2173       if (++depth >= 100)  break;
2174       Node* ifproj = dom;
2175       if (!ifproj->is_Proj())  continue;
2176       Node* iff = ifproj->in(0);
2177       if (!iff->is_If())  continue;
2178       Node* bol = iff->in(1);
2179       if (!bol->is_Bool())  continue;
2180       Node* cmp = bol->in(1);
2181       if (cmp == NULL)  continue;
2182       for (cmpn = 0; cmpn < NCMPS; cmpn++)
2183         if (cmps[cmpn] == cmp)  break;
2184       if (cmpn == NCMPS)  continue;
2185       BoolTest::mask btest = bol->as_Bool()->_test._test;
2186       if (ifproj->is_IfFalse())  btest = BoolTest(btest).negate();
2187       if (cmp->in(1) == ykey)    btest = BoolTest(btest).commute();
2188       // At this point, we know that 'x btest y' is true.
2189       switch (btest) {
2190       case BoolTest::eq:
2191         // They are proven equal, so we can collapse the min/max.
2192         // Either value is the answer.  Choose the simpler.
2193         if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2194           return yvalue;
2195         return xvalue;
2196       case BoolTest::lt:          // x < y
2197       case BoolTest::le:          // x <= y
2198         return (want_max ? yvalue : xvalue);
2199       case BoolTest::gt:          // x > y
2200       case BoolTest::ge:          // x >= y
2201         return (want_max ? xvalue : yvalue);
2202       }
2203     }
2204   }
2205 
2206   // We failed to find a dominating test.
2207   // Let's pick a test that might GVN with prior tests.
2208   Node*          best_bol   = NULL;
2209   BoolTest::mask best_btest = BoolTest::illegal;
2210   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2211     Node* cmp = cmps[cmpn];
2212     if (cmp == NULL)  continue;
2213     for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2214       Node* bol = cmp->fast_out(j);
2215       if (!bol->is_Bool())  continue;
2216       BoolTest::mask btest = bol->as_Bool()->_test._test;
2217       if (btest == BoolTest::eq || btest == BoolTest::ne)  continue;
2218       if (cmp->in(1) == ykey)   btest = BoolTest(btest).commute();
2219       if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2220         best_bol   = bol->as_Bool();
2221         best_btest = btest;
2222       }
2223     }
2224   }
2225 
2226   Node* answer_if_true  = NULL;
2227   Node* answer_if_false = NULL;
2228   switch (best_btest) {
2229   default:
2230     if (cmpxy == NULL)
2231       cmpxy = ideal_cmpxy;
2232     best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2233     // and fall through:
2234   case BoolTest::lt:          // x < y
2235   case BoolTest::le:          // x <= y
2236     answer_if_true  = (want_max ? yvalue : xvalue);
2237     answer_if_false = (want_max ? xvalue : yvalue);
2238     break;
2239   case BoolTest::gt:          // x > y
2240   case BoolTest::ge:          // x >= y
2241     answer_if_true  = (want_max ? xvalue : yvalue);
2242     answer_if_false = (want_max ? yvalue : xvalue);
2243     break;
2244   }
2245 
2246   jint hi, lo;
2247   if (want_max) {
2248     // We can sharpen the minimum.
2249     hi = MAX2(txvalue->_hi, tyvalue->_hi);
2250     lo = MAX2(txvalue->_lo, tyvalue->_lo);
2251   } else {
2252     // We can sharpen the maximum.
2253     hi = MIN2(txvalue->_hi, tyvalue->_hi);
2254     lo = MIN2(txvalue->_lo, tyvalue->_lo);
2255   }
2256 
2257   // Use a flow-free graph structure, to avoid creating excess control edges
2258   // which could hinder other optimizations.
2259   // Since Math.min/max is often used with arraycopy, we want
2260   // tightly_coupled_allocation to be able to see beyond min/max expressions.
2261   Node* cmov = CMoveNode::make(NULL, best_bol,
2262                                answer_if_false, answer_if_true,
2263                                TypeInt::make(lo, hi, widen));
2264 
2265   return _gvn.transform(cmov);
2266 
2267   /*
2268   // This is not as desirable as it may seem, since Min and Max
2269   // nodes do not have a full set of optimizations.
2270   // And they would interfere, anyway, with 'if' optimizations
2271   // and with CMoveI canonical forms.
2272   switch (id) {
2273   case vmIntrinsics::_min:
2274     result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2275   case vmIntrinsics::_max:
2276     result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2277   default:
2278     ShouldNotReachHere();
2279   }
2280   */
2281 }
2282 
2283 inline int
2284 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, bool decode_offset) {
2285   const TypePtr* base_type = TypePtr::NULL_PTR;
2286   if (base != NULL)  base_type = _gvn.type(base)->isa_ptr();
2287   if (base_type == NULL) {
2288     // Unknown type.
2289     return Type::AnyPtr;
2290   } else if (base_type == TypePtr::NULL_PTR) {
2291     // Since this is a NULL+long form, we have to switch to a rawptr.
2292     base   = _gvn.transform(new CastX2PNode(offset));
2293     offset = MakeConX(0);
2294     return Type::RawPtr;
2295   } else if (base_type->base() == Type::RawPtr) {
2296     return Type::RawPtr;
2297   } else if (base_type->isa_oopptr()) {
2298     Node* decoded_offset = offset;
2299     if (decode_offset) {
2300       decoded_offset = _gvn.transform(new RShiftXNode(offset, intcon(Unsafe::offset_shift)));
2301     }
2302     // Base is never null => always a heap address.
2303     if (base_type->ptr() == TypePtr::NotNull) {
2304       offset = decoded_offset;
2305       return Type::OopPtr;
2306     }
2307     // Offset is small => always a heap address.
2308     const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2309     if (offset_type != NULL &&
2310         base_type->offset() == 0 &&     // (should always be?)
2311         offset_type->_lo >= 0 &&
2312         !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2313       offset = decoded_offset;
2314       return Type::OopPtr;
2315     }
2316     // Otherwise, it might either be oop+off or NULL+addr.
2317     // For oop+off case the offset should be decoded first, but
2318     // NULL+addr can be used as is.
2319     IdealKit ideal(this);
2320 #define __ ideal.
2321     IdealVariable off(ideal);
2322     __ declarations_done();
2323     __ set(off, offset);
2324     __ if_then(base, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2325       __ set(off, decoded_offset);
2326     } __ end_if();
2327     // Final sync IdealKit and GraphKit.
2328     decoded_offset = __ value(off);
2329     final_sync(ideal);
2330 #undef __
2331     offset = decoded_offset;
2332     return Type::AnyPtr;
2333   } else {
2334     // No information:
2335     return Type::AnyPtr;
2336   }
2337 }
2338 
2339 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset, bool decode_offset) {
2340   int kind = classify_unsafe_addr(base, offset, decode_offset);
2341   if (kind == Type::RawPtr) {
2342     return basic_plus_adr(top(), base, offset);
2343   } else {
2344     return basic_plus_adr(base, offset);
2345   }
2346 }
2347 
2348 //--------------------------inline_number_methods-----------------------------
2349 // inline int     Integer.numberOfLeadingZeros(int)
2350 // inline int        Long.numberOfLeadingZeros(long)
2351 //
2352 // inline int     Integer.numberOfTrailingZeros(int)
2353 // inline int        Long.numberOfTrailingZeros(long)
2354 //
2355 // inline int     Integer.bitCount(int)
2356 // inline int        Long.bitCount(long)
2357 //
2358 // inline char  Character.reverseBytes(char)
2359 // inline short     Short.reverseBytes(short)
2360 // inline int     Integer.reverseBytes(int)
2361 // inline long       Long.reverseBytes(long)
2362 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2363   Node* arg = argument(0);
2364   Node* n;
2365   switch (id) {
2366   case vmIntrinsics::_numberOfLeadingZeros_i:   n = new CountLeadingZerosINode( arg);  break;
2367   case vmIntrinsics::_numberOfLeadingZeros_l:   n = new CountLeadingZerosLNode( arg);  break;
2368   case vmIntrinsics::_numberOfTrailingZeros_i:  n = new CountTrailingZerosINode(arg);  break;
2369   case vmIntrinsics::_numberOfTrailingZeros_l:  n = new CountTrailingZerosLNode(arg);  break;
2370   case vmIntrinsics::_bitCount_i:               n = new PopCountINode(          arg);  break;
2371   case vmIntrinsics::_bitCount_l:               n = new PopCountLNode(          arg);  break;
2372   case vmIntrinsics::_reverseBytes_c:           n = new ReverseBytesUSNode(0,   arg);  break;
2373   case vmIntrinsics::_reverseBytes_s:           n = new ReverseBytesSNode( 0,   arg);  break;
2374   case vmIntrinsics::_reverseBytes_i:           n = new ReverseBytesINode( 0,   arg);  break;
2375   case vmIntrinsics::_reverseBytes_l:           n = new ReverseBytesLNode( 0,   arg);  break;
2376   default:  fatal_unexpected_iid(id);  break;
2377   }
2378   set_result(_gvn.transform(n));
2379   return true;
2380 }
2381 
2382 //----------------------------inline_unsafe_access----------------------------
2383 
2384 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2385 
2386 // Helper that guards and inserts a pre-barrier.
2387 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2388                                         Node* pre_val, bool need_mem_bar) {
2389   // We could be accessing the referent field of a reference object. If so, when G1
2390   // is enabled, we need to log the value in the referent field in an SATB buffer.
2391   // This routine performs some compile time filters and generates suitable
2392   // runtime filters that guard the pre-barrier code.
2393   // Also add memory barrier for non volatile load from the referent field
2394   // to prevent commoning of loads across safepoint.
2395   if (!UseG1GC && !need_mem_bar)
2396     return;
2397 
2398   // Some compile time checks.
2399 
2400   // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2401   const TypeX* otype = offset->find_intptr_t_type();
2402   if (otype != NULL && otype->is_con() &&
2403       otype->get_con() != java_lang_ref_Reference::referent_offset) {
2404     // Constant offset but not the reference_offset so just return
2405     return;
2406   }
2407 
2408   // We only need to generate the runtime guards for instances.
2409   const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2410   if (btype != NULL) {
2411     if (btype->isa_aryptr()) {
2412       // Array type so nothing to do
2413       return;
2414     }
2415 
2416     const TypeInstPtr* itype = btype->isa_instptr();
2417     if (itype != NULL) {
2418       // Can the klass of base_oop be statically determined to be
2419       // _not_ a sub-class of Reference and _not_ Object?
2420       ciKlass* klass = itype->klass();
2421       if ( klass->is_loaded() &&
2422           !klass->is_subtype_of(env()->Reference_klass()) &&
2423           !env()->Object_klass()->is_subtype_of(klass)) {
2424         return;
2425       }
2426     }
2427   }
2428 
2429   // The compile time filters did not reject base_oop/offset so
2430   // we need to generate the following runtime filters
2431   //
2432   // if (offset == java_lang_ref_Reference::_reference_offset) {
2433   //   if (instance_of(base, java.lang.ref.Reference)) {
2434   //     pre_barrier(_, pre_val, ...);
2435   //   }
2436   // }
2437 
2438   float likely   = PROB_LIKELY(  0.999);
2439   float unlikely = PROB_UNLIKELY(0.999);
2440 
2441   IdealKit ideal(this);
2442 #define __ ideal.
2443 
2444   Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2445 
2446   __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2447       // Update graphKit memory and control from IdealKit.
2448       sync_kit(ideal);
2449 
2450       Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2451       Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2452 
2453       // Update IdealKit memory and control from graphKit.
2454       __ sync_kit(this);
2455 
2456       Node* one = __ ConI(1);
2457       // is_instof == 0 if base_oop == NULL
2458       __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2459 
2460         // Update graphKit from IdeakKit.
2461         sync_kit(ideal);
2462 
2463         // Use the pre-barrier to record the value in the referent field
2464         pre_barrier(false /* do_load */,
2465                     __ ctrl(),
2466                     NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2467                     pre_val /* pre_val */,
2468                     T_OBJECT);
2469         if (need_mem_bar) {
2470           // Add memory barrier to prevent commoning reads from this field
2471           // across safepoint since GC can change its value.
2472           insert_mem_bar(Op_MemBarCPUOrder);
2473         }
2474         // Update IdealKit from graphKit.
2475         __ sync_kit(this);
2476 
2477       } __ end_if(); // _ref_type != ref_none
2478   } __ end_if(); // offset == referent_offset
2479 
2480   // Final sync IdealKit and GraphKit.
2481   final_sync(ideal);
2482 #undef __
2483 }
2484 




2485 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2486   // Attempt to infer a sharper value type from the offset and base type.
2487   ciKlass* sharpened_klass = NULL;
2488 
2489   // See if it is an instance field, with an object type.
2490   if (alias_type->field() != NULL) {
2491     assert(!is_native_ptr, "native pointer op cannot use a java address");
2492     if (alias_type->field()->type()->is_klass()) {
2493       sharpened_klass = alias_type->field()->type()->as_klass();
2494     }
2495   }
2496 
2497   // See if it is a narrow oop array.
2498   if (adr_type->isa_aryptr()) {
2499     if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2500       const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2501       if (elem_type != NULL) {
2502         sharpened_klass = elem_type->klass();
2503       }
2504     }
2505   }
2506 
2507   // The sharpened class might be unloaded if there is no class loader
2508   // contraint in place.
2509   if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2510     const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2511 
2512 #ifndef PRODUCT
2513     if (C->print_intrinsics() || C->print_inlining()) {
2514       tty->print("  from base type: ");  adr_type->dump();
2515       tty->print("  sharpened value: ");  tjp->dump();
2516     }
2517 #endif
2518     // Sharpen the value type.
2519     return tjp;
2520   }
2521   return NULL;
2522 }
2523 
2524 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) {
2525   if (callee()->is_static())  return false;  // caller must have the capability!
2526 
2527 #ifndef PRODUCT
2528   {
2529     ResourceMark rm;
2530     // Check the signatures.
2531     ciSignature* sig = callee()->signature();
2532 #ifdef ASSERT
2533     if (!is_store) {
2534       // Object getObject(Object base, int/long offset), etc.
2535       BasicType rtype = sig->return_type()->basic_type();
2536       if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2537           rtype = T_ADDRESS;  // it is really a C void*
2538       assert(rtype == type, "getter must return the expected value");
2539       if (!is_native_ptr) {
2540         assert(sig->count() == 2, "oop getter has 2 arguments");
2541         assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2542         assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2543       } else {
2544         assert(sig->count() == 1, "native getter has 1 argument");
2545         assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2546       }
2547     } else {
2548       // void putObject(Object base, int/long offset, Object x), etc.
2549       assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2550       if (!is_native_ptr) {
2551         assert(sig->count() == 3, "oop putter has 3 arguments");
2552         assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2553         assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2554       } else {
2555         assert(sig->count() == 2, "native putter has 2 arguments");
2556         assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2557       }
2558       BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2559       if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2560         vtype = T_ADDRESS;  // it is really a C void*
2561       assert(vtype == type, "putter must accept the expected value");
2562     }
2563 #endif // ASSERT
2564  }
2565 #endif //PRODUCT
2566 
2567   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2568 
2569   Node* receiver = argument(0);  // type: oop
2570 
2571   // Build address expression.
2572   Node* adr;
2573   Node* heap_base_oop = top();
2574   Node* offset = top();
2575   Node* val;
2576 
2577   if (!is_native_ptr) {
2578     // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2579     Node* base = argument(1);  // type: oop
2580     // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2581     offset = argument(2);  // type: long





2582     // 32-bit machines ignore the high half!
2583     offset = ConvL2X(offset);
2584     adr = make_unsafe_address(base, offset, /*decode=*/!unaligned);
2585     heap_base_oop = base;
2586     val = is_store ? argument(4) : NULL;
2587   } else {
2588     Node* ptr = argument(1);  // type: long
2589     ptr = ConvL2X(ptr);  // adjust Java long to machine word
2590     adr = make_unsafe_address(NULL, ptr, /*decode=*/false);
2591     val = is_store ? argument(3) : NULL;
2592   }
2593 
2594   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2595 
2596   // First guess at the value type.
2597   const Type *value_type = Type::get_const_basic_type(type);
2598 
2599   // Try to categorize the address.  If it comes up as TypeJavaPtr::BOTTOM,
2600   // there was not enough information to nail it down.
2601   Compile::AliasType* alias_type = C->alias_type(adr_type);
2602   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2603 
2604   // We will need memory barriers unless we can determine a unique
2605   // alias category for this reference.  (Note:  If for some reason
2606   // the barriers get omitted and the unsafe reference begins to "pollute"
2607   // the alias analysis of the rest of the graph, either Compile::can_alias
2608   // or Compile::must_alias will throw a diagnostic assert.)
2609   bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2610 
2611   // If we are reading the value of the referent field of a Reference
2612   // object (either by using Unsafe directly or through reflection)
2613   // then, if G1 is enabled, we need to record the referent in an
2614   // SATB log buffer using the pre-barrier mechanism.
2615   // Also we need to add memory barrier to prevent commoning reads
2616   // from this field across safepoint since GC can change its value.
2617   bool need_read_barrier = !is_native_ptr && !is_store &&
2618                            offset != top() && heap_base_oop != top();
2619 
2620   if (!is_store && type == T_OBJECT) {
2621     const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2622     if (tjp != NULL) {
2623       value_type = tjp;
2624     }
2625   }
2626 
2627   receiver = null_check(receiver);
2628   if (stopped()) {
2629     return true;
2630   }
2631   // Heap pointers get a null-check from the interpreter,
2632   // as a courtesy.  However, this is not guaranteed by Unsafe,
2633   // and it is not possible to fully distinguish unintended nulls
2634   // from intended ones in this API.
2635 
2636   if (is_volatile) {
2637     // We need to emit leading and trailing CPU membars (see below) in
2638     // addition to memory membars when is_volatile. This is a little
2639     // too strong, but avoids the need to insert per-alias-type
2640     // volatile membars (for stores; compare Parse::do_put_xxx), which
2641     // we cannot do effectively here because we probably only have a
2642     // rough approximation of type.
2643     need_mem_bar = true;
2644     // For Stores, place a memory ordering barrier now.
2645     if (is_store) {
2646       insert_mem_bar(Op_MemBarRelease);
2647     } else {
2648       if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2649         insert_mem_bar(Op_MemBarVolatile);
2650       }
2651     }
2652   }
2653 
2654   // Memory barrier to prevent normal and 'unsafe' accesses from
2655   // bypassing each other.  Happens after null checks, so the
2656   // exception paths do not take memory state from the memory barrier,
2657   // so there's no problems making a strong assert about mixing users
2658   // of safe & unsafe memory.
2659   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2660 
2661    if (!is_store) {
2662     Node* p = NULL;
2663     // Try to constant fold a load from a constant field
2664     ciField* field = alias_type->field();
2665     if (heap_base_oop != top() &&
2666         field != NULL && field->is_constant() && field->layout_type() == type) {
2667       // final or stable field
2668       ciField* field = alias_type->field();
2669       const Type* con_type = Type::make_constant(field, heap_base_oop);
2670       if (con_type != NULL) {
2671         if (TrustFinalNonStaticFields &&
2672             !field->is_static() && heap_base_oop->is_Con()) {
2673           const TypeOopPtr* oop_ptr = heap_base_oop->bottom_type()->isa_oopptr();
2674           ciObject* constant_oop = oop_ptr->const_oop();
2675           C->dependencies()->assert_constant_field_value_instance(field, constant_oop);
2676         }
2677         p = makecon(con_type);
2678       }
2679     }
2680     if (p == NULL) {
2681       MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2682       // To be valid, unsafe loads may depend on other conditions than
2683       // the one that guards them: pin the Load node
2684       p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2685       // load value
2686       switch (type) {
2687       case T_BOOLEAN:
2688       case T_CHAR:
2689       case T_BYTE:
2690       case T_SHORT:
2691       case T_INT:
2692       case T_LONG:
2693       case T_FLOAT:
2694       case T_DOUBLE:
2695         break;
2696       case T_OBJECT:
2697         if (need_read_barrier) {
2698           insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2699         }
2700         break;
2701       case T_ADDRESS:
2702         // Cast to an int type.
2703         p = _gvn.transform(new CastP2XNode(NULL, p));
2704         p = ConvX2UL(p);
2705         break;
2706       default:
2707         fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2708         break;
2709       }
2710     }
2711     // The load node has the control of the preceding MemBarCPUOrder.  All
2712     // following nodes will have the control of the MemBarCPUOrder inserted at
2713     // the end of this method.  So, pushing the load onto the stack at a later
2714     // point is fine.
2715     set_result(p);
2716   } else {
2717     // place effect of store into memory
2718     switch (type) {
2719     case T_DOUBLE:
2720       val = dstore_rounding(val);
2721       break;
2722     case T_ADDRESS:
2723       // Repackage the long as a pointer.
2724       val = ConvL2X(val);
2725       val = _gvn.transform(new CastX2PNode(val));
2726       break;
2727     }
2728 
2729     { // Need to check all dependent nmethods when final field is updated through Unsafe.
2730       Node* final_bit = _gvn.transform(new AndXNode(/*offset*/argument(2), MakeConX(Unsafe::final_mask)));
2731       Node* cmp_final_bit = _gvn.transform(new CmpXNode(final_bit, MakeConX(0)));
2732       Node* bol_final_bit = _gvn.transform(new BoolNode(cmp_final_bit, BoolTest::eq));
2733 
2734       BuildCutout unless(this, bol_final_bit, PROB_MAX);
2735       uncommon_trap(Deoptimization::Reason_intrinsic,
2736                     Deoptimization::Action_none,
2737                     NULL, "final_field_unsafe_update");
2738     }
2739 
2740     MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2741     if (type != T_OBJECT ) {
2742       (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2743     } else {
2744       // Possibly an oop being stored to Java heap or native memory
2745       if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2746         // oop to Java heap.
2747         (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2748       } else {
2749         // We can't tell at compile time if we are storing in the Java heap or outside
2750         // of it. So we need to emit code to conditionally do the proper type of
2751         // store.
2752 
2753         IdealKit ideal(this);
2754 #define __ ideal.
2755         // QQQ who knows what probability is here??
2756         __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2757           // Sync IdealKit and graphKit.
2758           sync_kit(ideal);
2759           Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2760           // Update IdealKit memory.
2761           __ sync_kit(this);
2762         } __ else_(); {
2763           __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2764         } __ end_if();
2765         // Final sync IdealKit and GraphKit.
2766         final_sync(ideal);
2767 #undef __
2768       }
2769     }
2770   }
2771 
2772   if (is_volatile) {
2773     if (!is_store) {
2774       insert_mem_bar(Op_MemBarAcquire);
2775     } else {
2776       if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2777         insert_mem_bar(Op_MemBarVolatile);
2778       }
2779     }
2780   }
2781 
2782   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2783 
2784   return true;
2785 }
2786 
2787 //----------------------------inline_unsafe_load_store----------------------------
2788 // This method serves a couple of different customers (depending on LoadStoreKind):
2789 //
2790 // LS_cmpxchg:
2791 //   public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2792 //   public final native boolean compareAndSwapInt(   Object o, long offset, int    expected, int    x);
2793 //   public final native boolean compareAndSwapLong(  Object o, long offset, long   expected, long   x);
2794 //
2795 // LS_xadd:
2796 //   public int  getAndAddInt( Object o, long offset, int  delta)
2797 //   public long getAndAddLong(Object o, long offset, long delta)
2798 //
2799 // LS_xchg:
2800 //   int    getAndSet(Object o, long offset, int    newValue)
2801 //   long   getAndSet(Object o, long offset, long   newValue)
2802 //   Object getAndSet(Object o, long offset, Object newValue)
2803 //
2804 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2805   // This basic scheme here is the same as inline_unsafe_access, but
2806   // differs in enough details that combining them would make the code
2807   // overly confusing.  (This is a true fact! I originally combined
2808   // them, but even I was confused by it!) As much code/comments as
2809   // possible are retained from inline_unsafe_access though to make
2810   // the correspondences clearer. - dl
2811 
2812   if (callee()->is_static())  return false;  // caller must have the capability!
2813 
2814 #ifndef PRODUCT
2815   BasicType rtype;
2816   {
2817     ResourceMark rm;
2818     // Check the signatures.
2819     ciSignature* sig = callee()->signature();
2820     rtype = sig->return_type()->basic_type();
2821     if (kind == LS_xadd || kind == LS_xchg) {
2822       // Check the signatures.
2823 #ifdef ASSERT
2824       assert(rtype == type, "get and set must return the expected type");
2825       assert(sig->count() == 3, "get and set has 3 arguments");
2826       assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2827       assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2828       assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2829 #endif // ASSERT
2830     } else if (kind == LS_cmpxchg) {
2831       // Check the signatures.
2832 #ifdef ASSERT
2833       assert(rtype == T_BOOLEAN, "CAS must return boolean");
2834       assert(sig->count() == 4, "CAS has 4 arguments");
2835       assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2836       assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2837 #endif // ASSERT
2838     } else {
2839       ShouldNotReachHere();
2840     }
2841   }
2842 #endif //PRODUCT
2843 
2844   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2845 
2846   // Get arguments:
2847   Node* receiver = NULL;
2848   Node* base     = NULL;
2849   Node* offset   = NULL;
2850   Node* oldval   = NULL;
2851   Node* newval   = NULL;
2852   if (kind == LS_cmpxchg) {
2853     const bool two_slot_type = type2size[type] == 2;
2854     receiver = argument(0);  // type: oop
2855     base     = argument(1);  // type: oop
2856     offset   = argument(2);  // type: long
2857     oldval   = argument(4);  // type: oop, int, or long
2858     newval   = argument(two_slot_type ? 6 : 5);  // type: oop, int, or long
2859   } else if (kind == LS_xadd || kind == LS_xchg){
2860     receiver = argument(0);  // type: oop
2861     base     = argument(1);  // type: oop
2862     offset   = argument(2);  // type: long
2863     oldval   = NULL;
2864     newval   = argument(4);  // type: oop, int, or long
2865   }
2866 
2867   // Null check receiver.
2868   receiver = null_check(receiver);
2869   if (stopped()) {
2870     return true;
2871   }
2872 
2873   // Build field offset expression.




2874   // 32-bit machines ignore the high half of long offsets
2875   offset = ConvL2X(offset);
2876   Node* adr = make_unsafe_address(base, offset, /*decode=*/true);
2877   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2878 
2879   // For CAS, unlike inline_unsafe_access, there seems no point in
2880   // trying to refine types. Just use the coarse types here.
2881   const Type *value_type = Type::get_const_basic_type(type);
2882   Compile::AliasType* alias_type = C->alias_type(adr_type);
2883   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2884 
2885   if (kind == LS_xchg && type == T_OBJECT) {
2886     const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2887     if (tjp != NULL) {
2888       value_type = tjp;
2889     }
2890   }
2891 
2892   int alias_idx = C->get_alias_index(adr_type);
2893 
2894   // Memory-model-wise, a LoadStore acts like a little synchronized
2895   // block, so needs barriers on each side.  These don't translate
2896   // into actual barriers on most machines, but we still need rest of
2897   // compiler to respect ordering.
2898 
2899   insert_mem_bar(Op_MemBarRelease);
2900   insert_mem_bar(Op_MemBarCPUOrder);
2901 
2902   // 4984716: MemBars must be inserted before this
2903   //          memory node in order to avoid a false
2904   //          dependency which will confuse the scheduler.
2905   Node *mem = memory(alias_idx);
2906 
2907   // For now, we handle only those cases that actually exist: ints,
2908   // longs, and Object. Adding others should be straightforward.
2909   Node* load_store;
2910   switch(type) {
2911   case T_INT:
2912     if (kind == LS_xadd) {
2913       load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2914     } else if (kind == LS_xchg) {
2915       load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2916     } else if (kind == LS_cmpxchg) {
2917       load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2918     } else {
2919       ShouldNotReachHere();
2920     }
2921     break;
2922   case T_LONG:
2923     if (kind == LS_xadd) {
2924       load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2925     } else if (kind == LS_xchg) {
2926       load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2927     } else if (kind == LS_cmpxchg) {
2928       load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2929     } else {
2930       ShouldNotReachHere();
2931     }
2932     break;
2933   case T_OBJECT:
2934     // Transformation of a value which could be NULL pointer (CastPP #NULL)
2935     // could be delayed during Parse (for example, in adjust_map_after_if()).
2936     // Execute transformation here to avoid barrier generation in such case.
2937     if (_gvn.type(newval) == TypePtr::NULL_PTR)
2938       newval = _gvn.makecon(TypePtr::NULL_PTR);
2939 
2940     // Reference stores need a store barrier.
2941     if (kind == LS_xchg) {
2942       // If pre-barrier must execute before the oop store, old value will require do_load here.
2943       if (!can_move_pre_barrier()) {
2944         pre_barrier(true /* do_load*/,
2945                     control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2946                     NULL /* pre_val*/,
2947                     T_OBJECT);
2948       } // Else move pre_barrier to use load_store value, see below.
2949     } else if (kind == LS_cmpxchg) {
2950       // Same as for newval above:
2951       if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2952         oldval = _gvn.makecon(TypePtr::NULL_PTR);
2953       }
2954       // The only known value which might get overwritten is oldval.
2955       pre_barrier(false /* do_load */,
2956                   control(), NULL, NULL, max_juint, NULL, NULL,
2957                   oldval /* pre_val */,
2958                   T_OBJECT);
2959     } else {
2960       ShouldNotReachHere();
2961     }
2962 
2963 #ifdef _LP64
2964     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2965       Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2966       if (kind == LS_xchg) {
2967         load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2968                                                        newval_enc, adr_type, value_type->make_narrowoop()));
2969       } else {
2970         assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2971         Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2972         load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2973                                                                 newval_enc, oldval_enc));
2974       }
2975     } else
2976 #endif
2977     {
2978       if (kind == LS_xchg) {
2979         load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2980       } else {
2981         assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2982         load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2983       }
2984     }
2985     post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2986     break;
2987   default:
2988     fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2989     break;
2990   }
2991 
2992   // SCMemProjNodes represent the memory state of a LoadStore. Their
2993   // main role is to prevent LoadStore nodes from being optimized away
2994   // when their results aren't used.
2995   Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2996   set_memory(proj, alias_idx);
2997 
2998   if (type == T_OBJECT && kind == LS_xchg) {
2999 #ifdef _LP64
3000     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3001       load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3002     }
3003 #endif
3004     if (can_move_pre_barrier()) {
3005       // Don't need to load pre_val. The old value is returned by load_store.
3006       // The pre_barrier can execute after the xchg as long as no safepoint
3007       // gets inserted between them.
3008       pre_barrier(false /* do_load */,
3009                   control(), NULL, NULL, max_juint, NULL, NULL,
3010                   load_store /* pre_val */,
3011                   T_OBJECT);
3012     }
3013   }
3014 
3015   // Add the trailing membar surrounding the access
3016   insert_mem_bar(Op_MemBarCPUOrder);
3017   insert_mem_bar(Op_MemBarAcquire);
3018 
3019   assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3020   set_result(load_store);
3021   return true;
3022 }
3023 
3024 //----------------------------inline_unsafe_ordered_store----------------------
3025 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3026 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3027 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3028 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3029   // This is another variant of inline_unsafe_access, differing in
3030   // that it always issues store-store ("release") barrier and ensures
3031   // store-atomicity (which only matters for "long").
3032 
3033   if (callee()->is_static())  return false;  // caller must have the capability!
3034 
3035 #ifndef PRODUCT
3036   {
3037     ResourceMark rm;
3038     // Check the signatures.
3039     ciSignature* sig = callee()->signature();
3040 #ifdef ASSERT
3041     BasicType rtype = sig->return_type()->basic_type();
3042     assert(rtype == T_VOID, "must return void");
3043     assert(sig->count() == 3, "has 3 arguments");
3044     assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3045     assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3046 #endif // ASSERT
3047   }
3048 #endif //PRODUCT
3049 
3050   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
3051 
3052   // Get arguments:
3053   Node* receiver = argument(0);  // type: oop
3054   Node* base     = argument(1);  // type: oop
3055   Node* offset   = argument(2);  // type: long
3056   Node* val      = argument(4);  // type: oop, int, or long
3057 
3058   // Null check receiver.
3059   receiver = null_check(receiver);
3060   if (stopped()) {
3061     return true;
3062   }
3063 
3064   // Build field offset expression.

3065   // 32-bit machines ignore the high half of long offsets
3066   offset = ConvL2X(offset);
3067   Node* adr = make_unsafe_address(base, offset, /*decode=*/true);
3068   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3069   const Type *value_type = Type::get_const_basic_type(type);
3070   Compile::AliasType* alias_type = C->alias_type(adr_type);
3071 
3072   insert_mem_bar(Op_MemBarRelease);
3073   insert_mem_bar(Op_MemBarCPUOrder);
3074   // Ensure that the store is atomic for longs:
3075   const bool require_atomic_access = true;
3076   Node* store;
3077   if (type == T_OBJECT) // reference stores need a store barrier.
3078     store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3079   else {
3080     store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3081   }
3082   insert_mem_bar(Op_MemBarCPUOrder);
3083   return true;
3084 }
3085 
3086 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3087   // Regardless of form, don't allow previous ld/st to move down,
3088   // then issue acquire, release, or volatile mem_bar.
3089   insert_mem_bar(Op_MemBarCPUOrder);
3090   switch(id) {
3091     case vmIntrinsics::_loadFence:
3092       insert_mem_bar(Op_LoadFence);
3093       return true;
3094     case vmIntrinsics::_storeFence:
3095       insert_mem_bar(Op_StoreFence);
3096       return true;
3097     case vmIntrinsics::_fullFence:
3098       insert_mem_bar(Op_MemBarVolatile);
3099       return true;
3100     default:
3101       fatal_unexpected_iid(id);
3102       return false;
3103   }
3104 }
3105 
3106 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3107   if (!kls->is_Con()) {
3108     return true;
3109   }
3110   const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3111   if (klsptr == NULL) {
3112     return true;
3113   }
3114   ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3115   // don't need a guard for a klass that is already initialized
3116   return !ik->is_initialized();
3117 }
3118 
3119 //----------------------------inline_unsafe_allocate---------------------------
3120 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3121 bool LibraryCallKit::inline_unsafe_allocate() {
3122   if (callee()->is_static())  return false;  // caller must have the capability!
3123 
3124   null_check_receiver();  // null-check, then ignore
3125   Node* cls = null_check(argument(1));
3126   if (stopped())  return true;
3127 
3128   Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3129   kls = null_check(kls);
3130   if (stopped())  return true;  // argument was like int.class
3131 
3132   Node* test = NULL;
3133   if (LibraryCallKit::klass_needs_init_guard(kls)) {
3134     // Note:  The argument might still be an illegal value like
3135     // Serializable.class or Object[].class.   The runtime will handle it.
3136     // But we must make an explicit check for initialization.
3137     Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3138     // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3139     // can generate code to load it as unsigned byte.
3140     Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3141     Node* bits = intcon(InstanceKlass::fully_initialized);
3142     test = _gvn.transform(new SubINode(inst, bits));
3143     // The 'test' is non-zero if we need to take a slow path.
3144   }
3145 
3146   Node* obj = new_instance(kls, test);
3147   set_result(obj);
3148   return true;
3149 }
3150 
3151 #ifdef TRACE_HAVE_INTRINSICS
3152 /*
3153  * oop -> myklass
3154  * myklass->trace_id |= USED
3155  * return myklass->trace_id & ~0x3
3156  */
3157 bool LibraryCallKit::inline_native_classID() {
3158   null_check_receiver();  // null-check, then ignore
3159   Node* cls = null_check(argument(1), T_OBJECT);
3160   Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3161   kls = null_check(kls, T_OBJECT);
3162   ByteSize offset = TRACE_ID_OFFSET;
3163   Node* insp = basic_plus_adr(kls, in_bytes(offset));
3164   Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3165   Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3166   Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3167   Node* clsused = longcon(0x01l); // set the class bit
3168   Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3169 
3170   const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3171   store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3172   set_result(andl);
3173   return true;
3174 }
3175 
3176 bool LibraryCallKit::inline_native_threadID() {
3177   Node* tls_ptr = NULL;
3178   Node* cur_thr = generate_current_thread(tls_ptr);
3179   Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3180   Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3181   p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3182 
3183   Node* threadid = NULL;
3184   size_t thread_id_size = OSThread::thread_id_size();
3185   if (thread_id_size == (size_t) BytesPerLong) {
3186     threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3187   } else if (thread_id_size == (size_t) BytesPerInt) {
3188     threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3189   } else {
3190     ShouldNotReachHere();
3191   }
3192   set_result(threadid);
3193   return true;
3194 }
3195 #endif
3196 
3197 //------------------------inline_native_time_funcs--------------
3198 // inline code for System.currentTimeMillis() and System.nanoTime()
3199 // these have the same type and signature
3200 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3201   const TypeFunc* tf = OptoRuntime::void_long_Type();
3202   const TypePtr* no_memory_effects = NULL;
3203   Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3204   Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3205 #ifdef ASSERT
3206   Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3207   assert(value_top == top(), "second value must be top");
3208 #endif
3209   set_result(value);
3210   return true;
3211 }
3212 
3213 //------------------------inline_native_currentThread------------------
3214 bool LibraryCallKit::inline_native_currentThread() {
3215   Node* junk = NULL;
3216   set_result(generate_current_thread(junk));
3217   return true;
3218 }
3219 
3220 //------------------------inline_native_isInterrupted------------------
3221 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3222 bool LibraryCallKit::inline_native_isInterrupted() {
3223   // Add a fast path to t.isInterrupted(clear_int):
3224   //   (t == Thread.current() &&
3225   //    (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3226   //   ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3227   // So, in the common case that the interrupt bit is false,
3228   // we avoid making a call into the VM.  Even if the interrupt bit
3229   // is true, if the clear_int argument is false, we avoid the VM call.
3230   // However, if the receiver is not currentThread, we must call the VM,
3231   // because there must be some locking done around the operation.
3232 
3233   // We only go to the fast case code if we pass two guards.
3234   // Paths which do not pass are accumulated in the slow_region.
3235 
3236   enum {
3237     no_int_result_path   = 1, // t == Thread.current() && !TLS._osthread._interrupted
3238     no_clear_result_path = 2, // t == Thread.current() &&  TLS._osthread._interrupted && !clear_int
3239     slow_result_path     = 3, // slow path: t.isInterrupted(clear_int)
3240     PATH_LIMIT
3241   };
3242 
3243   // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3244   // out of the function.
3245   insert_mem_bar(Op_MemBarCPUOrder);
3246 
3247   RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3248   PhiNode*    result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3249 
3250   RegionNode* slow_region = new RegionNode(1);
3251   record_for_igvn(slow_region);
3252 
3253   // (a) Receiving thread must be the current thread.
3254   Node* rec_thr = argument(0);
3255   Node* tls_ptr = NULL;
3256   Node* cur_thr = generate_current_thread(tls_ptr);
3257   Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3258   Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3259 
3260   generate_slow_guard(bol_thr, slow_region);
3261 
3262   // (b) Interrupt bit on TLS must be false.
3263   Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3264   Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3265   p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3266 
3267   // Set the control input on the field _interrupted read to prevent it floating up.
3268   Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3269   Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3270   Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3271 
3272   IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3273 
3274   // First fast path:  if (!TLS._interrupted) return false;
3275   Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3276   result_rgn->init_req(no_int_result_path, false_bit);
3277   result_val->init_req(no_int_result_path, intcon(0));
3278 
3279   // drop through to next case
3280   set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3281 
3282 #ifndef TARGET_OS_FAMILY_windows
3283   // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3284   Node* clr_arg = argument(1);
3285   Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3286   Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3287   IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3288 
3289   // Second fast path:  ... else if (!clear_int) return true;
3290   Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3291   result_rgn->init_req(no_clear_result_path, false_arg);
3292   result_val->init_req(no_clear_result_path, intcon(1));
3293 
3294   // drop through to next case
3295   set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3296 #else
3297   // To return true on Windows you must read the _interrupted field
3298   // and check the the event state i.e. take the slow path.
3299 #endif // TARGET_OS_FAMILY_windows
3300 
3301   // (d) Otherwise, go to the slow path.
3302   slow_region->add_req(control());
3303   set_control( _gvn.transform(slow_region));
3304 
3305   if (stopped()) {
3306     // There is no slow path.
3307     result_rgn->init_req(slow_result_path, top());
3308     result_val->init_req(slow_result_path, top());
3309   } else {
3310     // non-virtual because it is a private non-static
3311     CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3312 
3313     Node* slow_val = set_results_for_java_call(slow_call);
3314     // this->control() comes from set_results_for_java_call
3315 
3316     Node* fast_io  = slow_call->in(TypeFunc::I_O);
3317     Node* fast_mem = slow_call->in(TypeFunc::Memory);
3318 
3319     // These two phis are pre-filled with copies of of the fast IO and Memory
3320     PhiNode* result_mem  = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3321     PhiNode* result_io   = PhiNode::make(result_rgn, fast_io,  Type::ABIO);
3322 
3323     result_rgn->init_req(slow_result_path, control());
3324     result_io ->init_req(slow_result_path, i_o());
3325     result_mem->init_req(slow_result_path, reset_memory());
3326     result_val->init_req(slow_result_path, slow_val);
3327 
3328     set_all_memory(_gvn.transform(result_mem));
3329     set_i_o(       _gvn.transform(result_io));
3330   }
3331 
3332   C->set_has_split_ifs(true); // Has chance for split-if optimization
3333   set_result(result_rgn, result_val);
3334   return true;
3335 }
3336 
3337 //---------------------------load_mirror_from_klass----------------------------
3338 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3339 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3340   Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3341   return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3342 }
3343 
3344 //-----------------------load_klass_from_mirror_common-------------------------
3345 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3346 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3347 // and branch to the given path on the region.
3348 // If never_see_null, take an uncommon trap on null, so we can optimistically
3349 // compile for the non-null case.
3350 // If the region is NULL, force never_see_null = true.
3351 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3352                                                     bool never_see_null,
3353                                                     RegionNode* region,
3354                                                     int null_path,
3355                                                     int offset) {
3356   if (region == NULL)  never_see_null = true;
3357   Node* p = basic_plus_adr(mirror, offset);
3358   const TypeKlassPtr*  kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3359   Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3360   Node* null_ctl = top();
3361   kls = null_check_oop(kls, &null_ctl, never_see_null);
3362   if (region != NULL) {
3363     // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3364     region->init_req(null_path, null_ctl);
3365   } else {
3366     assert(null_ctl == top(), "no loose ends");
3367   }
3368   return kls;
3369 }
3370 
3371 //--------------------(inline_native_Class_query helpers)---------------------
3372 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3373 // Fall through if (mods & mask) == bits, take the guard otherwise.
3374 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3375   // Branch around if the given klass has the given modifier bit set.
3376   // Like generate_guard, adds a new path onto the region.
3377   Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3378   Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3379   Node* mask = intcon(modifier_mask);
3380   Node* bits = intcon(modifier_bits);
3381   Node* mbit = _gvn.transform(new AndINode(mods, mask));
3382   Node* cmp  = _gvn.transform(new CmpINode(mbit, bits));
3383   Node* bol  = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3384   return generate_fair_guard(bol, region);
3385 }
3386 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3387   return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3388 }
3389 
3390 //-------------------------inline_native_Class_query-------------------
3391 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3392   const Type* return_type = TypeInt::BOOL;
3393   Node* prim_return_value = top();  // what happens if it's a primitive class?
3394   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3395   bool expect_prim = false;     // most of these guys expect to work on refs
3396 
3397   enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3398 
3399   Node* mirror = argument(0);
3400   Node* obj    = top();
3401 
3402   switch (id) {
3403   case vmIntrinsics::_isInstance:
3404     // nothing is an instance of a primitive type
3405     prim_return_value = intcon(0);
3406     obj = argument(1);
3407     break;
3408   case vmIntrinsics::_getModifiers:
3409     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3410     assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3411     return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3412     break;
3413   case vmIntrinsics::_isInterface:
3414     prim_return_value = intcon(0);
3415     break;
3416   case vmIntrinsics::_isArray:
3417     prim_return_value = intcon(0);
3418     expect_prim = true;  // cf. ObjectStreamClass.getClassSignature
3419     break;
3420   case vmIntrinsics::_isPrimitive:
3421     prim_return_value = intcon(1);
3422     expect_prim = true;  // obviously
3423     break;
3424   case vmIntrinsics::_getSuperclass:
3425     prim_return_value = null();
3426     return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3427     break;
3428   case vmIntrinsics::_getClassAccessFlags:
3429     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3430     return_type = TypeInt::INT;  // not bool!  6297094
3431     break;
3432   default:
3433     fatal_unexpected_iid(id);
3434     break;
3435   }
3436 
3437   const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3438   if (mirror_con == NULL)  return false;  // cannot happen?
3439 
3440 #ifndef PRODUCT
3441   if (C->print_intrinsics() || C->print_inlining()) {
3442     ciType* k = mirror_con->java_mirror_type();
3443     if (k) {
3444       tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3445       k->print_name();
3446       tty->cr();
3447     }
3448   }
3449 #endif
3450 
3451   // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3452   RegionNode* region = new RegionNode(PATH_LIMIT);
3453   record_for_igvn(region);
3454   PhiNode* phi = new PhiNode(region, return_type);
3455 
3456   // The mirror will never be null of Reflection.getClassAccessFlags, however
3457   // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3458   // if it is. See bug 4774291.
3459 
3460   // For Reflection.getClassAccessFlags(), the null check occurs in
3461   // the wrong place; see inline_unsafe_access(), above, for a similar
3462   // situation.
3463   mirror = null_check(mirror);
3464   // If mirror or obj is dead, only null-path is taken.
3465   if (stopped())  return true;
3466 
3467   if (expect_prim)  never_see_null = false;  // expect nulls (meaning prims)
3468 
3469   // Now load the mirror's klass metaobject, and null-check it.
3470   // Side-effects region with the control path if the klass is null.
3471   Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3472   // If kls is null, we have a primitive mirror.
3473   phi->init_req(_prim_path, prim_return_value);
3474   if (stopped()) { set_result(region, phi); return true; }
3475   bool safe_for_replace = (region->in(_prim_path) == top());
3476 
3477   Node* p;  // handy temp
3478   Node* null_ctl;
3479 
3480   // Now that we have the non-null klass, we can perform the real query.
3481   // For constant classes, the query will constant-fold in LoadNode::Value.
3482   Node* query_value = top();
3483   switch (id) {
3484   case vmIntrinsics::_isInstance:
3485     // nothing is an instance of a primitive type
3486     query_value = gen_instanceof(obj, kls, safe_for_replace);
3487     break;
3488 
3489   case vmIntrinsics::_getModifiers:
3490     p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3491     query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3492     break;
3493 
3494   case vmIntrinsics::_isInterface:
3495     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3496     if (generate_interface_guard(kls, region) != NULL)
3497       // A guard was added.  If the guard is taken, it was an interface.
3498       phi->add_req(intcon(1));
3499     // If we fall through, it's a plain class.
3500     query_value = intcon(0);
3501     break;
3502 
3503   case vmIntrinsics::_isArray:
3504     // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3505     if (generate_array_guard(kls, region) != NULL)
3506       // A guard was added.  If the guard is taken, it was an array.
3507       phi->add_req(intcon(1));
3508     // If we fall through, it's a plain class.
3509     query_value = intcon(0);
3510     break;
3511 
3512   case vmIntrinsics::_isPrimitive:
3513     query_value = intcon(0); // "normal" path produces false
3514     break;
3515 
3516   case vmIntrinsics::_getSuperclass:
3517     // The rules here are somewhat unfortunate, but we can still do better
3518     // with random logic than with a JNI call.
3519     // Interfaces store null or Object as _super, but must report null.
3520     // Arrays store an intermediate super as _super, but must report Object.
3521     // Other types can report the actual _super.
3522     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3523     if (generate_interface_guard(kls, region) != NULL)
3524       // A guard was added.  If the guard is taken, it was an interface.
3525       phi->add_req(null());
3526     if (generate_array_guard(kls, region) != NULL)
3527       // A guard was added.  If the guard is taken, it was an array.
3528       phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3529     // If we fall through, it's a plain class.  Get its _super.
3530     p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3531     kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3532     null_ctl = top();
3533     kls = null_check_oop(kls, &null_ctl);
3534     if (null_ctl != top()) {
3535       // If the guard is taken, Object.superClass is null (both klass and mirror).
3536       region->add_req(null_ctl);
3537       phi   ->add_req(null());
3538     }
3539     if (!stopped()) {
3540       query_value = load_mirror_from_klass(kls);
3541     }
3542     break;
3543 
3544   case vmIntrinsics::_getClassAccessFlags:
3545     p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3546     query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3547     break;
3548 
3549   default:
3550     fatal_unexpected_iid(id);
3551     break;
3552   }
3553 
3554   // Fall-through is the normal case of a query to a real class.
3555   phi->init_req(1, query_value);
3556   region->init_req(1, control());
3557 
3558   C->set_has_split_ifs(true); // Has chance for split-if optimization
3559   set_result(region, phi);
3560   return true;
3561 }
3562 
3563 //-------------------------inline_Class_cast-------------------
3564 bool LibraryCallKit::inline_Class_cast() {
3565   Node* mirror = argument(0); // Class
3566   Node* obj    = argument(1);
3567   const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3568   if (mirror_con == NULL) {
3569     return false;  // dead path (mirror->is_top()).
3570   }
3571   if (obj == NULL || obj->is_top()) {
3572     return false;  // dead path
3573   }
3574   const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3575 
3576   // First, see if Class.cast() can be folded statically.
3577   // java_mirror_type() returns non-null for compile-time Class constants.
3578   ciType* tm = mirror_con->java_mirror_type();
3579   if (tm != NULL && tm->is_klass() &&
3580       tp != NULL && tp->klass() != NULL) {
3581     if (!tp->klass()->is_loaded()) {
3582       // Don't use intrinsic when class is not loaded.
3583       return false;
3584     } else {
3585       int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3586       if (static_res == Compile::SSC_always_true) {
3587         // isInstance() is true - fold the code.
3588         set_result(obj);
3589         return true;
3590       } else if (static_res == Compile::SSC_always_false) {
3591         // Don't use intrinsic, have to throw ClassCastException.
3592         // If the reference is null, the non-intrinsic bytecode will
3593         // be optimized appropriately.
3594         return false;
3595       }
3596     }
3597   }
3598 
3599   // Bailout intrinsic and do normal inlining if exception path is frequent.
3600   if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3601     return false;
3602   }
3603 
3604   // Generate dynamic checks.
3605   // Class.cast() is java implementation of _checkcast bytecode.
3606   // Do checkcast (Parse::do_checkcast()) optimizations here.
3607 
3608   mirror = null_check(mirror);
3609   // If mirror is dead, only null-path is taken.
3610   if (stopped()) {
3611     return true;
3612   }
3613 
3614   // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3615   enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3616   RegionNode* region = new RegionNode(PATH_LIMIT);
3617   record_for_igvn(region);
3618 
3619   // Now load the mirror's klass metaobject, and null-check it.
3620   // If kls is null, we have a primitive mirror and
3621   // nothing is an instance of a primitive type.
3622   Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3623 
3624   Node* res = top();
3625   if (!stopped()) {
3626     Node* bad_type_ctrl = top();
3627     // Do checkcast optimizations.
3628     res = gen_checkcast(obj, kls, &bad_type_ctrl);
3629     region->init_req(_bad_type_path, bad_type_ctrl);
3630   }
3631   if (region->in(_prim_path) != top() ||
3632       region->in(_bad_type_path) != top()) {
3633     // Let Interpreter throw ClassCastException.
3634     PreserveJVMState pjvms(this);
3635     set_control(_gvn.transform(region));
3636     uncommon_trap(Deoptimization::Reason_intrinsic,
3637                   Deoptimization::Action_maybe_recompile);
3638   }
3639   if (!stopped()) {
3640     set_result(res);
3641   }
3642   return true;
3643 }
3644 
3645 
3646 //--------------------------inline_native_subtype_check------------------------
3647 // This intrinsic takes the JNI calls out of the heart of
3648 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3649 bool LibraryCallKit::inline_native_subtype_check() {
3650   // Pull both arguments off the stack.
3651   Node* args[2];                // two java.lang.Class mirrors: superc, subc
3652   args[0] = argument(0);
3653   args[1] = argument(1);
3654   Node* klasses[2];             // corresponding Klasses: superk, subk
3655   klasses[0] = klasses[1] = top();
3656 
3657   enum {
3658     // A full decision tree on {superc is prim, subc is prim}:
3659     _prim_0_path = 1,           // {P,N} => false
3660                                 // {P,P} & superc!=subc => false
3661     _prim_same_path,            // {P,P} & superc==subc => true
3662     _prim_1_path,               // {N,P} => false
3663     _ref_subtype_path,          // {N,N} & subtype check wins => true
3664     _both_ref_path,             // {N,N} & subtype check loses => false
3665     PATH_LIMIT
3666   };
3667 
3668   RegionNode* region = new RegionNode(PATH_LIMIT);
3669   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
3670   record_for_igvn(region);
3671 
3672   const TypePtr* adr_type = TypeRawPtr::BOTTOM;   // memory type of loads
3673   const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3674   int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3675 
3676   // First null-check both mirrors and load each mirror's klass metaobject.
3677   int which_arg;
3678   for (which_arg = 0; which_arg <= 1; which_arg++) {
3679     Node* arg = args[which_arg];
3680     arg = null_check(arg);
3681     if (stopped())  break;
3682     args[which_arg] = arg;
3683 
3684     Node* p = basic_plus_adr(arg, class_klass_offset);
3685     Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3686     klasses[which_arg] = _gvn.transform(kls);
3687   }
3688 
3689   // Having loaded both klasses, test each for null.
3690   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3691   for (which_arg = 0; which_arg <= 1; which_arg++) {
3692     Node* kls = klasses[which_arg];
3693     Node* null_ctl = top();
3694     kls = null_check_oop(kls, &null_ctl, never_see_null);
3695     int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3696     region->init_req(prim_path, null_ctl);
3697     if (stopped())  break;
3698     klasses[which_arg] = kls;
3699   }
3700 
3701   if (!stopped()) {
3702     // now we have two reference types, in klasses[0..1]
3703     Node* subk   = klasses[1];  // the argument to isAssignableFrom
3704     Node* superk = klasses[0];  // the receiver
3705     region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3706     // now we have a successful reference subtype check
3707     region->set_req(_ref_subtype_path, control());
3708   }
3709 
3710   // If both operands are primitive (both klasses null), then
3711   // we must return true when they are identical primitives.
3712   // It is convenient to test this after the first null klass check.
3713   set_control(region->in(_prim_0_path)); // go back to first null check
3714   if (!stopped()) {
3715     // Since superc is primitive, make a guard for the superc==subc case.
3716     Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3717     Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3718     generate_guard(bol_eq, region, PROB_FAIR);
3719     if (region->req() == PATH_LIMIT+1) {
3720       // A guard was added.  If the added guard is taken, superc==subc.
3721       region->swap_edges(PATH_LIMIT, _prim_same_path);
3722       region->del_req(PATH_LIMIT);
3723     }
3724     region->set_req(_prim_0_path, control()); // Not equal after all.
3725   }
3726 
3727   // these are the only paths that produce 'true':
3728   phi->set_req(_prim_same_path,   intcon(1));
3729   phi->set_req(_ref_subtype_path, intcon(1));
3730 
3731   // pull together the cases:
3732   assert(region->req() == PATH_LIMIT, "sane region");
3733   for (uint i = 1; i < region->req(); i++) {
3734     Node* ctl = region->in(i);
3735     if (ctl == NULL || ctl == top()) {
3736       region->set_req(i, top());
3737       phi   ->set_req(i, top());
3738     } else if (phi->in(i) == NULL) {
3739       phi->set_req(i, intcon(0)); // all other paths produce 'false'
3740     }
3741   }
3742 
3743   set_control(_gvn.transform(region));
3744   set_result(_gvn.transform(phi));
3745   return true;
3746 }
3747 
3748 //---------------------generate_array_guard_common------------------------
3749 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3750                                                   bool obj_array, bool not_array) {
3751 
3752   if (stopped()) {
3753     return NULL;
3754   }
3755 
3756   // If obj_array/non_array==false/false:
3757   // Branch around if the given klass is in fact an array (either obj or prim).
3758   // If obj_array/non_array==false/true:
3759   // Branch around if the given klass is not an array klass of any kind.
3760   // If obj_array/non_array==true/true:
3761   // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3762   // If obj_array/non_array==true/false:
3763   // Branch around if the kls is an oop array (Object[] or subtype)
3764   //
3765   // Like generate_guard, adds a new path onto the region.
3766   jint  layout_con = 0;
3767   Node* layout_val = get_layout_helper(kls, layout_con);
3768   if (layout_val == NULL) {
3769     bool query = (obj_array
3770                   ? Klass::layout_helper_is_objArray(layout_con)
3771                   : Klass::layout_helper_is_array(layout_con));
3772     if (query == not_array) {
3773       return NULL;                       // never a branch
3774     } else {                             // always a branch
3775       Node* always_branch = control();
3776       if (region != NULL)
3777         region->add_req(always_branch);
3778       set_control(top());
3779       return always_branch;
3780     }
3781   }
3782   // Now test the correct condition.
3783   jint  nval = (obj_array
3784                 ? ((jint)Klass::_lh_array_tag_type_value
3785                    <<    Klass::_lh_array_tag_shift)
3786                 : Klass::_lh_neutral_value);
3787   Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3788   BoolTest::mask btest = BoolTest::lt;  // correct for testing is_[obj]array
3789   // invert the test if we are looking for a non-array
3790   if (not_array)  btest = BoolTest(btest).negate();
3791   Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3792   return generate_fair_guard(bol, region);
3793 }
3794 
3795 
3796 //-----------------------inline_native_newArray--------------------------
3797 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3798 bool LibraryCallKit::inline_native_newArray() {
3799   Node* mirror    = argument(0);
3800   Node* count_val = argument(1);
3801 
3802   mirror = null_check(mirror);
3803   // If mirror or obj is dead, only null-path is taken.
3804   if (stopped())  return true;
3805 
3806   enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3807   RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3808   PhiNode*    result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3809   PhiNode*    result_io  = new PhiNode(result_reg, Type::ABIO);
3810   PhiNode*    result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3811 
3812   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3813   Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3814                                                   result_reg, _slow_path);
3815   Node* normal_ctl   = control();
3816   Node* no_array_ctl = result_reg->in(_slow_path);
3817 
3818   // Generate code for the slow case.  We make a call to newArray().
3819   set_control(no_array_ctl);
3820   if (!stopped()) {
3821     // Either the input type is void.class, or else the
3822     // array klass has not yet been cached.  Either the
3823     // ensuing call will throw an exception, or else it
3824     // will cache the array klass for next time.
3825     PreserveJVMState pjvms(this);
3826     CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3827     Node* slow_result = set_results_for_java_call(slow_call);
3828     // this->control() comes from set_results_for_java_call
3829     result_reg->set_req(_slow_path, control());
3830     result_val->set_req(_slow_path, slow_result);
3831     result_io ->set_req(_slow_path, i_o());
3832     result_mem->set_req(_slow_path, reset_memory());
3833   }
3834 
3835   set_control(normal_ctl);
3836   if (!stopped()) {
3837     // Normal case:  The array type has been cached in the java.lang.Class.
3838     // The following call works fine even if the array type is polymorphic.
3839     // It could be a dynamic mix of int[], boolean[], Object[], etc.
3840     Node* obj = new_array(klass_node, count_val, 0);  // no arguments to push
3841     result_reg->init_req(_normal_path, control());
3842     result_val->init_req(_normal_path, obj);
3843     result_io ->init_req(_normal_path, i_o());
3844     result_mem->init_req(_normal_path, reset_memory());
3845   }
3846 
3847   // Return the combined state.
3848   set_i_o(        _gvn.transform(result_io)  );
3849   set_all_memory( _gvn.transform(result_mem));
3850 
3851   C->set_has_split_ifs(true); // Has chance for split-if optimization
3852   set_result(result_reg, result_val);
3853   return true;
3854 }
3855 
3856 //----------------------inline_native_getLength--------------------------
3857 // public static native int java.lang.reflect.Array.getLength(Object array);
3858 bool LibraryCallKit::inline_native_getLength() {
3859   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3860 
3861   Node* array = null_check(argument(0));
3862   // If array is dead, only null-path is taken.
3863   if (stopped())  return true;
3864 
3865   // Deoptimize if it is a non-array.
3866   Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3867 
3868   if (non_array != NULL) {
3869     PreserveJVMState pjvms(this);
3870     set_control(non_array);
3871     uncommon_trap(Deoptimization::Reason_intrinsic,
3872                   Deoptimization::Action_maybe_recompile);
3873   }
3874 
3875   // If control is dead, only non-array-path is taken.
3876   if (stopped())  return true;
3877 
3878   // The works fine even if the array type is polymorphic.
3879   // It could be a dynamic mix of int[], boolean[], Object[], etc.
3880   Node* result = load_array_length(array);
3881 
3882   C->set_has_split_ifs(true);  // Has chance for split-if optimization
3883   set_result(result);
3884   return true;
3885 }
3886 
3887 //------------------------inline_array_copyOf----------------------------
3888 // public static <T,U> T[] java.util.Arrays.copyOf(     U[] original, int newLength,         Class<? extends T[]> newType);
3889 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from,      int to, Class<? extends T[]> newType);
3890 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3891   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3892 
3893   // Get the arguments.
3894   Node* original          = argument(0);
3895   Node* start             = is_copyOfRange? argument(1): intcon(0);
3896   Node* end               = is_copyOfRange? argument(2): argument(1);
3897   Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3898 
3899   Node* newcopy;
3900 
3901   // Set the original stack and the reexecute bit for the interpreter to reexecute
3902   // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3903   { PreserveReexecuteState preexecs(this);
3904     jvms()->set_should_reexecute(true);
3905 
3906     array_type_mirror = null_check(array_type_mirror);
3907     original          = null_check(original);
3908 
3909     // Check if a null path was taken unconditionally.
3910     if (stopped())  return true;
3911 
3912     Node* orig_length = load_array_length(original);
3913 
3914     Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3915     klass_node = null_check(klass_node);
3916 
3917     RegionNode* bailout = new RegionNode(1);
3918     record_for_igvn(bailout);
3919 
3920     // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3921     // Bail out if that is so.
3922     Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3923     if (not_objArray != NULL) {
3924       // Improve the klass node's type from the new optimistic assumption:
3925       ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3926       const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3927       Node* cast = new CastPPNode(klass_node, akls);
3928       cast->init_req(0, control());
3929       klass_node = _gvn.transform(cast);
3930     }
3931 
3932     // Bail out if either start or end is negative.
3933     generate_negative_guard(start, bailout, &start);
3934     generate_negative_guard(end,   bailout, &end);
3935 
3936     Node* length = end;
3937     if (_gvn.type(start) != TypeInt::ZERO) {
3938       length = _gvn.transform(new SubINode(end, start));
3939     }
3940 
3941     // Bail out if length is negative.
3942     // Without this the new_array would throw
3943     // NegativeArraySizeException but IllegalArgumentException is what
3944     // should be thrown
3945     generate_negative_guard(length, bailout, &length);
3946 
3947     if (bailout->req() > 1) {
3948       PreserveJVMState pjvms(this);
3949       set_control(_gvn.transform(bailout));
3950       uncommon_trap(Deoptimization::Reason_intrinsic,
3951                     Deoptimization::Action_maybe_recompile);
3952     }
3953 
3954     if (!stopped()) {
3955       // How many elements will we copy from the original?
3956       // The answer is MinI(orig_length - start, length).
3957       Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3958       Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3959 
3960       // Generate a direct call to the right arraycopy function(s).
3961       // We know the copy is disjoint but we might not know if the
3962       // oop stores need checking.
3963       // Extreme case:  Arrays.copyOf((Integer[])x, 10, String[].class).
3964       // This will fail a store-check if x contains any non-nulls.
3965 
3966       // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3967       // loads/stores but it is legal only if we're sure the
3968       // Arrays.copyOf would succeed. So we need all input arguments
3969       // to the copyOf to be validated, including that the copy to the
3970       // new array won't trigger an ArrayStoreException. That subtype
3971       // check can be optimized if we know something on the type of
3972       // the input array from type speculation.
3973       if (_gvn.type(klass_node)->singleton()) {
3974         ciKlass* subk   = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3975         ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3976 
3977         int test = C->static_subtype_check(superk, subk);
3978         if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3979           const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3980           if (t_original->speculative_type() != NULL) {
3981             original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3982           }
3983         }
3984       }
3985 
3986       bool validated = false;
3987       // Reason_class_check rather than Reason_intrinsic because we
3988       // want to intrinsify even if this traps.
3989       if (!too_many_traps(Deoptimization::Reason_class_check)) {
3990         Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3991                                                    klass_node);
3992 
3993         if (not_subtype_ctrl != top()) {
3994           PreserveJVMState pjvms(this);
3995           set_control(not_subtype_ctrl);
3996           uncommon_trap(Deoptimization::Reason_class_check,
3997                         Deoptimization::Action_make_not_entrant);
3998           assert(stopped(), "Should be stopped");
3999         }
4000         validated = true;
4001       }
4002 
4003       if (!stopped()) {
4004         newcopy = new_array(klass_node, length, 0);  // no arguments to push
4005 
4006         ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
4007                                                 load_object_klass(original), klass_node);
4008         if (!is_copyOfRange) {
4009           ac->set_copyof(validated);
4010         } else {
4011           ac->set_copyofrange(validated);
4012         }
4013         Node* n = _gvn.transform(ac);
4014         if (n == ac) {
4015           ac->connect_outputs(this);
4016         } else {
4017           assert(validated, "shouldn't transform if all arguments not validated");
4018           set_all_memory(n);
4019         }
4020       }
4021     }
4022   } // original reexecute is set back here
4023 
4024   C->set_has_split_ifs(true); // Has chance for split-if optimization
4025   if (!stopped()) {
4026     set_result(newcopy);
4027   }
4028   return true;
4029 }
4030 
4031 
4032 //----------------------generate_virtual_guard---------------------------
4033 // Helper for hashCode and clone.  Peeks inside the vtable to avoid a call.
4034 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4035                                              RegionNode* slow_region) {
4036   ciMethod* method = callee();
4037   int vtable_index = method->vtable_index();
4038   assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4039          err_msg_res("bad index %d", vtable_index));
4040   // Get the Method* out of the appropriate vtable entry.
4041   int entry_offset  = (InstanceKlass::vtable_start_offset() +
4042                      vtable_index*vtableEntry::size()) * wordSize +
4043                      vtableEntry::method_offset_in_bytes();
4044   Node* entry_addr  = basic_plus_adr(obj_klass, entry_offset);
4045   Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4046 
4047   // Compare the target method with the expected method (e.g., Object.hashCode).
4048   const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4049 
4050   Node* native_call = makecon(native_call_addr);
4051   Node* chk_native  = _gvn.transform(new CmpPNode(target_call, native_call));
4052   Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4053 
4054   return generate_slow_guard(test_native, slow_region);
4055 }
4056 
4057 //-----------------------generate_method_call----------------------------
4058 // Use generate_method_call to make a slow-call to the real
4059 // method if the fast path fails.  An alternative would be to
4060 // use a stub like OptoRuntime::slow_arraycopy_Java.
4061 // This only works for expanding the current library call,
4062 // not another intrinsic.  (E.g., don't use this for making an
4063 // arraycopy call inside of the copyOf intrinsic.)
4064 CallJavaNode*
4065 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4066   // When compiling the intrinsic method itself, do not use this technique.
4067   guarantee(callee() != C->method(), "cannot make slow-call to self");
4068 
4069   ciMethod* method = callee();
4070   // ensure the JVMS we have will be correct for this call
4071   guarantee(method_id == method->intrinsic_id(), "must match");
4072 
4073   const TypeFunc* tf = TypeFunc::make(method);
4074   CallJavaNode* slow_call;
4075   if (is_static) {
4076     assert(!is_virtual, "");
4077     slow_call = new CallStaticJavaNode(C, tf,
4078                            SharedRuntime::get_resolve_static_call_stub(),
4079                            method, bci());
4080   } else if (is_virtual) {
4081     null_check_receiver();
4082     int vtable_index = Method::invalid_vtable_index;
4083     if (UseInlineCaches) {
4084       // Suppress the vtable call
4085     } else {
4086       // hashCode and clone are not a miranda methods,
4087       // so the vtable index is fixed.
4088       // No need to use the linkResolver to get it.
4089        vtable_index = method->vtable_index();
4090        assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4091               err_msg_res("bad index %d", vtable_index));
4092     }
4093     slow_call = new CallDynamicJavaNode(tf,
4094                           SharedRuntime::get_resolve_virtual_call_stub(),
4095                           method, vtable_index, bci());
4096   } else {  // neither virtual nor static:  opt_virtual
4097     null_check_receiver();
4098     slow_call = new CallStaticJavaNode(C, tf,
4099                                 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4100                                 method, bci());
4101     slow_call->set_optimized_virtual(true);
4102   }
4103   set_arguments_for_java_call(slow_call);
4104   set_edges_for_java_call(slow_call);
4105   return slow_call;
4106 }
4107 
4108 
4109 /**
4110  * Build special case code for calls to hashCode on an object. This call may
4111  * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4112  * slightly different code.
4113  */
4114 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4115   assert(is_static == callee()->is_static(), "correct intrinsic selection");
4116   assert(!(is_virtual && is_static), "either virtual, special, or static");
4117 
4118   enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4119 
4120   RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4121   PhiNode*    result_val = new PhiNode(result_reg, TypeInt::INT);
4122   PhiNode*    result_io  = new PhiNode(result_reg, Type::ABIO);
4123   PhiNode*    result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4124   Node* obj = NULL;
4125   if (!is_static) {
4126     // Check for hashing null object
4127     obj = null_check_receiver();
4128     if (stopped())  return true;        // unconditionally null
4129     result_reg->init_req(_null_path, top());
4130     result_val->init_req(_null_path, top());
4131   } else {
4132     // Do a null check, and return zero if null.
4133     // System.identityHashCode(null) == 0
4134     obj = argument(0);
4135     Node* null_ctl = top();
4136     obj = null_check_oop(obj, &null_ctl);
4137     result_reg->init_req(_null_path, null_ctl);
4138     result_val->init_req(_null_path, _gvn.intcon(0));
4139   }
4140 
4141   // Unconditionally null?  Then return right away.
4142   if (stopped()) {
4143     set_control( result_reg->in(_null_path));
4144     if (!stopped())
4145       set_result(result_val->in(_null_path));
4146     return true;
4147   }
4148 
4149   // We only go to the fast case code if we pass a number of guards.  The
4150   // paths which do not pass are accumulated in the slow_region.
4151   RegionNode* slow_region = new RegionNode(1);
4152   record_for_igvn(slow_region);
4153 
4154   // If this is a virtual call, we generate a funny guard.  We pull out
4155   // the vtable entry corresponding to hashCode() from the target object.
4156   // If the target method which we are calling happens to be the native
4157   // Object hashCode() method, we pass the guard.  We do not need this
4158   // guard for non-virtual calls -- the caller is known to be the native
4159   // Object hashCode().
4160   if (is_virtual) {
4161     // After null check, get the object's klass.
4162     Node* obj_klass = load_object_klass(obj);
4163     generate_virtual_guard(obj_klass, slow_region);
4164   }
4165 
4166   // Get the header out of the object, use LoadMarkNode when available
4167   Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4168   // The control of the load must be NULL. Otherwise, the load can move before
4169   // the null check after castPP removal.
4170   Node* no_ctrl = NULL;
4171   Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4172 
4173   // Test the header to see if it is unlocked.
4174   Node *lock_mask      = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4175   Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4176   Node *unlocked_val   = _gvn.MakeConX(markOopDesc::unlocked_value);
4177   Node *chk_unlocked   = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4178   Node *test_unlocked  = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4179 
4180   generate_slow_guard(test_unlocked, slow_region);
4181 
4182   // Get the hash value and check to see that it has been properly assigned.
4183   // We depend on hash_mask being at most 32 bits and avoid the use of
4184   // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4185   // vm: see markOop.hpp.
4186   Node *hash_mask      = _gvn.intcon(markOopDesc::hash_mask);
4187   Node *hash_shift     = _gvn.intcon(markOopDesc::hash_shift);
4188   Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4189   // This hack lets the hash bits live anywhere in the mark object now, as long
4190   // as the shift drops the relevant bits into the low 32 bits.  Note that
4191   // Java spec says that HashCode is an int so there's no point in capturing
4192   // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4193   hshifted_header      = ConvX2I(hshifted_header);
4194   Node *hash_val       = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4195 
4196   Node *no_hash_val    = _gvn.intcon(markOopDesc::no_hash);
4197   Node *chk_assigned   = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4198   Node *test_assigned  = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4199 
4200   generate_slow_guard(test_assigned, slow_region);
4201 
4202   Node* init_mem = reset_memory();
4203   // fill in the rest of the null path:
4204   result_io ->init_req(_null_path, i_o());
4205   result_mem->init_req(_null_path, init_mem);
4206 
4207   result_val->init_req(_fast_path, hash_val);
4208   result_reg->init_req(_fast_path, control());
4209   result_io ->init_req(_fast_path, i_o());
4210   result_mem->init_req(_fast_path, init_mem);
4211 
4212   // Generate code for the slow case.  We make a call to hashCode().
4213   set_control(_gvn.transform(slow_region));
4214   if (!stopped()) {
4215     // No need for PreserveJVMState, because we're using up the present state.
4216     set_all_memory(init_mem);
4217     vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4218     CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4219     Node* slow_result = set_results_for_java_call(slow_call);
4220     // this->control() comes from set_results_for_java_call
4221     result_reg->init_req(_slow_path, control());
4222     result_val->init_req(_slow_path, slow_result);
4223     result_io  ->set_req(_slow_path, i_o());
4224     result_mem ->set_req(_slow_path, reset_memory());
4225   }
4226 
4227   // Return the combined state.
4228   set_i_o(        _gvn.transform(result_io)  );
4229   set_all_memory( _gvn.transform(result_mem));
4230 
4231   set_result(result_reg, result_val);
4232   return true;
4233 }
4234 
4235 //---------------------------inline_native_getClass----------------------------
4236 // public final native Class<?> java.lang.Object.getClass();
4237 //
4238 // Build special case code for calls to getClass on an object.
4239 bool LibraryCallKit::inline_native_getClass() {
4240   Node* obj = null_check_receiver();
4241   if (stopped())  return true;
4242   set_result(load_mirror_from_klass(load_object_klass(obj)));
4243   return true;
4244 }
4245 
4246 //-----------------inline_native_Reflection_getCallerClass---------------------
4247 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4248 //
4249 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4250 //
4251 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4252 // in that it must skip particular security frames and checks for
4253 // caller sensitive methods.
4254 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4255 #ifndef PRODUCT
4256   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4257     tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4258   }
4259 #endif
4260 
4261   if (!jvms()->has_method()) {
4262 #ifndef PRODUCT
4263     if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4264       tty->print_cr("  Bailing out because intrinsic was inlined at top level");
4265     }
4266 #endif
4267     return false;
4268   }
4269 
4270   // Walk back up the JVM state to find the caller at the required
4271   // depth.
4272   JVMState* caller_jvms = jvms();
4273 
4274   // Cf. JVM_GetCallerClass
4275   // NOTE: Start the loop at depth 1 because the current JVM state does
4276   // not include the Reflection.getCallerClass() frame.
4277   for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4278     ciMethod* m = caller_jvms->method();
4279     switch (n) {
4280     case 0:
4281       fatal("current JVM state does not include the Reflection.getCallerClass frame");
4282       break;
4283     case 1:
4284       // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4285       if (!m->caller_sensitive()) {
4286 #ifndef PRODUCT
4287         if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4288           tty->print_cr("  Bailing out: CallerSensitive annotation expected at frame %d", n);
4289         }
4290 #endif
4291         return false;  // bail-out; let JVM_GetCallerClass do the work
4292       }
4293       break;
4294     default:
4295       if (!m->is_ignored_by_security_stack_walk()) {
4296         // We have reached the desired frame; return the holder class.
4297         // Acquire method holder as java.lang.Class and push as constant.
4298         ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4299         ciInstance* caller_mirror = caller_klass->java_mirror();
4300         set_result(makecon(TypeInstPtr::make(caller_mirror)));
4301 
4302 #ifndef PRODUCT
4303         if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4304           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());
4305           tty->print_cr("  JVM state at this point:");
4306           for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4307             ciMethod* m = jvms()->of_depth(i)->method();
4308             tty->print_cr("   %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4309           }
4310         }
4311 #endif
4312         return true;
4313       }
4314       break;
4315     }
4316   }
4317 
4318 #ifndef PRODUCT
4319   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4320     tty->print_cr("  Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4321     tty->print_cr("  JVM state at this point:");
4322     for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4323       ciMethod* m = jvms()->of_depth(i)->method();
4324       tty->print_cr("   %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4325     }
4326   }
4327 #endif
4328 
4329   return false;  // bail-out; let JVM_GetCallerClass do the work
4330 }
4331 
4332 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4333   Node* arg = argument(0);
4334   Node* result;
4335 
4336   switch (id) {
4337   case vmIntrinsics::_floatToRawIntBits:    result = new MoveF2INode(arg);  break;
4338   case vmIntrinsics::_intBitsToFloat:       result = new MoveI2FNode(arg);  break;
4339   case vmIntrinsics::_doubleToRawLongBits:  result = new MoveD2LNode(arg);  break;
4340   case vmIntrinsics::_longBitsToDouble:     result = new MoveL2DNode(arg);  break;
4341 
4342   case vmIntrinsics::_doubleToLongBits: {
4343     // two paths (plus control) merge in a wood
4344     RegionNode *r = new RegionNode(3);
4345     Node *phi = new PhiNode(r, TypeLong::LONG);
4346 
4347     Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4348     // Build the boolean node
4349     Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4350 
4351     // Branch either way.
4352     // NaN case is less traveled, which makes all the difference.
4353     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4354     Node *opt_isnan = _gvn.transform(ifisnan);
4355     assert( opt_isnan->is_If(), "Expect an IfNode");
4356     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4357     Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4358 
4359     set_control(iftrue);
4360 
4361     static const jlong nan_bits = CONST64(0x7ff8000000000000);
4362     Node *slow_result = longcon(nan_bits); // return NaN
4363     phi->init_req(1, _gvn.transform( slow_result ));
4364     r->init_req(1, iftrue);
4365 
4366     // Else fall through
4367     Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4368     set_control(iffalse);
4369 
4370     phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4371     r->init_req(2, iffalse);
4372 
4373     // Post merge
4374     set_control(_gvn.transform(r));
4375     record_for_igvn(r);
4376 
4377     C->set_has_split_ifs(true); // Has chance for split-if optimization
4378     result = phi;
4379     assert(result->bottom_type()->isa_long(), "must be");
4380     break;
4381   }
4382 
4383   case vmIntrinsics::_floatToIntBits: {
4384     // two paths (plus control) merge in a wood
4385     RegionNode *r = new RegionNode(3);
4386     Node *phi = new PhiNode(r, TypeInt::INT);
4387 
4388     Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4389     // Build the boolean node
4390     Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4391 
4392     // Branch either way.
4393     // NaN case is less traveled, which makes all the difference.
4394     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4395     Node *opt_isnan = _gvn.transform(ifisnan);
4396     assert( opt_isnan->is_If(), "Expect an IfNode");
4397     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4398     Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4399 
4400     set_control(iftrue);
4401 
4402     static const jint nan_bits = 0x7fc00000;
4403     Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4404     phi->init_req(1, _gvn.transform( slow_result ));
4405     r->init_req(1, iftrue);
4406 
4407     // Else fall through
4408     Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4409     set_control(iffalse);
4410 
4411     phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4412     r->init_req(2, iffalse);
4413 
4414     // Post merge
4415     set_control(_gvn.transform(r));
4416     record_for_igvn(r);
4417 
4418     C->set_has_split_ifs(true); // Has chance for split-if optimization
4419     result = phi;
4420     assert(result->bottom_type()->isa_int(), "must be");
4421     break;
4422   }
4423 
4424   default:
4425     fatal_unexpected_iid(id);
4426     break;
4427   }
4428   set_result(_gvn.transform(result));
4429   return true;
4430 }
4431 
4432 #ifdef _LP64
4433 #define XTOP ,top() /*additional argument*/
4434 #else  //_LP64
4435 #define XTOP        /*no additional argument*/
4436 #endif //_LP64
4437 
4438 //----------------------inline_unsafe_copyMemory-------------------------
4439 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4440 bool LibraryCallKit::inline_unsafe_copyMemory() {
4441   if (callee()->is_static())  return false;  // caller must have the capability!
4442   null_check_receiver();  // null-check receiver
4443   if (stopped())  return true;
4444 
4445   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
4446 
4447   Node* src_ptr =         argument(1);   // type: oop
4448   Node* src_off = ConvL2X(argument(2));  // type: long
4449   Node* dst_ptr =         argument(4);   // type: oop
4450   Node* dst_off = ConvL2X(argument(5));  // type: long
4451   Node* size    = ConvL2X(argument(7));  // type: long
4452 
4453   Node* src = make_unsafe_address(src_ptr, src_off, /*decode=*/true);
4454   Node* dst = make_unsafe_address(dst_ptr, dst_off, /*decode=*/true);



4455 
4456   // Conservatively insert a memory barrier on all memory slices.
4457   // Do not let writes of the copy source or destination float below the copy.
4458   insert_mem_bar(Op_MemBarCPUOrder);
4459 
4460   // Call it.  Note that the length argument is not scaled.
4461   make_runtime_call(RC_LEAF|RC_NO_FP,
4462                     OptoRuntime::fast_arraycopy_Type(),
4463                     StubRoutines::unsafe_arraycopy(),
4464                     "unsafe_arraycopy",
4465                     TypeRawPtr::BOTTOM,
4466                     src, dst, size XTOP);
4467 
4468   // Do not let reads of the copy destination float above the copy.
4469   insert_mem_bar(Op_MemBarCPUOrder);
4470 
4471   return true;
4472 }
4473 
4474 //------------------------clone_coping-----------------------------------
4475 // Helper function for inline_native_clone.
4476 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4477   assert(obj_size != NULL, "");
4478   Node* raw_obj = alloc_obj->in(1);
4479   assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4480 
4481   AllocateNode* alloc = NULL;
4482   if (ReduceBulkZeroing) {
4483     // We will be completely responsible for initializing this object -
4484     // mark Initialize node as complete.
4485     alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4486     // The object was just allocated - there should be no any stores!
4487     guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4488     // Mark as complete_with_arraycopy so that on AllocateNode
4489     // expansion, we know this AllocateNode is initialized by an array
4490     // copy and a StoreStore barrier exists after the array copy.
4491     alloc->initialization()->set_complete_with_arraycopy();
4492   }
4493 
4494   // Copy the fastest available way.
4495   // TODO: generate fields copies for small objects instead.
4496   Node* src  = obj;
4497   Node* dest = alloc_obj;
4498   Node* size = _gvn.transform(obj_size);
4499 
4500   // Exclude the header but include array length to copy by 8 bytes words.
4501   // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4502   int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4503                             instanceOopDesc::base_offset_in_bytes();
4504   // base_off:
4505   // 8  - 32-bit VM
4506   // 12 - 64-bit VM, compressed klass
4507   // 16 - 64-bit VM, normal klass
4508   if (base_off % BytesPerLong != 0) {
4509     assert(UseCompressedClassPointers, "");
4510     if (is_array) {
4511       // Exclude length to copy by 8 bytes words.
4512       base_off += sizeof(int);
4513     } else {
4514       // Include klass to copy by 8 bytes words.
4515       base_off = instanceOopDesc::klass_offset_in_bytes();
4516     }
4517     assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4518   }
4519   src  = basic_plus_adr(src,  base_off);
4520   dest = basic_plus_adr(dest, base_off);
4521 
4522   // Compute the length also, if needed:
4523   Node* countx = size;
4524   countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4525   countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4526 
4527   const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4528 
4529   ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4530   ac->set_clonebasic();
4531   Node* n = _gvn.transform(ac);
4532   if (n == ac) {
4533     set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4534   } else {
4535     set_all_memory(n);
4536   }
4537 
4538   // If necessary, emit some card marks afterwards.  (Non-arrays only.)
4539   if (card_mark) {
4540     assert(!is_array, "");
4541     // Put in store barrier for any and all oops we are sticking
4542     // into this object.  (We could avoid this if we could prove
4543     // that the object type contains no oop fields at all.)
4544     Node* no_particular_value = NULL;
4545     Node* no_particular_field = NULL;
4546     int raw_adr_idx = Compile::AliasIdxRaw;
4547     post_barrier(control(),
4548                  memory(raw_adr_type),
4549                  alloc_obj,
4550                  no_particular_field,
4551                  raw_adr_idx,
4552                  no_particular_value,
4553                  T_OBJECT,
4554                  false);
4555   }
4556 
4557   // Do not let reads from the cloned object float above the arraycopy.
4558   if (alloc != NULL) {
4559     // Do not let stores that initialize this object be reordered with
4560     // a subsequent store that would make this object accessible by
4561     // other threads.
4562     // Record what AllocateNode this StoreStore protects so that
4563     // escape analysis can go from the MemBarStoreStoreNode to the
4564     // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4565     // based on the escape status of the AllocateNode.
4566     insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4567   } else {
4568     insert_mem_bar(Op_MemBarCPUOrder);
4569   }
4570 }
4571 
4572 //------------------------inline_native_clone----------------------------
4573 // protected native Object java.lang.Object.clone();
4574 //
4575 // Here are the simple edge cases:
4576 //  null receiver => normal trap
4577 //  virtual and clone was overridden => slow path to out-of-line clone
4578 //  not cloneable or finalizer => slow path to out-of-line Object.clone
4579 //
4580 // The general case has two steps, allocation and copying.
4581 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4582 //
4583 // Copying also has two cases, oop arrays and everything else.
4584 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4585 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4586 //
4587 // These steps fold up nicely if and when the cloned object's klass
4588 // can be sharply typed as an object array, a type array, or an instance.
4589 //
4590 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4591   PhiNode* result_val;
4592 
4593   // Set the reexecute bit for the interpreter to reexecute
4594   // the bytecode that invokes Object.clone if deoptimization happens.
4595   { PreserveReexecuteState preexecs(this);
4596     jvms()->set_should_reexecute(true);
4597 
4598     Node* obj = null_check_receiver();
4599     if (stopped())  return true;
4600 
4601     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4602 
4603     // If we are going to clone an instance, we need its exact type to
4604     // know the number and types of fields to convert the clone to
4605     // loads/stores. Maybe a speculative type can help us.
4606     if (!obj_type->klass_is_exact() &&
4607         obj_type->speculative_type() != NULL &&
4608         obj_type->speculative_type()->is_instance_klass()) {
4609       ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4610       if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4611           !spec_ik->has_injected_fields()) {
4612         ciKlass* k = obj_type->klass();
4613         if (!k->is_instance_klass() ||
4614             k->as_instance_klass()->is_interface() ||
4615             k->as_instance_klass()->has_subklass()) {
4616           obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4617         }
4618       }
4619     }
4620 
4621     Node* obj_klass = load_object_klass(obj);
4622     const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4623     const TypeOopPtr*   toop   = ((tklass != NULL)
4624                                 ? tklass->as_instance_type()
4625                                 : TypeInstPtr::NOTNULL);
4626 
4627     // Conservatively insert a memory barrier on all memory slices.
4628     // Do not let writes into the original float below the clone.
4629     insert_mem_bar(Op_MemBarCPUOrder);
4630 
4631     // paths into result_reg:
4632     enum {
4633       _slow_path = 1,     // out-of-line call to clone method (virtual or not)
4634       _objArray_path,     // plain array allocation, plus arrayof_oop_arraycopy
4635       _array_path,        // plain array allocation, plus arrayof_long_arraycopy
4636       _instance_path,     // plain instance allocation, plus arrayof_long_arraycopy
4637       PATH_LIMIT
4638     };
4639     RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4640     result_val             = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4641     PhiNode*    result_i_o = new PhiNode(result_reg, Type::ABIO);
4642     PhiNode*    result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4643     record_for_igvn(result_reg);
4644 
4645     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4646     int raw_adr_idx = Compile::AliasIdxRaw;
4647 
4648     Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4649     if (array_ctl != NULL) {
4650       // It's an array.
4651       PreserveJVMState pjvms(this);
4652       set_control(array_ctl);
4653       Node* obj_length = load_array_length(obj);
4654       Node* obj_size  = NULL;
4655       Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);  // no arguments to push
4656 
4657       if (!use_ReduceInitialCardMarks()) {
4658         // If it is an oop array, it requires very special treatment,
4659         // because card marking is required on each card of the array.
4660         Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4661         if (is_obja != NULL) {
4662           PreserveJVMState pjvms2(this);
4663           set_control(is_obja);
4664           // Generate a direct call to the right arraycopy function(s).
4665           Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4666           ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4667           ac->set_cloneoop();
4668           Node* n = _gvn.transform(ac);
4669           assert(n == ac, "cannot disappear");
4670           ac->connect_outputs(this);
4671 
4672           result_reg->init_req(_objArray_path, control());
4673           result_val->init_req(_objArray_path, alloc_obj);
4674           result_i_o ->set_req(_objArray_path, i_o());
4675           result_mem ->set_req(_objArray_path, reset_memory());
4676         }
4677       }
4678       // Otherwise, there are no card marks to worry about.
4679       // (We can dispense with card marks if we know the allocation
4680       //  comes out of eden (TLAB)...  In fact, ReduceInitialCardMarks
4681       //  causes the non-eden paths to take compensating steps to
4682       //  simulate a fresh allocation, so that no further
4683       //  card marks are required in compiled code to initialize
4684       //  the object.)
4685 
4686       if (!stopped()) {
4687         copy_to_clone(obj, alloc_obj, obj_size, true, false);
4688 
4689         // Present the results of the copy.
4690         result_reg->init_req(_array_path, control());
4691         result_val->init_req(_array_path, alloc_obj);
4692         result_i_o ->set_req(_array_path, i_o());
4693         result_mem ->set_req(_array_path, reset_memory());
4694       }
4695     }
4696 
4697     // We only go to the instance fast case code if we pass a number of guards.
4698     // The paths which do not pass are accumulated in the slow_region.
4699     RegionNode* slow_region = new RegionNode(1);
4700     record_for_igvn(slow_region);
4701     if (!stopped()) {
4702       // It's an instance (we did array above).  Make the slow-path tests.
4703       // If this is a virtual call, we generate a funny guard.  We grab
4704       // the vtable entry corresponding to clone() from the target object.
4705       // If the target method which we are calling happens to be the
4706       // Object clone() method, we pass the guard.  We do not need this
4707       // guard for non-virtual calls; the caller is known to be the native
4708       // Object clone().
4709       if (is_virtual) {
4710         generate_virtual_guard(obj_klass, slow_region);
4711       }
4712 
4713       // The object must be cloneable and must not have a finalizer.
4714       // Both of these conditions may be checked in a single test.
4715       // We could optimize the cloneable test further, but we don't care.
4716       generate_access_flags_guard(obj_klass,
4717                                   // Test both conditions:
4718                                   JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4719                                   // Must be cloneable but not finalizer:
4720                                   JVM_ACC_IS_CLONEABLE,
4721                                   slow_region);
4722     }
4723 
4724     if (!stopped()) {
4725       // It's an instance, and it passed the slow-path tests.
4726       PreserveJVMState pjvms(this);
4727       Node* obj_size  = NULL;
4728       // Need to deoptimize on exception from allocation since Object.clone intrinsic
4729       // is reexecuted if deoptimization occurs and there could be problems when merging
4730       // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4731       Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4732 
4733       copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4734 
4735       // Present the results of the slow call.
4736       result_reg->init_req(_instance_path, control());
4737       result_val->init_req(_instance_path, alloc_obj);
4738       result_i_o ->set_req(_instance_path, i_o());
4739       result_mem ->set_req(_instance_path, reset_memory());
4740     }
4741 
4742     // Generate code for the slow case.  We make a call to clone().
4743     set_control(_gvn.transform(slow_region));
4744     if (!stopped()) {
4745       PreserveJVMState pjvms(this);
4746       CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4747       Node* slow_result = set_results_for_java_call(slow_call);
4748       // this->control() comes from set_results_for_java_call
4749       result_reg->init_req(_slow_path, control());
4750       result_val->init_req(_slow_path, slow_result);
4751       result_i_o ->set_req(_slow_path, i_o());
4752       result_mem ->set_req(_slow_path, reset_memory());
4753     }
4754 
4755     // Return the combined state.
4756     set_control(    _gvn.transform(result_reg));
4757     set_i_o(        _gvn.transform(result_i_o));
4758     set_all_memory( _gvn.transform(result_mem));
4759   } // original reexecute is set back here
4760 
4761   set_result(_gvn.transform(result_val));
4762   return true;
4763 }
4764 
4765 // If we have a tighly coupled allocation, the arraycopy may take care
4766 // of the array initialization. If one of the guards we insert between
4767 // the allocation and the arraycopy causes a deoptimization, an
4768 // unitialized array will escape the compiled method. To prevent that
4769 // we set the JVM state for uncommon traps between the allocation and
4770 // the arraycopy to the state before the allocation so, in case of
4771 // deoptimization, we'll reexecute the allocation and the
4772 // initialization.
4773 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4774   if (alloc != NULL) {
4775     ciMethod* trap_method = alloc->jvms()->method();
4776     int trap_bci = alloc->jvms()->bci();
4777 
4778     if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4779           !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4780       // Make sure there's no store between the allocation and the
4781       // arraycopy otherwise visible side effects could be rexecuted
4782       // in case of deoptimization and cause incorrect execution.
4783       bool no_interfering_store = true;
4784       Node* mem = alloc->in(TypeFunc::Memory);
4785       if (mem->is_MergeMem()) {
4786         for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4787           Node* n = mms.memory();
4788           if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4789             assert(n->is_Store(), "what else?");
4790             no_interfering_store = false;
4791             break;
4792           }
4793         }
4794       } else {
4795         for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4796           Node* n = mms.memory();
4797           if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4798             assert(n->is_Store(), "what else?");
4799             no_interfering_store = false;
4800             break;
4801           }
4802         }
4803       }
4804 
4805       if (no_interfering_store) {
4806         JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4807         uint size = alloc->req();
4808         SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4809         old_jvms->set_map(sfpt);
4810         for (uint i = 0; i < size; i++) {
4811           sfpt->init_req(i, alloc->in(i));
4812         }
4813         // re-push array length for deoptimization
4814         sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4815         old_jvms->set_sp(old_jvms->sp()+1);
4816         old_jvms->set_monoff(old_jvms->monoff()+1);
4817         old_jvms->set_scloff(old_jvms->scloff()+1);
4818         old_jvms->set_endoff(old_jvms->endoff()+1);
4819         old_jvms->set_should_reexecute(true);
4820 
4821         sfpt->set_i_o(map()->i_o());
4822         sfpt->set_memory(map()->memory());
4823         sfpt->set_control(map()->control());
4824 
4825         JVMState* saved_jvms = jvms();
4826         saved_reexecute_sp = _reexecute_sp;
4827 
4828         set_jvms(sfpt->jvms());
4829         _reexecute_sp = jvms()->sp();
4830 
4831         return saved_jvms;
4832       }
4833     }
4834   }
4835   return NULL;
4836 }
4837 
4838 // In case of a deoptimization, we restart execution at the
4839 // allocation, allocating a new array. We would leave an uninitialized
4840 // array in the heap that GCs wouldn't expect. Move the allocation
4841 // after the traps so we don't allocate the array if we
4842 // deoptimize. This is possible because tightly_coupled_allocation()
4843 // guarantees there's no observer of the allocated array at this point
4844 // and the control flow is simple enough.
4845 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4846   if (saved_jvms != NULL && !stopped()) {
4847     assert(alloc != NULL, "only with a tightly coupled allocation");
4848     // restore JVM state to the state at the arraycopy
4849     saved_jvms->map()->set_control(map()->control());
4850     assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4851     assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4852     // If we've improved the types of some nodes (null check) while
4853     // emitting the guards, propagate them to the current state
4854     map()->replaced_nodes().apply(saved_jvms->map());
4855     set_jvms(saved_jvms);
4856     _reexecute_sp = saved_reexecute_sp;
4857 
4858     // Remove the allocation from above the guards
4859     CallProjections callprojs;
4860     alloc->extract_projections(&callprojs, true);
4861     InitializeNode* init = alloc->initialization();
4862     Node* alloc_mem = alloc->in(TypeFunc::Memory);
4863     C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4864     C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4865     C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4866 
4867     // move the allocation here (after the guards)
4868     _gvn.hash_delete(alloc);
4869     alloc->set_req(TypeFunc::Control, control());
4870     alloc->set_req(TypeFunc::I_O, i_o());
4871     Node *mem = reset_memory();
4872     set_all_memory(mem);
4873     alloc->set_req(TypeFunc::Memory, mem);
4874     set_control(init->proj_out(TypeFunc::Control));
4875     set_i_o(callprojs.fallthrough_ioproj);
4876 
4877     // Update memory as done in GraphKit::set_output_for_allocation()
4878     const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4879     const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4880     if (ary_type->isa_aryptr() && length_type != NULL) {
4881       ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4882     }
4883     const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4884     int            elemidx  = C->get_alias_index(telemref);
4885     set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4886     set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4887 
4888     Node* allocx = _gvn.transform(alloc);
4889     assert(allocx == alloc, "where has the allocation gone?");
4890     assert(dest->is_CheckCastPP(), "not an allocation result?");
4891 
4892     _gvn.hash_delete(dest);
4893     dest->set_req(0, control());
4894     Node* destx = _gvn.transform(dest);
4895     assert(destx == dest, "where has the allocation result gone?");
4896   }
4897 }
4898 
4899 
4900 //------------------------------inline_arraycopy-----------------------
4901 // public static native void java.lang.System.arraycopy(Object src,  int  srcPos,
4902 //                                                      Object dest, int destPos,
4903 //                                                      int length);
4904 bool LibraryCallKit::inline_arraycopy() {
4905   // Get the arguments.
4906   Node* src         = argument(0);  // type: oop
4907   Node* src_offset  = argument(1);  // type: int
4908   Node* dest        = argument(2);  // type: oop
4909   Node* dest_offset = argument(3);  // type: int
4910   Node* length      = argument(4);  // type: int
4911 
4912 
4913   // Check for allocation before we add nodes that would confuse
4914   // tightly_coupled_allocation()
4915   AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4916 
4917   int saved_reexecute_sp = -1;
4918   JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4919   // See arraycopy_restore_alloc_state() comment
4920   // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4921   // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4922   // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4923   bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4924 
4925   // The following tests must be performed
4926   // (1) src and dest are arrays.
4927   // (2) src and dest arrays must have elements of the same BasicType
4928   // (3) src and dest must not be null.
4929   // (4) src_offset must not be negative.
4930   // (5) dest_offset must not be negative.
4931   // (6) length must not be negative.
4932   // (7) src_offset + length must not exceed length of src.
4933   // (8) dest_offset + length must not exceed length of dest.
4934   // (9) each element of an oop array must be assignable
4935 
4936   // (3) src and dest must not be null.
4937   // always do this here because we need the JVM state for uncommon traps
4938   Node* null_ctl = top();
4939   src  = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src,  T_ARRAY);
4940   assert(null_ctl->is_top(), "no null control here");
4941   dest = null_check(dest, T_ARRAY);
4942 
4943   if (!can_emit_guards) {
4944     // if saved_jvms == NULL and alloc != NULL, we don't emit any
4945     // guards but the arraycopy node could still take advantage of a
4946     // tightly allocated allocation. tightly_coupled_allocation() is
4947     // called again to make sure it takes the null check above into
4948     // account: the null check is mandatory and if it caused an
4949     // uncommon trap to be emitted then the allocation can't be
4950     // considered tightly coupled in this context.
4951     alloc = tightly_coupled_allocation(dest, NULL);
4952   }
4953 
4954   bool validated = false;
4955 
4956   const Type* src_type  = _gvn.type(src);
4957   const Type* dest_type = _gvn.type(dest);
4958   const TypeAryPtr* top_src  = src_type->isa_aryptr();
4959   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4960 
4961   // Do we have the type of src?
4962   bool has_src = (top_src != NULL && top_src->klass() != NULL);
4963   // Do we have the type of dest?
4964   bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4965   // Is the type for src from speculation?
4966   bool src_spec = false;
4967   // Is the type for dest from speculation?
4968   bool dest_spec = false;
4969 
4970   if ((!has_src || !has_dest) && can_emit_guards) {
4971     // We don't have sufficient type information, let's see if
4972     // speculative types can help. We need to have types for both src
4973     // and dest so that it pays off.
4974 
4975     // Do we already have or could we have type information for src
4976     bool could_have_src = has_src;
4977     // Do we already have or could we have type information for dest
4978     bool could_have_dest = has_dest;
4979 
4980     ciKlass* src_k = NULL;
4981     if (!has_src) {
4982       src_k = src_type->speculative_type_not_null();
4983       if (src_k != NULL && src_k->is_array_klass()) {
4984         could_have_src = true;
4985       }
4986     }
4987 
4988     ciKlass* dest_k = NULL;
4989     if (!has_dest) {
4990       dest_k = dest_type->speculative_type_not_null();
4991       if (dest_k != NULL && dest_k->is_array_klass()) {
4992         could_have_dest = true;
4993       }
4994     }
4995 
4996     if (could_have_src && could_have_dest) {
4997       // This is going to pay off so emit the required guards
4998       if (!has_src) {
4999         src = maybe_cast_profiled_obj(src, src_k, true);
5000         src_type  = _gvn.type(src);
5001         top_src  = src_type->isa_aryptr();
5002         has_src = (top_src != NULL && top_src->klass() != NULL);
5003         src_spec = true;
5004       }
5005       if (!has_dest) {
5006         dest = maybe_cast_profiled_obj(dest, dest_k, true);
5007         dest_type  = _gvn.type(dest);
5008         top_dest  = dest_type->isa_aryptr();
5009         has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5010         dest_spec = true;
5011       }
5012     }
5013   }
5014 
5015   if (has_src && has_dest && can_emit_guards) {
5016     BasicType src_elem  = top_src->klass()->as_array_klass()->element_type()->basic_type();
5017     BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5018     if (src_elem  == T_ARRAY)  src_elem  = T_OBJECT;
5019     if (dest_elem == T_ARRAY)  dest_elem = T_OBJECT;
5020 
5021     if (src_elem == dest_elem && src_elem == T_OBJECT) {
5022       // If both arrays are object arrays then having the exact types
5023       // for both will remove the need for a subtype check at runtime
5024       // before the call and may make it possible to pick a faster copy
5025       // routine (without a subtype check on every element)
5026       // Do we have the exact type of src?
5027       bool could_have_src = src_spec;
5028       // Do we have the exact type of dest?
5029       bool could_have_dest = dest_spec;
5030       ciKlass* src_k = top_src->klass();
5031       ciKlass* dest_k = top_dest->klass();
5032       if (!src_spec) {
5033         src_k = src_type->speculative_type_not_null();
5034         if (src_k != NULL && src_k->is_array_klass()) {
5035           could_have_src = true;
5036         }
5037       }
5038       if (!dest_spec) {
5039         dest_k = dest_type->speculative_type_not_null();
5040         if (dest_k != NULL && dest_k->is_array_klass()) {
5041           could_have_dest = true;
5042         }
5043       }
5044       if (could_have_src && could_have_dest) {
5045         // If we can have both exact types, emit the missing guards
5046         if (could_have_src && !src_spec) {
5047           src = maybe_cast_profiled_obj(src, src_k, true);
5048         }
5049         if (could_have_dest && !dest_spec) {
5050           dest = maybe_cast_profiled_obj(dest, dest_k, true);
5051         }
5052       }
5053     }
5054   }
5055 
5056   ciMethod* trap_method = method();
5057   int trap_bci = bci();
5058   if (saved_jvms != NULL) {
5059     trap_method = alloc->jvms()->method();
5060     trap_bci = alloc->jvms()->bci();
5061   }
5062 
5063   if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5064       can_emit_guards &&
5065       !src->is_top() && !dest->is_top()) {
5066     // validate arguments: enables transformation the ArrayCopyNode
5067     validated = true;
5068 
5069     RegionNode* slow_region = new RegionNode(1);
5070     record_for_igvn(slow_region);
5071 
5072     // (1) src and dest are arrays.
5073     generate_non_array_guard(load_object_klass(src), slow_region);
5074     generate_non_array_guard(load_object_klass(dest), slow_region);
5075 
5076     // (2) src and dest arrays must have elements of the same BasicType
5077     // done at macro expansion or at Ideal transformation time
5078 
5079     // (4) src_offset must not be negative.
5080     generate_negative_guard(src_offset, slow_region);
5081 
5082     // (5) dest_offset must not be negative.
5083     generate_negative_guard(dest_offset, slow_region);
5084 
5085     // (7) src_offset + length must not exceed length of src.
5086     generate_limit_guard(src_offset, length,
5087                          load_array_length(src),
5088                          slow_region);
5089 
5090     // (8) dest_offset + length must not exceed length of dest.
5091     generate_limit_guard(dest_offset, length,
5092                          load_array_length(dest),
5093                          slow_region);
5094 
5095     // (9) each element of an oop array must be assignable
5096     Node* src_klass  = load_object_klass(src);
5097     Node* dest_klass = load_object_klass(dest);
5098     Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5099 
5100     if (not_subtype_ctrl != top()) {
5101       PreserveJVMState pjvms(this);
5102       set_control(not_subtype_ctrl);
5103       uncommon_trap(Deoptimization::Reason_intrinsic,
5104                     Deoptimization::Action_make_not_entrant);
5105       assert(stopped(), "Should be stopped");
5106     }
5107     {
5108       PreserveJVMState pjvms(this);
5109       set_control(_gvn.transform(slow_region));
5110       uncommon_trap(Deoptimization::Reason_intrinsic,
5111                     Deoptimization::Action_make_not_entrant);
5112       assert(stopped(), "Should be stopped");
5113     }
5114   }
5115 
5116   arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5117 
5118   if (stopped()) {
5119     return true;
5120   }
5121 
5122   ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
5123                                           // Create LoadRange and LoadKlass nodes for use during macro expansion here
5124                                           // so the compiler has a chance to eliminate them: during macro expansion,
5125                                           // we have to set their control (CastPP nodes are eliminated).
5126                                           load_object_klass(src), load_object_klass(dest),
5127                                           load_array_length(src), load_array_length(dest));
5128 
5129   ac->set_arraycopy(validated);
5130 
5131   Node* n = _gvn.transform(ac);
5132   if (n == ac) {
5133     ac->connect_outputs(this);
5134   } else {
5135     assert(validated, "shouldn't transform if all arguments not validated");
5136     set_all_memory(n);
5137   }
5138 
5139   return true;
5140 }
5141 
5142 
5143 // Helper function which determines if an arraycopy immediately follows
5144 // an allocation, with no intervening tests or other escapes for the object.
5145 AllocateArrayNode*
5146 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5147                                            RegionNode* slow_region) {
5148   if (stopped())             return NULL;  // no fast path
5149   if (C->AliasLevel() == 0)  return NULL;  // no MergeMems around
5150 
5151   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5152   if (alloc == NULL)  return NULL;
5153 
5154   Node* rawmem = memory(Compile::AliasIdxRaw);
5155   // Is the allocation's memory state untouched?
5156   if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5157     // Bail out if there have been raw-memory effects since the allocation.
5158     // (Example:  There might have been a call or safepoint.)
5159     return NULL;
5160   }
5161   rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5162   if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5163     return NULL;
5164   }
5165 
5166   // There must be no unexpected observers of this allocation.
5167   for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5168     Node* obs = ptr->fast_out(i);
5169     if (obs != this->map()) {
5170       return NULL;
5171     }
5172   }
5173 
5174   // This arraycopy must unconditionally follow the allocation of the ptr.
5175   Node* alloc_ctl = ptr->in(0);
5176   assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5177 
5178   Node* ctl = control();
5179   while (ctl != alloc_ctl) {
5180     // There may be guards which feed into the slow_region.
5181     // Any other control flow means that we might not get a chance
5182     // to finish initializing the allocated object.
5183     if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5184       IfNode* iff = ctl->in(0)->as_If();
5185       Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5186       assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5187       if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5188         ctl = iff->in(0);       // This test feeds the known slow_region.
5189         continue;
5190       }
5191       // One more try:  Various low-level checks bottom out in
5192       // uncommon traps.  If the debug-info of the trap omits
5193       // any reference to the allocation, as we've already
5194       // observed, then there can be no objection to the trap.
5195       bool found_trap = false;
5196       for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5197         Node* obs = not_ctl->fast_out(j);
5198         if (obs->in(0) == not_ctl && obs->is_Call() &&
5199             (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5200           found_trap = true; break;
5201         }
5202       }
5203       if (found_trap) {
5204         ctl = iff->in(0);       // This test feeds a harmless uncommon trap.
5205         continue;
5206       }
5207     }
5208     return NULL;
5209   }
5210 
5211   // If we get this far, we have an allocation which immediately
5212   // precedes the arraycopy, and we can take over zeroing the new object.
5213   // The arraycopy will finish the initialization, and provide
5214   // a new control state to which we will anchor the destination pointer.
5215 
5216   return alloc;
5217 }
5218 
5219 //-------------inline_encodeISOArray-----------------------------------
5220 // encode char[] to byte[] in ISO_8859_1
5221 bool LibraryCallKit::inline_encodeISOArray() {
5222   assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5223   // no receiver since it is static method
5224   Node *src         = argument(0);
5225   Node *src_offset  = argument(1);
5226   Node *dst         = argument(2);
5227   Node *dst_offset  = argument(3);
5228   Node *length      = argument(4);
5229 
5230   const Type* src_type = src->Value(&_gvn);
5231   const Type* dst_type = dst->Value(&_gvn);
5232   const TypeAryPtr* top_src = src_type->isa_aryptr();
5233   const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5234   if (top_src  == NULL || top_src->klass()  == NULL ||
5235       top_dest == NULL || top_dest->klass() == NULL) {
5236     // failed array check
5237     return false;
5238   }
5239 
5240   // Figure out the size and type of the elements we will be copying.
5241   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5242   BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5243   if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5244     return false;
5245   }
5246   Node* src_start = array_element_address(src, src_offset, src_elem);
5247   Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5248   // 'src_start' points to src array + scaled offset
5249   // 'dst_start' points to dst array + scaled offset
5250 
5251   const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5252   Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5253   enc = _gvn.transform(enc);
5254   Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5255   set_memory(res_mem, mtype);
5256   set_result(enc);
5257   return true;
5258 }
5259 
5260 //-------------inline_multiplyToLen-----------------------------------
5261 bool LibraryCallKit::inline_multiplyToLen() {
5262   assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5263 
5264   address stubAddr = StubRoutines::multiplyToLen();
5265   if (stubAddr == NULL) {
5266     return false; // Intrinsic's stub is not implemented on this platform
5267   }
5268   const char* stubName = "multiplyToLen";
5269 
5270   assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5271 
5272   Node* x    = argument(1);
5273   Node* xlen = argument(2);
5274   Node* y    = argument(3);
5275   Node* ylen = argument(4);
5276   Node* z    = argument(5);
5277 
5278   const Type* x_type = x->Value(&_gvn);
5279   const Type* y_type = y->Value(&_gvn);
5280   const TypeAryPtr* top_x = x_type->isa_aryptr();
5281   const TypeAryPtr* top_y = y_type->isa_aryptr();
5282   if (top_x  == NULL || top_x->klass()  == NULL ||
5283       top_y == NULL || top_y->klass() == NULL) {
5284     // failed array check
5285     return false;
5286   }
5287 
5288   BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5289   BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5290   if (x_elem != T_INT || y_elem != T_INT) {
5291     return false;
5292   }
5293 
5294   // Set the original stack and the reexecute bit for the interpreter to reexecute
5295   // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5296   // on the return from z array allocation in runtime.
5297   { PreserveReexecuteState preexecs(this);
5298     jvms()->set_should_reexecute(true);
5299 
5300     Node* x_start = array_element_address(x, intcon(0), x_elem);
5301     Node* y_start = array_element_address(y, intcon(0), y_elem);
5302     // 'x_start' points to x array + scaled xlen
5303     // 'y_start' points to y array + scaled ylen
5304 
5305     // Allocate the result array
5306     Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5307     ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5308     Node* klass_node = makecon(TypeKlassPtr::make(klass));
5309 
5310     IdealKit ideal(this);
5311 
5312 #define __ ideal.
5313      Node* one = __ ConI(1);
5314      Node* zero = __ ConI(0);
5315      IdealVariable need_alloc(ideal), z_alloc(ideal);  __ declarations_done();
5316      __ set(need_alloc, zero);
5317      __ set(z_alloc, z);
5318      __ if_then(z, BoolTest::eq, null()); {
5319        __ increment (need_alloc, one);
5320      } __ else_(); {
5321        // Update graphKit memory and control from IdealKit.
5322        sync_kit(ideal);
5323        Node* zlen_arg = load_array_length(z);
5324        // Update IdealKit memory and control from graphKit.
5325        __ sync_kit(this);
5326        __ if_then(zlen_arg, BoolTest::lt, zlen); {
5327          __ increment (need_alloc, one);
5328        } __ end_if();
5329      } __ end_if();
5330 
5331      __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5332        // Update graphKit memory and control from IdealKit.
5333        sync_kit(ideal);
5334        Node * narr = new_array(klass_node, zlen, 1);
5335        // Update IdealKit memory and control from graphKit.
5336        __ sync_kit(this);
5337        __ set(z_alloc, narr);
5338      } __ end_if();
5339 
5340      sync_kit(ideal);
5341      z = __ value(z_alloc);
5342      // Can't use TypeAryPtr::INTS which uses Bottom offset.
5343      _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5344      // Final sync IdealKit and GraphKit.
5345      final_sync(ideal);
5346 #undef __
5347 
5348     Node* z_start = array_element_address(z, intcon(0), T_INT);
5349 
5350     Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5351                                    OptoRuntime::multiplyToLen_Type(),
5352                                    stubAddr, stubName, TypePtr::BOTTOM,
5353                                    x_start, xlen, y_start, ylen, z_start, zlen);
5354   } // original reexecute is set back here
5355 
5356   C->set_has_split_ifs(true); // Has chance for split-if optimization
5357   set_result(z);
5358   return true;
5359 }
5360 
5361 //-------------inline_squareToLen------------------------------------
5362 bool LibraryCallKit::inline_squareToLen() {
5363   assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5364 
5365   address stubAddr = StubRoutines::squareToLen();
5366   if (stubAddr == NULL) {
5367     return false; // Intrinsic's stub is not implemented on this platform
5368   }
5369   const char* stubName = "squareToLen";
5370 
5371   assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5372 
5373   Node* x    = argument(0);
5374   Node* len  = argument(1);
5375   Node* z    = argument(2);
5376   Node* zlen = argument(3);
5377 
5378   const Type* x_type = x->Value(&_gvn);
5379   const Type* z_type = z->Value(&_gvn);
5380   const TypeAryPtr* top_x = x_type->isa_aryptr();
5381   const TypeAryPtr* top_z = z_type->isa_aryptr();
5382   if (top_x  == NULL || top_x->klass()  == NULL ||
5383       top_z  == NULL || top_z->klass()  == NULL) {
5384     // failed array check
5385     return false;
5386   }
5387 
5388   BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5389   BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5390   if (x_elem != T_INT || z_elem != T_INT) {
5391     return false;
5392   }
5393 
5394 
5395   Node* x_start = array_element_address(x, intcon(0), x_elem);
5396   Node* z_start = array_element_address(z, intcon(0), z_elem);
5397 
5398   Node*  call = make_runtime_call(RC_LEAF|RC_NO_FP,
5399                                   OptoRuntime::squareToLen_Type(),
5400                                   stubAddr, stubName, TypePtr::BOTTOM,
5401                                   x_start, len, z_start, zlen);
5402 
5403   set_result(z);
5404   return true;
5405 }
5406 
5407 //-------------inline_mulAdd------------------------------------------
5408 bool LibraryCallKit::inline_mulAdd() {
5409   assert(UseMulAddIntrinsic, "not implementated on this platform");
5410 
5411   address stubAddr = StubRoutines::mulAdd();
5412   if (stubAddr == NULL) {
5413     return false; // Intrinsic's stub is not implemented on this platform
5414   }
5415   const char* stubName = "mulAdd";
5416 
5417   assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5418 
5419   Node* out      = argument(0);
5420   Node* in       = argument(1);
5421   Node* offset   = argument(2);
5422   Node* len      = argument(3);
5423   Node* k        = argument(4);
5424 
5425   const Type* out_type = out->Value(&_gvn);
5426   const Type* in_type = in->Value(&_gvn);
5427   const TypeAryPtr* top_out = out_type->isa_aryptr();
5428   const TypeAryPtr* top_in = in_type->isa_aryptr();
5429   if (top_out  == NULL || top_out->klass()  == NULL ||
5430       top_in == NULL || top_in->klass() == NULL) {
5431     // failed array check
5432     return false;
5433   }
5434 
5435   BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5436   BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5437   if (out_elem != T_INT || in_elem != T_INT) {
5438     return false;
5439   }
5440 
5441   Node* outlen = load_array_length(out);
5442   Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5443   Node* out_start = array_element_address(out, intcon(0), out_elem);
5444   Node* in_start = array_element_address(in, intcon(0), in_elem);
5445 
5446   Node*  call = make_runtime_call(RC_LEAF|RC_NO_FP,
5447                                   OptoRuntime::mulAdd_Type(),
5448                                   stubAddr, stubName, TypePtr::BOTTOM,
5449                                   out_start,in_start, new_offset, len, k);
5450   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5451   set_result(result);
5452   return true;
5453 }
5454 
5455 
5456 /**
5457  * Calculate CRC32 for byte.
5458  * int java.util.zip.CRC32.update(int crc, int b)
5459  */
5460 bool LibraryCallKit::inline_updateCRC32() {
5461   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5462   assert(callee()->signature()->size() == 2, "update has 2 parameters");
5463   // no receiver since it is static method
5464   Node* crc  = argument(0); // type: int
5465   Node* b    = argument(1); // type: int
5466 
5467   /*
5468    *    int c = ~ crc;
5469    *    b = timesXtoThe32[(b ^ c) & 0xFF];
5470    *    b = b ^ (c >>> 8);
5471    *    crc = ~b;
5472    */
5473 
5474   Node* M1 = intcon(-1);
5475   crc = _gvn.transform(new XorINode(crc, M1));
5476   Node* result = _gvn.transform(new XorINode(crc, b));
5477   result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5478 
5479   Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5480   Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5481   Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5482   result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5483 
5484   crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5485   result = _gvn.transform(new XorINode(crc, result));
5486   result = _gvn.transform(new XorINode(result, M1));
5487   set_result(result);
5488   return true;
5489 }
5490 
5491 /**
5492  * Calculate CRC32 for byte[] array.
5493  * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5494  */
5495 bool LibraryCallKit::inline_updateBytesCRC32() {
5496   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5497   assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5498   // no receiver since it is static method
5499   Node* crc     = argument(0); // type: int
5500   Node* src     = argument(1); // type: oop
5501   Node* offset  = argument(2); // type: int
5502   Node* length  = argument(3); // type: int
5503 
5504   const Type* src_type = src->Value(&_gvn);
5505   const TypeAryPtr* top_src = src_type->isa_aryptr();
5506   if (top_src  == NULL || top_src->klass()  == NULL) {
5507     // failed array check
5508     return false;
5509   }
5510 
5511   // Figure out the size and type of the elements we will be copying.
5512   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5513   if (src_elem != T_BYTE) {
5514     return false;
5515   }
5516 
5517   // 'src_start' points to src array + scaled offset
5518   Node* src_start = array_element_address(src, offset, src_elem);
5519 
5520   // We assume that range check is done by caller.
5521   // TODO: generate range check (offset+length < src.length) in debug VM.
5522 
5523   // Call the stub.
5524   address stubAddr = StubRoutines::updateBytesCRC32();
5525   const char *stubName = "updateBytesCRC32";
5526 
5527   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5528                                  stubAddr, stubName, TypePtr::BOTTOM,
5529                                  crc, src_start, length);
5530   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5531   set_result(result);
5532   return true;
5533 }
5534 
5535 /**
5536  * Calculate CRC32 for ByteBuffer.
5537  * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5538  */
5539 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5540   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5541   assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5542   // no receiver since it is static method
5543   Node* crc     = argument(0); // type: int
5544   Node* src     = argument(1); // type: long
5545   Node* offset  = argument(3); // type: int
5546   Node* length  = argument(4); // type: int
5547 
5548   src = ConvL2X(src);  // adjust Java long to machine word
5549   Node* base = _gvn.transform(new CastX2PNode(src));
5550   offset = ConvI2X(offset);
5551 
5552   // 'src_start' points to src array + scaled offset
5553   Node* src_start = basic_plus_adr(top(), base, offset);
5554 
5555   // Call the stub.
5556   address stubAddr = StubRoutines::updateBytesCRC32();
5557   const char *stubName = "updateBytesCRC32";
5558 
5559   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5560                                  stubAddr, stubName, TypePtr::BOTTOM,
5561                                  crc, src_start, length);
5562   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5563   set_result(result);
5564   return true;
5565 }
5566 
5567 //----------------------------inline_reference_get----------------------------
5568 // public T java.lang.ref.Reference.get();
5569 bool LibraryCallKit::inline_reference_get() {
5570   const int referent_offset = java_lang_ref_Reference::referent_offset;
5571   guarantee(referent_offset > 0, "should have already been set");
5572 
5573   // Get the argument:
5574   Node* reference_obj = null_check_receiver();
5575   if (stopped()) return true;
5576 
5577   Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5578 
5579   ciInstanceKlass* klass = env()->Object_klass();
5580   const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5581 
5582   Node* no_ctrl = NULL;
5583   Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5584 
5585   // Use the pre-barrier to record the value in the referent field
5586   pre_barrier(false /* do_load */,
5587               control(),
5588               NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5589               result /* pre_val */,
5590               T_OBJECT);
5591 
5592   // Add memory barrier to prevent commoning reads from this field
5593   // across safepoint since GC can change its value.
5594   insert_mem_bar(Op_MemBarCPUOrder);
5595 
5596   set_result(result);
5597   return true;
5598 }
5599 
5600 
5601 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5602                                               bool is_exact=true, bool is_static=false) {
5603 
5604   const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5605   assert(tinst != NULL, "obj is null");
5606   assert(tinst->klass()->is_loaded(), "obj is not loaded");
5607   assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5608 
5609   ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5610                                                                           ciSymbol::make(fieldTypeString),
5611                                                                           is_static);
5612   if (field == NULL) return (Node *) NULL;
5613   assert (field != NULL, "undefined field");
5614 
5615   // Next code  copied from Parse::do_get_xxx():
5616 
5617   // Compute address and memory type.
5618   int offset  = field->offset_in_bytes();
5619   bool is_vol = field->is_volatile();
5620   ciType* field_klass = field->type();
5621   assert(field_klass->is_loaded(), "should be loaded");
5622   const TypePtr* adr_type = C->alias_type(field)->adr_type();
5623   Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5624   BasicType bt = field->layout_type();
5625 
5626   // Build the resultant type of the load
5627   const Type *type;
5628   if (bt == T_OBJECT) {
5629     type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5630   } else {
5631     type = Type::get_const_basic_type(bt);
5632   }
5633 
5634   if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5635     insert_mem_bar(Op_MemBarVolatile);   // StoreLoad barrier
5636   }
5637   // Build the load.
5638   MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5639   Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5640   // If reference is volatile, prevent following memory ops from
5641   // floating up past the volatile read.  Also prevents commoning
5642   // another volatile read.
5643   if (is_vol) {
5644     // Memory barrier includes bogus read of value to force load BEFORE membar
5645     insert_mem_bar(Op_MemBarAcquire, loadedField);
5646   }
5647   return loadedField;
5648 }
5649 
5650 
5651 //------------------------------inline_aescrypt_Block-----------------------
5652 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5653   address stubAddr;
5654   const char *stubName;
5655   assert(UseAES, "need AES instruction support");
5656 
5657   switch(id) {
5658   case vmIntrinsics::_aescrypt_encryptBlock:
5659     stubAddr = StubRoutines::aescrypt_encryptBlock();
5660     stubName = "aescrypt_encryptBlock";
5661     break;
5662   case vmIntrinsics::_aescrypt_decryptBlock:
5663     stubAddr = StubRoutines::aescrypt_decryptBlock();
5664     stubName = "aescrypt_decryptBlock";
5665     break;
5666   }
5667   if (stubAddr == NULL) return false;
5668 
5669   Node* aescrypt_object = argument(0);
5670   Node* src             = argument(1);
5671   Node* src_offset      = argument(2);
5672   Node* dest            = argument(3);
5673   Node* dest_offset     = argument(4);
5674 
5675   // (1) src and dest are arrays.
5676   const Type* src_type = src->Value(&_gvn);
5677   const Type* dest_type = dest->Value(&_gvn);
5678   const TypeAryPtr* top_src = src_type->isa_aryptr();
5679   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5680   assert (top_src  != NULL && top_src->klass()  != NULL &&  top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5681 
5682   // for the quick and dirty code we will skip all the checks.
5683   // we are just trying to get the call to be generated.
5684   Node* src_start  = src;
5685   Node* dest_start = dest;
5686   if (src_offset != NULL || dest_offset != NULL) {
5687     assert(src_offset != NULL && dest_offset != NULL, "");
5688     src_start  = array_element_address(src,  src_offset,  T_BYTE);
5689     dest_start = array_element_address(dest, dest_offset, T_BYTE);
5690   }
5691 
5692   // now need to get the start of its expanded key array
5693   // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5694   Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5695   if (k_start == NULL) return false;
5696 
5697   if (Matcher::pass_original_key_for_aes()) {
5698     // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5699     // compatibility issues between Java key expansion and SPARC crypto instructions
5700     Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5701     if (original_k_start == NULL) return false;
5702 
5703     // Call the stub.
5704     make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5705                       stubAddr, stubName, TypePtr::BOTTOM,
5706                       src_start, dest_start, k_start, original_k_start);
5707   } else {
5708     // Call the stub.
5709     make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5710                       stubAddr, stubName, TypePtr::BOTTOM,
5711                       src_start, dest_start, k_start);
5712   }
5713 
5714   return true;
5715 }
5716 
5717 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5718 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5719   address stubAddr;
5720   const char *stubName;
5721 
5722   assert(UseAES, "need AES instruction support");
5723 
5724   switch(id) {
5725   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5726     stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5727     stubName = "cipherBlockChaining_encryptAESCrypt";
5728     break;
5729   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5730     stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5731     stubName = "cipherBlockChaining_decryptAESCrypt";
5732     break;
5733   }
5734   if (stubAddr == NULL) return false;
5735 
5736   Node* cipherBlockChaining_object = argument(0);
5737   Node* src                        = argument(1);
5738   Node* src_offset                 = argument(2);
5739   Node* len                        = argument(3);
5740   Node* dest                       = argument(4);
5741   Node* dest_offset                = argument(5);
5742 
5743   // (1) src and dest are arrays.
5744   const Type* src_type = src->Value(&_gvn);
5745   const Type* dest_type = dest->Value(&_gvn);
5746   const TypeAryPtr* top_src = src_type->isa_aryptr();
5747   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5748   assert (top_src  != NULL && top_src->klass()  != NULL
5749           &&  top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5750 
5751   // checks are the responsibility of the caller
5752   Node* src_start  = src;
5753   Node* dest_start = dest;
5754   if (src_offset != NULL || dest_offset != NULL) {
5755     assert(src_offset != NULL && dest_offset != NULL, "");
5756     src_start  = array_element_address(src,  src_offset,  T_BYTE);
5757     dest_start = array_element_address(dest, dest_offset, T_BYTE);
5758   }
5759 
5760   // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5761   // (because of the predicated logic executed earlier).
5762   // so we cast it here safely.
5763   // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5764 
5765   Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5766   if (embeddedCipherObj == NULL) return false;
5767 
5768   // cast it to what we know it will be at runtime
5769   const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5770   assert(tinst != NULL, "CBC obj is null");
5771   assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5772   ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5773   assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5774 
5775   ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5776   const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5777   const TypeOopPtr* xtype = aklass->as_instance_type();
5778   Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5779   aescrypt_object = _gvn.transform(aescrypt_object);
5780 
5781   // we need to get the start of the aescrypt_object's expanded key array
5782   Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5783   if (k_start == NULL) return false;
5784 
5785   // similarly, get the start address of the r vector
5786   Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5787   if (objRvec == NULL) return false;
5788   Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5789 
5790   Node* cbcCrypt;
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, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5798     cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5799                                  OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5800                                  stubAddr, stubName, TypePtr::BOTTOM,
5801                                  src_start, dest_start, k_start, r_start, len, original_k_start);
5802   } else {
5803     // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5804     cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5805                                  OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5806                                  stubAddr, stubName, TypePtr::BOTTOM,
5807                                  src_start, dest_start, k_start, r_start, len);
5808   }
5809 
5810   // return cipher length (int)
5811   Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5812   set_result(retvalue);
5813   return true;
5814 }
5815 
5816 //------------------------------get_key_start_from_aescrypt_object-----------------------
5817 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5818   Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5819   assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5820   if (objAESCryptKey == NULL) return (Node *) NULL;
5821 
5822   // now have the array, need to get the start address of the K array
5823   Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5824   return k_start;
5825 }
5826 
5827 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5828 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5829   Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5830   assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5831   if (objAESCryptKey == NULL) return (Node *) NULL;
5832 
5833   // now have the array, need to get the start address of the lastKey array
5834   Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5835   return original_k_start;
5836 }
5837 
5838 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5839 // Return node representing slow path of predicate check.
5840 // the pseudo code we want to emulate with this predicate is:
5841 // for encryption:
5842 //    if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5843 // for decryption:
5844 //    if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5845 //    note cipher==plain is more conservative than the original java code but that's OK
5846 //
5847 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5848   // The receiver was checked for NULL already.
5849   Node* objCBC = argument(0);
5850 
5851   // Load embeddedCipher field of CipherBlockChaining object.
5852   Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5853 
5854   // get AESCrypt klass for instanceOf check
5855   // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5856   // will have same classloader as CipherBlockChaining object
5857   const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5858   assert(tinst != NULL, "CBCobj is null");
5859   assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5860 
5861   // we want to do an instanceof comparison against the AESCrypt class
5862   ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5863   if (!klass_AESCrypt->is_loaded()) {
5864     // if AESCrypt is not even loaded, we never take the intrinsic fast path
5865     Node* ctrl = control();
5866     set_control(top()); // no regular fast path
5867     return ctrl;
5868   }
5869   ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5870 
5871   Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5872   Node* cmp_instof  = _gvn.transform(new CmpINode(instof, intcon(1)));
5873   Node* bool_instof  = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5874 
5875   Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5876 
5877   // for encryption, we are done
5878   if (!decrypting)
5879     return instof_false;  // even if it is NULL
5880 
5881   // for decryption, we need to add a further check to avoid
5882   // taking the intrinsic path when cipher and plain are the same
5883   // see the original java code for why.
5884   RegionNode* region = new RegionNode(3);
5885   region->init_req(1, instof_false);
5886   Node* src = argument(1);
5887   Node* dest = argument(4);
5888   Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5889   Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5890   Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5891   region->init_req(2, src_dest_conjoint);
5892 
5893   record_for_igvn(region);
5894   return _gvn.transform(region);
5895 }
5896 
5897 //------------------------------inline_sha_implCompress-----------------------
5898 //
5899 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
5900 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
5901 //
5902 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
5903 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
5904 //
5905 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
5906 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
5907 //
5908 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
5909   assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
5910 
5911   Node* sha_obj = argument(0);
5912   Node* src     = argument(1); // type oop
5913   Node* ofs     = argument(2); // type int
5914 
5915   const Type* src_type = src->Value(&_gvn);
5916   const TypeAryPtr* top_src = src_type->isa_aryptr();
5917   if (top_src  == NULL || top_src->klass()  == NULL) {
5918     // failed array check
5919     return false;
5920   }
5921   // Figure out the size and type of the elements we will be copying.
5922   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5923   if (src_elem != T_BYTE) {
5924     return false;
5925   }
5926   // 'src_start' points to src array + offset
5927   Node* src_start = array_element_address(src, ofs, src_elem);
5928   Node* state = NULL;
5929   address stubAddr;
5930   const char *stubName;
5931 
5932   switch(id) {
5933   case vmIntrinsics::_sha_implCompress:
5934     assert(UseSHA1Intrinsics, "need SHA1 instruction support");
5935     state = get_state_from_sha_object(sha_obj);
5936     stubAddr = StubRoutines::sha1_implCompress();
5937     stubName = "sha1_implCompress";
5938     break;
5939   case vmIntrinsics::_sha2_implCompress:
5940     assert(UseSHA256Intrinsics, "need SHA256 instruction support");
5941     state = get_state_from_sha_object(sha_obj);
5942     stubAddr = StubRoutines::sha256_implCompress();
5943     stubName = "sha256_implCompress";
5944     break;
5945   case vmIntrinsics::_sha5_implCompress:
5946     assert(UseSHA512Intrinsics, "need SHA512 instruction support");
5947     state = get_state_from_sha5_object(sha_obj);
5948     stubAddr = StubRoutines::sha512_implCompress();
5949     stubName = "sha512_implCompress";
5950     break;
5951   default:
5952     fatal_unexpected_iid(id);
5953     return false;
5954   }
5955   if (state == NULL) return false;
5956 
5957   // Call the stub.
5958   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
5959                                  stubAddr, stubName, TypePtr::BOTTOM,
5960                                  src_start, state);
5961 
5962   return true;
5963 }
5964 
5965 //------------------------------inline_digestBase_implCompressMB-----------------------
5966 //
5967 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
5968 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
5969 //
5970 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
5971   assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
5972          "need SHA1/SHA256/SHA512 instruction support");
5973   assert((uint)predicate < 3, "sanity");
5974   assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
5975 
5976   Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
5977   Node* src            = argument(1); // byte[] array
5978   Node* ofs            = argument(2); // type int
5979   Node* limit          = argument(3); // type int
5980 
5981   const Type* src_type = src->Value(&_gvn);
5982   const TypeAryPtr* top_src = src_type->isa_aryptr();
5983   if (top_src  == NULL || top_src->klass()  == NULL) {
5984     // failed array check
5985     return false;
5986   }
5987   // Figure out the size and type of the elements we will be copying.
5988   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5989   if (src_elem != T_BYTE) {
5990     return false;
5991   }
5992   // 'src_start' points to src array + offset
5993   Node* src_start = array_element_address(src, ofs, src_elem);
5994 
5995   const char* klass_SHA_name = NULL;
5996   const char* stub_name = NULL;
5997   address     stub_addr = NULL;
5998   bool        long_state = false;
5999 
6000   switch (predicate) {
6001   case 0:
6002     if (UseSHA1Intrinsics) {
6003       klass_SHA_name = "sun/security/provider/SHA";
6004       stub_name = "sha1_implCompressMB";
6005       stub_addr = StubRoutines::sha1_implCompressMB();
6006     }
6007     break;
6008   case 1:
6009     if (UseSHA256Intrinsics) {
6010       klass_SHA_name = "sun/security/provider/SHA2";
6011       stub_name = "sha256_implCompressMB";
6012       stub_addr = StubRoutines::sha256_implCompressMB();
6013     }
6014     break;
6015   case 2:
6016     if (UseSHA512Intrinsics) {
6017       klass_SHA_name = "sun/security/provider/SHA5";
6018       stub_name = "sha512_implCompressMB";
6019       stub_addr = StubRoutines::sha512_implCompressMB();
6020       long_state = true;
6021     }
6022     break;
6023   default:
6024     fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6025   }
6026   if (klass_SHA_name != NULL) {
6027     // get DigestBase klass to lookup for SHA klass
6028     const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6029     assert(tinst != NULL, "digestBase_obj is not instance???");
6030     assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6031 
6032     ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6033     assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6034     ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6035     return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6036   }
6037   return false;
6038 }
6039 //------------------------------inline_sha_implCompressMB-----------------------
6040 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6041                                                bool long_state, address stubAddr, const char *stubName,
6042                                                Node* src_start, Node* ofs, Node* limit) {
6043   const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6044   const TypeOopPtr* xtype = aklass->as_instance_type();
6045   Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6046   sha_obj = _gvn.transform(sha_obj);
6047 
6048   Node* state;
6049   if (long_state) {
6050     state = get_state_from_sha5_object(sha_obj);
6051   } else {
6052     state = get_state_from_sha_object(sha_obj);
6053   }
6054   if (state == NULL) return false;
6055 
6056   // Call the stub.
6057   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6058                                  OptoRuntime::digestBase_implCompressMB_Type(),
6059                                  stubAddr, stubName, TypePtr::BOTTOM,
6060                                  src_start, state, ofs, limit);
6061   // return ofs (int)
6062   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6063   set_result(result);
6064 
6065   return true;
6066 }
6067 
6068 //------------------------------get_state_from_sha_object-----------------------
6069 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6070   Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6071   assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6072   if (sha_state == NULL) return (Node *) NULL;
6073 
6074   // now have the array, need to get the start address of the state array
6075   Node* state = array_element_address(sha_state, intcon(0), T_INT);
6076   return state;
6077 }
6078 
6079 //------------------------------get_state_from_sha5_object-----------------------
6080 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6081   Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6082   assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6083   if (sha_state == NULL) return (Node *) NULL;
6084 
6085   // now have the array, need to get the start address of the state array
6086   Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6087   return state;
6088 }
6089 
6090 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6091 // Return node representing slow path of predicate check.
6092 // the pseudo code we want to emulate with this predicate is:
6093 //    if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6094 //
6095 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6096   assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6097          "need SHA1/SHA256/SHA512 instruction support");
6098   assert((uint)predicate < 3, "sanity");
6099 
6100   // The receiver was checked for NULL already.
6101   Node* digestBaseObj = argument(0);
6102 
6103   // get DigestBase klass for instanceOf check
6104   const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6105   assert(tinst != NULL, "digestBaseObj is null");
6106   assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6107 
6108   const char* klass_SHA_name = NULL;
6109   switch (predicate) {
6110   case 0:
6111     if (UseSHA1Intrinsics) {
6112       // we want to do an instanceof comparison against the SHA class
6113       klass_SHA_name = "sun/security/provider/SHA";
6114     }
6115     break;
6116   case 1:
6117     if (UseSHA256Intrinsics) {
6118       // we want to do an instanceof comparison against the SHA2 class
6119       klass_SHA_name = "sun/security/provider/SHA2";
6120     }
6121     break;
6122   case 2:
6123     if (UseSHA512Intrinsics) {
6124       // we want to do an instanceof comparison against the SHA5 class
6125       klass_SHA_name = "sun/security/provider/SHA5";
6126     }
6127     break;
6128   default:
6129     fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6130   }
6131 
6132   ciKlass* klass_SHA = NULL;
6133   if (klass_SHA_name != NULL) {
6134     klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6135   }
6136   if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6137     // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6138     Node* ctrl = control();
6139     set_control(top()); // no intrinsic path
6140     return ctrl;
6141   }
6142   ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6143 
6144   Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6145   Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6146   Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6147   Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6148 
6149   return instof_false;  // even if it is NULL
6150 }
6151 
6152 bool LibraryCallKit::inline_profileBoolean() {
6153   Node* counts = argument(1);
6154   const TypeAryPtr* ary = NULL;
6155   ciArray* aobj = NULL;
6156   if (counts->is_Con()
6157       && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6158       && (aobj = ary->const_oop()->as_array()) != NULL
6159       && (aobj->length() == 2)) {
6160     // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6161     jint false_cnt = aobj->element_value(0).as_int();
6162     jint  true_cnt = aobj->element_value(1).as_int();
6163 
6164     if (C->log() != NULL) {
6165       C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6166                      false_cnt, true_cnt);
6167     }
6168 
6169     if (false_cnt + true_cnt == 0) {
6170       // According to profile, never executed.
6171       uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6172                           Deoptimization::Action_reinterpret);
6173       return true;
6174     }
6175 
6176     // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6177     // is a number of each value occurrences.
6178     Node* result = argument(0);
6179     if (false_cnt == 0 || true_cnt == 0) {
6180       // According to profile, one value has been never seen.
6181       int expected_val = (false_cnt == 0) ? 1 : 0;
6182 
6183       Node* cmp  = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6184       Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6185 
6186       IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6187       Node* fast_path = _gvn.transform(new IfTrueNode(check));
6188       Node* slow_path = _gvn.transform(new IfFalseNode(check));
6189 
6190       { // Slow path: uncommon trap for never seen value and then reexecute
6191         // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6192         // the value has been seen at least once.
6193         PreserveJVMState pjvms(this);
6194         PreserveReexecuteState preexecs(this);
6195         jvms()->set_should_reexecute(true);
6196 
6197         set_control(slow_path);
6198         set_i_o(i_o());
6199 
6200         uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6201                             Deoptimization::Action_reinterpret);
6202       }
6203       // The guard for never seen value enables sharpening of the result and
6204       // returning a constant. It allows to eliminate branches on the same value
6205       // later on.
6206       set_control(fast_path);
6207       result = intcon(expected_val);
6208     }
6209     // Stop profiling.
6210     // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6211     // By replacing method body with profile data (represented as ProfileBooleanNode
6212     // on IR level) we effectively disable profiling.
6213     // It enables full speed execution once optimized code is generated.
6214     Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6215     C->record_for_igvn(profile);
6216     set_result(profile);
6217     return true;
6218   } else {
6219     // Continue profiling.
6220     // Profile data isn't available at the moment. So, execute method's bytecode version.
6221     // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6222     // is compiled and counters aren't available since corresponding MethodHandle
6223     // isn't a compile-time constant.
6224     return false;
6225   }
6226 }
6227 
6228 bool LibraryCallKit::inline_isCompileConstant() {
6229   Node* n = argument(0);
6230   set_result(n->is_Con() ? intcon(1) : intcon(0));
6231   return true;
6232 }
--- EOF ---