1 /*
   2  * Copyright (c) 1999, 2013, 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 "classfile/systemDictionary.hpp"
  27 #include "classfile/vmSymbols.hpp"
  28 #include "compiler/compileBroker.hpp"
  29 #include "compiler/compileLog.hpp"
  30 #include "oops/objArrayKlass.hpp"
  31 #include "opto/addnode.hpp"
  32 #include "opto/callGenerator.hpp"
  33 #include "opto/cfgnode.hpp"
  34 #include "opto/idealKit.hpp"
  35 #include "opto/mathexactnode.hpp"
  36 #include "opto/mulnode.hpp"
  37 #include "opto/parse.hpp"
  38 #include "opto/runtime.hpp"
  39 #include "opto/subnode.hpp"
  40 #include "prims/nativeLookup.hpp"
  41 #include "runtime/sharedRuntime.hpp"
  42 #include "trace/traceMacros.hpp"
  43 
  44 class LibraryIntrinsic : public InlineCallGenerator {
  45   // Extend the set of intrinsics known to the runtime:
  46  public:
  47  private:
  48   bool             _is_virtual;
  49   bool             _is_predicted;
  50   bool             _does_virtual_dispatch;
  51   vmIntrinsics::ID _intrinsic_id;
  52 
  53  public:
  54   LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id)
  55     : InlineCallGenerator(m),
  56       _is_virtual(is_virtual),
  57       _is_predicted(is_predicted),
  58       _does_virtual_dispatch(does_virtual_dispatch),
  59       _intrinsic_id(id)
  60   {
  61   }
  62   virtual bool is_intrinsic() const { return true; }
  63   virtual bool is_virtual()   const { return _is_virtual; }
  64   virtual bool is_predicted()   const { return _is_predicted; }
  65   virtual bool does_virtual_dispatch()   const { return _does_virtual_dispatch; }
  66   virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
  67   virtual Node* generate_predicate(JVMState* jvms);
  68   vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
  69 };
  70 
  71 
  72 // Local helper class for LibraryIntrinsic:
  73 class LibraryCallKit : public GraphKit {
  74  private:
  75   LibraryIntrinsic* _intrinsic;     // the library intrinsic being called
  76   Node*             _result;        // the result node, if any
  77   int               _reexecute_sp;  // the stack pointer when bytecode needs to be reexecuted
  78 
  79   const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
  80 
  81  public:
  82   LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
  83     : GraphKit(jvms),
  84       _intrinsic(intrinsic),
  85       _result(NULL)
  86   {
  87     // Check if this is a root compile.  In that case we don't have a caller.
  88     if (!jvms->has_method()) {
  89       _reexecute_sp = sp();
  90     } else {
  91       // Find out how many arguments the interpreter needs when deoptimizing
  92       // and save the stack pointer value so it can used by uncommon_trap.
  93       // We find the argument count by looking at the declared signature.
  94       bool ignored_will_link;
  95       ciSignature* declared_signature = NULL;
  96       ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
  97       const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
  98       _reexecute_sp = sp() + nargs;  // "push" arguments back on stack
  99     }
 100   }
 101 
 102   virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
 103 
 104   ciMethod*         caller()    const    { return jvms()->method(); }
 105   int               bci()       const    { return jvms()->bci(); }
 106   LibraryIntrinsic* intrinsic() const    { return _intrinsic; }
 107   vmIntrinsics::ID  intrinsic_id() const { return _intrinsic->intrinsic_id(); }
 108   ciMethod*         callee()    const    { return _intrinsic->method(); }
 109 
 110   bool try_to_inline();
 111   Node* try_to_predicate();
 112 
 113   void push_result() {
 114     // Push the result onto the stack.
 115     if (!stopped() && result() != NULL) {
 116       BasicType bt = result()->bottom_type()->basic_type();
 117       push_node(bt, result());
 118     }
 119   }
 120 
 121  private:
 122   void fatal_unexpected_iid(vmIntrinsics::ID iid) {
 123     fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
 124   }
 125 
 126   void  set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
 127   void  set_result(RegionNode* region, PhiNode* value);
 128   Node*     result() { return _result; }
 129 
 130   virtual int reexecute_sp() { return _reexecute_sp; }
 131 
 132   // Helper functions to inline natives
 133   Node* generate_guard(Node* test, RegionNode* region, float true_prob);
 134   Node* generate_slow_guard(Node* test, RegionNode* region);
 135   Node* generate_fair_guard(Node* test, RegionNode* region);
 136   Node* generate_negative_guard(Node* index, RegionNode* region,
 137                                 // resulting CastII of index:
 138                                 Node* *pos_index = NULL);
 139   Node* generate_nonpositive_guard(Node* index, bool never_negative,
 140                                    // resulting CastII of index:
 141                                    Node* *pos_index = NULL);
 142   Node* generate_limit_guard(Node* offset, Node* subseq_length,
 143                              Node* array_length,
 144                              RegionNode* region);
 145   Node* generate_current_thread(Node* &tls_output);
 146   address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
 147                               bool disjoint_bases, const char* &name, bool dest_uninitialized);
 148   Node* load_mirror_from_klass(Node* klass);
 149   Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
 150                                       RegionNode* region, int null_path,
 151                                       int offset);
 152   Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
 153                                RegionNode* region, int null_path) {
 154     int offset = java_lang_Class::klass_offset_in_bytes();
 155     return load_klass_from_mirror_common(mirror, never_see_null,
 156                                          region, null_path,
 157                                          offset);
 158   }
 159   Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
 160                                      RegionNode* region, int null_path) {
 161     int offset = java_lang_Class::array_klass_offset_in_bytes();
 162     return load_klass_from_mirror_common(mirror, never_see_null,
 163                                          region, null_path,
 164                                          offset);
 165   }
 166   Node* generate_access_flags_guard(Node* kls,
 167                                     int modifier_mask, int modifier_bits,
 168                                     RegionNode* region);
 169   Node* generate_interface_guard(Node* kls, RegionNode* region);
 170   Node* generate_array_guard(Node* kls, RegionNode* region) {
 171     return generate_array_guard_common(kls, region, false, false);
 172   }
 173   Node* generate_non_array_guard(Node* kls, RegionNode* region) {
 174     return generate_array_guard_common(kls, region, false, true);
 175   }
 176   Node* generate_objArray_guard(Node* kls, RegionNode* region) {
 177     return generate_array_guard_common(kls, region, true, false);
 178   }
 179   Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
 180     return generate_array_guard_common(kls, region, true, true);
 181   }
 182   Node* generate_array_guard_common(Node* kls, RegionNode* region,
 183                                     bool obj_array, bool not_array);
 184   Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
 185   CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
 186                                      bool is_virtual = false, bool is_static = false);
 187   CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
 188     return generate_method_call(method_id, false, true);
 189   }
 190   CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
 191     return generate_method_call(method_id, true, false);
 192   }
 193   Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
 194 
 195   Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
 196   Node* make_string_method_node(int opcode, Node* str1, Node* str2);
 197   bool inline_string_compareTo();
 198   bool inline_string_indexOf();
 199   Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
 200   bool inline_string_equals();
 201   Node* round_double_node(Node* n);
 202   bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
 203   bool inline_math_native(vmIntrinsics::ID id);
 204   bool inline_trig(vmIntrinsics::ID id);
 205   bool inline_math(vmIntrinsics::ID id);
 206   void inline_math_mathExact(Node* math);
 207   bool inline_math_addExactI(bool is_increment);
 208   bool inline_math_addExactL(bool is_increment);
 209   bool inline_math_multiplyExactI();
 210   bool inline_math_multiplyExactL();
 211   bool inline_math_negateExactI();
 212   bool inline_math_negateExactL();
 213   bool inline_math_subtractExactI(bool is_decrement);
 214   bool inline_math_subtractExactL(bool is_decrement);
 215   bool inline_exp();
 216   bool inline_pow();
 217   void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
 218   bool inline_min_max(vmIntrinsics::ID id);
 219   Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
 220   // This returns Type::AnyPtr, RawPtr, or OopPtr.
 221   int classify_unsafe_addr(Node* &base, Node* &offset);
 222   Node* make_unsafe_address(Node* base, Node* offset);
 223   // Helper for inline_unsafe_access.
 224   // Generates the guards that check whether the result of
 225   // Unsafe.getObject should be recorded in an SATB log buffer.
 226   void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
 227   bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
 228   bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
 229   static bool klass_needs_init_guard(Node* kls);
 230   bool inline_unsafe_allocate();
 231   bool inline_unsafe_copyMemory();
 232   bool inline_native_currentThread();
 233 #ifdef TRACE_HAVE_INTRINSICS
 234   bool inline_native_classID();
 235   bool inline_native_threadID();
 236 #endif
 237   bool inline_native_time_funcs(address method, const char* funcName);
 238   bool inline_native_isInterrupted();
 239   bool inline_native_Class_query(vmIntrinsics::ID id);
 240   bool inline_native_subtype_check();
 241 
 242   bool inline_native_newArray();
 243   bool inline_native_getLength();
 244   bool inline_array_copyOf(bool is_copyOfRange);
 245   bool inline_array_equals();
 246   void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
 247   bool inline_native_clone(bool is_virtual);
 248   bool inline_native_Reflection_getCallerClass();
 249   // Helper function for inlining native object hash method
 250   bool inline_native_hashcode(bool is_virtual, bool is_static);
 251   bool inline_native_getClass();
 252 
 253   // Helper functions for inlining arraycopy
 254   bool inline_arraycopy();
 255   void generate_arraycopy(const TypePtr* adr_type,
 256                           BasicType basic_elem_type,
 257                           Node* src,  Node* src_offset,
 258                           Node* dest, Node* dest_offset,
 259                           Node* copy_length,
 260                           bool disjoint_bases = false,
 261                           bool length_never_negative = false,
 262                           RegionNode* slow_region = NULL);
 263   AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
 264                                                 RegionNode* slow_region);
 265   void generate_clear_array(const TypePtr* adr_type,
 266                             Node* dest,
 267                             BasicType basic_elem_type,
 268                             Node* slice_off,
 269                             Node* slice_len,
 270                             Node* slice_end);
 271   bool generate_block_arraycopy(const TypePtr* adr_type,
 272                                 BasicType basic_elem_type,
 273                                 AllocateNode* alloc,
 274                                 Node* src,  Node* src_offset,
 275                                 Node* dest, Node* dest_offset,
 276                                 Node* dest_size, bool dest_uninitialized);
 277   void generate_slow_arraycopy(const TypePtr* adr_type,
 278                                Node* src,  Node* src_offset,
 279                                Node* dest, Node* dest_offset,
 280                                Node* copy_length, bool dest_uninitialized);
 281   Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
 282                                      Node* dest_elem_klass,
 283                                      Node* src,  Node* src_offset,
 284                                      Node* dest, Node* dest_offset,
 285                                      Node* copy_length, bool dest_uninitialized);
 286   Node* generate_generic_arraycopy(const TypePtr* adr_type,
 287                                    Node* src,  Node* src_offset,
 288                                    Node* dest, Node* dest_offset,
 289                                    Node* copy_length, bool dest_uninitialized);
 290   void generate_unchecked_arraycopy(const TypePtr* adr_type,
 291                                     BasicType basic_elem_type,
 292                                     bool disjoint_bases,
 293                                     Node* src,  Node* src_offset,
 294                                     Node* dest, Node* dest_offset,
 295                                     Node* copy_length, bool dest_uninitialized);
 296   typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
 297   bool inline_unsafe_load_store(BasicType type,  LoadStoreKind kind);
 298   bool inline_unsafe_ordered_store(BasicType type);
 299   bool inline_unsafe_fence(vmIntrinsics::ID id);
 300   bool inline_fp_conversions(vmIntrinsics::ID id);
 301   bool inline_number_methods(vmIntrinsics::ID id);
 302   bool inline_reference_get();
 303   bool inline_aescrypt_Block(vmIntrinsics::ID id);
 304   bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
 305   Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
 306   Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
 307   bool inline_encodeISOArray();
 308   bool inline_updateCRC32();
 309   bool inline_updateBytesCRC32();
 310   bool inline_updateByteBufferCRC32();
 311 };
 312 
 313 
 314 //---------------------------make_vm_intrinsic----------------------------
 315 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
 316   vmIntrinsics::ID id = m->intrinsic_id();
 317   assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
 318 
 319   if (DisableIntrinsic[0] != '\0'
 320       && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
 321     // disabled by a user request on the command line:
 322     // example: -XX:DisableIntrinsic=_hashCode,_getClass
 323     return NULL;
 324   }
 325 
 326   if (!m->is_loaded()) {
 327     // do not attempt to inline unloaded methods
 328     return NULL;
 329   }
 330 
 331   // Only a few intrinsics implement a virtual dispatch.
 332   // They are expensive calls which are also frequently overridden.
 333   if (is_virtual) {
 334     switch (id) {
 335     case vmIntrinsics::_hashCode:
 336     case vmIntrinsics::_clone:
 337       // OK, Object.hashCode and Object.clone intrinsics come in both flavors
 338       break;
 339     default:
 340       return NULL;
 341     }
 342   }
 343 
 344   // -XX:-InlineNatives disables nearly all intrinsics:
 345   if (!InlineNatives) {
 346     switch (id) {
 347     case vmIntrinsics::_indexOf:
 348     case vmIntrinsics::_compareTo:
 349     case vmIntrinsics::_equals:
 350     case vmIntrinsics::_equalsC:
 351     case vmIntrinsics::_getAndAddInt:
 352     case vmIntrinsics::_getAndAddLong:
 353     case vmIntrinsics::_getAndSetInt:
 354     case vmIntrinsics::_getAndSetLong:
 355     case vmIntrinsics::_getAndSetObject:
 356     case vmIntrinsics::_loadFence:
 357     case vmIntrinsics::_storeFence:
 358     case vmIntrinsics::_fullFence:
 359       break;  // InlineNatives does not control String.compareTo
 360     case vmIntrinsics::_Reference_get:
 361       break;  // InlineNatives does not control Reference.get
 362     default:
 363       return NULL;
 364     }
 365   }
 366 
 367   bool is_predicted = false;
 368   bool does_virtual_dispatch = false;
 369 
 370   switch (id) {
 371   case vmIntrinsics::_compareTo:
 372     if (!SpecialStringCompareTo)  return NULL;
 373     if (!Matcher::match_rule_supported(Op_StrComp))  return NULL;
 374     break;
 375   case vmIntrinsics::_indexOf:
 376     if (!SpecialStringIndexOf)  return NULL;
 377     break;
 378   case vmIntrinsics::_equals:
 379     if (!SpecialStringEquals)  return NULL;
 380     if (!Matcher::match_rule_supported(Op_StrEquals))  return NULL;
 381     break;
 382   case vmIntrinsics::_equalsC:
 383     if (!SpecialArraysEquals)  return NULL;
 384     if (!Matcher::match_rule_supported(Op_AryEq))  return NULL;
 385     break;
 386   case vmIntrinsics::_arraycopy:
 387     if (!InlineArrayCopy)  return NULL;
 388     break;
 389   case vmIntrinsics::_copyMemory:
 390     if (StubRoutines::unsafe_arraycopy() == NULL)  return NULL;
 391     if (!InlineArrayCopy)  return NULL;
 392     break;
 393   case vmIntrinsics::_hashCode:
 394     if (!InlineObjectHash)  return NULL;
 395     does_virtual_dispatch = true;
 396     break;
 397   case vmIntrinsics::_clone:
 398     does_virtual_dispatch = true;
 399   case vmIntrinsics::_copyOf:
 400   case vmIntrinsics::_copyOfRange:
 401     if (!InlineObjectCopy)  return NULL;
 402     // These also use the arraycopy intrinsic mechanism:
 403     if (!InlineArrayCopy)  return NULL;
 404     break;
 405   case vmIntrinsics::_encodeISOArray:
 406     if (!SpecialEncodeISOArray)  return NULL;
 407     if (!Matcher::match_rule_supported(Op_EncodeISOArray))  return NULL;
 408     break;
 409   case vmIntrinsics::_checkIndex:
 410     // We do not intrinsify this.  The optimizer does fine with it.
 411     return NULL;
 412 
 413   case vmIntrinsics::_getCallerClass:
 414     if (!UseNewReflection)  return NULL;
 415     if (!InlineReflectionGetCallerClass)  return NULL;
 416     if (SystemDictionary::reflect_CallerSensitive_klass() == NULL)  return NULL;
 417     break;
 418 
 419   case vmIntrinsics::_bitCount_i:
 420     if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
 421     break;
 422 
 423   case vmIntrinsics::_bitCount_l:
 424     if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
 425     break;
 426 
 427   case vmIntrinsics::_numberOfLeadingZeros_i:
 428     if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
 429     break;
 430 
 431   case vmIntrinsics::_numberOfLeadingZeros_l:
 432     if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
 433     break;
 434 
 435   case vmIntrinsics::_numberOfTrailingZeros_i:
 436     if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
 437     break;
 438 
 439   case vmIntrinsics::_numberOfTrailingZeros_l:
 440     if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
 441     break;
 442 
 443   case vmIntrinsics::_reverseBytes_c:
 444     if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
 445     break;
 446   case vmIntrinsics::_reverseBytes_s:
 447     if (!Matcher::match_rule_supported(Op_ReverseBytesS))  return NULL;
 448     break;
 449   case vmIntrinsics::_reverseBytes_i:
 450     if (!Matcher::match_rule_supported(Op_ReverseBytesI))  return NULL;
 451     break;
 452   case vmIntrinsics::_reverseBytes_l:
 453     if (!Matcher::match_rule_supported(Op_ReverseBytesL))  return NULL;
 454     break;
 455 
 456   case vmIntrinsics::_Reference_get:
 457     // Use the intrinsic version of Reference.get() so that the value in
 458     // the referent field can be registered by the G1 pre-barrier code.
 459     // Also add memory barrier to prevent commoning reads from this field
 460     // across safepoint since GC can change it value.
 461     break;
 462 
 463   case vmIntrinsics::_compareAndSwapObject:
 464 #ifdef _LP64
 465     if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
 466 #endif
 467     break;
 468 
 469   case vmIntrinsics::_compareAndSwapLong:
 470     if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
 471     break;
 472 
 473   case vmIntrinsics::_getAndAddInt:
 474     if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
 475     break;
 476 
 477   case vmIntrinsics::_getAndAddLong:
 478     if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
 479     break;
 480 
 481   case vmIntrinsics::_getAndSetInt:
 482     if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
 483     break;
 484 
 485   case vmIntrinsics::_getAndSetLong:
 486     if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
 487     break;
 488 
 489   case vmIntrinsics::_getAndSetObject:
 490 #ifdef _LP64
 491     if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
 492     if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
 493     break;
 494 #else
 495     if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
 496     break;
 497 #endif
 498 
 499   case vmIntrinsics::_aescrypt_encryptBlock:
 500   case vmIntrinsics::_aescrypt_decryptBlock:
 501     if (!UseAESIntrinsics) return NULL;
 502     break;
 503 
 504   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 505   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 506     if (!UseAESIntrinsics) return NULL;
 507     // these two require the predicated logic
 508     is_predicted = true;
 509     break;
 510 
 511   case vmIntrinsics::_updateCRC32:
 512   case vmIntrinsics::_updateBytesCRC32:
 513   case vmIntrinsics::_updateByteBufferCRC32:
 514     if (!UseCRC32Intrinsics) return NULL;
 515     break;
 516 
 517   case vmIntrinsics::_incrementExactI:
 518   case vmIntrinsics::_addExactI:
 519     if (!Matcher::match_rule_supported(Op_AddExactI) || !UseMathExactIntrinsics) return NULL;
 520     break;
 521   case vmIntrinsics::_incrementExactL:
 522   case vmIntrinsics::_addExactL:
 523     if (!Matcher::match_rule_supported(Op_AddExactL) || !UseMathExactIntrinsics) return NULL;
 524     break;
 525   case vmIntrinsics::_decrementExactI:
 526   case vmIntrinsics::_subtractExactI:
 527     if (!Matcher::match_rule_supported(Op_SubExactI) || !UseMathExactIntrinsics) return NULL;
 528     break;
 529   case vmIntrinsics::_decrementExactL:
 530   case vmIntrinsics::_subtractExactL:
 531     if (!Matcher::match_rule_supported(Op_SubExactL) || !UseMathExactIntrinsics) return NULL;
 532     break;
 533   case vmIntrinsics::_negateExactI:
 534     if (!Matcher::match_rule_supported(Op_NegExactI) || !UseMathExactIntrinsics) return NULL;
 535     break;
 536   case vmIntrinsics::_negateExactL:
 537     if (!Matcher::match_rule_supported(Op_NegExactL) || !UseMathExactIntrinsics) return NULL;
 538     break;
 539   case vmIntrinsics::_multiplyExactI:
 540     if (!Matcher::match_rule_supported(Op_MulExactI) || !UseMathExactIntrinsics) return NULL;
 541     break;
 542   case vmIntrinsics::_multiplyExactL:
 543     if (!Matcher::match_rule_supported(Op_MulExactL) || !UseMathExactIntrinsics) return NULL;
 544     break;
 545 
 546  default:
 547     assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
 548     assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
 549     break;
 550   }
 551 
 552   // -XX:-InlineClassNatives disables natives from the Class class.
 553   // The flag applies to all reflective calls, notably Array.newArray
 554   // (visible to Java programmers as Array.newInstance).
 555   if (m->holder()->name() == ciSymbol::java_lang_Class() ||
 556       m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
 557     if (!InlineClassNatives)  return NULL;
 558   }
 559 
 560   // -XX:-InlineThreadNatives disables natives from the Thread class.
 561   if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
 562     if (!InlineThreadNatives)  return NULL;
 563   }
 564 
 565   // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
 566   if (m->holder()->name() == ciSymbol::java_lang_Math() ||
 567       m->holder()->name() == ciSymbol::java_lang_Float() ||
 568       m->holder()->name() == ciSymbol::java_lang_Double()) {
 569     if (!InlineMathNatives)  return NULL;
 570   }
 571 
 572   // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
 573   if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
 574     if (!InlineUnsafeOps)  return NULL;
 575   }
 576 
 577   return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
 578 }
 579 
 580 //----------------------register_library_intrinsics-----------------------
 581 // Initialize this file's data structures, for each Compile instance.
 582 void Compile::register_library_intrinsics() {
 583   // Nothing to do here.
 584 }
 585 
 586 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
 587   LibraryCallKit kit(jvms, this);
 588   Compile* C = kit.C;
 589   int nodes = C->unique();
 590 #ifndef PRODUCT
 591   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
 592     char buf[1000];
 593     const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
 594     tty->print_cr("Intrinsic %s", str);
 595   }
 596 #endif
 597   ciMethod* callee = kit.callee();
 598   const int bci    = kit.bci();
 599 
 600   // Try to inline the intrinsic.
 601   if (kit.try_to_inline()) {
 602     if (C->print_intrinsics() || C->print_inlining()) {
 603       C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
 604     }
 605     C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
 606     if (C->log()) {
 607       C->log()->elem("intrinsic id='%s'%s nodes='%d'",
 608                      vmIntrinsics::name_at(intrinsic_id()),
 609                      (is_virtual() ? " virtual='1'" : ""),
 610                      C->unique() - nodes);
 611     }
 612     // Push the result from the inlined method onto the stack.
 613     kit.push_result();
 614     return kit.transfer_exceptions_into_jvms();
 615   }
 616 
 617   // The intrinsic bailed out
 618   if (C->print_intrinsics() || C->print_inlining()) {
 619     if (jvms->has_method()) {
 620       // Not a root compile.
 621       const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
 622       C->print_inlining(callee, jvms->depth() - 1, bci, msg);
 623     } else {
 624       // Root compile
 625       tty->print("Did not generate intrinsic %s%s at bci:%d in",
 626                vmIntrinsics::name_at(intrinsic_id()),
 627                (is_virtual() ? " (virtual)" : ""), bci);
 628     }
 629   }
 630   C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
 631   return NULL;
 632 }
 633 
 634 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
 635   LibraryCallKit kit(jvms, this);
 636   Compile* C = kit.C;
 637   int nodes = C->unique();
 638 #ifndef PRODUCT
 639   assert(is_predicted(), "sanity");
 640   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
 641     char buf[1000];
 642     const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
 643     tty->print_cr("Predicate for intrinsic %s", str);
 644   }
 645 #endif
 646   ciMethod* callee = kit.callee();
 647   const int bci    = kit.bci();
 648 
 649   Node* slow_ctl = kit.try_to_predicate();
 650   if (!kit.failing()) {
 651     if (C->print_intrinsics() || C->print_inlining()) {
 652       C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
 653     }
 654     C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
 655     if (C->log()) {
 656       C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
 657                      vmIntrinsics::name_at(intrinsic_id()),
 658                      (is_virtual() ? " virtual='1'" : ""),
 659                      C->unique() - nodes);
 660     }
 661     return slow_ctl; // Could be NULL if the check folds.
 662   }
 663 
 664   // The intrinsic bailed out
 665   if (C->print_intrinsics() || C->print_inlining()) {
 666     if (jvms->has_method()) {
 667       // Not a root compile.
 668       const char* msg = "failed to generate predicate for intrinsic";
 669       C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
 670     } else {
 671       // Root compile
 672       C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
 673                                         vmIntrinsics::name_at(intrinsic_id()),
 674                                         (is_virtual() ? " (virtual)" : ""), bci);
 675     }
 676   }
 677   C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
 678   return NULL;
 679 }
 680 
 681 bool LibraryCallKit::try_to_inline() {
 682   // Handle symbolic names for otherwise undistinguished boolean switches:
 683   const bool is_store       = true;
 684   const bool is_native_ptr  = true;
 685   const bool is_static      = true;
 686   const bool is_volatile    = true;
 687 
 688   if (!jvms()->has_method()) {
 689     // Root JVMState has a null method.
 690     assert(map()->memory()->Opcode() == Op_Parm, "");
 691     // Insert the memory aliasing node
 692     set_all_memory(reset_memory());
 693   }
 694   assert(merged_memory(), "");
 695 
 696 
 697   switch (intrinsic_id()) {
 698   case vmIntrinsics::_hashCode:                 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
 699   case vmIntrinsics::_identityHashCode:         return inline_native_hashcode(/*!virtual*/ false,         is_static);
 700   case vmIntrinsics::_getClass:                 return inline_native_getClass();
 701 
 702   case vmIntrinsics::_dsin:
 703   case vmIntrinsics::_dcos:
 704   case vmIntrinsics::_dtan:
 705   case vmIntrinsics::_dabs:
 706   case vmIntrinsics::_datan2:
 707   case vmIntrinsics::_dsqrt:
 708   case vmIntrinsics::_dexp:
 709   case vmIntrinsics::_dlog:
 710   case vmIntrinsics::_dlog10:
 711   case vmIntrinsics::_dpow:                     return inline_math_native(intrinsic_id());
 712 
 713   case vmIntrinsics::_min:
 714   case vmIntrinsics::_max:                      return inline_min_max(intrinsic_id());
 715 
 716   case vmIntrinsics::_addExactI:                return inline_math_addExactI(false /* add */);
 717   case vmIntrinsics::_addExactL:                return inline_math_addExactL(false /* add */);
 718   case vmIntrinsics::_decrementExactI:          return inline_math_subtractExactI(true /* decrement */);
 719   case vmIntrinsics::_decrementExactL:          return inline_math_subtractExactL(true /* decrement */);
 720   case vmIntrinsics::_incrementExactI:          return inline_math_addExactI(true /* increment */);
 721   case vmIntrinsics::_incrementExactL:          return inline_math_addExactL(true /* increment */);
 722   case vmIntrinsics::_multiplyExactI:           return inline_math_multiplyExactI();
 723   case vmIntrinsics::_multiplyExactL:           return inline_math_multiplyExactL();
 724   case vmIntrinsics::_negateExactI:             return inline_math_negateExactI();
 725   case vmIntrinsics::_negateExactL:             return inline_math_negateExactL();
 726   case vmIntrinsics::_subtractExactI:           return inline_math_subtractExactI(false /* subtract */);
 727   case vmIntrinsics::_subtractExactL:           return inline_math_subtractExactL(false /* subtract */);
 728 
 729   case vmIntrinsics::_arraycopy:                return inline_arraycopy();
 730 
 731   case vmIntrinsics::_compareTo:                return inline_string_compareTo();
 732   case vmIntrinsics::_indexOf:                  return inline_string_indexOf();
 733   case vmIntrinsics::_equals:                   return inline_string_equals();
 734 
 735   case vmIntrinsics::_getObject:                return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT,  !is_volatile);
 736   case vmIntrinsics::_getBoolean:               return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
 737   case vmIntrinsics::_getByte:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE,    !is_volatile);
 738   case vmIntrinsics::_getShort:                 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,   !is_volatile);
 739   case vmIntrinsics::_getChar:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,    !is_volatile);
 740   case vmIntrinsics::_getInt:                   return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,     !is_volatile);
 741   case vmIntrinsics::_getLong:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,    !is_volatile);
 742   case vmIntrinsics::_getFloat:                 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT,   !is_volatile);
 743   case vmIntrinsics::_getDouble:                return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE,  !is_volatile);
 744 
 745   case vmIntrinsics::_putObject:                return inline_unsafe_access(!is_native_ptr,  is_store, T_OBJECT,  !is_volatile);
 746   case vmIntrinsics::_putBoolean:               return inline_unsafe_access(!is_native_ptr,  is_store, T_BOOLEAN, !is_volatile);
 747   case vmIntrinsics::_putByte:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_BYTE,    !is_volatile);
 748   case vmIntrinsics::_putShort:                 return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,   !is_volatile);
 749   case vmIntrinsics::_putChar:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,    !is_volatile);
 750   case vmIntrinsics::_putInt:                   return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,     !is_volatile);
 751   case vmIntrinsics::_putLong:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,    !is_volatile);
 752   case vmIntrinsics::_putFloat:                 return inline_unsafe_access(!is_native_ptr,  is_store, T_FLOAT,   !is_volatile);
 753   case vmIntrinsics::_putDouble:                return inline_unsafe_access(!is_native_ptr,  is_store, T_DOUBLE,  !is_volatile);
 754 
 755   case vmIntrinsics::_getByte_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE,    !is_volatile);
 756   case vmIntrinsics::_getShort_raw:             return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT,   !is_volatile);
 757   case vmIntrinsics::_getChar_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR,    !is_volatile);
 758   case vmIntrinsics::_getInt_raw:               return inline_unsafe_access( is_native_ptr, !is_store, T_INT,     !is_volatile);
 759   case vmIntrinsics::_getLong_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_LONG,    !is_volatile);
 760   case vmIntrinsics::_getFloat_raw:             return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT,   !is_volatile);
 761   case vmIntrinsics::_getDouble_raw:            return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE,  !is_volatile);
 762   case vmIntrinsics::_getAddress_raw:           return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
 763 
 764   case vmIntrinsics::_putByte_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_BYTE,    !is_volatile);
 765   case vmIntrinsics::_putShort_raw:             return inline_unsafe_access( is_native_ptr,  is_store, T_SHORT,   !is_volatile);
 766   case vmIntrinsics::_putChar_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_CHAR,    !is_volatile);
 767   case vmIntrinsics::_putInt_raw:               return inline_unsafe_access( is_native_ptr,  is_store, T_INT,     !is_volatile);
 768   case vmIntrinsics::_putLong_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_LONG,    !is_volatile);
 769   case vmIntrinsics::_putFloat_raw:             return inline_unsafe_access( is_native_ptr,  is_store, T_FLOAT,   !is_volatile);
 770   case vmIntrinsics::_putDouble_raw:            return inline_unsafe_access( is_native_ptr,  is_store, T_DOUBLE,  !is_volatile);
 771   case vmIntrinsics::_putAddress_raw:           return inline_unsafe_access( is_native_ptr,  is_store, T_ADDRESS, !is_volatile);
 772 
 773   case vmIntrinsics::_getObjectVolatile:        return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT,   is_volatile);
 774   case vmIntrinsics::_getBooleanVolatile:       return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN,  is_volatile);
 775   case vmIntrinsics::_getByteVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE,     is_volatile);
 776   case vmIntrinsics::_getShortVolatile:         return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,    is_volatile);
 777   case vmIntrinsics::_getCharVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,     is_volatile);
 778   case vmIntrinsics::_getIntVolatile:           return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,      is_volatile);
 779   case vmIntrinsics::_getLongVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,     is_volatile);
 780   case vmIntrinsics::_getFloatVolatile:         return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT,    is_volatile);
 781   case vmIntrinsics::_getDoubleVolatile:        return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE,   is_volatile);
 782 
 783   case vmIntrinsics::_putObjectVolatile:        return inline_unsafe_access(!is_native_ptr,  is_store, T_OBJECT,   is_volatile);
 784   case vmIntrinsics::_putBooleanVolatile:       return inline_unsafe_access(!is_native_ptr,  is_store, T_BOOLEAN,  is_volatile);
 785   case vmIntrinsics::_putByteVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_BYTE,     is_volatile);
 786   case vmIntrinsics::_putShortVolatile:         return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,    is_volatile);
 787   case vmIntrinsics::_putCharVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,     is_volatile);
 788   case vmIntrinsics::_putIntVolatile:           return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,      is_volatile);
 789   case vmIntrinsics::_putLongVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,     is_volatile);
 790   case vmIntrinsics::_putFloatVolatile:         return inline_unsafe_access(!is_native_ptr,  is_store, T_FLOAT,    is_volatile);
 791   case vmIntrinsics::_putDoubleVolatile:        return inline_unsafe_access(!is_native_ptr,  is_store, T_DOUBLE,   is_volatile);
 792 
 793   case vmIntrinsics::_prefetchRead:             return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
 794   case vmIntrinsics::_prefetchWrite:            return inline_unsafe_prefetch(!is_native_ptr,  is_store, !is_static);
 795   case vmIntrinsics::_prefetchReadStatic:       return inline_unsafe_prefetch(!is_native_ptr, !is_store,  is_static);
 796   case vmIntrinsics::_prefetchWriteStatic:      return inline_unsafe_prefetch(!is_native_ptr,  is_store,  is_static);
 797 
 798   case vmIntrinsics::_compareAndSwapObject:     return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
 799   case vmIntrinsics::_compareAndSwapInt:        return inline_unsafe_load_store(T_INT,    LS_cmpxchg);
 800   case vmIntrinsics::_compareAndSwapLong:       return inline_unsafe_load_store(T_LONG,   LS_cmpxchg);
 801 
 802   case vmIntrinsics::_putOrderedObject:         return inline_unsafe_ordered_store(T_OBJECT);
 803   case vmIntrinsics::_putOrderedInt:            return inline_unsafe_ordered_store(T_INT);
 804   case vmIntrinsics::_putOrderedLong:           return inline_unsafe_ordered_store(T_LONG);
 805 
 806   case vmIntrinsics::_getAndAddInt:             return inline_unsafe_load_store(T_INT,    LS_xadd);
 807   case vmIntrinsics::_getAndAddLong:            return inline_unsafe_load_store(T_LONG,   LS_xadd);
 808   case vmIntrinsics::_getAndSetInt:             return inline_unsafe_load_store(T_INT,    LS_xchg);
 809   case vmIntrinsics::_getAndSetLong:            return inline_unsafe_load_store(T_LONG,   LS_xchg);
 810   case vmIntrinsics::_getAndSetObject:          return inline_unsafe_load_store(T_OBJECT, LS_xchg);
 811 
 812   case vmIntrinsics::_loadFence:
 813   case vmIntrinsics::_storeFence:
 814   case vmIntrinsics::_fullFence:                return inline_unsafe_fence(intrinsic_id());
 815 
 816   case vmIntrinsics::_currentThread:            return inline_native_currentThread();
 817   case vmIntrinsics::_isInterrupted:            return inline_native_isInterrupted();
 818 
 819 #ifdef TRACE_HAVE_INTRINSICS
 820   case vmIntrinsics::_classID:                  return inline_native_classID();
 821   case vmIntrinsics::_threadID:                 return inline_native_threadID();
 822   case vmIntrinsics::_counterTime:              return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
 823 #endif
 824   case vmIntrinsics::_currentTimeMillis:        return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
 825   case vmIntrinsics::_nanoTime:                 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
 826   case vmIntrinsics::_allocateInstance:         return inline_unsafe_allocate();
 827   case vmIntrinsics::_copyMemory:               return inline_unsafe_copyMemory();
 828   case vmIntrinsics::_newArray:                 return inline_native_newArray();
 829   case vmIntrinsics::_getLength:                return inline_native_getLength();
 830   case vmIntrinsics::_copyOf:                   return inline_array_copyOf(false);
 831   case vmIntrinsics::_copyOfRange:              return inline_array_copyOf(true);
 832   case vmIntrinsics::_equalsC:                  return inline_array_equals();
 833   case vmIntrinsics::_clone:                    return inline_native_clone(intrinsic()->is_virtual());
 834 
 835   case vmIntrinsics::_isAssignableFrom:         return inline_native_subtype_check();
 836 
 837   case vmIntrinsics::_isInstance:
 838   case vmIntrinsics::_getModifiers:
 839   case vmIntrinsics::_isInterface:
 840   case vmIntrinsics::_isArray:
 841   case vmIntrinsics::_isPrimitive:
 842   case vmIntrinsics::_getSuperclass:
 843   case vmIntrinsics::_getComponentType:
 844   case vmIntrinsics::_getClassAccessFlags:      return inline_native_Class_query(intrinsic_id());
 845 
 846   case vmIntrinsics::_floatToRawIntBits:
 847   case vmIntrinsics::_floatToIntBits:
 848   case vmIntrinsics::_intBitsToFloat:
 849   case vmIntrinsics::_doubleToRawLongBits:
 850   case vmIntrinsics::_doubleToLongBits:
 851   case vmIntrinsics::_longBitsToDouble:         return inline_fp_conversions(intrinsic_id());
 852 
 853   case vmIntrinsics::_numberOfLeadingZeros_i:
 854   case vmIntrinsics::_numberOfLeadingZeros_l:
 855   case vmIntrinsics::_numberOfTrailingZeros_i:
 856   case vmIntrinsics::_numberOfTrailingZeros_l:
 857   case vmIntrinsics::_bitCount_i:
 858   case vmIntrinsics::_bitCount_l:
 859   case vmIntrinsics::_reverseBytes_i:
 860   case vmIntrinsics::_reverseBytes_l:
 861   case vmIntrinsics::_reverseBytes_s:
 862   case vmIntrinsics::_reverseBytes_c:           return inline_number_methods(intrinsic_id());
 863 
 864   case vmIntrinsics::_getCallerClass:           return inline_native_Reflection_getCallerClass();
 865 
 866   case vmIntrinsics::_Reference_get:            return inline_reference_get();
 867 
 868   case vmIntrinsics::_aescrypt_encryptBlock:
 869   case vmIntrinsics::_aescrypt_decryptBlock:    return inline_aescrypt_Block(intrinsic_id());
 870 
 871   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 872   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 873     return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
 874 
 875   case vmIntrinsics::_encodeISOArray:
 876     return inline_encodeISOArray();
 877 
 878   case vmIntrinsics::_updateCRC32:
 879     return inline_updateCRC32();
 880   case vmIntrinsics::_updateBytesCRC32:
 881     return inline_updateBytesCRC32();
 882   case vmIntrinsics::_updateByteBufferCRC32:
 883     return inline_updateByteBufferCRC32();
 884 
 885   default:
 886     // If you get here, it may be that someone has added a new intrinsic
 887     // to the list in vmSymbols.hpp without implementing it here.
 888 #ifndef PRODUCT
 889     if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
 890       tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
 891                     vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
 892     }
 893 #endif
 894     return false;
 895   }
 896 }
 897 
 898 Node* LibraryCallKit::try_to_predicate() {
 899   if (!jvms()->has_method()) {
 900     // Root JVMState has a null method.
 901     assert(map()->memory()->Opcode() == Op_Parm, "");
 902     // Insert the memory aliasing node
 903     set_all_memory(reset_memory());
 904   }
 905   assert(merged_memory(), "");
 906 
 907   switch (intrinsic_id()) {
 908   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 909     return inline_cipherBlockChaining_AESCrypt_predicate(false);
 910   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 911     return inline_cipherBlockChaining_AESCrypt_predicate(true);
 912 
 913   default:
 914     // If you get here, it may be that someone has added a new intrinsic
 915     // to the list in vmSymbols.hpp without implementing it here.
 916 #ifndef PRODUCT
 917     if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
 918       tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
 919                     vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
 920     }
 921 #endif
 922     Node* slow_ctl = control();
 923     set_control(top()); // No fast path instrinsic
 924     return slow_ctl;
 925   }
 926 }
 927 
 928 //------------------------------set_result-------------------------------
 929 // Helper function for finishing intrinsics.
 930 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
 931   record_for_igvn(region);
 932   set_control(_gvn.transform(region));
 933   set_result( _gvn.transform(value));
 934   assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
 935 }
 936 
 937 //------------------------------generate_guard---------------------------
 938 // Helper function for generating guarded fast-slow graph structures.
 939 // The given 'test', if true, guards a slow path.  If the test fails
 940 // then a fast path can be taken.  (We generally hope it fails.)
 941 // In all cases, GraphKit::control() is updated to the fast path.
 942 // The returned value represents the control for the slow path.
 943 // The return value is never 'top'; it is either a valid control
 944 // or NULL if it is obvious that the slow path can never be taken.
 945 // Also, if region and the slow control are not NULL, the slow edge
 946 // is appended to the region.
 947 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
 948   if (stopped()) {
 949     // Already short circuited.
 950     return NULL;
 951   }
 952 
 953   // Build an if node and its projections.
 954   // If test is true we take the slow path, which we assume is uncommon.
 955   if (_gvn.type(test) == TypeInt::ZERO) {
 956     // The slow branch is never taken.  No need to build this guard.
 957     return NULL;
 958   }
 959 
 960   IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
 961 
 962   Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
 963   if (if_slow == top()) {
 964     // The slow branch is never taken.  No need to build this guard.
 965     return NULL;
 966   }
 967 
 968   if (region != NULL)
 969     region->add_req(if_slow);
 970 
 971   Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
 972   set_control(if_fast);
 973 
 974   return if_slow;
 975 }
 976 
 977 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
 978   return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
 979 }
 980 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
 981   return generate_guard(test, region, PROB_FAIR);
 982 }
 983 
 984 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
 985                                                      Node* *pos_index) {
 986   if (stopped())
 987     return NULL;                // already stopped
 988   if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
 989     return NULL;                // index is already adequately typed
 990   Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
 991   Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
 992   Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
 993   if (is_neg != NULL && pos_index != NULL) {
 994     // Emulate effect of Parse::adjust_map_after_if.
 995     Node* ccast = new (C) CastIINode(index, TypeInt::POS);
 996     ccast->set_req(0, control());
 997     (*pos_index) = _gvn.transform(ccast);
 998   }
 999   return is_neg;
1000 }
1001 
1002 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1003                                                         Node* *pos_index) {
1004   if (stopped())
1005     return NULL;                // already stopped
1006   if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1007     return NULL;                // index is already adequately typed
1008   Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1009   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1010   Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1011   Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1012   if (is_notp != NULL && pos_index != NULL) {
1013     // Emulate effect of Parse::adjust_map_after_if.
1014     Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1015     ccast->set_req(0, control());
1016     (*pos_index) = _gvn.transform(ccast);
1017   }
1018   return is_notp;
1019 }
1020 
1021 // Make sure that 'position' is a valid limit index, in [0..length].
1022 // There are two equivalent plans for checking this:
1023 //   A. (offset + copyLength)  unsigned<=  arrayLength
1024 //   B. offset  <=  (arrayLength - copyLength)
1025 // We require that all of the values above, except for the sum and
1026 // difference, are already known to be non-negative.
1027 // Plan A is robust in the face of overflow, if offset and copyLength
1028 // are both hugely positive.
1029 //
1030 // Plan B is less direct and intuitive, but it does not overflow at
1031 // all, since the difference of two non-negatives is always
1032 // representable.  Whenever Java methods must perform the equivalent
1033 // check they generally use Plan B instead of Plan A.
1034 // For the moment we use Plan A.
1035 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1036                                                   Node* subseq_length,
1037                                                   Node* array_length,
1038                                                   RegionNode* region) {
1039   if (stopped())
1040     return NULL;                // already stopped
1041   bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1042   if (zero_offset && subseq_length->eqv_uncast(array_length))
1043     return NULL;                // common case of whole-array copy
1044   Node* last = subseq_length;
1045   if (!zero_offset)             // last += offset
1046     last = _gvn.transform(new (C) AddINode(last, offset));
1047   Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1048   Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1049   Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1050   return is_over;
1051 }
1052 
1053 
1054 //--------------------------generate_current_thread--------------------
1055 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1056   ciKlass*    thread_klass = env()->Thread_klass();
1057   const Type* thread_type  = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1058   Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1059   Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1060   Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
1061   tls_output = thread;
1062   return threadObj;
1063 }
1064 
1065 
1066 //------------------------------make_string_method_node------------------------
1067 // Helper method for String intrinsic functions. This version is called
1068 // with str1 and str2 pointing to String object nodes.
1069 //
1070 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1071   Node* no_ctrl = NULL;
1072 
1073   // Get start addr of string
1074   Node* str1_value   = load_String_value(no_ctrl, str1);
1075   Node* str1_offset  = load_String_offset(no_ctrl, str1);
1076   Node* str1_start   = array_element_address(str1_value, str1_offset, T_CHAR);
1077 
1078   // Get length of string 1
1079   Node* str1_len  = load_String_length(no_ctrl, str1);
1080 
1081   Node* str2_value   = load_String_value(no_ctrl, str2);
1082   Node* str2_offset  = load_String_offset(no_ctrl, str2);
1083   Node* str2_start   = array_element_address(str2_value, str2_offset, T_CHAR);
1084 
1085   Node* str2_len = NULL;
1086   Node* result = NULL;
1087 
1088   switch (opcode) {
1089   case Op_StrIndexOf:
1090     // Get length of string 2
1091     str2_len = load_String_length(no_ctrl, str2);
1092 
1093     result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1094                                  str1_start, str1_len, str2_start, str2_len);
1095     break;
1096   case Op_StrComp:
1097     // Get length of string 2
1098     str2_len = load_String_length(no_ctrl, str2);
1099 
1100     result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1101                                  str1_start, str1_len, str2_start, str2_len);
1102     break;
1103   case Op_StrEquals:
1104     result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1105                                str1_start, str2_start, str1_len);
1106     break;
1107   default:
1108     ShouldNotReachHere();
1109     return NULL;
1110   }
1111 
1112   // All these intrinsics have checks.
1113   C->set_has_split_ifs(true); // Has chance for split-if optimization
1114 
1115   return _gvn.transform(result);
1116 }
1117 
1118 // Helper method for String intrinsic functions. This version is called
1119 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1120 // to Int nodes containing the lenghts of str1 and str2.
1121 //
1122 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1123   Node* result = NULL;
1124   switch (opcode) {
1125   case Op_StrIndexOf:
1126     result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1127                                  str1_start, cnt1, str2_start, cnt2);
1128     break;
1129   case Op_StrComp:
1130     result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1131                                  str1_start, cnt1, str2_start, cnt2);
1132     break;
1133   case Op_StrEquals:
1134     result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1135                                  str1_start, str2_start, cnt1);
1136     break;
1137   default:
1138     ShouldNotReachHere();
1139     return NULL;
1140   }
1141 
1142   // All these intrinsics have checks.
1143   C->set_has_split_ifs(true); // Has chance for split-if optimization
1144 
1145   return _gvn.transform(result);
1146 }
1147 
1148 //------------------------------inline_string_compareTo------------------------
1149 // public int java.lang.String.compareTo(String anotherString);
1150 bool LibraryCallKit::inline_string_compareTo() {
1151   Node* receiver = null_check(argument(0));
1152   Node* arg      = null_check(argument(1));
1153   if (stopped()) {
1154     return true;
1155   }
1156   set_result(make_string_method_node(Op_StrComp, receiver, arg));
1157   return true;
1158 }
1159 
1160 //------------------------------inline_string_equals------------------------
1161 bool LibraryCallKit::inline_string_equals() {
1162   Node* receiver = null_check_receiver();
1163   // NOTE: Do not null check argument for String.equals() because spec
1164   // allows to specify NULL as argument.
1165   Node* argument = this->argument(1);
1166   if (stopped()) {
1167     return true;
1168   }
1169 
1170   // paths (plus control) merge
1171   RegionNode* region = new (C) RegionNode(5);
1172   Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1173 
1174   // does source == target string?
1175   Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1176   Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1177 
1178   Node* if_eq = generate_slow_guard(bol, NULL);
1179   if (if_eq != NULL) {
1180     // receiver == argument
1181     phi->init_req(2, intcon(1));
1182     region->init_req(2, if_eq);
1183   }
1184 
1185   // get String klass for instanceOf
1186   ciInstanceKlass* klass = env()->String_klass();
1187 
1188   if (!stopped()) {
1189     Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1190     Node* cmp  = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1191     Node* bol  = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1192 
1193     Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1194     //instanceOf == true, fallthrough
1195 
1196     if (inst_false != NULL) {
1197       phi->init_req(3, intcon(0));
1198       region->init_req(3, inst_false);
1199     }
1200   }
1201 
1202   if (!stopped()) {
1203     const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1204 
1205     // Properly cast the argument to String
1206     argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1207     // This path is taken only when argument's type is String:NotNull.
1208     argument = cast_not_null(argument, false);
1209 
1210     Node* no_ctrl = NULL;
1211 
1212     // Get start addr of receiver
1213     Node* receiver_val    = load_String_value(no_ctrl, receiver);
1214     Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1215     Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1216 
1217     // Get length of receiver
1218     Node* receiver_cnt  = load_String_length(no_ctrl, receiver);
1219 
1220     // Get start addr of argument
1221     Node* argument_val    = load_String_value(no_ctrl, argument);
1222     Node* argument_offset = load_String_offset(no_ctrl, argument);
1223     Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1224 
1225     // Get length of argument
1226     Node* argument_cnt  = load_String_length(no_ctrl, argument);
1227 
1228     // Check for receiver count != argument count
1229     Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1230     Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1231     Node* if_ne = generate_slow_guard(bol, NULL);
1232     if (if_ne != NULL) {
1233       phi->init_req(4, intcon(0));
1234       region->init_req(4, if_ne);
1235     }
1236 
1237     // Check for count == 0 is done by assembler code for StrEquals.
1238 
1239     if (!stopped()) {
1240       Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1241       phi->init_req(1, equals);
1242       region->init_req(1, control());
1243     }
1244   }
1245 
1246   // post merge
1247   set_control(_gvn.transform(region));
1248   record_for_igvn(region);
1249 
1250   set_result(_gvn.transform(phi));
1251   return true;
1252 }
1253 
1254 //------------------------------inline_array_equals----------------------------
1255 bool LibraryCallKit::inline_array_equals() {
1256   Node* arg1 = argument(0);
1257   Node* arg2 = argument(1);
1258   set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1259   return true;
1260 }
1261 
1262 // Java version of String.indexOf(constant string)
1263 // class StringDecl {
1264 //   StringDecl(char[] ca) {
1265 //     offset = 0;
1266 //     count = ca.length;
1267 //     value = ca;
1268 //   }
1269 //   int offset;
1270 //   int count;
1271 //   char[] value;
1272 // }
1273 //
1274 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1275 //                             int targetOffset, int cache_i, int md2) {
1276 //   int cache = cache_i;
1277 //   int sourceOffset = string_object.offset;
1278 //   int sourceCount = string_object.count;
1279 //   int targetCount = target_object.length;
1280 //
1281 //   int targetCountLess1 = targetCount - 1;
1282 //   int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1283 //
1284 //   char[] source = string_object.value;
1285 //   char[] target = target_object;
1286 //   int lastChar = target[targetCountLess1];
1287 //
1288 //  outer_loop:
1289 //   for (int i = sourceOffset; i < sourceEnd; ) {
1290 //     int src = source[i + targetCountLess1];
1291 //     if (src == lastChar) {
1292 //       // With random strings and a 4-character alphabet,
1293 //       // reverse matching at this point sets up 0.8% fewer
1294 //       // frames, but (paradoxically) makes 0.3% more probes.
1295 //       // Since those probes are nearer the lastChar probe,
1296 //       // there is may be a net D$ win with reverse matching.
1297 //       // But, reversing loop inhibits unroll of inner loop
1298 //       // for unknown reason.  So, does running outer loop from
1299 //       // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1300 //       for (int j = 0; j < targetCountLess1; j++) {
1301 //         if (target[targetOffset + j] != source[i+j]) {
1302 //           if ((cache & (1 << source[i+j])) == 0) {
1303 //             if (md2 < j+1) {
1304 //               i += j+1;
1305 //               continue outer_loop;
1306 //             }
1307 //           }
1308 //           i += md2;
1309 //           continue outer_loop;
1310 //         }
1311 //       }
1312 //       return i - sourceOffset;
1313 //     }
1314 //     if ((cache & (1 << src)) == 0) {
1315 //       i += targetCountLess1;
1316 //     } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1317 //     i++;
1318 //   }
1319 //   return -1;
1320 // }
1321 
1322 //------------------------------string_indexOf------------------------
1323 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1324                                      jint cache_i, jint md2_i) {
1325 
1326   Node* no_ctrl  = NULL;
1327   float likely   = PROB_LIKELY(0.9);
1328   float unlikely = PROB_UNLIKELY(0.9);
1329 
1330   const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1331 
1332   Node* source        = load_String_value(no_ctrl, string_object);
1333   Node* sourceOffset  = load_String_offset(no_ctrl, string_object);
1334   Node* sourceCount   = load_String_length(no_ctrl, string_object);
1335 
1336   Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1337   jint target_length = target_array->length();
1338   const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1339   const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1340 
1341   // String.value field is known to be @Stable.
1342   if (UseImplicitStableValues) {
1343     target = cast_array_to_stable(target, target_type);
1344   }
1345 
1346   IdealKit kit(this, false, true);
1347 #define __ kit.
1348   Node* zero             = __ ConI(0);
1349   Node* one              = __ ConI(1);
1350   Node* cache            = __ ConI(cache_i);
1351   Node* md2              = __ ConI(md2_i);
1352   Node* lastChar         = __ ConI(target_array->char_at(target_length - 1));
1353   Node* targetCount      = __ ConI(target_length);
1354   Node* targetCountLess1 = __ ConI(target_length - 1);
1355   Node* targetOffset     = __ ConI(targetOffset_i);
1356   Node* sourceEnd        = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1357 
1358   IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1359   Node* outer_loop = __ make_label(2 /* goto */);
1360   Node* return_    = __ make_label(1);
1361 
1362   __ set(rtn,__ ConI(-1));
1363   __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1364        Node* i2  = __ AddI(__ value(i), targetCountLess1);
1365        // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1366        Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1367        __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1368          __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1369               Node* tpj = __ AddI(targetOffset, __ value(j));
1370               Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1371               Node* ipj  = __ AddI(__ value(i), __ value(j));
1372               Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1373               __ if_then(targ, BoolTest::ne, src2); {
1374                 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1375                   __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1376                     __ increment(i, __ AddI(__ value(j), one));
1377                     __ goto_(outer_loop);
1378                   } __ end_if(); __ dead(j);
1379                 }__ end_if(); __ dead(j);
1380                 __ increment(i, md2);
1381                 __ goto_(outer_loop);
1382               }__ end_if();
1383               __ increment(j, one);
1384          }__ end_loop(); __ dead(j);
1385          __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1386          __ goto_(return_);
1387        }__ end_if();
1388        __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1389          __ increment(i, targetCountLess1);
1390        }__ end_if();
1391        __ increment(i, one);
1392        __ bind(outer_loop);
1393   }__ end_loop(); __ dead(i);
1394   __ bind(return_);
1395 
1396   // Final sync IdealKit and GraphKit.
1397   final_sync(kit);
1398   Node* result = __ value(rtn);
1399 #undef __
1400   C->set_has_loops(true);
1401   return result;
1402 }
1403 
1404 //------------------------------inline_string_indexOf------------------------
1405 bool LibraryCallKit::inline_string_indexOf() {
1406   Node* receiver = argument(0);
1407   Node* arg      = argument(1);
1408 
1409   Node* result;
1410   // Disable the use of pcmpestri until it can be guaranteed that
1411   // the load doesn't cross into the uncommited space.
1412   if (Matcher::has_match_rule(Op_StrIndexOf) &&
1413       UseSSE42Intrinsics) {
1414     // Generate SSE4.2 version of indexOf
1415     // We currently only have match rules that use SSE4.2
1416 
1417     receiver = null_check(receiver);
1418     arg      = null_check(arg);
1419     if (stopped()) {
1420       return true;
1421     }
1422 
1423     ciInstanceKlass* str_klass = env()->String_klass();
1424     const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1425 
1426     // Make the merge point
1427     RegionNode* result_rgn = new (C) RegionNode(4);
1428     Node*       result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1429     Node* no_ctrl  = NULL;
1430 
1431     // Get start addr of source string
1432     Node* source = load_String_value(no_ctrl, receiver);
1433     Node* source_offset = load_String_offset(no_ctrl, receiver);
1434     Node* source_start = array_element_address(source, source_offset, T_CHAR);
1435 
1436     // Get length of source string
1437     Node* source_cnt  = load_String_length(no_ctrl, receiver);
1438 
1439     // Get start addr of substring
1440     Node* substr = load_String_value(no_ctrl, arg);
1441     Node* substr_offset = load_String_offset(no_ctrl, arg);
1442     Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1443 
1444     // Get length of source string
1445     Node* substr_cnt  = load_String_length(no_ctrl, arg);
1446 
1447     // Check for substr count > string count
1448     Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1449     Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1450     Node* if_gt = generate_slow_guard(bol, NULL);
1451     if (if_gt != NULL) {
1452       result_phi->init_req(2, intcon(-1));
1453       result_rgn->init_req(2, if_gt);
1454     }
1455 
1456     if (!stopped()) {
1457       // Check for substr count == 0
1458       cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1459       bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1460       Node* if_zero = generate_slow_guard(bol, NULL);
1461       if (if_zero != NULL) {
1462         result_phi->init_req(3, intcon(0));
1463         result_rgn->init_req(3, if_zero);
1464       }
1465     }
1466 
1467     if (!stopped()) {
1468       result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1469       result_phi->init_req(1, result);
1470       result_rgn->init_req(1, control());
1471     }
1472     set_control(_gvn.transform(result_rgn));
1473     record_for_igvn(result_rgn);
1474     result = _gvn.transform(result_phi);
1475 
1476   } else { // Use LibraryCallKit::string_indexOf
1477     // don't intrinsify if argument isn't a constant string.
1478     if (!arg->is_Con()) {
1479      return false;
1480     }
1481     const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1482     if (str_type == NULL) {
1483       return false;
1484     }
1485     ciInstanceKlass* klass = env()->String_klass();
1486     ciObject* str_const = str_type->const_oop();
1487     if (str_const == NULL || str_const->klass() != klass) {
1488       return false;
1489     }
1490     ciInstance* str = str_const->as_instance();
1491     assert(str != NULL, "must be instance");
1492 
1493     ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1494     ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1495 
1496     int o;
1497     int c;
1498     if (java_lang_String::has_offset_field()) {
1499       o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1500       c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1501     } else {
1502       o = 0;
1503       c = pat->length();
1504     }
1505 
1506     // constant strings have no offset and count == length which
1507     // simplifies the resulting code somewhat so lets optimize for that.
1508     if (o != 0 || c != pat->length()) {
1509      return false;
1510     }
1511 
1512     receiver = null_check(receiver, T_OBJECT);
1513     // NOTE: No null check on the argument is needed since it's a constant String oop.
1514     if (stopped()) {
1515       return true;
1516     }
1517 
1518     // The null string as a pattern always returns 0 (match at beginning of string)
1519     if (c == 0) {
1520       set_result(intcon(0));
1521       return true;
1522     }
1523 
1524     // Generate default indexOf
1525     jchar lastChar = pat->char_at(o + (c - 1));
1526     int cache = 0;
1527     int i;
1528     for (i = 0; i < c - 1; i++) {
1529       assert(i < pat->length(), "out of range");
1530       cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1531     }
1532 
1533     int md2 = c;
1534     for (i = 0; i < c - 1; i++) {
1535       assert(i < pat->length(), "out of range");
1536       if (pat->char_at(o + i) == lastChar) {
1537         md2 = (c - 1) - i;
1538       }
1539     }
1540 
1541     result = string_indexOf(receiver, pat, o, cache, md2);
1542   }
1543   set_result(result);
1544   return true;
1545 }
1546 
1547 //--------------------------round_double_node--------------------------------
1548 // Round a double node if necessary.
1549 Node* LibraryCallKit::round_double_node(Node* n) {
1550   if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1551     n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1552   return n;
1553 }
1554 
1555 //------------------------------inline_math-----------------------------------
1556 // public static double Math.abs(double)
1557 // public static double Math.sqrt(double)
1558 // public static double Math.log(double)
1559 // public static double Math.log10(double)
1560 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1561   Node* arg = round_double_node(argument(0));
1562   Node* n;
1563   switch (id) {
1564   case vmIntrinsics::_dabs:   n = new (C) AbsDNode(                arg);  break;
1565   case vmIntrinsics::_dsqrt:  n = new (C) SqrtDNode(C, control(),  arg);  break;
1566   case vmIntrinsics::_dlog:   n = new (C) LogDNode(C, control(),   arg);  break;
1567   case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg);  break;
1568   default:  fatal_unexpected_iid(id);  break;
1569   }
1570   set_result(_gvn.transform(n));
1571   return true;
1572 }
1573 
1574 //------------------------------inline_trig----------------------------------
1575 // Inline sin/cos/tan instructions, if possible.  If rounding is required, do
1576 // argument reduction which will turn into a fast/slow diamond.
1577 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1578   Node* arg = round_double_node(argument(0));
1579   Node* n = NULL;
1580 
1581   switch (id) {
1582   case vmIntrinsics::_dsin:  n = new (C) SinDNode(C, control(), arg);  break;
1583   case vmIntrinsics::_dcos:  n = new (C) CosDNode(C, control(), arg);  break;
1584   case vmIntrinsics::_dtan:  n = new (C) TanDNode(C, control(), arg);  break;
1585   default:  fatal_unexpected_iid(id);  break;
1586   }
1587   n = _gvn.transform(n);
1588 
1589   // Rounding required?  Check for argument reduction!
1590   if (Matcher::strict_fp_requires_explicit_rounding) {
1591     static const double     pi_4 =  0.7853981633974483;
1592     static const double neg_pi_4 = -0.7853981633974483;
1593     // pi/2 in 80-bit extended precision
1594     // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1595     // -pi/2 in 80-bit extended precision
1596     // 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};
1597     // Cutoff value for using this argument reduction technique
1598     //static const double    pi_2_minus_epsilon =  1.564660403643354;
1599     //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1600 
1601     // Pseudocode for sin:
1602     // if (x <= Math.PI / 4.0) {
1603     //   if (x >= -Math.PI / 4.0) return  fsin(x);
1604     //   if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1605     // } else {
1606     //   if (x <=  Math.PI / 2.0) return  fcos(x - Math.PI / 2.0);
1607     // }
1608     // return StrictMath.sin(x);
1609 
1610     // Pseudocode for cos:
1611     // if (x <= Math.PI / 4.0) {
1612     //   if (x >= -Math.PI / 4.0) return  fcos(x);
1613     //   if (x >= -Math.PI / 2.0) return  fsin(x + Math.PI / 2.0);
1614     // } else {
1615     //   if (x <=  Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1616     // }
1617     // return StrictMath.cos(x);
1618 
1619     // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1620     // requires a special machine instruction to load it.  Instead we'll try
1621     // the 'easy' case.  If we really need the extra range +/- PI/2 we'll
1622     // probably do the math inside the SIN encoding.
1623 
1624     // Make the merge point
1625     RegionNode* r = new (C) RegionNode(3);
1626     Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1627 
1628     // Flatten arg so we need only 1 test
1629     Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1630     // Node for PI/4 constant
1631     Node *pi4 = makecon(TypeD::make(pi_4));
1632     // Check PI/4 : abs(arg)
1633     Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1634     // Check: If PI/4 < abs(arg) then go slow
1635     Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1636     // Branch either way
1637     IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1638     set_control(opt_iff(r,iff));
1639 
1640     // Set fast path result
1641     phi->init_req(2, n);
1642 
1643     // Slow path - non-blocking leaf call
1644     Node* call = NULL;
1645     switch (id) {
1646     case vmIntrinsics::_dsin:
1647       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1648                                CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1649                                "Sin", NULL, arg, top());
1650       break;
1651     case vmIntrinsics::_dcos:
1652       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1653                                CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1654                                "Cos", NULL, arg, top());
1655       break;
1656     case vmIntrinsics::_dtan:
1657       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1658                                CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1659                                "Tan", NULL, arg, top());
1660       break;
1661     }
1662     assert(control()->in(0) == call, "");
1663     Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1664     r->init_req(1, control());
1665     phi->init_req(1, slow_result);
1666 
1667     // Post-merge
1668     set_control(_gvn.transform(r));
1669     record_for_igvn(r);
1670     n = _gvn.transform(phi);
1671 
1672     C->set_has_split_ifs(true); // Has chance for split-if optimization
1673   }
1674   set_result(n);
1675   return true;
1676 }
1677 
1678 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1679   //-------------------
1680   //result=(result.isNaN())? funcAddr():result;
1681   // Check: If isNaN() by checking result!=result? then either trap
1682   // or go to runtime
1683   Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1684   // Build the boolean node
1685   Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1686 
1687   if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1688     { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1689       // The pow or exp intrinsic returned a NaN, which requires a call
1690       // to the runtime.  Recompile with the runtime call.
1691       uncommon_trap(Deoptimization::Reason_intrinsic,
1692                     Deoptimization::Action_make_not_entrant);
1693     }
1694     set_result(result);
1695   } else {
1696     // If this inlining ever returned NaN in the past, we compile a call
1697     // to the runtime to properly handle corner cases
1698 
1699     IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1700     Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1701     Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1702 
1703     if (!if_slow->is_top()) {
1704       RegionNode* result_region = new (C) RegionNode(3);
1705       PhiNode*    result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1706 
1707       result_region->init_req(1, if_fast);
1708       result_val->init_req(1, result);
1709 
1710       set_control(if_slow);
1711 
1712       const TypePtr* no_memory_effects = NULL;
1713       Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1714                                    no_memory_effects,
1715                                    x, top(), y, y ? top() : NULL);
1716       Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1717 #ifdef ASSERT
1718       Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1719       assert(value_top == top(), "second value must be top");
1720 #endif
1721 
1722       result_region->init_req(2, control());
1723       result_val->init_req(2, value);
1724       set_result(result_region, result_val);
1725     } else {
1726       set_result(result);
1727     }
1728   }
1729 }
1730 
1731 //------------------------------inline_exp-------------------------------------
1732 // Inline exp instructions, if possible.  The Intel hardware only misses
1733 // really odd corner cases (+/- Infinity).  Just uncommon-trap them.
1734 bool LibraryCallKit::inline_exp() {
1735   Node* arg = round_double_node(argument(0));
1736   Node* n   = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1737 
1738   finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1739 
1740   C->set_has_split_ifs(true); // Has chance for split-if optimization
1741   return true;
1742 }
1743 
1744 //------------------------------inline_pow-------------------------------------
1745 // Inline power instructions, if possible.
1746 bool LibraryCallKit::inline_pow() {
1747   // Pseudocode for pow
1748   // if (x <= 0.0) {
1749   //   long longy = (long)y;
1750   //   if ((double)longy == y) { // if y is long
1751   //     if (y + 1 == y) longy = 0; // huge number: even
1752   //     result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1753   //   } else {
1754   //     result = NaN;
1755   //   }
1756   // } else {
1757   //   result = DPow(x,y);
1758   // }
1759   // if (result != result)?  {
1760   //   result = uncommon_trap() or runtime_call();
1761   // }
1762   // return result;
1763 
1764   Node* x = round_double_node(argument(0));
1765   Node* y = round_double_node(argument(2));
1766 
1767   Node* result = NULL;
1768 
1769   if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1770     // Short form: skip the fancy tests and just check for NaN result.
1771     result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1772   } else {
1773     // If this inlining ever returned NaN in the past, include all
1774     // checks + call to the runtime.
1775 
1776     // Set the merge point for If node with condition of (x <= 0.0)
1777     // There are four possible paths to region node and phi node
1778     RegionNode *r = new (C) RegionNode(4);
1779     Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1780 
1781     // Build the first if node: if (x <= 0.0)
1782     // Node for 0 constant
1783     Node *zeronode = makecon(TypeD::ZERO);
1784     // Check x:0
1785     Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1786     // Check: If (x<=0) then go complex path
1787     Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1788     // Branch either way
1789     IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1790     // Fast path taken; set region slot 3
1791     Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1792     r->init_req(3,fast_taken); // Capture fast-control
1793 
1794     // Fast path not-taken, i.e. slow path
1795     Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1796 
1797     // Set fast path result
1798     Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1799     phi->init_req(3, fast_result);
1800 
1801     // Complex path
1802     // Build the second if node (if y is long)
1803     // Node for (long)y
1804     Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1805     // Node for (double)((long) y)
1806     Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1807     // Check (double)((long) y) : y
1808     Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1809     // Check if (y isn't long) then go to slow path
1810 
1811     Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1812     // Branch either way
1813     IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1814     Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1815 
1816     Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1817 
1818     // Calculate DPow(abs(x), y)*(1 & (long)y)
1819     // Node for constant 1
1820     Node *conone = longcon(1);
1821     // 1& (long)y
1822     Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1823 
1824     // A huge number is always even. Detect a huge number by checking
1825     // if y + 1 == y and set integer to be tested for parity to 0.
1826     // Required for corner case:
1827     // (long)9.223372036854776E18 = max_jlong
1828     // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1829     // max_jlong is odd but 9.223372036854776E18 is even
1830     Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1831     Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1832     Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1833     Node* correctedsign = NULL;
1834     if (ConditionalMoveLimit != 0) {
1835       correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1836     } else {
1837       IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1838       RegionNode *r = new (C) RegionNode(3);
1839       Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1840       r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1841       r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1842       phi->init_req(1, signnode);
1843       phi->init_req(2, longcon(0));
1844       correctedsign = _gvn.transform(phi);
1845       ylong_path = _gvn.transform(r);
1846       record_for_igvn(r);
1847     }
1848 
1849     // zero node
1850     Node *conzero = longcon(0);
1851     // Check (1&(long)y)==0?
1852     Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1853     // Check if (1&(long)y)!=0?, if so the result is negative
1854     Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1855     // abs(x)
1856     Node *absx=_gvn.transform(new (C) AbsDNode(x));
1857     // abs(x)^y
1858     Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1859     // -abs(x)^y
1860     Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1861     // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1862     Node *signresult = NULL;
1863     if (ConditionalMoveLimit != 0) {
1864       signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1865     } else {
1866       IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1867       RegionNode *r = new (C) RegionNode(3);
1868       Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1869       r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1870       r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1871       phi->init_req(1, absxpowy);
1872       phi->init_req(2, negabsxpowy);
1873       signresult = _gvn.transform(phi);
1874       ylong_path = _gvn.transform(r);
1875       record_for_igvn(r);
1876     }
1877     // Set complex path fast result
1878     r->init_req(2, ylong_path);
1879     phi->init_req(2, signresult);
1880 
1881     static const jlong nan_bits = CONST64(0x7ff8000000000000);
1882     Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1883     r->init_req(1,slow_path);
1884     phi->init_req(1,slow_result);
1885 
1886     // Post merge
1887     set_control(_gvn.transform(r));
1888     record_for_igvn(r);
1889     result = _gvn.transform(phi);
1890   }
1891 
1892   finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1893 
1894   C->set_has_split_ifs(true); // Has chance for split-if optimization
1895   return true;
1896 }
1897 
1898 //------------------------------runtime_math-----------------------------
1899 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1900   assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1901          "must be (DD)D or (D)D type");
1902 
1903   // Inputs
1904   Node* a = round_double_node(argument(0));
1905   Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1906 
1907   const TypePtr* no_memory_effects = NULL;
1908   Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1909                                  no_memory_effects,
1910                                  a, top(), b, b ? top() : NULL);
1911   Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1912 #ifdef ASSERT
1913   Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1914   assert(value_top == top(), "second value must be top");
1915 #endif
1916 
1917   set_result(value);
1918   return true;
1919 }
1920 
1921 //------------------------------inline_math_native-----------------------------
1922 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1923 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1924   switch (id) {
1925     // These intrinsics are not properly supported on all hardware
1926   case vmIntrinsics::_dcos:   return Matcher::has_match_rule(Op_CosD)   ? inline_trig(id) :
1927     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos),   "COS");
1928   case vmIntrinsics::_dsin:   return Matcher::has_match_rule(Op_SinD)   ? inline_trig(id) :
1929     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin),   "SIN");
1930   case vmIntrinsics::_dtan:   return Matcher::has_match_rule(Op_TanD)   ? inline_trig(id) :
1931     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan),   "TAN");
1932 
1933   case vmIntrinsics::_dlog:   return Matcher::has_match_rule(Op_LogD)   ? inline_math(id) :
1934     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog),   "LOG");
1935   case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1936     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1937 
1938     // These intrinsics are supported on all hardware
1939   case vmIntrinsics::_dsqrt:  return Matcher::has_match_rule(Op_SqrtD)  ? inline_math(id) : false;
1940   case vmIntrinsics::_dabs:   return Matcher::has_match_rule(Op_AbsD)   ? inline_math(id) : false;
1941 
1942   case vmIntrinsics::_dexp:   return Matcher::has_match_rule(Op_ExpD)   ? inline_exp()    :
1943     runtime_math(OptoRuntime::Math_D_D_Type(),  FN_PTR(SharedRuntime::dexp),  "EXP");
1944   case vmIntrinsics::_dpow:   return Matcher::has_match_rule(Op_PowD)   ? inline_pow()    :
1945     runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow),  "POW");
1946 #undef FN_PTR
1947 
1948    // These intrinsics are not yet correctly implemented
1949   case vmIntrinsics::_datan2:
1950     return false;
1951 
1952   default:
1953     fatal_unexpected_iid(id);
1954     return false;
1955   }
1956 }
1957 
1958 static bool is_simple_name(Node* n) {
1959   return (n->req() == 1         // constant
1960           || (n->is_Type() && n->as_Type()->type()->singleton())
1961           || n->is_Proj()       // parameter or return value
1962           || n->is_Phi()        // local of some sort
1963           );
1964 }
1965 
1966 //----------------------------inline_min_max-----------------------------------
1967 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1968   set_result(generate_min_max(id, argument(0), argument(1)));
1969   return true;
1970 }
1971 
1972 void LibraryCallKit::inline_math_mathExact(Node* math) {
1973   // If we didn't get the expected opcode it means we have optimized
1974   // the node to something else and don't need the exception edge.
1975   if (!math->is_MathExact()) {
1976     set_result(math);
1977     return;
1978   }
1979 
1980   Node* result = _gvn.transform( new(C) ProjNode(math, MathExactNode::result_proj_node));
1981   Node* flags = _gvn.transform( new(C) FlagsProjNode(math, MathExactNode::flags_proj_node));
1982 
1983   Node* bol = _gvn.transform( new (C) BoolNode(flags, BoolTest::overflow) );
1984   IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1985   Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
1986   Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
1987 
1988   {
1989     PreserveJVMState pjvms(this);
1990     PreserveReexecuteState preexecs(this);
1991     jvms()->set_should_reexecute(true);
1992 
1993     set_control(slow_path);
1994     set_i_o(i_o());
1995 
1996     uncommon_trap(Deoptimization::Reason_intrinsic,
1997                   Deoptimization::Action_none);
1998   }
1999 
2000   set_control(fast_path);
2001   set_result(result);
2002 }
2003 
2004 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2005   Node* arg1 = argument(0);
2006   Node* arg2 = NULL;
2007 
2008   if (is_increment) {
2009       arg2 = intcon(1);
2010   } else {
2011       arg2 = argument(1);
2012   }
2013 
2014   Node* add = _gvn.transform( new(C) AddExactINode(NULL, arg1, arg2) );
2015   inline_math_mathExact(add);
2016   return true;
2017 }
2018 
2019 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2020   Node* arg1 = argument(0); // type long
2021   // argument(1) == TOP
2022   Node* arg2 = NULL;
2023 
2024   if (is_increment) {
2025     arg2 = longcon(1);
2026   } else {
2027     arg2 = argument(2); // type long
2028     // argument(3) == TOP
2029   }
2030 
2031   Node* add = _gvn.transform(new(C) AddExactLNode(NULL, arg1, arg2));
2032   inline_math_mathExact(add);
2033   return true;
2034 }
2035 
2036 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2037   Node* arg1 = argument(0);
2038   Node* arg2 = NULL;
2039 
2040   if (is_decrement) {
2041     arg2 = intcon(1);
2042   } else {
2043     arg2 = argument(1);
2044   }
2045 
2046   Node* sub = _gvn.transform(new(C) SubExactINode(NULL, arg1, arg2));
2047   inline_math_mathExact(sub);
2048   return true;
2049 }
2050 
2051 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2052   Node* arg1 = argument(0); // type long
2053   // argument(1) == TOP
2054   Node* arg2 = NULL;
2055 
2056   if (is_decrement) {
2057     arg2 = longcon(1);
2058   } else {
2059     Node* arg2 = argument(2); // type long
2060     // argument(3) == TOP
2061   }
2062 
2063   Node* sub = _gvn.transform(new(C) SubExactLNode(NULL, arg1, arg2));
2064   inline_math_mathExact(sub);
2065   return true;
2066 }
2067 
2068 bool LibraryCallKit::inline_math_negateExactI() {
2069   Node* arg1 = argument(0);
2070 
2071   Node* neg = _gvn.transform(new(C) NegExactINode(NULL, arg1));
2072   inline_math_mathExact(neg);
2073   return true;
2074 }
2075 
2076 bool LibraryCallKit::inline_math_negateExactL() {
2077   Node* arg1 = argument(0);
2078   // argument(1) == TOP
2079 
2080   Node* neg = _gvn.transform(new(C) NegExactLNode(NULL, arg1));
2081   inline_math_mathExact(neg);
2082   return true;
2083 }
2084 
2085 bool LibraryCallKit::inline_math_multiplyExactI() {
2086   Node* arg1 = argument(0);
2087   Node* arg2 = argument(1);
2088 
2089   Node* mul = _gvn.transform(new(C) MulExactINode(NULL, arg1, arg2));
2090   inline_math_mathExact(mul);
2091   return true;
2092 }
2093 
2094 bool LibraryCallKit::inline_math_multiplyExactL() {
2095   Node* arg1 = argument(0);
2096   // argument(1) == TOP
2097   Node* arg2 = argument(2);
2098   // argument(3) == TOP
2099 
2100   Node* mul = _gvn.transform(new(C) MulExactLNode(NULL, arg1, arg2));
2101   inline_math_mathExact(mul);
2102   return true;
2103 }
2104 
2105 Node*
2106 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2107   // These are the candidate return value:
2108   Node* xvalue = x0;
2109   Node* yvalue = y0;
2110 
2111   if (xvalue == yvalue) {
2112     return xvalue;
2113   }
2114 
2115   bool want_max = (id == vmIntrinsics::_max);
2116 
2117   const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2118   const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2119   if (txvalue == NULL || tyvalue == NULL)  return top();
2120   // This is not really necessary, but it is consistent with a
2121   // hypothetical MaxINode::Value method:
2122   int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2123 
2124   // %%% This folding logic should (ideally) be in a different place.
2125   // Some should be inside IfNode, and there to be a more reliable
2126   // transformation of ?: style patterns into cmoves.  We also want
2127   // more powerful optimizations around cmove and min/max.
2128 
2129   // Try to find a dominating comparison of these guys.
2130   // It can simplify the index computation for Arrays.copyOf
2131   // and similar uses of System.arraycopy.
2132   // First, compute the normalized version of CmpI(x, y).
2133   int   cmp_op = Op_CmpI;
2134   Node* xkey = xvalue;
2135   Node* ykey = yvalue;
2136   Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2137   if (ideal_cmpxy->is_Cmp()) {
2138     // E.g., if we have CmpI(length - offset, count),
2139     // it might idealize to CmpI(length, count + offset)
2140     cmp_op = ideal_cmpxy->Opcode();
2141     xkey = ideal_cmpxy->in(1);
2142     ykey = ideal_cmpxy->in(2);
2143   }
2144 
2145   // Start by locating any relevant comparisons.
2146   Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2147   Node* cmpxy = NULL;
2148   Node* cmpyx = NULL;
2149   for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2150     Node* cmp = start_from->fast_out(k);
2151     if (cmp->outcnt() > 0 &&            // must have prior uses
2152         cmp->in(0) == NULL &&           // must be context-independent
2153         cmp->Opcode() == cmp_op) {      // right kind of compare
2154       if (cmp->in(1) == xkey && cmp->in(2) == ykey)  cmpxy = cmp;
2155       if (cmp->in(1) == ykey && cmp->in(2) == xkey)  cmpyx = cmp;
2156     }
2157   }
2158 
2159   const int NCMPS = 2;
2160   Node* cmps[NCMPS] = { cmpxy, cmpyx };
2161   int cmpn;
2162   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2163     if (cmps[cmpn] != NULL)  break;     // find a result
2164   }
2165   if (cmpn < NCMPS) {
2166     // Look for a dominating test that tells us the min and max.
2167     int depth = 0;                // Limit search depth for speed
2168     Node* dom = control();
2169     for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2170       if (++depth >= 100)  break;
2171       Node* ifproj = dom;
2172       if (!ifproj->is_Proj())  continue;
2173       Node* iff = ifproj->in(0);
2174       if (!iff->is_If())  continue;
2175       Node* bol = iff->in(1);
2176       if (!bol->is_Bool())  continue;
2177       Node* cmp = bol->in(1);
2178       if (cmp == NULL)  continue;
2179       for (cmpn = 0; cmpn < NCMPS; cmpn++)
2180         if (cmps[cmpn] == cmp)  break;
2181       if (cmpn == NCMPS)  continue;
2182       BoolTest::mask btest = bol->as_Bool()->_test._test;
2183       if (ifproj->is_IfFalse())  btest = BoolTest(btest).negate();
2184       if (cmp->in(1) == ykey)    btest = BoolTest(btest).commute();
2185       // At this point, we know that 'x btest y' is true.
2186       switch (btest) {
2187       case BoolTest::eq:
2188         // They are proven equal, so we can collapse the min/max.
2189         // Either value is the answer.  Choose the simpler.
2190         if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2191           return yvalue;
2192         return xvalue;
2193       case BoolTest::lt:          // x < y
2194       case BoolTest::le:          // x <= y
2195         return (want_max ? yvalue : xvalue);
2196       case BoolTest::gt:          // x > y
2197       case BoolTest::ge:          // x >= y
2198         return (want_max ? xvalue : yvalue);
2199       }
2200     }
2201   }
2202 
2203   // We failed to find a dominating test.
2204   // Let's pick a test that might GVN with prior tests.
2205   Node*          best_bol   = NULL;
2206   BoolTest::mask best_btest = BoolTest::illegal;
2207   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2208     Node* cmp = cmps[cmpn];
2209     if (cmp == NULL)  continue;
2210     for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2211       Node* bol = cmp->fast_out(j);
2212       if (!bol->is_Bool())  continue;
2213       BoolTest::mask btest = bol->as_Bool()->_test._test;
2214       if (btest == BoolTest::eq || btest == BoolTest::ne)  continue;
2215       if (cmp->in(1) == ykey)   btest = BoolTest(btest).commute();
2216       if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2217         best_bol   = bol->as_Bool();
2218         best_btest = btest;
2219       }
2220     }
2221   }
2222 
2223   Node* answer_if_true  = NULL;
2224   Node* answer_if_false = NULL;
2225   switch (best_btest) {
2226   default:
2227     if (cmpxy == NULL)
2228       cmpxy = ideal_cmpxy;
2229     best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2230     // and fall through:
2231   case BoolTest::lt:          // x < y
2232   case BoolTest::le:          // x <= y
2233     answer_if_true  = (want_max ? yvalue : xvalue);
2234     answer_if_false = (want_max ? xvalue : yvalue);
2235     break;
2236   case BoolTest::gt:          // x > y
2237   case BoolTest::ge:          // x >= y
2238     answer_if_true  = (want_max ? xvalue : yvalue);
2239     answer_if_false = (want_max ? yvalue : xvalue);
2240     break;
2241   }
2242 
2243   jint hi, lo;
2244   if (want_max) {
2245     // We can sharpen the minimum.
2246     hi = MAX2(txvalue->_hi, tyvalue->_hi);
2247     lo = MAX2(txvalue->_lo, tyvalue->_lo);
2248   } else {
2249     // We can sharpen the maximum.
2250     hi = MIN2(txvalue->_hi, tyvalue->_hi);
2251     lo = MIN2(txvalue->_lo, tyvalue->_lo);
2252   }
2253 
2254   // Use a flow-free graph structure, to avoid creating excess control edges
2255   // which could hinder other optimizations.
2256   // Since Math.min/max is often used with arraycopy, we want
2257   // tightly_coupled_allocation to be able to see beyond min/max expressions.
2258   Node* cmov = CMoveNode::make(C, NULL, best_bol,
2259                                answer_if_false, answer_if_true,
2260                                TypeInt::make(lo, hi, widen));
2261 
2262   return _gvn.transform(cmov);
2263 
2264   /*
2265   // This is not as desirable as it may seem, since Min and Max
2266   // nodes do not have a full set of optimizations.
2267   // And they would interfere, anyway, with 'if' optimizations
2268   // and with CMoveI canonical forms.
2269   switch (id) {
2270   case vmIntrinsics::_min:
2271     result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2272   case vmIntrinsics::_max:
2273     result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2274   default:
2275     ShouldNotReachHere();
2276   }
2277   */
2278 }
2279 
2280 inline int
2281 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2282   const TypePtr* base_type = TypePtr::NULL_PTR;
2283   if (base != NULL)  base_type = _gvn.type(base)->isa_ptr();
2284   if (base_type == NULL) {
2285     // Unknown type.
2286     return Type::AnyPtr;
2287   } else if (base_type == TypePtr::NULL_PTR) {
2288     // Since this is a NULL+long form, we have to switch to a rawptr.
2289     base   = _gvn.transform(new (C) CastX2PNode(offset));
2290     offset = MakeConX(0);
2291     return Type::RawPtr;
2292   } else if (base_type->base() == Type::RawPtr) {
2293     return Type::RawPtr;
2294   } else if (base_type->isa_oopptr()) {
2295     // Base is never null => always a heap address.
2296     if (base_type->ptr() == TypePtr::NotNull) {
2297       return Type::OopPtr;
2298     }
2299     // Offset is small => always a heap address.
2300     const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2301     if (offset_type != NULL &&
2302         base_type->offset() == 0 &&     // (should always be?)
2303         offset_type->_lo >= 0 &&
2304         !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2305       return Type::OopPtr;
2306     }
2307     // Otherwise, it might either be oop+off or NULL+addr.
2308     return Type::AnyPtr;
2309   } else {
2310     // No information:
2311     return Type::AnyPtr;
2312   }
2313 }
2314 
2315 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2316   int kind = classify_unsafe_addr(base, offset);
2317   if (kind == Type::RawPtr) {
2318     return basic_plus_adr(top(), base, offset);
2319   } else {
2320     return basic_plus_adr(base, offset);
2321   }
2322 }
2323 
2324 //--------------------------inline_number_methods-----------------------------
2325 // inline int     Integer.numberOfLeadingZeros(int)
2326 // inline int        Long.numberOfLeadingZeros(long)
2327 //
2328 // inline int     Integer.numberOfTrailingZeros(int)
2329 // inline int        Long.numberOfTrailingZeros(long)
2330 //
2331 // inline int     Integer.bitCount(int)
2332 // inline int        Long.bitCount(long)
2333 //
2334 // inline char  Character.reverseBytes(char)
2335 // inline short     Short.reverseBytes(short)
2336 // inline int     Integer.reverseBytes(int)
2337 // inline long       Long.reverseBytes(long)
2338 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2339   Node* arg = argument(0);
2340   Node* n;
2341   switch (id) {
2342   case vmIntrinsics::_numberOfLeadingZeros_i:   n = new (C) CountLeadingZerosINode( arg);  break;
2343   case vmIntrinsics::_numberOfLeadingZeros_l:   n = new (C) CountLeadingZerosLNode( arg);  break;
2344   case vmIntrinsics::_numberOfTrailingZeros_i:  n = new (C) CountTrailingZerosINode(arg);  break;
2345   case vmIntrinsics::_numberOfTrailingZeros_l:  n = new (C) CountTrailingZerosLNode(arg);  break;
2346   case vmIntrinsics::_bitCount_i:               n = new (C) PopCountINode(          arg);  break;
2347   case vmIntrinsics::_bitCount_l:               n = new (C) PopCountLNode(          arg);  break;
2348   case vmIntrinsics::_reverseBytes_c:           n = new (C) ReverseBytesUSNode(0,   arg);  break;
2349   case vmIntrinsics::_reverseBytes_s:           n = new (C) ReverseBytesSNode( 0,   arg);  break;
2350   case vmIntrinsics::_reverseBytes_i:           n = new (C) ReverseBytesINode( 0,   arg);  break;
2351   case vmIntrinsics::_reverseBytes_l:           n = new (C) ReverseBytesLNode( 0,   arg);  break;
2352   default:  fatal_unexpected_iid(id);  break;
2353   }
2354   set_result(_gvn.transform(n));
2355   return true;
2356 }
2357 
2358 //----------------------------inline_unsafe_access----------------------------
2359 
2360 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2361 
2362 // Helper that guards and inserts a pre-barrier.
2363 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2364                                         Node* pre_val, bool need_mem_bar) {
2365   // We could be accessing the referent field of a reference object. If so, when G1
2366   // is enabled, we need to log the value in the referent field in an SATB buffer.
2367   // This routine performs some compile time filters and generates suitable
2368   // runtime filters that guard the pre-barrier code.
2369   // Also add memory barrier for non volatile load from the referent field
2370   // to prevent commoning of loads across safepoint.
2371   if (!UseG1GC && !need_mem_bar)
2372     return;
2373 
2374   // Some compile time checks.
2375 
2376   // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2377   const TypeX* otype = offset->find_intptr_t_type();
2378   if (otype != NULL && otype->is_con() &&
2379       otype->get_con() != java_lang_ref_Reference::referent_offset) {
2380     // Constant offset but not the reference_offset so just return
2381     return;
2382   }
2383 
2384   // We only need to generate the runtime guards for instances.
2385   const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2386   if (btype != NULL) {
2387     if (btype->isa_aryptr()) {
2388       // Array type so nothing to do
2389       return;
2390     }
2391 
2392     const TypeInstPtr* itype = btype->isa_instptr();
2393     if (itype != NULL) {
2394       // Can the klass of base_oop be statically determined to be
2395       // _not_ a sub-class of Reference and _not_ Object?
2396       ciKlass* klass = itype->klass();
2397       if ( klass->is_loaded() &&
2398           !klass->is_subtype_of(env()->Reference_klass()) &&
2399           !env()->Object_klass()->is_subtype_of(klass)) {
2400         return;
2401       }
2402     }
2403   }
2404 
2405   // The compile time filters did not reject base_oop/offset so
2406   // we need to generate the following runtime filters
2407   //
2408   // if (offset == java_lang_ref_Reference::_reference_offset) {
2409   //   if (instance_of(base, java.lang.ref.Reference)) {
2410   //     pre_barrier(_, pre_val, ...);
2411   //   }
2412   // }
2413 
2414   float likely   = PROB_LIKELY(  0.999);
2415   float unlikely = PROB_UNLIKELY(0.999);
2416 
2417   IdealKit ideal(this);
2418 #define __ ideal.
2419 
2420   Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2421 
2422   __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2423       // Update graphKit memory and control from IdealKit.
2424       sync_kit(ideal);
2425 
2426       Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2427       Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2428 
2429       // Update IdealKit memory and control from graphKit.
2430       __ sync_kit(this);
2431 
2432       Node* one = __ ConI(1);
2433       // is_instof == 0 if base_oop == NULL
2434       __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2435 
2436         // Update graphKit from IdeakKit.
2437         sync_kit(ideal);
2438 
2439         // Use the pre-barrier to record the value in the referent field
2440         pre_barrier(false /* do_load */,
2441                     __ ctrl(),
2442                     NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2443                     pre_val /* pre_val */,
2444                     T_OBJECT);
2445         if (need_mem_bar) {
2446           // Add memory barrier to prevent commoning reads from this field
2447           // across safepoint since GC can change its value.
2448           insert_mem_bar(Op_MemBarCPUOrder);
2449         }
2450         // Update IdealKit from graphKit.
2451         __ sync_kit(this);
2452 
2453       } __ end_if(); // _ref_type != ref_none
2454   } __ end_if(); // offset == referent_offset
2455 
2456   // Final sync IdealKit and GraphKit.
2457   final_sync(ideal);
2458 #undef __
2459 }
2460 
2461 
2462 // Interpret Unsafe.fieldOffset cookies correctly:
2463 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2464 
2465 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2466   // Attempt to infer a sharper value type from the offset and base type.
2467   ciKlass* sharpened_klass = NULL;
2468 
2469   // See if it is an instance field, with an object type.
2470   if (alias_type->field() != NULL) {
2471     assert(!is_native_ptr, "native pointer op cannot use a java address");
2472     if (alias_type->field()->type()->is_klass()) {
2473       sharpened_klass = alias_type->field()->type()->as_klass();
2474     }
2475   }
2476 
2477   // See if it is a narrow oop array.
2478   if (adr_type->isa_aryptr()) {
2479     if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2480       const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2481       if (elem_type != NULL) {
2482         sharpened_klass = elem_type->klass();
2483       }
2484     }
2485   }
2486 
2487   // The sharpened class might be unloaded if there is no class loader
2488   // contraint in place.
2489   if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2490     const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2491 
2492 #ifndef PRODUCT
2493     if (C->print_intrinsics() || C->print_inlining()) {
2494       tty->print("  from base type: ");  adr_type->dump();
2495       tty->print("  sharpened value: ");  tjp->dump();
2496     }
2497 #endif
2498     // Sharpen the value type.
2499     return tjp;
2500   }
2501   return NULL;
2502 }
2503 
2504 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2505   if (callee()->is_static())  return false;  // caller must have the capability!
2506 
2507 #ifndef PRODUCT
2508   {
2509     ResourceMark rm;
2510     // Check the signatures.
2511     ciSignature* sig = callee()->signature();
2512 #ifdef ASSERT
2513     if (!is_store) {
2514       // Object getObject(Object base, int/long offset), etc.
2515       BasicType rtype = sig->return_type()->basic_type();
2516       if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2517           rtype = T_ADDRESS;  // it is really a C void*
2518       assert(rtype == type, "getter must return the expected value");
2519       if (!is_native_ptr) {
2520         assert(sig->count() == 2, "oop getter has 2 arguments");
2521         assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2522         assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2523       } else {
2524         assert(sig->count() == 1, "native getter has 1 argument");
2525         assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2526       }
2527     } else {
2528       // void putObject(Object base, int/long offset, Object x), etc.
2529       assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2530       if (!is_native_ptr) {
2531         assert(sig->count() == 3, "oop putter has 3 arguments");
2532         assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2533         assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2534       } else {
2535         assert(sig->count() == 2, "native putter has 2 arguments");
2536         assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2537       }
2538       BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2539       if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2540         vtype = T_ADDRESS;  // it is really a C void*
2541       assert(vtype == type, "putter must accept the expected value");
2542     }
2543 #endif // ASSERT
2544  }
2545 #endif //PRODUCT
2546 
2547   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2548 
2549   Node* receiver = argument(0);  // type: oop
2550 
2551   // Build address expression.  See the code in inline_unsafe_prefetch.
2552   Node* adr;
2553   Node* heap_base_oop = top();
2554   Node* offset = top();
2555   Node* val;
2556 
2557   if (!is_native_ptr) {
2558     // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2559     Node* base = argument(1);  // type: oop
2560     // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2561     offset = argument(2);  // type: long
2562     // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2563     // to be plain byte offsets, which are also the same as those accepted
2564     // by oopDesc::field_base.
2565     assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2566            "fieldOffset must be byte-scaled");
2567     // 32-bit machines ignore the high half!
2568     offset = ConvL2X(offset);
2569     adr = make_unsafe_address(base, offset);
2570     heap_base_oop = base;
2571     val = is_store ? argument(4) : NULL;
2572   } else {
2573     Node* ptr = argument(1);  // type: long
2574     ptr = ConvL2X(ptr);  // adjust Java long to machine word
2575     adr = make_unsafe_address(NULL, ptr);
2576     val = is_store ? argument(3) : NULL;
2577   }
2578 
2579   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2580 
2581   // First guess at the value type.
2582   const Type *value_type = Type::get_const_basic_type(type);
2583 
2584   // Try to categorize the address.  If it comes up as TypeJavaPtr::BOTTOM,
2585   // there was not enough information to nail it down.
2586   Compile::AliasType* alias_type = C->alias_type(adr_type);
2587   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2588 
2589   // We will need memory barriers unless we can determine a unique
2590   // alias category for this reference.  (Note:  If for some reason
2591   // the barriers get omitted and the unsafe reference begins to "pollute"
2592   // the alias analysis of the rest of the graph, either Compile::can_alias
2593   // or Compile::must_alias will throw a diagnostic assert.)
2594   bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2595 
2596   // If we are reading the value of the referent field of a Reference
2597   // object (either by using Unsafe directly or through reflection)
2598   // then, if G1 is enabled, we need to record the referent in an
2599   // SATB log buffer using the pre-barrier mechanism.
2600   // Also we need to add memory barrier to prevent commoning reads
2601   // from this field across safepoint since GC can change its value.
2602   bool need_read_barrier = !is_native_ptr && !is_store &&
2603                            offset != top() && heap_base_oop != top();
2604 
2605   if (!is_store && type == T_OBJECT) {
2606     const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2607     if (tjp != NULL) {
2608       value_type = tjp;
2609     }
2610   }
2611 
2612   receiver = null_check(receiver);
2613   if (stopped()) {
2614     return true;
2615   }
2616   // Heap pointers get a null-check from the interpreter,
2617   // as a courtesy.  However, this is not guaranteed by Unsafe,
2618   // and it is not possible to fully distinguish unintended nulls
2619   // from intended ones in this API.
2620 
2621   if (is_volatile) {
2622     // We need to emit leading and trailing CPU membars (see below) in
2623     // addition to memory membars when is_volatile. This is a little
2624     // too strong, but avoids the need to insert per-alias-type
2625     // volatile membars (for stores; compare Parse::do_put_xxx), which
2626     // we cannot do effectively here because we probably only have a
2627     // rough approximation of type.
2628     need_mem_bar = true;
2629     // For Stores, place a memory ordering barrier now.
2630     if (is_store)
2631       insert_mem_bar(Op_MemBarRelease);
2632   }
2633 
2634   // Memory barrier to prevent normal and 'unsafe' accesses from
2635   // bypassing each other.  Happens after null checks, so the
2636   // exception paths do not take memory state from the memory barrier,
2637   // so there's no problems making a strong assert about mixing users
2638   // of safe & unsafe memory.  Otherwise fails in a CTW of rt.jar
2639   // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2640   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2641 
2642   if (!is_store) {
2643     Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2644     // load value
2645     switch (type) {
2646     case T_BOOLEAN:
2647     case T_CHAR:
2648     case T_BYTE:
2649     case T_SHORT:
2650     case T_INT:
2651     case T_LONG:
2652     case T_FLOAT:
2653     case T_DOUBLE:
2654       break;
2655     case T_OBJECT:
2656       if (need_read_barrier) {
2657         insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2658       }
2659       break;
2660     case T_ADDRESS:
2661       // Cast to an int type.
2662       p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2663       p = ConvX2L(p);
2664       break;
2665     default:
2666       fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2667       break;
2668     }
2669     // The load node has the control of the preceding MemBarCPUOrder.  All
2670     // following nodes will have the control of the MemBarCPUOrder inserted at
2671     // the end of this method.  So, pushing the load onto the stack at a later
2672     // point is fine.
2673     set_result(p);
2674   } else {
2675     // place effect of store into memory
2676     switch (type) {
2677     case T_DOUBLE:
2678       val = dstore_rounding(val);
2679       break;
2680     case T_ADDRESS:
2681       // Repackage the long as a pointer.
2682       val = ConvL2X(val);
2683       val = _gvn.transform(new (C) CastX2PNode(val));
2684       break;
2685     }
2686 
2687     if (type != T_OBJECT ) {
2688       (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2689     } else {
2690       // Possibly an oop being stored to Java heap or native memory
2691       if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2692         // oop to Java heap.
2693         (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2694       } else {
2695         // We can't tell at compile time if we are storing in the Java heap or outside
2696         // of it. So we need to emit code to conditionally do the proper type of
2697         // store.
2698 
2699         IdealKit ideal(this);
2700 #define __ ideal.
2701         // QQQ who knows what probability is here??
2702         __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2703           // Sync IdealKit and graphKit.
2704           sync_kit(ideal);
2705           Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2706           // Update IdealKit memory.
2707           __ sync_kit(this);
2708         } __ else_(); {
2709           __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2710         } __ end_if();
2711         // Final sync IdealKit and GraphKit.
2712         final_sync(ideal);
2713 #undef __
2714       }
2715     }
2716   }
2717 
2718   if (is_volatile) {
2719     if (!is_store)
2720       insert_mem_bar(Op_MemBarAcquire);
2721     else
2722       insert_mem_bar(Op_MemBarVolatile);
2723   }
2724 
2725   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2726 
2727   return true;
2728 }
2729 
2730 //----------------------------inline_unsafe_prefetch----------------------------
2731 
2732 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2733 #ifndef PRODUCT
2734   {
2735     ResourceMark rm;
2736     // Check the signatures.
2737     ciSignature* sig = callee()->signature();
2738 #ifdef ASSERT
2739     // Object getObject(Object base, int/long offset), etc.
2740     BasicType rtype = sig->return_type()->basic_type();
2741     if (!is_native_ptr) {
2742       assert(sig->count() == 2, "oop prefetch has 2 arguments");
2743       assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2744       assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2745     } else {
2746       assert(sig->count() == 1, "native prefetch has 1 argument");
2747       assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2748     }
2749 #endif // ASSERT
2750   }
2751 #endif // !PRODUCT
2752 
2753   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2754 
2755   const int idx = is_static ? 0 : 1;
2756   if (!is_static) {
2757     null_check_receiver();
2758     if (stopped()) {
2759       return true;
2760     }
2761   }
2762 
2763   // Build address expression.  See the code in inline_unsafe_access.
2764   Node *adr;
2765   if (!is_native_ptr) {
2766     // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2767     Node* base   = argument(idx + 0);  // type: oop
2768     // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2769     Node* offset = argument(idx + 1);  // type: long
2770     // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2771     // to be plain byte offsets, which are also the same as those accepted
2772     // by oopDesc::field_base.
2773     assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2774            "fieldOffset must be byte-scaled");
2775     // 32-bit machines ignore the high half!
2776     offset = ConvL2X(offset);
2777     adr = make_unsafe_address(base, offset);
2778   } else {
2779     Node* ptr = argument(idx + 0);  // type: long
2780     ptr = ConvL2X(ptr);  // adjust Java long to machine word
2781     adr = make_unsafe_address(NULL, ptr);
2782   }
2783 
2784   // Generate the read or write prefetch
2785   Node *prefetch;
2786   if (is_store) {
2787     prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2788   } else {
2789     prefetch = new (C) PrefetchReadNode(i_o(), adr);
2790   }
2791   prefetch->init_req(0, control());
2792   set_i_o(_gvn.transform(prefetch));
2793 
2794   return true;
2795 }
2796 
2797 //----------------------------inline_unsafe_load_store----------------------------
2798 // This method serves a couple of different customers (depending on LoadStoreKind):
2799 //
2800 // LS_cmpxchg:
2801 //   public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2802 //   public final native boolean compareAndSwapInt(   Object o, long offset, int    expected, int    x);
2803 //   public final native boolean compareAndSwapLong(  Object o, long offset, long   expected, long   x);
2804 //
2805 // LS_xadd:
2806 //   public int  getAndAddInt( Object o, long offset, int  delta)
2807 //   public long getAndAddLong(Object o, long offset, long delta)
2808 //
2809 // LS_xchg:
2810 //   int    getAndSet(Object o, long offset, int    newValue)
2811 //   long   getAndSet(Object o, long offset, long   newValue)
2812 //   Object getAndSet(Object o, long offset, Object newValue)
2813 //
2814 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2815   // This basic scheme here is the same as inline_unsafe_access, but
2816   // differs in enough details that combining them would make the code
2817   // overly confusing.  (This is a true fact! I originally combined
2818   // them, but even I was confused by it!) As much code/comments as
2819   // possible are retained from inline_unsafe_access though to make
2820   // the correspondences clearer. - dl
2821 
2822   if (callee()->is_static())  return false;  // caller must have the capability!
2823 
2824 #ifndef PRODUCT
2825   BasicType rtype;
2826   {
2827     ResourceMark rm;
2828     // Check the signatures.
2829     ciSignature* sig = callee()->signature();
2830     rtype = sig->return_type()->basic_type();
2831     if (kind == LS_xadd || kind == LS_xchg) {
2832       // Check the signatures.
2833 #ifdef ASSERT
2834       assert(rtype == type, "get and set must return the expected type");
2835       assert(sig->count() == 3, "get and set has 3 arguments");
2836       assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2837       assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2838       assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2839 #endif // ASSERT
2840     } else if (kind == LS_cmpxchg) {
2841       // Check the signatures.
2842 #ifdef ASSERT
2843       assert(rtype == T_BOOLEAN, "CAS must return boolean");
2844       assert(sig->count() == 4, "CAS has 4 arguments");
2845       assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2846       assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2847 #endif // ASSERT
2848     } else {
2849       ShouldNotReachHere();
2850     }
2851   }
2852 #endif //PRODUCT
2853 
2854   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2855 
2856   // Get arguments:
2857   Node* receiver = NULL;
2858   Node* base     = NULL;
2859   Node* offset   = NULL;
2860   Node* oldval   = NULL;
2861   Node* newval   = NULL;
2862   if (kind == LS_cmpxchg) {
2863     const bool two_slot_type = type2size[type] == 2;
2864     receiver = argument(0);  // type: oop
2865     base     = argument(1);  // type: oop
2866     offset   = argument(2);  // type: long
2867     oldval   = argument(4);  // type: oop, int, or long
2868     newval   = argument(two_slot_type ? 6 : 5);  // type: oop, int, or long
2869   } else if (kind == LS_xadd || kind == LS_xchg){
2870     receiver = argument(0);  // type: oop
2871     base     = argument(1);  // type: oop
2872     offset   = argument(2);  // type: long
2873     oldval   = NULL;
2874     newval   = argument(4);  // type: oop, int, or long
2875   }
2876 
2877   // Null check receiver.
2878   receiver = null_check(receiver);
2879   if (stopped()) {
2880     return true;
2881   }
2882 
2883   // Build field offset expression.
2884   // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2885   // to be plain byte offsets, which are also the same as those accepted
2886   // by oopDesc::field_base.
2887   assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2888   // 32-bit machines ignore the high half of long offsets
2889   offset = ConvL2X(offset);
2890   Node* adr = make_unsafe_address(base, offset);
2891   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2892 
2893   // For CAS, unlike inline_unsafe_access, there seems no point in
2894   // trying to refine types. Just use the coarse types here.
2895   const Type *value_type = Type::get_const_basic_type(type);
2896   Compile::AliasType* alias_type = C->alias_type(adr_type);
2897   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2898 
2899   if (kind == LS_xchg && type == T_OBJECT) {
2900     const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2901     if (tjp != NULL) {
2902       value_type = tjp;
2903     }
2904   }
2905 
2906   int alias_idx = C->get_alias_index(adr_type);
2907 
2908   // Memory-model-wise, a LoadStore acts like a little synchronized
2909   // block, so needs barriers on each side.  These don't translate
2910   // into actual barriers on most machines, but we still need rest of
2911   // compiler to respect ordering.
2912 
2913   insert_mem_bar(Op_MemBarRelease);
2914   insert_mem_bar(Op_MemBarCPUOrder);
2915 
2916   // 4984716: MemBars must be inserted before this
2917   //          memory node in order to avoid a false
2918   //          dependency which will confuse the scheduler.
2919   Node *mem = memory(alias_idx);
2920 
2921   // For now, we handle only those cases that actually exist: ints,
2922   // longs, and Object. Adding others should be straightforward.
2923   Node* load_store;
2924   switch(type) {
2925   case T_INT:
2926     if (kind == LS_xadd) {
2927       load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2928     } else if (kind == LS_xchg) {
2929       load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2930     } else if (kind == LS_cmpxchg) {
2931       load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2932     } else {
2933       ShouldNotReachHere();
2934     }
2935     break;
2936   case T_LONG:
2937     if (kind == LS_xadd) {
2938       load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2939     } else if (kind == LS_xchg) {
2940       load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2941     } else if (kind == LS_cmpxchg) {
2942       load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2943     } else {
2944       ShouldNotReachHere();
2945     }
2946     break;
2947   case T_OBJECT:
2948     // Transformation of a value which could be NULL pointer (CastPP #NULL)
2949     // could be delayed during Parse (for example, in adjust_map_after_if()).
2950     // Execute transformation here to avoid barrier generation in such case.
2951     if (_gvn.type(newval) == TypePtr::NULL_PTR)
2952       newval = _gvn.makecon(TypePtr::NULL_PTR);
2953 
2954     // Reference stores need a store barrier.
2955     if (kind == LS_xchg) {
2956       // If pre-barrier must execute before the oop store, old value will require do_load here.
2957       if (!can_move_pre_barrier()) {
2958         pre_barrier(true /* do_load*/,
2959                     control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2960                     NULL /* pre_val*/,
2961                     T_OBJECT);
2962       } // Else move pre_barrier to use load_store value, see below.
2963     } else if (kind == LS_cmpxchg) {
2964       // Same as for newval above:
2965       if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2966         oldval = _gvn.makecon(TypePtr::NULL_PTR);
2967       }
2968       // The only known value which might get overwritten is oldval.
2969       pre_barrier(false /* do_load */,
2970                   control(), NULL, NULL, max_juint, NULL, NULL,
2971                   oldval /* pre_val */,
2972                   T_OBJECT);
2973     } else {
2974       ShouldNotReachHere();
2975     }
2976 
2977 #ifdef _LP64
2978     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2979       Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2980       if (kind == LS_xchg) {
2981         load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2982                                                               newval_enc, adr_type, value_type->make_narrowoop()));
2983       } else {
2984         assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2985         Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2986         load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2987                                                                    newval_enc, oldval_enc));
2988       }
2989     } else
2990 #endif
2991     {
2992       if (kind == LS_xchg) {
2993         load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2994       } else {
2995         assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2996         load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2997       }
2998     }
2999     post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3000     break;
3001   default:
3002     fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3003     break;
3004   }
3005 
3006   // SCMemProjNodes represent the memory state of a LoadStore. Their
3007   // main role is to prevent LoadStore nodes from being optimized away
3008   // when their results aren't used.
3009   Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3010   set_memory(proj, alias_idx);
3011 
3012   if (type == T_OBJECT && kind == LS_xchg) {
3013 #ifdef _LP64
3014     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3015       load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3016     }
3017 #endif
3018     if (can_move_pre_barrier()) {
3019       // Don't need to load pre_val. The old value is returned by load_store.
3020       // The pre_barrier can execute after the xchg as long as no safepoint
3021       // gets inserted between them.
3022       pre_barrier(false /* do_load */,
3023                   control(), NULL, NULL, max_juint, NULL, NULL,
3024                   load_store /* pre_val */,
3025                   T_OBJECT);
3026     }
3027   }
3028 
3029   // Add the trailing membar surrounding the access
3030   insert_mem_bar(Op_MemBarCPUOrder);
3031   insert_mem_bar(Op_MemBarAcquire);
3032 
3033   assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3034   set_result(load_store);
3035   return true;
3036 }
3037 
3038 //----------------------------inline_unsafe_ordered_store----------------------
3039 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3040 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3041 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3042 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3043   // This is another variant of inline_unsafe_access, differing in
3044   // that it always issues store-store ("release") barrier and ensures
3045   // store-atomicity (which only matters for "long").
3046 
3047   if (callee()->is_static())  return false;  // caller must have the capability!
3048 
3049 #ifndef PRODUCT
3050   {
3051     ResourceMark rm;
3052     // Check the signatures.
3053     ciSignature* sig = callee()->signature();
3054 #ifdef ASSERT
3055     BasicType rtype = sig->return_type()->basic_type();
3056     assert(rtype == T_VOID, "must return void");
3057     assert(sig->count() == 3, "has 3 arguments");
3058     assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3059     assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3060 #endif // ASSERT
3061   }
3062 #endif //PRODUCT
3063 
3064   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
3065 
3066   // Get arguments:
3067   Node* receiver = argument(0);  // type: oop
3068   Node* base     = argument(1);  // type: oop
3069   Node* offset   = argument(2);  // type: long
3070   Node* val      = argument(4);  // type: oop, int, or long
3071 
3072   // Null check receiver.
3073   receiver = null_check(receiver);
3074   if (stopped()) {
3075     return true;
3076   }
3077 
3078   // Build field offset expression.
3079   assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3080   // 32-bit machines ignore the high half of long offsets
3081   offset = ConvL2X(offset);
3082   Node* adr = make_unsafe_address(base, offset);
3083   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3084   const Type *value_type = Type::get_const_basic_type(type);
3085   Compile::AliasType* alias_type = C->alias_type(adr_type);
3086 
3087   insert_mem_bar(Op_MemBarRelease);
3088   insert_mem_bar(Op_MemBarCPUOrder);
3089   // Ensure that the store is atomic for longs:
3090   const bool require_atomic_access = true;
3091   Node* store;
3092   if (type == T_OBJECT) // reference stores need a store barrier.
3093     store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
3094   else {
3095     store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
3096   }
3097   insert_mem_bar(Op_MemBarCPUOrder);
3098   return true;
3099 }
3100 
3101 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3102   // Regardless of form, don't allow previous ld/st to move down,
3103   // then issue acquire, release, or volatile mem_bar.
3104   insert_mem_bar(Op_MemBarCPUOrder);
3105   switch(id) {
3106     case vmIntrinsics::_loadFence:
3107       insert_mem_bar(Op_MemBarAcquire);
3108       return true;
3109     case vmIntrinsics::_storeFence:
3110       insert_mem_bar(Op_MemBarRelease);
3111       return true;
3112     case vmIntrinsics::_fullFence:
3113       insert_mem_bar(Op_MemBarVolatile);
3114       return true;
3115     default:
3116       fatal_unexpected_iid(id);
3117       return false;
3118   }
3119 }
3120 
3121 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3122   if (!kls->is_Con()) {
3123     return true;
3124   }
3125   const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3126   if (klsptr == NULL) {
3127     return true;
3128   }
3129   ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3130   // don't need a guard for a klass that is already initialized
3131   return !ik->is_initialized();
3132 }
3133 
3134 //----------------------------inline_unsafe_allocate---------------------------
3135 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3136 bool LibraryCallKit::inline_unsafe_allocate() {
3137   if (callee()->is_static())  return false;  // caller must have the capability!
3138 
3139   null_check_receiver();  // null-check, then ignore
3140   Node* cls = null_check(argument(1));
3141   if (stopped())  return true;
3142 
3143   Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3144   kls = null_check(kls);
3145   if (stopped())  return true;  // argument was like int.class
3146 
3147   Node* test = NULL;
3148   if (LibraryCallKit::klass_needs_init_guard(kls)) {
3149     // Note:  The argument might still be an illegal value like
3150     // Serializable.class or Object[].class.   The runtime will handle it.
3151     // But we must make an explicit check for initialization.
3152     Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3153     // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3154     // can generate code to load it as unsigned byte.
3155     Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
3156     Node* bits = intcon(InstanceKlass::fully_initialized);
3157     test = _gvn.transform(new (C) SubINode(inst, bits));
3158     // The 'test' is non-zero if we need to take a slow path.
3159   }
3160 
3161   Node* obj = new_instance(kls, test);
3162   set_result(obj);
3163   return true;
3164 }
3165 
3166 #ifdef TRACE_HAVE_INTRINSICS
3167 /*
3168  * oop -> myklass
3169  * myklass->trace_id |= USED
3170  * return myklass->trace_id & ~0x3
3171  */
3172 bool LibraryCallKit::inline_native_classID() {
3173   null_check_receiver();  // null-check, then ignore
3174   Node* cls = null_check(argument(1), T_OBJECT);
3175   Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3176   kls = null_check(kls, T_OBJECT);
3177   ByteSize offset = TRACE_ID_OFFSET;
3178   Node* insp = basic_plus_adr(kls, in_bytes(offset));
3179   Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
3180   Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3181   Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3182   Node* clsused = longcon(0x01l); // set the class bit
3183   Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3184 
3185   const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3186   store_to_memory(control(), insp, orl, T_LONG, adr_type);
3187   set_result(andl);
3188   return true;
3189 }
3190 
3191 bool LibraryCallKit::inline_native_threadID() {
3192   Node* tls_ptr = NULL;
3193   Node* cur_thr = generate_current_thread(tls_ptr);
3194   Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3195   Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3196   p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3197 
3198   Node* threadid = NULL;
3199   size_t thread_id_size = OSThread::thread_id_size();
3200   if (thread_id_size == (size_t) BytesPerLong) {
3201     threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
3202   } else if (thread_id_size == (size_t) BytesPerInt) {
3203     threadid = make_load(control(), p, TypeInt::INT, T_INT);
3204   } else {
3205     ShouldNotReachHere();
3206   }
3207   set_result(threadid);
3208   return true;
3209 }
3210 #endif
3211 
3212 //------------------------inline_native_time_funcs--------------
3213 // inline code for System.currentTimeMillis() and System.nanoTime()
3214 // these have the same type and signature
3215 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3216   const TypeFunc* tf = OptoRuntime::void_long_Type();
3217   const TypePtr* no_memory_effects = NULL;
3218   Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3219   Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3220 #ifdef ASSERT
3221   Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3222   assert(value_top == top(), "second value must be top");
3223 #endif
3224   set_result(value);
3225   return true;
3226 }
3227 
3228 //------------------------inline_native_currentThread------------------
3229 bool LibraryCallKit::inline_native_currentThread() {
3230   Node* junk = NULL;
3231   set_result(generate_current_thread(junk));
3232   return true;
3233 }
3234 
3235 //------------------------inline_native_isInterrupted------------------
3236 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3237 bool LibraryCallKit::inline_native_isInterrupted() {
3238   // Add a fast path to t.isInterrupted(clear_int):
3239   //   (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
3240   //   ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3241   // So, in the common case that the interrupt bit is false,
3242   // we avoid making a call into the VM.  Even if the interrupt bit
3243   // is true, if the clear_int argument is false, we avoid the VM call.
3244   // However, if the receiver is not currentThread, we must call the VM,
3245   // because there must be some locking done around the operation.
3246 
3247   // We only go to the fast case code if we pass two guards.
3248   // Paths which do not pass are accumulated in the slow_region.
3249 
3250   enum {
3251     no_int_result_path   = 1, // t == Thread.current() && !TLS._osthread._interrupted
3252     no_clear_result_path = 2, // t == Thread.current() &&  TLS._osthread._interrupted && !clear_int
3253     slow_result_path     = 3, // slow path: t.isInterrupted(clear_int)
3254     PATH_LIMIT
3255   };
3256 
3257   // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3258   // out of the function.
3259   insert_mem_bar(Op_MemBarCPUOrder);
3260 
3261   RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3262   PhiNode*    result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3263 
3264   RegionNode* slow_region = new (C) RegionNode(1);
3265   record_for_igvn(slow_region);
3266 
3267   // (a) Receiving thread must be the current thread.
3268   Node* rec_thr = argument(0);
3269   Node* tls_ptr = NULL;
3270   Node* cur_thr = generate_current_thread(tls_ptr);
3271   Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3272   Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3273 
3274   generate_slow_guard(bol_thr, slow_region);
3275 
3276   // (b) Interrupt bit on TLS must be false.
3277   Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3278   Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3279   p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3280 
3281   // Set the control input on the field _interrupted read to prevent it floating up.
3282   Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3283   Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3284   Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3285 
3286   IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3287 
3288   // First fast path:  if (!TLS._interrupted) return false;
3289   Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3290   result_rgn->init_req(no_int_result_path, false_bit);
3291   result_val->init_req(no_int_result_path, intcon(0));
3292 
3293   // drop through to next case
3294   set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3295 
3296   // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3297   Node* clr_arg = argument(1);
3298   Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3299   Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3300   IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3301 
3302   // Second fast path:  ... else if (!clear_int) return true;
3303   Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3304   result_rgn->init_req(no_clear_result_path, false_arg);
3305   result_val->init_req(no_clear_result_path, intcon(1));
3306 
3307   // drop through to next case
3308   set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3309 
3310   // (d) Otherwise, go to the slow path.
3311   slow_region->add_req(control());
3312   set_control( _gvn.transform(slow_region));
3313 
3314   if (stopped()) {
3315     // There is no slow path.
3316     result_rgn->init_req(slow_result_path, top());
3317     result_val->init_req(slow_result_path, top());
3318   } else {
3319     // non-virtual because it is a private non-static
3320     CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3321 
3322     Node* slow_val = set_results_for_java_call(slow_call);
3323     // this->control() comes from set_results_for_java_call
3324 
3325     Node* fast_io  = slow_call->in(TypeFunc::I_O);
3326     Node* fast_mem = slow_call->in(TypeFunc::Memory);
3327 
3328     // These two phis are pre-filled with copies of of the fast IO and Memory
3329     PhiNode* result_mem  = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3330     PhiNode* result_io   = PhiNode::make(result_rgn, fast_io,  Type::ABIO);
3331 
3332     result_rgn->init_req(slow_result_path, control());
3333     result_io ->init_req(slow_result_path, i_o());
3334     result_mem->init_req(slow_result_path, reset_memory());
3335     result_val->init_req(slow_result_path, slow_val);
3336 
3337     set_all_memory(_gvn.transform(result_mem));
3338     set_i_o(       _gvn.transform(result_io));
3339   }
3340 
3341   C->set_has_split_ifs(true); // Has chance for split-if optimization
3342   set_result(result_rgn, result_val);
3343   return true;
3344 }
3345 
3346 //---------------------------load_mirror_from_klass----------------------------
3347 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3348 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3349   Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3350   return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3351 }
3352 
3353 //-----------------------load_klass_from_mirror_common-------------------------
3354 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3355 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3356 // and branch to the given path on the region.
3357 // If never_see_null, take an uncommon trap on null, so we can optimistically
3358 // compile for the non-null case.
3359 // If the region is NULL, force never_see_null = true.
3360 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3361                                                     bool never_see_null,
3362                                                     RegionNode* region,
3363                                                     int null_path,
3364                                                     int offset) {
3365   if (region == NULL)  never_see_null = true;
3366   Node* p = basic_plus_adr(mirror, offset);
3367   const TypeKlassPtr*  kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3368   Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3369   Node* null_ctl = top();
3370   kls = null_check_oop(kls, &null_ctl, never_see_null);
3371   if (region != NULL) {
3372     // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3373     region->init_req(null_path, null_ctl);
3374   } else {
3375     assert(null_ctl == top(), "no loose ends");
3376   }
3377   return kls;
3378 }
3379 
3380 //--------------------(inline_native_Class_query helpers)---------------------
3381 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3382 // Fall through if (mods & mask) == bits, take the guard otherwise.
3383 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3384   // Branch around if the given klass has the given modifier bit set.
3385   // Like generate_guard, adds a new path onto the region.
3386   Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3387   Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3388   Node* mask = intcon(modifier_mask);
3389   Node* bits = intcon(modifier_bits);
3390   Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3391   Node* cmp  = _gvn.transform(new (C) CmpINode(mbit, bits));
3392   Node* bol  = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3393   return generate_fair_guard(bol, region);
3394 }
3395 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3396   return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3397 }
3398 
3399 //-------------------------inline_native_Class_query-------------------
3400 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3401   const Type* return_type = TypeInt::BOOL;
3402   Node* prim_return_value = top();  // what happens if it's a primitive class?
3403   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3404   bool expect_prim = false;     // most of these guys expect to work on refs
3405 
3406   enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3407 
3408   Node* mirror = argument(0);
3409   Node* obj    = top();
3410 
3411   switch (id) {
3412   case vmIntrinsics::_isInstance:
3413     // nothing is an instance of a primitive type
3414     prim_return_value = intcon(0);
3415     obj = argument(1);
3416     break;
3417   case vmIntrinsics::_getModifiers:
3418     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3419     assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3420     return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3421     break;
3422   case vmIntrinsics::_isInterface:
3423     prim_return_value = intcon(0);
3424     break;
3425   case vmIntrinsics::_isArray:
3426     prim_return_value = intcon(0);
3427     expect_prim = true;  // cf. ObjectStreamClass.getClassSignature
3428     break;
3429   case vmIntrinsics::_isPrimitive:
3430     prim_return_value = intcon(1);
3431     expect_prim = true;  // obviously
3432     break;
3433   case vmIntrinsics::_getSuperclass:
3434     prim_return_value = null();
3435     return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3436     break;
3437   case vmIntrinsics::_getComponentType:
3438     prim_return_value = null();
3439     return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3440     break;
3441   case vmIntrinsics::_getClassAccessFlags:
3442     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3443     return_type = TypeInt::INT;  // not bool!  6297094
3444     break;
3445   default:
3446     fatal_unexpected_iid(id);
3447     break;
3448   }
3449 
3450   const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3451   if (mirror_con == NULL)  return false;  // cannot happen?
3452 
3453 #ifndef PRODUCT
3454   if (C->print_intrinsics() || C->print_inlining()) {
3455     ciType* k = mirror_con->java_mirror_type();
3456     if (k) {
3457       tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3458       k->print_name();
3459       tty->cr();
3460     }
3461   }
3462 #endif
3463 
3464   // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3465   RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3466   record_for_igvn(region);
3467   PhiNode* phi = new (C) PhiNode(region, return_type);
3468 
3469   // The mirror will never be null of Reflection.getClassAccessFlags, however
3470   // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3471   // if it is. See bug 4774291.
3472 
3473   // For Reflection.getClassAccessFlags(), the null check occurs in
3474   // the wrong place; see inline_unsafe_access(), above, for a similar
3475   // situation.
3476   mirror = null_check(mirror);
3477   // If mirror or obj is dead, only null-path is taken.
3478   if (stopped())  return true;
3479 
3480   if (expect_prim)  never_see_null = false;  // expect nulls (meaning prims)
3481 
3482   // Now load the mirror's klass metaobject, and null-check it.
3483   // Side-effects region with the control path if the klass is null.
3484   Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3485   // If kls is null, we have a primitive mirror.
3486   phi->init_req(_prim_path, prim_return_value);
3487   if (stopped()) { set_result(region, phi); return true; }
3488 
3489   Node* p;  // handy temp
3490   Node* null_ctl;
3491 
3492   // Now that we have the non-null klass, we can perform the real query.
3493   // For constant classes, the query will constant-fold in LoadNode::Value.
3494   Node* query_value = top();
3495   switch (id) {
3496   case vmIntrinsics::_isInstance:
3497     // nothing is an instance of a primitive type
3498     query_value = gen_instanceof(obj, kls);
3499     break;
3500 
3501   case vmIntrinsics::_getModifiers:
3502     p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3503     query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3504     break;
3505 
3506   case vmIntrinsics::_isInterface:
3507     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3508     if (generate_interface_guard(kls, region) != NULL)
3509       // A guard was added.  If the guard is taken, it was an interface.
3510       phi->add_req(intcon(1));
3511     // If we fall through, it's a plain class.
3512     query_value = intcon(0);
3513     break;
3514 
3515   case vmIntrinsics::_isArray:
3516     // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3517     if (generate_array_guard(kls, region) != NULL)
3518       // A guard was added.  If the guard is taken, it was an array.
3519       phi->add_req(intcon(1));
3520     // If we fall through, it's a plain class.
3521     query_value = intcon(0);
3522     break;
3523 
3524   case vmIntrinsics::_isPrimitive:
3525     query_value = intcon(0); // "normal" path produces false
3526     break;
3527 
3528   case vmIntrinsics::_getSuperclass:
3529     // The rules here are somewhat unfortunate, but we can still do better
3530     // with random logic than with a JNI call.
3531     // Interfaces store null or Object as _super, but must report null.
3532     // Arrays store an intermediate super as _super, but must report Object.
3533     // Other types can report the actual _super.
3534     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3535     if (generate_interface_guard(kls, region) != NULL)
3536       // A guard was added.  If the guard is taken, it was an interface.
3537       phi->add_req(null());
3538     if (generate_array_guard(kls, region) != NULL)
3539       // A guard was added.  If the guard is taken, it was an array.
3540       phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3541     // If we fall through, it's a plain class.  Get its _super.
3542     p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3543     kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3544     null_ctl = top();
3545     kls = null_check_oop(kls, &null_ctl);
3546     if (null_ctl != top()) {
3547       // If the guard is taken, Object.superClass is null (both klass and mirror).
3548       region->add_req(null_ctl);
3549       phi   ->add_req(null());
3550     }
3551     if (!stopped()) {
3552       query_value = load_mirror_from_klass(kls);
3553     }
3554     break;
3555 
3556   case vmIntrinsics::_getComponentType:
3557     if (generate_array_guard(kls, region) != NULL) {
3558       // Be sure to pin the oop load to the guard edge just created:
3559       Node* is_array_ctrl = region->in(region->req()-1);
3560       Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3561       Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3562       phi->add_req(cmo);
3563     }
3564     query_value = null();  // non-array case is null
3565     break;
3566 
3567   case vmIntrinsics::_getClassAccessFlags:
3568     p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3569     query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3570     break;
3571 
3572   default:
3573     fatal_unexpected_iid(id);
3574     break;
3575   }
3576 
3577   // Fall-through is the normal case of a query to a real class.
3578   phi->init_req(1, query_value);
3579   region->init_req(1, control());
3580 
3581   C->set_has_split_ifs(true); // Has chance for split-if optimization
3582   set_result(region, phi);
3583   return true;
3584 }
3585 
3586 //--------------------------inline_native_subtype_check------------------------
3587 // This intrinsic takes the JNI calls out of the heart of
3588 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3589 bool LibraryCallKit::inline_native_subtype_check() {
3590   // Pull both arguments off the stack.
3591   Node* args[2];                // two java.lang.Class mirrors: superc, subc
3592   args[0] = argument(0);
3593   args[1] = argument(1);
3594   Node* klasses[2];             // corresponding Klasses: superk, subk
3595   klasses[0] = klasses[1] = top();
3596 
3597   enum {
3598     // A full decision tree on {superc is prim, subc is prim}:
3599     _prim_0_path = 1,           // {P,N} => false
3600                                 // {P,P} & superc!=subc => false
3601     _prim_same_path,            // {P,P} & superc==subc => true
3602     _prim_1_path,               // {N,P} => false
3603     _ref_subtype_path,          // {N,N} & subtype check wins => true
3604     _both_ref_path,             // {N,N} & subtype check loses => false
3605     PATH_LIMIT
3606   };
3607 
3608   RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3609   Node*       phi    = new (C) PhiNode(region, TypeInt::BOOL);
3610   record_for_igvn(region);
3611 
3612   const TypePtr* adr_type = TypeRawPtr::BOTTOM;   // memory type of loads
3613   const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3614   int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3615 
3616   // First null-check both mirrors and load each mirror's klass metaobject.
3617   int which_arg;
3618   for (which_arg = 0; which_arg <= 1; which_arg++) {
3619     Node* arg = args[which_arg];
3620     arg = null_check(arg);
3621     if (stopped())  break;
3622     args[which_arg] = arg;
3623 
3624     Node* p = basic_plus_adr(arg, class_klass_offset);
3625     Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3626     klasses[which_arg] = _gvn.transform(kls);
3627   }
3628 
3629   // Having loaded both klasses, test each for null.
3630   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3631   for (which_arg = 0; which_arg <= 1; which_arg++) {
3632     Node* kls = klasses[which_arg];
3633     Node* null_ctl = top();
3634     kls = null_check_oop(kls, &null_ctl, never_see_null);
3635     int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3636     region->init_req(prim_path, null_ctl);
3637     if (stopped())  break;
3638     klasses[which_arg] = kls;
3639   }
3640 
3641   if (!stopped()) {
3642     // now we have two reference types, in klasses[0..1]
3643     Node* subk   = klasses[1];  // the argument to isAssignableFrom
3644     Node* superk = klasses[0];  // the receiver
3645     region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3646     // now we have a successful reference subtype check
3647     region->set_req(_ref_subtype_path, control());
3648   }
3649 
3650   // If both operands are primitive (both klasses null), then
3651   // we must return true when they are identical primitives.
3652   // It is convenient to test this after the first null klass check.
3653   set_control(region->in(_prim_0_path)); // go back to first null check
3654   if (!stopped()) {
3655     // Since superc is primitive, make a guard for the superc==subc case.
3656     Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3657     Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3658     generate_guard(bol_eq, region, PROB_FAIR);
3659     if (region->req() == PATH_LIMIT+1) {
3660       // A guard was added.  If the added guard is taken, superc==subc.
3661       region->swap_edges(PATH_LIMIT, _prim_same_path);
3662       region->del_req(PATH_LIMIT);
3663     }
3664     region->set_req(_prim_0_path, control()); // Not equal after all.
3665   }
3666 
3667   // these are the only paths that produce 'true':
3668   phi->set_req(_prim_same_path,   intcon(1));
3669   phi->set_req(_ref_subtype_path, intcon(1));
3670 
3671   // pull together the cases:
3672   assert(region->req() == PATH_LIMIT, "sane region");
3673   for (uint i = 1; i < region->req(); i++) {
3674     Node* ctl = region->in(i);
3675     if (ctl == NULL || ctl == top()) {
3676       region->set_req(i, top());
3677       phi   ->set_req(i, top());
3678     } else if (phi->in(i) == NULL) {
3679       phi->set_req(i, intcon(0)); // all other paths produce 'false'
3680     }
3681   }
3682 
3683   set_control(_gvn.transform(region));
3684   set_result(_gvn.transform(phi));
3685   return true;
3686 }
3687 
3688 //---------------------generate_array_guard_common------------------------
3689 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3690                                                   bool obj_array, bool not_array) {
3691   // If obj_array/non_array==false/false:
3692   // Branch around if the given klass is in fact an array (either obj or prim).
3693   // If obj_array/non_array==false/true:
3694   // Branch around if the given klass is not an array klass of any kind.
3695   // If obj_array/non_array==true/true:
3696   // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3697   // If obj_array/non_array==true/false:
3698   // Branch around if the kls is an oop array (Object[] or subtype)
3699   //
3700   // Like generate_guard, adds a new path onto the region.
3701   jint  layout_con = 0;
3702   Node* layout_val = get_layout_helper(kls, layout_con);
3703   if (layout_val == NULL) {
3704     bool query = (obj_array
3705                   ? Klass::layout_helper_is_objArray(layout_con)
3706                   : Klass::layout_helper_is_array(layout_con));
3707     if (query == not_array) {
3708       return NULL;                       // never a branch
3709     } else {                             // always a branch
3710       Node* always_branch = control();
3711       if (region != NULL)
3712         region->add_req(always_branch);
3713       set_control(top());
3714       return always_branch;
3715     }
3716   }
3717   // Now test the correct condition.
3718   jint  nval = (obj_array
3719                 ? ((jint)Klass::_lh_array_tag_type_value
3720                    <<    Klass::_lh_array_tag_shift)
3721                 : Klass::_lh_neutral_value);
3722   Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3723   BoolTest::mask btest = BoolTest::lt;  // correct for testing is_[obj]array
3724   // invert the test if we are looking for a non-array
3725   if (not_array)  btest = BoolTest(btest).negate();
3726   Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3727   return generate_fair_guard(bol, region);
3728 }
3729 
3730 
3731 //-----------------------inline_native_newArray--------------------------
3732 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3733 bool LibraryCallKit::inline_native_newArray() {
3734   Node* mirror    = argument(0);
3735   Node* count_val = argument(1);
3736 
3737   mirror = null_check(mirror);
3738   // If mirror or obj is dead, only null-path is taken.
3739   if (stopped())  return true;
3740 
3741   enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3742   RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3743   PhiNode*    result_val = new(C) PhiNode(result_reg,
3744                                           TypeInstPtr::NOTNULL);
3745   PhiNode*    result_io  = new(C) PhiNode(result_reg, Type::ABIO);
3746   PhiNode*    result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3747                                           TypePtr::BOTTOM);
3748 
3749   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3750   Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3751                                                   result_reg, _slow_path);
3752   Node* normal_ctl   = control();
3753   Node* no_array_ctl = result_reg->in(_slow_path);
3754 
3755   // Generate code for the slow case.  We make a call to newArray().
3756   set_control(no_array_ctl);
3757   if (!stopped()) {
3758     // Either the input type is void.class, or else the
3759     // array klass has not yet been cached.  Either the
3760     // ensuing call will throw an exception, or else it
3761     // will cache the array klass for next time.
3762     PreserveJVMState pjvms(this);
3763     CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3764     Node* slow_result = set_results_for_java_call(slow_call);
3765     // this->control() comes from set_results_for_java_call
3766     result_reg->set_req(_slow_path, control());
3767     result_val->set_req(_slow_path, slow_result);
3768     result_io ->set_req(_slow_path, i_o());
3769     result_mem->set_req(_slow_path, reset_memory());
3770   }
3771 
3772   set_control(normal_ctl);
3773   if (!stopped()) {
3774     // Normal case:  The array type has been cached in the java.lang.Class.
3775     // The following call works fine even if the array type is polymorphic.
3776     // It could be a dynamic mix of int[], boolean[], Object[], etc.
3777     Node* obj = new_array(klass_node, count_val, 0);  // no arguments to push
3778     result_reg->init_req(_normal_path, control());
3779     result_val->init_req(_normal_path, obj);
3780     result_io ->init_req(_normal_path, i_o());
3781     result_mem->init_req(_normal_path, reset_memory());
3782   }
3783 
3784   // Return the combined state.
3785   set_i_o(        _gvn.transform(result_io)  );
3786   set_all_memory( _gvn.transform(result_mem));
3787 
3788   C->set_has_split_ifs(true); // Has chance for split-if optimization
3789   set_result(result_reg, result_val);
3790   return true;
3791 }
3792 
3793 //----------------------inline_native_getLength--------------------------
3794 // public static native int java.lang.reflect.Array.getLength(Object array);
3795 bool LibraryCallKit::inline_native_getLength() {
3796   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3797 
3798   Node* array = null_check(argument(0));
3799   // If array is dead, only null-path is taken.
3800   if (stopped())  return true;
3801 
3802   // Deoptimize if it is a non-array.
3803   Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3804 
3805   if (non_array != NULL) {
3806     PreserveJVMState pjvms(this);
3807     set_control(non_array);
3808     uncommon_trap(Deoptimization::Reason_intrinsic,
3809                   Deoptimization::Action_maybe_recompile);
3810   }
3811 
3812   // If control is dead, only non-array-path is taken.
3813   if (stopped())  return true;
3814 
3815   // The works fine even if the array type is polymorphic.
3816   // It could be a dynamic mix of int[], boolean[], Object[], etc.
3817   Node* result = load_array_length(array);
3818 
3819   C->set_has_split_ifs(true);  // Has chance for split-if optimization
3820   set_result(result);
3821   return true;
3822 }
3823 
3824 //------------------------inline_array_copyOf----------------------------
3825 // public static <T,U> T[] java.util.Arrays.copyOf(     U[] original, int newLength,         Class<? extends T[]> newType);
3826 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from,      int to, Class<? extends T[]> newType);
3827 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3828   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3829 
3830   // Get the arguments.
3831   Node* original          = argument(0);
3832   Node* start             = is_copyOfRange? argument(1): intcon(0);
3833   Node* end               = is_copyOfRange? argument(2): argument(1);
3834   Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3835 
3836   Node* newcopy;
3837 
3838   // Set the original stack and the reexecute bit for the interpreter to reexecute
3839   // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3840   { PreserveReexecuteState preexecs(this);
3841     jvms()->set_should_reexecute(true);
3842 
3843     array_type_mirror = null_check(array_type_mirror);
3844     original          = null_check(original);
3845 
3846     // Check if a null path was taken unconditionally.
3847     if (stopped())  return true;
3848 
3849     Node* orig_length = load_array_length(original);
3850 
3851     Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3852     klass_node = null_check(klass_node);
3853 
3854     RegionNode* bailout = new (C) RegionNode(1);
3855     record_for_igvn(bailout);
3856 
3857     // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3858     // Bail out if that is so.
3859     Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3860     if (not_objArray != NULL) {
3861       // Improve the klass node's type from the new optimistic assumption:
3862       ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3863       const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3864       Node* cast = new (C) CastPPNode(klass_node, akls);
3865       cast->init_req(0, control());
3866       klass_node = _gvn.transform(cast);
3867     }
3868 
3869     // Bail out if either start or end is negative.
3870     generate_negative_guard(start, bailout, &start);
3871     generate_negative_guard(end,   bailout, &end);
3872 
3873     Node* length = end;
3874     if (_gvn.type(start) != TypeInt::ZERO) {
3875       length = _gvn.transform(new (C) SubINode(end, start));
3876     }
3877 
3878     // Bail out if length is negative.
3879     // Without this the new_array would throw
3880     // NegativeArraySizeException but IllegalArgumentException is what
3881     // should be thrown
3882     generate_negative_guard(length, bailout, &length);
3883 
3884     if (bailout->req() > 1) {
3885       PreserveJVMState pjvms(this);
3886       set_control(_gvn.transform(bailout));
3887       uncommon_trap(Deoptimization::Reason_intrinsic,
3888                     Deoptimization::Action_maybe_recompile);
3889     }
3890 
3891     if (!stopped()) {
3892       // How many elements will we copy from the original?
3893       // The answer is MinI(orig_length - start, length).
3894       Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3895       Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3896 
3897       newcopy = new_array(klass_node, length, 0);  // no argments to push
3898 
3899       // Generate a direct call to the right arraycopy function(s).
3900       // We know the copy is disjoint but we might not know if the
3901       // oop stores need checking.
3902       // Extreme case:  Arrays.copyOf((Integer[])x, 10, String[].class).
3903       // This will fail a store-check if x contains any non-nulls.
3904       bool disjoint_bases = true;
3905       // if start > orig_length then the length of the copy may be
3906       // negative.
3907       bool length_never_negative = !is_copyOfRange;
3908       generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3909                          original, start, newcopy, intcon(0), moved,
3910                          disjoint_bases, length_never_negative);
3911     }
3912   } // original reexecute is set back here
3913 
3914   C->set_has_split_ifs(true); // Has chance for split-if optimization
3915   if (!stopped()) {
3916     set_result(newcopy);
3917   }
3918   return true;
3919 }
3920 
3921 
3922 //----------------------generate_virtual_guard---------------------------
3923 // Helper for hashCode and clone.  Peeks inside the vtable to avoid a call.
3924 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3925                                              RegionNode* slow_region) {
3926   ciMethod* method = callee();
3927   int vtable_index = method->vtable_index();
3928   assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3929          err_msg_res("bad index %d", vtable_index));
3930   // Get the Method* out of the appropriate vtable entry.
3931   int entry_offset  = (InstanceKlass::vtable_start_offset() +
3932                      vtable_index*vtableEntry::size()) * wordSize +
3933                      vtableEntry::method_offset_in_bytes();
3934   Node* entry_addr  = basic_plus_adr(obj_klass, entry_offset);
3935   Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3936 
3937   // Compare the target method with the expected method (e.g., Object.hashCode).
3938   const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3939 
3940   Node* native_call = makecon(native_call_addr);
3941   Node* chk_native  = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3942   Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3943 
3944   return generate_slow_guard(test_native, slow_region);
3945 }
3946 
3947 //-----------------------generate_method_call----------------------------
3948 // Use generate_method_call to make a slow-call to the real
3949 // method if the fast path fails.  An alternative would be to
3950 // use a stub like OptoRuntime::slow_arraycopy_Java.
3951 // This only works for expanding the current library call,
3952 // not another intrinsic.  (E.g., don't use this for making an
3953 // arraycopy call inside of the copyOf intrinsic.)
3954 CallJavaNode*
3955 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3956   // When compiling the intrinsic method itself, do not use this technique.
3957   guarantee(callee() != C->method(), "cannot make slow-call to self");
3958 
3959   ciMethod* method = callee();
3960   // ensure the JVMS we have will be correct for this call
3961   guarantee(method_id == method->intrinsic_id(), "must match");
3962 
3963   const TypeFunc* tf = TypeFunc::make(method);
3964   CallJavaNode* slow_call;
3965   if (is_static) {
3966     assert(!is_virtual, "");
3967     slow_call = new(C) CallStaticJavaNode(C, tf,
3968                            SharedRuntime::get_resolve_static_call_stub(),
3969                            method, bci());
3970   } else if (is_virtual) {
3971     null_check_receiver();
3972     int vtable_index = Method::invalid_vtable_index;
3973     if (UseInlineCaches) {
3974       // Suppress the vtable call
3975     } else {
3976       // hashCode and clone are not a miranda methods,
3977       // so the vtable index is fixed.
3978       // No need to use the linkResolver to get it.
3979        vtable_index = method->vtable_index();
3980        assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3981               err_msg_res("bad index %d", vtable_index));
3982     }
3983     slow_call = new(C) CallDynamicJavaNode(tf,
3984                           SharedRuntime::get_resolve_virtual_call_stub(),
3985                           method, vtable_index, bci());
3986   } else {  // neither virtual nor static:  opt_virtual
3987     null_check_receiver();
3988     slow_call = new(C) CallStaticJavaNode(C, tf,
3989                                 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3990                                 method, bci());
3991     slow_call->set_optimized_virtual(true);
3992   }
3993   set_arguments_for_java_call(slow_call);
3994   set_edges_for_java_call(slow_call);
3995   return slow_call;
3996 }
3997 
3998 
3999 //------------------------------inline_native_hashcode--------------------
4000 // Build special case code for calls to hashCode on an object.
4001 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4002   assert(is_static == callee()->is_static(), "correct intrinsic selection");
4003   assert(!(is_virtual && is_static), "either virtual, special, or static");
4004 
4005   enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4006 
4007   RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4008   PhiNode*    result_val = new(C) PhiNode(result_reg,
4009                                           TypeInt::INT);
4010   PhiNode*    result_io  = new(C) PhiNode(result_reg, Type::ABIO);
4011   PhiNode*    result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4012                                           TypePtr::BOTTOM);
4013   Node* obj = NULL;
4014   if (!is_static) {
4015     // Check for hashing null object
4016     obj = null_check_receiver();
4017     if (stopped())  return true;        // unconditionally null
4018     result_reg->init_req(_null_path, top());
4019     result_val->init_req(_null_path, top());
4020   } else {
4021     // Do a null check, and return zero if null.
4022     // System.identityHashCode(null) == 0
4023     obj = argument(0);
4024     Node* null_ctl = top();
4025     obj = null_check_oop(obj, &null_ctl);
4026     result_reg->init_req(_null_path, null_ctl);
4027     result_val->init_req(_null_path, _gvn.intcon(0));
4028   }
4029 
4030   // Unconditionally null?  Then return right away.
4031   if (stopped()) {
4032     set_control( result_reg->in(_null_path));
4033     if (!stopped())
4034       set_result(result_val->in(_null_path));
4035     return true;
4036   }
4037 
4038   // After null check, get the object's klass.
4039   Node* obj_klass = load_object_klass(obj);
4040 
4041   // This call may be virtual (invokevirtual) or bound (invokespecial).
4042   // For each case we generate slightly different code.
4043 
4044   // We only go to the fast case code if we pass a number of guards.  The
4045   // paths which do not pass are accumulated in the slow_region.
4046   RegionNode* slow_region = new (C) RegionNode(1);
4047   record_for_igvn(slow_region);
4048 
4049   // If this is a virtual call, we generate a funny guard.  We pull out
4050   // the vtable entry corresponding to hashCode() from the target object.
4051   // If the target method which we are calling happens to be the native
4052   // Object hashCode() method, we pass the guard.  We do not need this
4053   // guard for non-virtual calls -- the caller is known to be the native
4054   // Object hashCode().
4055   if (is_virtual) {
4056     generate_virtual_guard(obj_klass, slow_region);
4057   }
4058 
4059   // Get the header out of the object, use LoadMarkNode when available
4060   Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4061   Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
4062 
4063   // Test the header to see if it is unlocked.
4064   Node *lock_mask      = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4065   Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4066   Node *unlocked_val   = _gvn.MakeConX(markOopDesc::unlocked_value);
4067   Node *chk_unlocked   = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4068   Node *test_unlocked  = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4069 
4070   generate_slow_guard(test_unlocked, slow_region);
4071 
4072   // Get the hash value and check to see that it has been properly assigned.
4073   // We depend on hash_mask being at most 32 bits and avoid the use of
4074   // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4075   // vm: see markOop.hpp.
4076   Node *hash_mask      = _gvn.intcon(markOopDesc::hash_mask);
4077   Node *hash_shift     = _gvn.intcon(markOopDesc::hash_shift);
4078   Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4079   // This hack lets the hash bits live anywhere in the mark object now, as long
4080   // as the shift drops the relevant bits into the low 32 bits.  Note that
4081   // Java spec says that HashCode is an int so there's no point in capturing
4082   // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4083   hshifted_header      = ConvX2I(hshifted_header);
4084   Node *hash_val       = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4085 
4086   Node *no_hash_val    = _gvn.intcon(markOopDesc::no_hash);
4087   Node *chk_assigned   = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4088   Node *test_assigned  = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4089 
4090   generate_slow_guard(test_assigned, slow_region);
4091 
4092   Node* init_mem = reset_memory();
4093   // fill in the rest of the null path:
4094   result_io ->init_req(_null_path, i_o());
4095   result_mem->init_req(_null_path, init_mem);
4096 
4097   result_val->init_req(_fast_path, hash_val);
4098   result_reg->init_req(_fast_path, control());
4099   result_io ->init_req(_fast_path, i_o());
4100   result_mem->init_req(_fast_path, init_mem);
4101 
4102   // Generate code for the slow case.  We make a call to hashCode().
4103   set_control(_gvn.transform(slow_region));
4104   if (!stopped()) {
4105     // No need for PreserveJVMState, because we're using up the present state.
4106     set_all_memory(init_mem);
4107     vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4108     CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4109     Node* slow_result = set_results_for_java_call(slow_call);
4110     // this->control() comes from set_results_for_java_call
4111     result_reg->init_req(_slow_path, control());
4112     result_val->init_req(_slow_path, slow_result);
4113     result_io  ->set_req(_slow_path, i_o());
4114     result_mem ->set_req(_slow_path, reset_memory());
4115   }
4116 
4117   // Return the combined state.
4118   set_i_o(        _gvn.transform(result_io)  );
4119   set_all_memory( _gvn.transform(result_mem));
4120 
4121   set_result(result_reg, result_val);
4122   return true;
4123 }
4124 
4125 //---------------------------inline_native_getClass----------------------------
4126 // public final native Class<?> java.lang.Object.getClass();
4127 //
4128 // Build special case code for calls to getClass on an object.
4129 bool LibraryCallKit::inline_native_getClass() {
4130   Node* obj = null_check_receiver();
4131   if (stopped())  return true;
4132   set_result(load_mirror_from_klass(load_object_klass(obj)));
4133   return true;
4134 }
4135 
4136 //-----------------inline_native_Reflection_getCallerClass---------------------
4137 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4138 //
4139 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4140 //
4141 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4142 // in that it must skip particular security frames and checks for
4143 // caller sensitive methods.
4144 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4145 #ifndef PRODUCT
4146   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4147     tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4148   }
4149 #endif
4150 
4151   if (!jvms()->has_method()) {
4152 #ifndef PRODUCT
4153     if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4154       tty->print_cr("  Bailing out because intrinsic was inlined at top level");
4155     }
4156 #endif
4157     return false;
4158   }
4159 
4160   // Walk back up the JVM state to find the caller at the required
4161   // depth.
4162   JVMState* caller_jvms = jvms();
4163 
4164   // Cf. JVM_GetCallerClass
4165   // NOTE: Start the loop at depth 1 because the current JVM state does
4166   // not include the Reflection.getCallerClass() frame.
4167   for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4168     ciMethod* m = caller_jvms->method();
4169     switch (n) {
4170     case 0:
4171       fatal("current JVM state does not include the Reflection.getCallerClass frame");
4172       break;
4173     case 1:
4174       // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4175       if (!m->caller_sensitive()) {
4176 #ifndef PRODUCT
4177         if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4178           tty->print_cr("  Bailing out: CallerSensitive annotation expected at frame %d", n);
4179         }
4180 #endif
4181         return false;  // bail-out; let JVM_GetCallerClass do the work
4182       }
4183       break;
4184     default:
4185       if (!m->is_ignored_by_security_stack_walk()) {
4186         // We have reached the desired frame; return the holder class.
4187         // Acquire method holder as java.lang.Class and push as constant.
4188         ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4189         ciInstance* caller_mirror = caller_klass->java_mirror();
4190         set_result(makecon(TypeInstPtr::make(caller_mirror)));
4191 
4192 #ifndef PRODUCT
4193         if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4194           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());
4195           tty->print_cr("  JVM state at this point:");
4196           for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4197             ciMethod* m = jvms()->of_depth(i)->method();
4198             tty->print_cr("   %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4199           }
4200         }
4201 #endif
4202         return true;
4203       }
4204       break;
4205     }
4206   }
4207 
4208 #ifndef PRODUCT
4209   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4210     tty->print_cr("  Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4211     tty->print_cr("  JVM state at this point:");
4212     for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4213       ciMethod* m = jvms()->of_depth(i)->method();
4214       tty->print_cr("   %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4215     }
4216   }
4217 #endif
4218 
4219   return false;  // bail-out; let JVM_GetCallerClass do the work
4220 }
4221 
4222 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4223   Node* arg = argument(0);
4224   Node* result;
4225 
4226   switch (id) {
4227   case vmIntrinsics::_floatToRawIntBits:    result = new (C) MoveF2INode(arg);  break;
4228   case vmIntrinsics::_intBitsToFloat:       result = new (C) MoveI2FNode(arg);  break;
4229   case vmIntrinsics::_doubleToRawLongBits:  result = new (C) MoveD2LNode(arg);  break;
4230   case vmIntrinsics::_longBitsToDouble:     result = new (C) MoveL2DNode(arg);  break;
4231 
4232   case vmIntrinsics::_doubleToLongBits: {
4233     // two paths (plus control) merge in a wood
4234     RegionNode *r = new (C) RegionNode(3);
4235     Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4236 
4237     Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4238     // Build the boolean node
4239     Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4240 
4241     // Branch either way.
4242     // NaN case is less traveled, which makes all the difference.
4243     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4244     Node *opt_isnan = _gvn.transform(ifisnan);
4245     assert( opt_isnan->is_If(), "Expect an IfNode");
4246     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4247     Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4248 
4249     set_control(iftrue);
4250 
4251     static const jlong nan_bits = CONST64(0x7ff8000000000000);
4252     Node *slow_result = longcon(nan_bits); // return NaN
4253     phi->init_req(1, _gvn.transform( slow_result ));
4254     r->init_req(1, iftrue);
4255 
4256     // Else fall through
4257     Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4258     set_control(iffalse);
4259 
4260     phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4261     r->init_req(2, iffalse);
4262 
4263     // Post merge
4264     set_control(_gvn.transform(r));
4265     record_for_igvn(r);
4266 
4267     C->set_has_split_ifs(true); // Has chance for split-if optimization
4268     result = phi;
4269     assert(result->bottom_type()->isa_long(), "must be");
4270     break;
4271   }
4272 
4273   case vmIntrinsics::_floatToIntBits: {
4274     // two paths (plus control) merge in a wood
4275     RegionNode *r = new (C) RegionNode(3);
4276     Node *phi = new (C) PhiNode(r, TypeInt::INT);
4277 
4278     Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4279     // Build the boolean node
4280     Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4281 
4282     // Branch either way.
4283     // NaN case is less traveled, which makes all the difference.
4284     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4285     Node *opt_isnan = _gvn.transform(ifisnan);
4286     assert( opt_isnan->is_If(), "Expect an IfNode");
4287     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4288     Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4289 
4290     set_control(iftrue);
4291 
4292     static const jint nan_bits = 0x7fc00000;
4293     Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4294     phi->init_req(1, _gvn.transform( slow_result ));
4295     r->init_req(1, iftrue);
4296 
4297     // Else fall through
4298     Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4299     set_control(iffalse);
4300 
4301     phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4302     r->init_req(2, iffalse);
4303 
4304     // Post merge
4305     set_control(_gvn.transform(r));
4306     record_for_igvn(r);
4307 
4308     C->set_has_split_ifs(true); // Has chance for split-if optimization
4309     result = phi;
4310     assert(result->bottom_type()->isa_int(), "must be");
4311     break;
4312   }
4313 
4314   default:
4315     fatal_unexpected_iid(id);
4316     break;
4317   }
4318   set_result(_gvn.transform(result));
4319   return true;
4320 }
4321 
4322 #ifdef _LP64
4323 #define XTOP ,top() /*additional argument*/
4324 #else  //_LP64
4325 #define XTOP        /*no additional argument*/
4326 #endif //_LP64
4327 
4328 //----------------------inline_unsafe_copyMemory-------------------------
4329 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4330 bool LibraryCallKit::inline_unsafe_copyMemory() {
4331   if (callee()->is_static())  return false;  // caller must have the capability!
4332   null_check_receiver();  // null-check receiver
4333   if (stopped())  return true;
4334 
4335   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
4336 
4337   Node* src_ptr =         argument(1);   // type: oop
4338   Node* src_off = ConvL2X(argument(2));  // type: long
4339   Node* dst_ptr =         argument(4);   // type: oop
4340   Node* dst_off = ConvL2X(argument(5));  // type: long
4341   Node* size    = ConvL2X(argument(7));  // type: long
4342 
4343   assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4344          "fieldOffset must be byte-scaled");
4345 
4346   Node* src = make_unsafe_address(src_ptr, src_off);
4347   Node* dst = make_unsafe_address(dst_ptr, dst_off);
4348 
4349   // Conservatively insert a memory barrier on all memory slices.
4350   // Do not let writes of the copy source or destination float below the copy.
4351   insert_mem_bar(Op_MemBarCPUOrder);
4352 
4353   // Call it.  Note that the length argument is not scaled.
4354   make_runtime_call(RC_LEAF|RC_NO_FP,
4355                     OptoRuntime::fast_arraycopy_Type(),
4356                     StubRoutines::unsafe_arraycopy(),
4357                     "unsafe_arraycopy",
4358                     TypeRawPtr::BOTTOM,
4359                     src, dst, size XTOP);
4360 
4361   // Do not let reads of the copy destination float above the copy.
4362   insert_mem_bar(Op_MemBarCPUOrder);
4363 
4364   return true;
4365 }
4366 
4367 //------------------------clone_coping-----------------------------------
4368 // Helper function for inline_native_clone.
4369 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4370   assert(obj_size != NULL, "");
4371   Node* raw_obj = alloc_obj->in(1);
4372   assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4373 
4374   AllocateNode* alloc = NULL;
4375   if (ReduceBulkZeroing) {
4376     // We will be completely responsible for initializing this object -
4377     // mark Initialize node as complete.
4378     alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4379     // The object was just allocated - there should be no any stores!
4380     guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4381     // Mark as complete_with_arraycopy so that on AllocateNode
4382     // expansion, we know this AllocateNode is initialized by an array
4383     // copy and a StoreStore barrier exists after the array copy.
4384     alloc->initialization()->set_complete_with_arraycopy();
4385   }
4386 
4387   // Copy the fastest available way.
4388   // TODO: generate fields copies for small objects instead.
4389   Node* src  = obj;
4390   Node* dest = alloc_obj;
4391   Node* size = _gvn.transform(obj_size);
4392 
4393   // Exclude the header but include array length to copy by 8 bytes words.
4394   // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4395   int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4396                             instanceOopDesc::base_offset_in_bytes();
4397   // base_off:
4398   // 8  - 32-bit VM
4399   // 12 - 64-bit VM, compressed klass
4400   // 16 - 64-bit VM, normal klass
4401   if (base_off % BytesPerLong != 0) {
4402     assert(UseCompressedClassPointers, "");
4403     if (is_array) {
4404       // Exclude length to copy by 8 bytes words.
4405       base_off += sizeof(int);
4406     } else {
4407       // Include klass to copy by 8 bytes words.
4408       base_off = instanceOopDesc::klass_offset_in_bytes();
4409     }
4410     assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4411   }
4412   src  = basic_plus_adr(src,  base_off);
4413   dest = basic_plus_adr(dest, base_off);
4414 
4415   // Compute the length also, if needed:
4416   Node* countx = size;
4417   countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4418   countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4419 
4420   const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4421   bool disjoint_bases = true;
4422   generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4423                                src, NULL, dest, NULL, countx,
4424                                /*dest_uninitialized*/true);
4425 
4426   // If necessary, emit some card marks afterwards.  (Non-arrays only.)
4427   if (card_mark) {
4428     assert(!is_array, "");
4429     // Put in store barrier for any and all oops we are sticking
4430     // into this object.  (We could avoid this if we could prove
4431     // that the object type contains no oop fields at all.)
4432     Node* no_particular_value = NULL;
4433     Node* no_particular_field = NULL;
4434     int raw_adr_idx = Compile::AliasIdxRaw;
4435     post_barrier(control(),
4436                  memory(raw_adr_type),
4437                  alloc_obj,
4438                  no_particular_field,
4439                  raw_adr_idx,
4440                  no_particular_value,
4441                  T_OBJECT,
4442                  false);
4443   }
4444 
4445   // Do not let reads from the cloned object float above the arraycopy.
4446   if (alloc != NULL) {
4447     // Do not let stores that initialize this object be reordered with
4448     // a subsequent store that would make this object accessible by
4449     // other threads.
4450     // Record what AllocateNode this StoreStore protects so that
4451     // escape analysis can go from the MemBarStoreStoreNode to the
4452     // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4453     // based on the escape status of the AllocateNode.
4454     insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4455   } else {
4456     insert_mem_bar(Op_MemBarCPUOrder);
4457   }
4458 }
4459 
4460 //------------------------inline_native_clone----------------------------
4461 // protected native Object java.lang.Object.clone();
4462 //
4463 // Here are the simple edge cases:
4464 //  null receiver => normal trap
4465 //  virtual and clone was overridden => slow path to out-of-line clone
4466 //  not cloneable or finalizer => slow path to out-of-line Object.clone
4467 //
4468 // The general case has two steps, allocation and copying.
4469 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4470 //
4471 // Copying also has two cases, oop arrays and everything else.
4472 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4473 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4474 //
4475 // These steps fold up nicely if and when the cloned object's klass
4476 // can be sharply typed as an object array, a type array, or an instance.
4477 //
4478 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4479   PhiNode* result_val;
4480 
4481   // Set the reexecute bit for the interpreter to reexecute
4482   // the bytecode that invokes Object.clone if deoptimization happens.
4483   { PreserveReexecuteState preexecs(this);
4484     jvms()->set_should_reexecute(true);
4485 
4486     Node* obj = null_check_receiver();
4487     if (stopped())  return true;
4488 
4489     Node* obj_klass = load_object_klass(obj);
4490     const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4491     const TypeOopPtr*   toop   = ((tklass != NULL)
4492                                 ? tklass->as_instance_type()
4493                                 : TypeInstPtr::NOTNULL);
4494 
4495     // Conservatively insert a memory barrier on all memory slices.
4496     // Do not let writes into the original float below the clone.
4497     insert_mem_bar(Op_MemBarCPUOrder);
4498 
4499     // paths into result_reg:
4500     enum {
4501       _slow_path = 1,     // out-of-line call to clone method (virtual or not)
4502       _objArray_path,     // plain array allocation, plus arrayof_oop_arraycopy
4503       _array_path,        // plain array allocation, plus arrayof_long_arraycopy
4504       _instance_path,     // plain instance allocation, plus arrayof_long_arraycopy
4505       PATH_LIMIT
4506     };
4507     RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4508     result_val             = new(C) PhiNode(result_reg,
4509                                             TypeInstPtr::NOTNULL);
4510     PhiNode*    result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4511     PhiNode*    result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4512                                             TypePtr::BOTTOM);
4513     record_for_igvn(result_reg);
4514 
4515     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4516     int raw_adr_idx = Compile::AliasIdxRaw;
4517 
4518     Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4519     if (array_ctl != NULL) {
4520       // It's an array.
4521       PreserveJVMState pjvms(this);
4522       set_control(array_ctl);
4523       Node* obj_length = load_array_length(obj);
4524       Node* obj_size  = NULL;
4525       Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);  // no arguments to push
4526 
4527       if (!use_ReduceInitialCardMarks()) {
4528         // If it is an oop array, it requires very special treatment,
4529         // because card marking is required on each card of the array.
4530         Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4531         if (is_obja != NULL) {
4532           PreserveJVMState pjvms2(this);
4533           set_control(is_obja);
4534           // Generate a direct call to the right arraycopy function(s).
4535           bool disjoint_bases = true;
4536           bool length_never_negative = true;
4537           generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4538                              obj, intcon(0), alloc_obj, intcon(0),
4539                              obj_length,
4540                              disjoint_bases, length_never_negative);
4541           result_reg->init_req(_objArray_path, control());
4542           result_val->init_req(_objArray_path, alloc_obj);
4543           result_i_o ->set_req(_objArray_path, i_o());
4544           result_mem ->set_req(_objArray_path, reset_memory());
4545         }
4546       }
4547       // Otherwise, there are no card marks to worry about.
4548       // (We can dispense with card marks if we know the allocation
4549       //  comes out of eden (TLAB)...  In fact, ReduceInitialCardMarks
4550       //  causes the non-eden paths to take compensating steps to
4551       //  simulate a fresh allocation, so that no further
4552       //  card marks are required in compiled code to initialize
4553       //  the object.)
4554 
4555       if (!stopped()) {
4556         copy_to_clone(obj, alloc_obj, obj_size, true, false);
4557 
4558         // Present the results of the copy.
4559         result_reg->init_req(_array_path, control());
4560         result_val->init_req(_array_path, alloc_obj);
4561         result_i_o ->set_req(_array_path, i_o());
4562         result_mem ->set_req(_array_path, reset_memory());
4563       }
4564     }
4565 
4566     // We only go to the instance fast case code if we pass a number of guards.
4567     // The paths which do not pass are accumulated in the slow_region.
4568     RegionNode* slow_region = new (C) RegionNode(1);
4569     record_for_igvn(slow_region);
4570     if (!stopped()) {
4571       // It's an instance (we did array above).  Make the slow-path tests.
4572       // If this is a virtual call, we generate a funny guard.  We grab
4573       // the vtable entry corresponding to clone() from the target object.
4574       // If the target method which we are calling happens to be the
4575       // Object clone() method, we pass the guard.  We do not need this
4576       // guard for non-virtual calls; the caller is known to be the native
4577       // Object clone().
4578       if (is_virtual) {
4579         generate_virtual_guard(obj_klass, slow_region);
4580       }
4581 
4582       // The object must be cloneable and must not have a finalizer.
4583       // Both of these conditions may be checked in a single test.
4584       // We could optimize the cloneable test further, but we don't care.
4585       generate_access_flags_guard(obj_klass,
4586                                   // Test both conditions:
4587                                   JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4588                                   // Must be cloneable but not finalizer:
4589                                   JVM_ACC_IS_CLONEABLE,
4590                                   slow_region);
4591     }
4592 
4593     if (!stopped()) {
4594       // It's an instance, and it passed the slow-path tests.
4595       PreserveJVMState pjvms(this);
4596       Node* obj_size  = NULL;
4597       Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4598 
4599       copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4600 
4601       // Present the results of the slow call.
4602       result_reg->init_req(_instance_path, control());
4603       result_val->init_req(_instance_path, alloc_obj);
4604       result_i_o ->set_req(_instance_path, i_o());
4605       result_mem ->set_req(_instance_path, reset_memory());
4606     }
4607 
4608     // Generate code for the slow case.  We make a call to clone().
4609     set_control(_gvn.transform(slow_region));
4610     if (!stopped()) {
4611       PreserveJVMState pjvms(this);
4612       CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4613       Node* slow_result = set_results_for_java_call(slow_call);
4614       // this->control() comes from set_results_for_java_call
4615       result_reg->init_req(_slow_path, control());
4616       result_val->init_req(_slow_path, slow_result);
4617       result_i_o ->set_req(_slow_path, i_o());
4618       result_mem ->set_req(_slow_path, reset_memory());
4619     }
4620 
4621     // Return the combined state.
4622     set_control(    _gvn.transform(result_reg));
4623     set_i_o(        _gvn.transform(result_i_o));
4624     set_all_memory( _gvn.transform(result_mem));
4625   } // original reexecute is set back here
4626 
4627   set_result(_gvn.transform(result_val));
4628   return true;
4629 }
4630 
4631 //------------------------------basictype2arraycopy----------------------------
4632 address LibraryCallKit::basictype2arraycopy(BasicType t,
4633                                             Node* src_offset,
4634                                             Node* dest_offset,
4635                                             bool disjoint_bases,
4636                                             const char* &name,
4637                                             bool dest_uninitialized) {
4638   const TypeInt* src_offset_inttype  = gvn().find_int_type(src_offset);;
4639   const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4640 
4641   bool aligned = false;
4642   bool disjoint = disjoint_bases;
4643 
4644   // if the offsets are the same, we can treat the memory regions as
4645   // disjoint, because either the memory regions are in different arrays,
4646   // or they are identical (which we can treat as disjoint.)  We can also
4647   // treat a copy with a destination index  less that the source index
4648   // as disjoint since a low->high copy will work correctly in this case.
4649   if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4650       dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4651     // both indices are constants
4652     int s_offs = src_offset_inttype->get_con();
4653     int d_offs = dest_offset_inttype->get_con();
4654     int element_size = type2aelembytes(t);
4655     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4656               ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4657     if (s_offs >= d_offs)  disjoint = true;
4658   } else if (src_offset == dest_offset && src_offset != NULL) {
4659     // This can occur if the offsets are identical non-constants.
4660     disjoint = true;
4661   }
4662 
4663   return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4664 }
4665 
4666 
4667 //------------------------------inline_arraycopy-----------------------
4668 // public static native void java.lang.System.arraycopy(Object src,  int  srcPos,
4669 //                                                      Object dest, int destPos,
4670 //                                                      int length);
4671 bool LibraryCallKit::inline_arraycopy() {
4672   // Get the arguments.
4673   Node* src         = argument(0);  // type: oop
4674   Node* src_offset  = argument(1);  // type: int
4675   Node* dest        = argument(2);  // type: oop
4676   Node* dest_offset = argument(3);  // type: int
4677   Node* length      = argument(4);  // type: int
4678 
4679   // Compile time checks.  If any of these checks cannot be verified at compile time,
4680   // we do not make a fast path for this call.  Instead, we let the call remain as it
4681   // is.  The checks we choose to mandate at compile time are:
4682   //
4683   // (1) src and dest are arrays.
4684   const Type* src_type  = src->Value(&_gvn);
4685   const Type* dest_type = dest->Value(&_gvn);
4686   const TypeAryPtr* top_src  = src_type->isa_aryptr();
4687   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4688   if (top_src  == NULL || top_src->klass()  == NULL ||
4689       top_dest == NULL || top_dest->klass() == NULL) {
4690     // Conservatively insert a memory barrier on all memory slices.
4691     // Do not let writes into the source float below the arraycopy.
4692     insert_mem_bar(Op_MemBarCPUOrder);
4693 
4694     // Call StubRoutines::generic_arraycopy stub.
4695     generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4696                        src, src_offset, dest, dest_offset, length);
4697 
4698     // Do not let reads from the destination float above the arraycopy.
4699     // Since we cannot type the arrays, we don't know which slices
4700     // might be affected.  We could restrict this barrier only to those
4701     // memory slices which pertain to array elements--but don't bother.
4702     if (!InsertMemBarAfterArraycopy)
4703       // (If InsertMemBarAfterArraycopy, there is already one in place.)
4704       insert_mem_bar(Op_MemBarCPUOrder);
4705     return true;
4706   }
4707 
4708   // (2) src and dest arrays must have elements of the same BasicType
4709   // Figure out the size and type of the elements we will be copying.
4710   BasicType src_elem  =  top_src->klass()->as_array_klass()->element_type()->basic_type();
4711   BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4712   if (src_elem  == T_ARRAY)  src_elem  = T_OBJECT;
4713   if (dest_elem == T_ARRAY)  dest_elem = T_OBJECT;
4714 
4715   if (src_elem != dest_elem || dest_elem == T_VOID) {
4716     // The component types are not the same or are not recognized.  Punt.
4717     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4718     generate_slow_arraycopy(TypePtr::BOTTOM,
4719                             src, src_offset, dest, dest_offset, length,
4720                             /*dest_uninitialized*/false);
4721     return true;
4722   }
4723 
4724   //---------------------------------------------------------------------------
4725   // We will make a fast path for this call to arraycopy.
4726 
4727   // We have the following tests left to perform:
4728   //
4729   // (3) src and dest must not be null.
4730   // (4) src_offset must not be negative.
4731   // (5) dest_offset must not be negative.
4732   // (6) length must not be negative.
4733   // (7) src_offset + length must not exceed length of src.
4734   // (8) dest_offset + length must not exceed length of dest.
4735   // (9) each element of an oop array must be assignable
4736 
4737   RegionNode* slow_region = new (C) RegionNode(1);
4738   record_for_igvn(slow_region);
4739 
4740   // (3) operands must not be null
4741   // We currently perform our null checks with the null_check routine.
4742   // This means that the null exceptions will be reported in the caller
4743   // rather than (correctly) reported inside of the native arraycopy call.
4744   // This should be corrected, given time.  We do our null check with the
4745   // stack pointer restored.
4746   src  = null_check(src,  T_ARRAY);
4747   dest = null_check(dest, T_ARRAY);
4748 
4749   // (4) src_offset must not be negative.
4750   generate_negative_guard(src_offset, slow_region);
4751 
4752   // (5) dest_offset must not be negative.
4753   generate_negative_guard(dest_offset, slow_region);
4754 
4755   // (6) length must not be negative (moved to generate_arraycopy()).
4756   // generate_negative_guard(length, slow_region);
4757 
4758   // (7) src_offset + length must not exceed length of src.
4759   generate_limit_guard(src_offset, length,
4760                        load_array_length(src),
4761                        slow_region);
4762 
4763   // (8) dest_offset + length must not exceed length of dest.
4764   generate_limit_guard(dest_offset, length,
4765                        load_array_length(dest),
4766                        slow_region);
4767 
4768   // (9) each element of an oop array must be assignable
4769   // The generate_arraycopy subroutine checks this.
4770 
4771   // This is where the memory effects are placed:
4772   const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4773   generate_arraycopy(adr_type, dest_elem,
4774                      src, src_offset, dest, dest_offset, length,
4775                      false, false, slow_region);
4776 
4777   return true;
4778 }
4779 
4780 //-----------------------------generate_arraycopy----------------------
4781 // Generate an optimized call to arraycopy.
4782 // Caller must guard against non-arrays.
4783 // Caller must determine a common array basic-type for both arrays.
4784 // Caller must validate offsets against array bounds.
4785 // The slow_region has already collected guard failure paths
4786 // (such as out of bounds length or non-conformable array types).
4787 // The generated code has this shape, in general:
4788 //
4789 //     if (length == 0)  return   // via zero_path
4790 //     slowval = -1
4791 //     if (types unknown) {
4792 //       slowval = call generic copy loop
4793 //       if (slowval == 0)  return  // via checked_path
4794 //     } else if (indexes in bounds) {
4795 //       if ((is object array) && !(array type check)) {
4796 //         slowval = call checked copy loop
4797 //         if (slowval == 0)  return  // via checked_path
4798 //       } else {
4799 //         call bulk copy loop
4800 //         return  // via fast_path
4801 //       }
4802 //     }
4803 //     // adjust params for remaining work:
4804 //     if (slowval != -1) {
4805 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4806 //     }
4807 //   slow_region:
4808 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
4809 //     return  // via slow_call_path
4810 //
4811 // This routine is used from several intrinsics:  System.arraycopy,
4812 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4813 //
4814 void
4815 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4816                                    BasicType basic_elem_type,
4817                                    Node* src,  Node* src_offset,
4818                                    Node* dest, Node* dest_offset,
4819                                    Node* copy_length,
4820                                    bool disjoint_bases,
4821                                    bool length_never_negative,
4822                                    RegionNode* slow_region) {
4823 
4824   if (slow_region == NULL) {
4825     slow_region = new(C) RegionNode(1);
4826     record_for_igvn(slow_region);
4827   }
4828 
4829   Node* original_dest      = dest;
4830   AllocateArrayNode* alloc = NULL;  // used for zeroing, if needed
4831   bool  dest_uninitialized = false;
4832 
4833   // See if this is the initialization of a newly-allocated array.
4834   // If so, we will take responsibility here for initializing it to zero.
4835   // (Note:  Because tightly_coupled_allocation performs checks on the
4836   // out-edges of the dest, we need to avoid making derived pointers
4837   // from it until we have checked its uses.)
4838   if (ReduceBulkZeroing
4839       && !ZeroTLAB              // pointless if already zeroed
4840       && basic_elem_type != T_CONFLICT // avoid corner case
4841       && !src->eqv_uncast(dest)
4842       && ((alloc = tightly_coupled_allocation(dest, slow_region))
4843           != NULL)
4844       && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4845       && alloc->maybe_set_complete(&_gvn)) {
4846     // "You break it, you buy it."
4847     InitializeNode* init = alloc->initialization();
4848     assert(init->is_complete(), "we just did this");
4849     init->set_complete_with_arraycopy();
4850     assert(dest->is_CheckCastPP(), "sanity");
4851     assert(dest->in(0)->in(0) == init, "dest pinned");
4852     adr_type = TypeRawPtr::BOTTOM;  // all initializations are into raw memory
4853     // From this point on, every exit path is responsible for
4854     // initializing any non-copied parts of the object to zero.
4855     // Also, if this flag is set we make sure that arraycopy interacts properly
4856     // with G1, eliding pre-barriers. See CR 6627983.
4857     dest_uninitialized = true;
4858   } else {
4859     // No zeroing elimination here.
4860     alloc             = NULL;
4861     //original_dest   = dest;
4862     //dest_uninitialized = false;
4863   }
4864 
4865   // Results are placed here:
4866   enum { fast_path        = 1,  // normal void-returning assembly stub
4867          checked_path     = 2,  // special assembly stub with cleanup
4868          slow_call_path   = 3,  // something went wrong; call the VM
4869          zero_path        = 4,  // bypass when length of copy is zero
4870          bcopy_path       = 5,  // copy primitive array by 64-bit blocks
4871          PATH_LIMIT       = 6
4872   };
4873   RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4874   PhiNode*    result_i_o    = new(C) PhiNode(result_region, Type::ABIO);
4875   PhiNode*    result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4876   record_for_igvn(result_region);
4877   _gvn.set_type_bottom(result_i_o);
4878   _gvn.set_type_bottom(result_memory);
4879   assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4880 
4881   // The slow_control path:
4882   Node* slow_control;
4883   Node* slow_i_o = i_o();
4884   Node* slow_mem = memory(adr_type);
4885   debug_only(slow_control = (Node*) badAddress);
4886 
4887   // Checked control path:
4888   Node* checked_control = top();
4889   Node* checked_mem     = NULL;
4890   Node* checked_i_o     = NULL;
4891   Node* checked_value   = NULL;
4892 
4893   if (basic_elem_type == T_CONFLICT) {
4894     assert(!dest_uninitialized, "");
4895     Node* cv = generate_generic_arraycopy(adr_type,
4896                                           src, src_offset, dest, dest_offset,
4897                                           copy_length, dest_uninitialized);
4898     if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
4899     checked_control = control();
4900     checked_i_o     = i_o();
4901     checked_mem     = memory(adr_type);
4902     checked_value   = cv;
4903     set_control(top());         // no fast path
4904   }
4905 
4906   Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4907   if (not_pos != NULL) {
4908     PreserveJVMState pjvms(this);
4909     set_control(not_pos);
4910 
4911     // (6) length must not be negative.
4912     if (!length_never_negative) {
4913       generate_negative_guard(copy_length, slow_region);
4914     }
4915 
4916     // copy_length is 0.
4917     if (!stopped() && dest_uninitialized) {
4918       Node* dest_length = alloc->in(AllocateNode::ALength);
4919       if (copy_length->eqv_uncast(dest_length)
4920           || _gvn.find_int_con(dest_length, 1) <= 0) {
4921         // There is no zeroing to do. No need for a secondary raw memory barrier.
4922       } else {
4923         // Clear the whole thing since there are no source elements to copy.
4924         generate_clear_array(adr_type, dest, basic_elem_type,
4925                              intcon(0), NULL,
4926                              alloc->in(AllocateNode::AllocSize));
4927         // Use a secondary InitializeNode as raw memory barrier.
4928         // Currently it is needed only on this path since other
4929         // paths have stub or runtime calls as raw memory barriers.
4930         InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4931                                                        Compile::AliasIdxRaw,
4932                                                        top())->as_Initialize();
4933         init->set_complete(&_gvn);  // (there is no corresponding AllocateNode)
4934       }
4935     }
4936 
4937     // Present the results of the fast call.
4938     result_region->init_req(zero_path, control());
4939     result_i_o   ->init_req(zero_path, i_o());
4940     result_memory->init_req(zero_path, memory(adr_type));
4941   }
4942 
4943   if (!stopped() && dest_uninitialized) {
4944     // We have to initialize the *uncopied* part of the array to zero.
4945     // The copy destination is the slice dest[off..off+len].  The other slices
4946     // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4947     Node* dest_size   = alloc->in(AllocateNode::AllocSize);
4948     Node* dest_length = alloc->in(AllocateNode::ALength);
4949     Node* dest_tail   = _gvn.transform(new(C) AddINode(dest_offset,
4950                                                           copy_length));
4951 
4952     // If there is a head section that needs zeroing, do it now.
4953     if (find_int_con(dest_offset, -1) != 0) {
4954       generate_clear_array(adr_type, dest, basic_elem_type,
4955                            intcon(0), dest_offset,
4956                            NULL);
4957     }
4958 
4959     // Next, perform a dynamic check on the tail length.
4960     // It is often zero, and we can win big if we prove this.
4961     // There are two wins:  Avoid generating the ClearArray
4962     // with its attendant messy index arithmetic, and upgrade
4963     // the copy to a more hardware-friendly word size of 64 bits.
4964     Node* tail_ctl = NULL;
4965     if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4966       Node* cmp_lt   = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
4967       Node* bol_lt   = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
4968       tail_ctl = generate_slow_guard(bol_lt, NULL);
4969       assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4970     }
4971 
4972     // At this point, let's assume there is no tail.
4973     if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4974       // There is no tail.  Try an upgrade to a 64-bit copy.
4975       bool didit = false;
4976       { PreserveJVMState pjvms(this);
4977         didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4978                                          src, src_offset, dest, dest_offset,
4979                                          dest_size, dest_uninitialized);
4980         if (didit) {
4981           // Present the results of the block-copying fast call.
4982           result_region->init_req(bcopy_path, control());
4983           result_i_o   ->init_req(bcopy_path, i_o());
4984           result_memory->init_req(bcopy_path, memory(adr_type));
4985         }
4986       }
4987       if (didit)
4988         set_control(top());     // no regular fast path
4989     }
4990 
4991     // Clear the tail, if any.
4992     if (tail_ctl != NULL) {
4993       Node* notail_ctl = stopped() ? NULL : control();
4994       set_control(tail_ctl);
4995       if (notail_ctl == NULL) {
4996         generate_clear_array(adr_type, dest, basic_elem_type,
4997                              dest_tail, NULL,
4998                              dest_size);
4999       } else {
5000         // Make a local merge.
5001         Node* done_ctl = new(C) RegionNode(3);
5002         Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5003         done_ctl->init_req(1, notail_ctl);
5004         done_mem->init_req(1, memory(adr_type));
5005         generate_clear_array(adr_type, dest, basic_elem_type,
5006                              dest_tail, NULL,
5007                              dest_size);
5008         done_ctl->init_req(2, control());
5009         done_mem->init_req(2, memory(adr_type));
5010         set_control( _gvn.transform(done_ctl));
5011         set_memory(  _gvn.transform(done_mem), adr_type );
5012       }
5013     }
5014   }
5015 
5016   BasicType copy_type = basic_elem_type;
5017   assert(basic_elem_type != T_ARRAY, "caller must fix this");
5018   if (!stopped() && copy_type == T_OBJECT) {
5019     // If src and dest have compatible element types, we can copy bits.
5020     // Types S[] and D[] are compatible if D is a supertype of S.
5021     //
5022     // If they are not, we will use checked_oop_disjoint_arraycopy,
5023     // which performs a fast optimistic per-oop check, and backs off
5024     // further to JVM_ArrayCopy on the first per-oop check that fails.
5025     // (Actually, we don't move raw bits only; the GC requires card marks.)
5026 
5027     // Get the Klass* for both src and dest
5028     Node* src_klass  = load_object_klass(src);
5029     Node* dest_klass = load_object_klass(dest);
5030 
5031     // Generate the subtype check.
5032     // This might fold up statically, or then again it might not.
5033     //
5034     // Non-static example:  Copying List<String>.elements to a new String[].
5035     // The backing store for a List<String> is always an Object[],
5036     // but its elements are always type String, if the generic types
5037     // are correct at the source level.
5038     //
5039     // Test S[] against D[], not S against D, because (probably)
5040     // the secondary supertype cache is less busy for S[] than S.
5041     // This usually only matters when D is an interface.
5042     Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5043     // Plug failing path into checked_oop_disjoint_arraycopy
5044     if (not_subtype_ctrl != top()) {
5045       PreserveJVMState pjvms(this);
5046       set_control(not_subtype_ctrl);
5047       // (At this point we can assume disjoint_bases, since types differ.)
5048       int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5049       Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5050       Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5051       Node* dest_elem_klass = _gvn.transform(n1);
5052       Node* cv = generate_checkcast_arraycopy(adr_type,
5053                                               dest_elem_klass,
5054                                               src, src_offset, dest, dest_offset,
5055                                               ConvI2X(copy_length), dest_uninitialized);
5056       if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
5057       checked_control = control();
5058       checked_i_o     = i_o();
5059       checked_mem     = memory(adr_type);
5060       checked_value   = cv;
5061     }
5062     // At this point we know we do not need type checks on oop stores.
5063 
5064     // Let's see if we need card marks:
5065     if (alloc != NULL && use_ReduceInitialCardMarks()) {
5066       // If we do not need card marks, copy using the jint or jlong stub.
5067       copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5068       assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5069              "sizes agree");
5070     }
5071   }
5072 
5073   if (!stopped()) {
5074     // Generate the fast path, if possible.
5075     PreserveJVMState pjvms(this);
5076     generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5077                                  src, src_offset, dest, dest_offset,
5078                                  ConvI2X(copy_length), dest_uninitialized);
5079 
5080     // Present the results of the fast call.
5081     result_region->init_req(fast_path, control());
5082     result_i_o   ->init_req(fast_path, i_o());
5083     result_memory->init_req(fast_path, memory(adr_type));
5084   }
5085 
5086   // Here are all the slow paths up to this point, in one bundle:
5087   slow_control = top();
5088   if (slow_region != NULL)
5089     slow_control = _gvn.transform(slow_region);
5090   DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5091 
5092   set_control(checked_control);
5093   if (!stopped()) {
5094     // Clean up after the checked call.
5095     // The returned value is either 0 or -1^K,
5096     // where K = number of partially transferred array elements.
5097     Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5098     Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5099     IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5100 
5101     // If it is 0, we are done, so transfer to the end.
5102     Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5103     result_region->init_req(checked_path, checks_done);
5104     result_i_o   ->init_req(checked_path, checked_i_o);
5105     result_memory->init_req(checked_path, checked_mem);
5106 
5107     // If it is not zero, merge into the slow call.
5108     set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5109     RegionNode* slow_reg2 = new(C) RegionNode(3);
5110     PhiNode*    slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5111     PhiNode*    slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5112     record_for_igvn(slow_reg2);
5113     slow_reg2  ->init_req(1, slow_control);
5114     slow_i_o2  ->init_req(1, slow_i_o);
5115     slow_mem2  ->init_req(1, slow_mem);
5116     slow_reg2  ->init_req(2, control());
5117     slow_i_o2  ->init_req(2, checked_i_o);
5118     slow_mem2  ->init_req(2, checked_mem);
5119 
5120     slow_control = _gvn.transform(slow_reg2);
5121     slow_i_o     = _gvn.transform(slow_i_o2);
5122     slow_mem     = _gvn.transform(slow_mem2);
5123 
5124     if (alloc != NULL) {
5125       // We'll restart from the very beginning, after zeroing the whole thing.
5126       // This can cause double writes, but that's OK since dest is brand new.
5127       // So we ignore the low 31 bits of the value returned from the stub.
5128     } else {
5129       // We must continue the copy exactly where it failed, or else
5130       // another thread might see the wrong number of writes to dest.
5131       Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5132       Node* slow_offset    = new(C) PhiNode(slow_reg2, TypeInt::INT);
5133       slow_offset->init_req(1, intcon(0));
5134       slow_offset->init_req(2, checked_offset);
5135       slow_offset  = _gvn.transform(slow_offset);
5136 
5137       // Adjust the arguments by the conditionally incoming offset.
5138       Node* src_off_plus  = _gvn.transform(new(C) AddINode(src_offset,  slow_offset));
5139       Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5140       Node* length_minus  = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5141 
5142       // Tweak the node variables to adjust the code produced below:
5143       src_offset  = src_off_plus;
5144       dest_offset = dest_off_plus;
5145       copy_length = length_minus;
5146     }
5147   }
5148 
5149   set_control(slow_control);
5150   if (!stopped()) {
5151     // Generate the slow path, if needed.
5152     PreserveJVMState pjvms(this);   // replace_in_map may trash the map
5153 
5154     set_memory(slow_mem, adr_type);
5155     set_i_o(slow_i_o);
5156 
5157     if (dest_uninitialized) {
5158       generate_clear_array(adr_type, dest, basic_elem_type,
5159                            intcon(0), NULL,
5160                            alloc->in(AllocateNode::AllocSize));
5161     }
5162 
5163     generate_slow_arraycopy(adr_type,
5164                             src, src_offset, dest, dest_offset,
5165                             copy_length, /*dest_uninitialized*/false);
5166 
5167     result_region->init_req(slow_call_path, control());
5168     result_i_o   ->init_req(slow_call_path, i_o());
5169     result_memory->init_req(slow_call_path, memory(adr_type));
5170   }
5171 
5172   // Remove unused edges.
5173   for (uint i = 1; i < result_region->req(); i++) {
5174     if (result_region->in(i) == NULL)
5175       result_region->init_req(i, top());
5176   }
5177 
5178   // Finished; return the combined state.
5179   set_control( _gvn.transform(result_region));
5180   set_i_o(     _gvn.transform(result_i_o)    );
5181   set_memory(  _gvn.transform(result_memory), adr_type );
5182 
5183   // The memory edges above are precise in order to model effects around
5184   // array copies accurately to allow value numbering of field loads around
5185   // arraycopy.  Such field loads, both before and after, are common in Java
5186   // collections and similar classes involving header/array data structures.
5187   //
5188   // But with low number of register or when some registers are used or killed
5189   // by arraycopy calls it causes registers spilling on stack. See 6544710.
5190   // The next memory barrier is added to avoid it. If the arraycopy can be
5191   // optimized away (which it can, sometimes) then we can manually remove
5192   // the membar also.
5193   //
5194   // Do not let reads from the cloned object float above the arraycopy.
5195   if (alloc != NULL) {
5196     // Do not let stores that initialize this object be reordered with
5197     // a subsequent store that would make this object accessible by
5198     // other threads.
5199     // Record what AllocateNode this StoreStore protects so that
5200     // escape analysis can go from the MemBarStoreStoreNode to the
5201     // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5202     // based on the escape status of the AllocateNode.
5203     insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5204   } else if (InsertMemBarAfterArraycopy)
5205     insert_mem_bar(Op_MemBarCPUOrder);
5206 }
5207 
5208 
5209 // Helper function which determines if an arraycopy immediately follows
5210 // an allocation, with no intervening tests or other escapes for the object.
5211 AllocateArrayNode*
5212 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5213                                            RegionNode* slow_region) {
5214   if (stopped())             return NULL;  // no fast path
5215   if (C->AliasLevel() == 0)  return NULL;  // no MergeMems around
5216 
5217   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5218   if (alloc == NULL)  return NULL;
5219 
5220   Node* rawmem = memory(Compile::AliasIdxRaw);
5221   // Is the allocation's memory state untouched?
5222   if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5223     // Bail out if there have been raw-memory effects since the allocation.
5224     // (Example:  There might have been a call or safepoint.)
5225     return NULL;
5226   }
5227   rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5228   if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5229     return NULL;
5230   }
5231 
5232   // There must be no unexpected observers of this allocation.
5233   for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5234     Node* obs = ptr->fast_out(i);
5235     if (obs != this->map()) {
5236       return NULL;
5237     }
5238   }
5239 
5240   // This arraycopy must unconditionally follow the allocation of the ptr.
5241   Node* alloc_ctl = ptr->in(0);
5242   assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5243 
5244   Node* ctl = control();
5245   while (ctl != alloc_ctl) {
5246     // There may be guards which feed into the slow_region.
5247     // Any other control flow means that we might not get a chance
5248     // to finish initializing the allocated object.
5249     if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5250       IfNode* iff = ctl->in(0)->as_If();
5251       Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5252       assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5253       if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5254         ctl = iff->in(0);       // This test feeds the known slow_region.
5255         continue;
5256       }
5257       // One more try:  Various low-level checks bottom out in
5258       // uncommon traps.  If the debug-info of the trap omits
5259       // any reference to the allocation, as we've already
5260       // observed, then there can be no objection to the trap.
5261       bool found_trap = false;
5262       for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5263         Node* obs = not_ctl->fast_out(j);
5264         if (obs->in(0) == not_ctl && obs->is_Call() &&
5265             (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5266           found_trap = true; break;
5267         }
5268       }
5269       if (found_trap) {
5270         ctl = iff->in(0);       // This test feeds a harmless uncommon trap.
5271         continue;
5272       }
5273     }
5274     return NULL;
5275   }
5276 
5277   // If we get this far, we have an allocation which immediately
5278   // precedes the arraycopy, and we can take over zeroing the new object.
5279   // The arraycopy will finish the initialization, and provide
5280   // a new control state to which we will anchor the destination pointer.
5281 
5282   return alloc;
5283 }
5284 
5285 // Helper for initialization of arrays, creating a ClearArray.
5286 // It writes zero bits in [start..end), within the body of an array object.
5287 // The memory effects are all chained onto the 'adr_type' alias category.
5288 //
5289 // Since the object is otherwise uninitialized, we are free
5290 // to put a little "slop" around the edges of the cleared area,
5291 // as long as it does not go back into the array's header,
5292 // or beyond the array end within the heap.
5293 //
5294 // The lower edge can be rounded down to the nearest jint and the
5295 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5296 //
5297 // Arguments:
5298 //   adr_type           memory slice where writes are generated
5299 //   dest               oop of the destination array
5300 //   basic_elem_type    element type of the destination
5301 //   slice_idx          array index of first element to store
5302 //   slice_len          number of elements to store (or NULL)
5303 //   dest_size          total size in bytes of the array object
5304 //
5305 // Exactly one of slice_len or dest_size must be non-NULL.
5306 // If dest_size is non-NULL, zeroing extends to the end of the object.
5307 // If slice_len is non-NULL, the slice_idx value must be a constant.
5308 void
5309 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5310                                      Node* dest,
5311                                      BasicType basic_elem_type,
5312                                      Node* slice_idx,
5313                                      Node* slice_len,
5314                                      Node* dest_size) {
5315   // one or the other but not both of slice_len and dest_size:
5316   assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5317   if (slice_len == NULL)  slice_len = top();
5318   if (dest_size == NULL)  dest_size = top();
5319 
5320   // operate on this memory slice:
5321   Node* mem = memory(adr_type); // memory slice to operate on
5322 
5323   // scaling and rounding of indexes:
5324   int scale = exact_log2(type2aelembytes(basic_elem_type));
5325   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5326   int clear_low = (-1 << scale) & (BytesPerInt  - 1);
5327   int bump_bit  = (-1 << scale) & BytesPerInt;
5328 
5329   // determine constant starts and ends
5330   const intptr_t BIG_NEG = -128;
5331   assert(BIG_NEG + 2*abase < 0, "neg enough");
5332   intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5333   intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5334   if (slice_len_con == 0) {
5335     return;                     // nothing to do here
5336   }
5337   intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5338   intptr_t end_con   = find_intptr_t_con(dest_size, -1);
5339   if (slice_idx_con >= 0 && slice_len_con >= 0) {
5340     assert(end_con < 0, "not two cons");
5341     end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5342                        BytesPerLong);
5343   }
5344 
5345   if (start_con >= 0 && end_con >= 0) {
5346     // Constant start and end.  Simple.
5347     mem = ClearArrayNode::clear_memory(control(), mem, dest,
5348                                        start_con, end_con, &_gvn);
5349   } else if (start_con >= 0 && dest_size != top()) {
5350     // Constant start, pre-rounded end after the tail of the array.
5351     Node* end = dest_size;
5352     mem = ClearArrayNode::clear_memory(control(), mem, dest,
5353                                        start_con, end, &_gvn);
5354   } else if (start_con >= 0 && slice_len != top()) {
5355     // Constant start, non-constant end.  End needs rounding up.
5356     // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5357     intptr_t end_base  = abase + (slice_idx_con << scale);
5358     int      end_round = (-1 << scale) & (BytesPerLong  - 1);
5359     Node*    end       = ConvI2X(slice_len);
5360     if (scale != 0)
5361       end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5362     end_base += end_round;
5363     end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5364     end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5365     mem = ClearArrayNode::clear_memory(control(), mem, dest,
5366                                        start_con, end, &_gvn);
5367   } else if (start_con < 0 && dest_size != top()) {
5368     // Non-constant start, pre-rounded end after the tail of the array.
5369     // This is almost certainly a "round-to-end" operation.
5370     Node* start = slice_idx;
5371     start = ConvI2X(start);
5372     if (scale != 0)
5373       start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5374     start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5375     if ((bump_bit | clear_low) != 0) {
5376       int to_clear = (bump_bit | clear_low);
5377       // Align up mod 8, then store a jint zero unconditionally
5378       // just before the mod-8 boundary.
5379       if (((abase + bump_bit) & ~to_clear) - bump_bit
5380           < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5381         bump_bit = 0;
5382         assert((abase & to_clear) == 0, "array base must be long-aligned");
5383       } else {
5384         // Bump 'start' up to (or past) the next jint boundary:
5385         start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5386         assert((abase & clear_low) == 0, "array base must be int-aligned");
5387       }
5388       // Round bumped 'start' down to jlong boundary in body of array.
5389       start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5390       if (bump_bit != 0) {
5391         // Store a zero to the immediately preceding jint:
5392         Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5393         Node* p1 = basic_plus_adr(dest, x1);
5394         mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5395         mem = _gvn.transform(mem);
5396       }
5397     }
5398     Node* end = dest_size; // pre-rounded
5399     mem = ClearArrayNode::clear_memory(control(), mem, dest,
5400                                        start, end, &_gvn);
5401   } else {
5402     // Non-constant start, unrounded non-constant end.
5403     // (Nobody zeroes a random midsection of an array using this routine.)
5404     ShouldNotReachHere();       // fix caller
5405   }
5406 
5407   // Done.
5408   set_memory(mem, adr_type);
5409 }
5410 
5411 
5412 bool
5413 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5414                                          BasicType basic_elem_type,
5415                                          AllocateNode* alloc,
5416                                          Node* src,  Node* src_offset,
5417                                          Node* dest, Node* dest_offset,
5418                                          Node* dest_size, bool dest_uninitialized) {
5419   // See if there is an advantage from block transfer.
5420   int scale = exact_log2(type2aelembytes(basic_elem_type));
5421   if (scale >= LogBytesPerLong)
5422     return false;               // it is already a block transfer
5423 
5424   // Look at the alignment of the starting offsets.
5425   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5426 
5427   intptr_t src_off_con  = (intptr_t) find_int_con(src_offset, -1);
5428   intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5429   if (src_off_con < 0 || dest_off_con < 0)
5430     // At present, we can only understand constants.
5431     return false;
5432 
5433   intptr_t src_off  = abase + (src_off_con  << scale);
5434   intptr_t dest_off = abase + (dest_off_con << scale);
5435 
5436   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5437     // Non-aligned; too bad.
5438     // One more chance:  Pick off an initial 32-bit word.
5439     // This is a common case, since abase can be odd mod 8.
5440     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5441         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5442       Node* sptr = basic_plus_adr(src,  src_off);
5443       Node* dptr = basic_plus_adr(dest, dest_off);
5444       Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5445       store_to_memory(control(), dptr, sval, T_INT, adr_type);
5446       src_off += BytesPerInt;
5447       dest_off += BytesPerInt;
5448     } else {
5449       return false;
5450     }
5451   }
5452   assert(src_off % BytesPerLong == 0, "");
5453   assert(dest_off % BytesPerLong == 0, "");
5454 
5455   // Do this copy by giant steps.
5456   Node* sptr  = basic_plus_adr(src,  src_off);
5457   Node* dptr  = basic_plus_adr(dest, dest_off);
5458   Node* countx = dest_size;
5459   countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5460   countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5461 
5462   bool disjoint_bases = true;   // since alloc != NULL
5463   generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5464                                sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5465 
5466   return true;
5467 }
5468 
5469 
5470 // Helper function; generates code for the slow case.
5471 // We make a call to a runtime method which emulates the native method,
5472 // but without the native wrapper overhead.
5473 void
5474 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5475                                         Node* src,  Node* src_offset,
5476                                         Node* dest, Node* dest_offset,
5477                                         Node* copy_length, bool dest_uninitialized) {
5478   assert(!dest_uninitialized, "Invariant");
5479   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5480                                  OptoRuntime::slow_arraycopy_Type(),
5481                                  OptoRuntime::slow_arraycopy_Java(),
5482                                  "slow_arraycopy", adr_type,
5483                                  src, src_offset, dest, dest_offset,
5484                                  copy_length);
5485 
5486   // Handle exceptions thrown by this fellow:
5487   make_slow_call_ex(call, env()->Throwable_klass(), false);
5488 }
5489 
5490 // Helper function; generates code for cases requiring runtime checks.
5491 Node*
5492 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5493                                              Node* dest_elem_klass,
5494                                              Node* src,  Node* src_offset,
5495                                              Node* dest, Node* dest_offset,
5496                                              Node* copy_length, bool dest_uninitialized) {
5497   if (stopped())  return NULL;
5498 
5499   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5500   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5501     return NULL;
5502   }
5503 
5504   // Pick out the parameters required to perform a store-check
5505   // for the target array.  This is an optimistic check.  It will
5506   // look in each non-null element's class, at the desired klass's
5507   // super_check_offset, for the desired klass.
5508   int sco_offset = in_bytes(Klass::super_check_offset_offset());
5509   Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5510   Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5511   Node* check_offset = ConvI2X(_gvn.transform(n3));
5512   Node* check_value  = dest_elem_klass;
5513 
5514   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT);
5515   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5516 
5517   // (We know the arrays are never conjoint, because their types differ.)
5518   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5519                                  OptoRuntime::checkcast_arraycopy_Type(),
5520                                  copyfunc_addr, "checkcast_arraycopy", adr_type,
5521                                  // five arguments, of which two are
5522                                  // intptr_t (jlong in LP64)
5523                                  src_start, dest_start,
5524                                  copy_length XTOP,
5525                                  check_offset XTOP,
5526                                  check_value);
5527 
5528   return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5529 }
5530 
5531 
5532 // Helper function; generates code for cases requiring runtime checks.
5533 Node*
5534 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5535                                            Node* src,  Node* src_offset,
5536                                            Node* dest, Node* dest_offset,
5537                                            Node* copy_length, bool dest_uninitialized) {
5538   assert(!dest_uninitialized, "Invariant");
5539   if (stopped())  return NULL;
5540   address copyfunc_addr = StubRoutines::generic_arraycopy();
5541   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5542     return NULL;
5543   }
5544 
5545   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5546                     OptoRuntime::generic_arraycopy_Type(),
5547                     copyfunc_addr, "generic_arraycopy", adr_type,
5548                     src, src_offset, dest, dest_offset, copy_length);
5549 
5550   return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5551 }
5552 
5553 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5554 void
5555 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5556                                              BasicType basic_elem_type,
5557                                              bool disjoint_bases,
5558                                              Node* src,  Node* src_offset,
5559                                              Node* dest, Node* dest_offset,
5560                                              Node* copy_length, bool dest_uninitialized) {
5561   if (stopped())  return;               // nothing to do
5562 
5563   Node* src_start  = src;
5564   Node* dest_start = dest;
5565   if (src_offset != NULL || dest_offset != NULL) {
5566     assert(src_offset != NULL && dest_offset != NULL, "");
5567     src_start  = array_element_address(src,  src_offset,  basic_elem_type);
5568     dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5569   }
5570 
5571   // Figure out which arraycopy runtime method to call.
5572   const char* copyfunc_name = "arraycopy";
5573   address     copyfunc_addr =
5574       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5575                           disjoint_bases, copyfunc_name, dest_uninitialized);
5576 
5577   // Call it.  Note that the count_ix value is not scaled to a byte-size.
5578   make_runtime_call(RC_LEAF|RC_NO_FP,
5579                     OptoRuntime::fast_arraycopy_Type(),
5580                     copyfunc_addr, copyfunc_name, adr_type,
5581                     src_start, dest_start, copy_length XTOP);
5582 }
5583 
5584 //-------------inline_encodeISOArray-----------------------------------
5585 // encode char[] to byte[] in ISO_8859_1
5586 bool LibraryCallKit::inline_encodeISOArray() {
5587   assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5588   // no receiver since it is static method
5589   Node *src         = argument(0);
5590   Node *src_offset  = argument(1);
5591   Node *dst         = argument(2);
5592   Node *dst_offset  = argument(3);
5593   Node *length      = argument(4);
5594 
5595   const Type* src_type = src->Value(&_gvn);
5596   const Type* dst_type = dst->Value(&_gvn);
5597   const TypeAryPtr* top_src = src_type->isa_aryptr();
5598   const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5599   if (top_src  == NULL || top_src->klass()  == NULL ||
5600       top_dest == NULL || top_dest->klass() == NULL) {
5601     // failed array check
5602     return false;
5603   }
5604 
5605   // Figure out the size and type of the elements we will be copying.
5606   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5607   BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5608   if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5609     return false;
5610   }
5611   Node* src_start = array_element_address(src, src_offset, src_elem);
5612   Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5613   // 'src_start' points to src array + scaled offset
5614   // 'dst_start' points to dst array + scaled offset
5615 
5616   const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5617   Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5618   enc = _gvn.transform(enc);
5619   Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5620   set_memory(res_mem, mtype);
5621   set_result(enc);
5622   return true;
5623 }
5624 
5625 /**
5626  * Calculate CRC32 for byte.
5627  * int java.util.zip.CRC32.update(int crc, int b)
5628  */
5629 bool LibraryCallKit::inline_updateCRC32() {
5630   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5631   assert(callee()->signature()->size() == 2, "update has 2 parameters");
5632   // no receiver since it is static method
5633   Node* crc  = argument(0); // type: int
5634   Node* b    = argument(1); // type: int
5635 
5636   /*
5637    *    int c = ~ crc;
5638    *    b = timesXtoThe32[(b ^ c) & 0xFF];
5639    *    b = b ^ (c >>> 8);
5640    *    crc = ~b;
5641    */
5642 
5643   Node* M1 = intcon(-1);
5644   crc = _gvn.transform(new (C) XorINode(crc, M1));
5645   Node* result = _gvn.transform(new (C) XorINode(crc, b));
5646   result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5647 
5648   Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5649   Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5650   Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5651   result = make_load(control(), adr, TypeInt::INT, T_INT);
5652 
5653   crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5654   result = _gvn.transform(new (C) XorINode(crc, result));
5655   result = _gvn.transform(new (C) XorINode(result, M1));
5656   set_result(result);
5657   return true;
5658 }
5659 
5660 /**
5661  * Calculate CRC32 for byte[] array.
5662  * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5663  */
5664 bool LibraryCallKit::inline_updateBytesCRC32() {
5665   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5666   assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5667   // no receiver since it is static method
5668   Node* crc     = argument(0); // type: int
5669   Node* src     = argument(1); // type: oop
5670   Node* offset  = argument(2); // type: int
5671   Node* length  = argument(3); // type: int
5672 
5673   const Type* src_type = src->Value(&_gvn);
5674   const TypeAryPtr* top_src = src_type->isa_aryptr();
5675   if (top_src  == NULL || top_src->klass()  == NULL) {
5676     // failed array check
5677     return false;
5678   }
5679 
5680   // Figure out the size and type of the elements we will be copying.
5681   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5682   if (src_elem != T_BYTE) {
5683     return false;
5684   }
5685 
5686   // 'src_start' points to src array + scaled offset
5687   Node* src_start = array_element_address(src, offset, src_elem);
5688 
5689   // We assume that range check is done by caller.
5690   // TODO: generate range check (offset+length < src.length) in debug VM.
5691 
5692   // Call the stub.
5693   address stubAddr = StubRoutines::updateBytesCRC32();
5694   const char *stubName = "updateBytesCRC32";
5695 
5696   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5697                                  stubAddr, stubName, TypePtr::BOTTOM,
5698                                  crc, src_start, length);
5699   Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5700   set_result(result);
5701   return true;
5702 }
5703 
5704 /**
5705  * Calculate CRC32 for ByteBuffer.
5706  * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5707  */
5708 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5709   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5710   assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5711   // no receiver since it is static method
5712   Node* crc     = argument(0); // type: int
5713   Node* src     = argument(1); // type: long
5714   Node* offset  = argument(3); // type: int
5715   Node* length  = argument(4); // type: int
5716 
5717   src = ConvL2X(src);  // adjust Java long to machine word
5718   Node* base = _gvn.transform(new (C) CastX2PNode(src));
5719   offset = ConvI2X(offset);
5720 
5721   // 'src_start' points to src array + scaled offset
5722   Node* src_start = basic_plus_adr(top(), base, offset);
5723 
5724   // Call the stub.
5725   address stubAddr = StubRoutines::updateBytesCRC32();
5726   const char *stubName = "updateBytesCRC32";
5727 
5728   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5729                                  stubAddr, stubName, TypePtr::BOTTOM,
5730                                  crc, src_start, length);
5731   Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5732   set_result(result);
5733   return true;
5734 }
5735 
5736 //----------------------------inline_reference_get----------------------------
5737 // public T java.lang.ref.Reference.get();
5738 bool LibraryCallKit::inline_reference_get() {
5739   const int referent_offset = java_lang_ref_Reference::referent_offset;
5740   guarantee(referent_offset > 0, "should have already been set");
5741 
5742   // Get the argument:
5743   Node* reference_obj = null_check_receiver();
5744   if (stopped()) return true;
5745 
5746   Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5747 
5748   ciInstanceKlass* klass = env()->Object_klass();
5749   const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5750 
5751   Node* no_ctrl = NULL;
5752   Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5753 
5754   // Use the pre-barrier to record the value in the referent field
5755   pre_barrier(false /* do_load */,
5756               control(),
5757               NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5758               result /* pre_val */,
5759               T_OBJECT);
5760 
5761   // Add memory barrier to prevent commoning reads from this field
5762   // across safepoint since GC can change its value.
5763   insert_mem_bar(Op_MemBarCPUOrder);
5764 
5765   set_result(result);
5766   return true;
5767 }
5768 
5769 
5770 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5771                                               bool is_exact=true, bool is_static=false) {
5772 
5773   const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5774   assert(tinst != NULL, "obj is null");
5775   assert(tinst->klass()->is_loaded(), "obj is not loaded");
5776   assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5777 
5778   ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5779                                                                           ciSymbol::make(fieldTypeString),
5780                                                                           is_static);
5781   if (field == NULL) return (Node *) NULL;
5782   assert (field != NULL, "undefined field");
5783 
5784   // Next code  copied from Parse::do_get_xxx():
5785 
5786   // Compute address and memory type.
5787   int offset  = field->offset_in_bytes();
5788   bool is_vol = field->is_volatile();
5789   ciType* field_klass = field->type();
5790   assert(field_klass->is_loaded(), "should be loaded");
5791   const TypePtr* adr_type = C->alias_type(field)->adr_type();
5792   Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5793   BasicType bt = field->layout_type();
5794 
5795   // Build the resultant type of the load
5796   const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5797 
5798   // Build the load.
5799   Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5800   return loadedField;
5801 }
5802 
5803 
5804 //------------------------------inline_aescrypt_Block-----------------------
5805 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5806   address stubAddr;
5807   const char *stubName;
5808   assert(UseAES, "need AES instruction support");
5809 
5810   switch(id) {
5811   case vmIntrinsics::_aescrypt_encryptBlock:
5812     stubAddr = StubRoutines::aescrypt_encryptBlock();
5813     stubName = "aescrypt_encryptBlock";
5814     break;
5815   case vmIntrinsics::_aescrypt_decryptBlock:
5816     stubAddr = StubRoutines::aescrypt_decryptBlock();
5817     stubName = "aescrypt_decryptBlock";
5818     break;
5819   }
5820   if (stubAddr == NULL) return false;
5821 
5822   Node* aescrypt_object = argument(0);
5823   Node* src             = argument(1);
5824   Node* src_offset      = argument(2);
5825   Node* dest            = argument(3);
5826   Node* dest_offset     = argument(4);
5827 
5828   // (1) src and dest are arrays.
5829   const Type* src_type = src->Value(&_gvn);
5830   const Type* dest_type = dest->Value(&_gvn);
5831   const TypeAryPtr* top_src = src_type->isa_aryptr();
5832   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5833   assert (top_src  != NULL && top_src->klass()  != NULL &&  top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5834 
5835   // for the quick and dirty code we will skip all the checks.
5836   // we are just trying to get the call to be generated.
5837   Node* src_start  = src;
5838   Node* dest_start = dest;
5839   if (src_offset != NULL || dest_offset != NULL) {
5840     assert(src_offset != NULL && dest_offset != NULL, "");
5841     src_start  = array_element_address(src,  src_offset,  T_BYTE);
5842     dest_start = array_element_address(dest, dest_offset, T_BYTE);
5843   }
5844 
5845   // now need to get the start of its expanded key array
5846   // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5847   Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5848   if (k_start == NULL) return false;
5849 
5850   // Call the stub.
5851   make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5852                     stubAddr, stubName, TypePtr::BOTTOM,
5853                     src_start, dest_start, k_start);
5854 
5855   return true;
5856 }
5857 
5858 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5859 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5860   address stubAddr;
5861   const char *stubName;
5862 
5863   assert(UseAES, "need AES instruction support");
5864 
5865   switch(id) {
5866   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5867     stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5868     stubName = "cipherBlockChaining_encryptAESCrypt";
5869     break;
5870   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5871     stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5872     stubName = "cipherBlockChaining_decryptAESCrypt";
5873     break;
5874   }
5875   if (stubAddr == NULL) return false;
5876 
5877   Node* cipherBlockChaining_object = argument(0);
5878   Node* src                        = argument(1);
5879   Node* src_offset                 = argument(2);
5880   Node* len                        = argument(3);
5881   Node* dest                       = argument(4);
5882   Node* dest_offset                = argument(5);
5883 
5884   // (1) src and dest are arrays.
5885   const Type* src_type = src->Value(&_gvn);
5886   const Type* dest_type = dest->Value(&_gvn);
5887   const TypeAryPtr* top_src = src_type->isa_aryptr();
5888   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5889   assert (top_src  != NULL && top_src->klass()  != NULL
5890           &&  top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5891 
5892   // checks are the responsibility of the caller
5893   Node* src_start  = src;
5894   Node* dest_start = dest;
5895   if (src_offset != NULL || dest_offset != NULL) {
5896     assert(src_offset != NULL && dest_offset != NULL, "");
5897     src_start  = array_element_address(src,  src_offset,  T_BYTE);
5898     dest_start = array_element_address(dest, dest_offset, T_BYTE);
5899   }
5900 
5901   // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5902   // (because of the predicated logic executed earlier).
5903   // so we cast it here safely.
5904   // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5905 
5906   Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5907   if (embeddedCipherObj == NULL) return false;
5908 
5909   // cast it to what we know it will be at runtime
5910   const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5911   assert(tinst != NULL, "CBC obj is null");
5912   assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5913   ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5914   if (!klass_AESCrypt->is_loaded()) return false;
5915 
5916   ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5917   const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5918   const TypeOopPtr* xtype = aklass->as_instance_type();
5919   Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5920   aescrypt_object = _gvn.transform(aescrypt_object);
5921 
5922   // we need to get the start of the aescrypt_object's expanded key array
5923   Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5924   if (k_start == NULL) return false;
5925 
5926   // similarly, get the start address of the r vector
5927   Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5928   if (objRvec == NULL) return false;
5929   Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5930 
5931   // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5932   make_runtime_call(RC_LEAF|RC_NO_FP,
5933                     OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5934                     stubAddr, stubName, TypePtr::BOTTOM,
5935                     src_start, dest_start, k_start, r_start, len);
5936 
5937   // return is void so no result needs to be pushed
5938 
5939   return true;
5940 }
5941 
5942 //------------------------------get_key_start_from_aescrypt_object-----------------------
5943 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5944   Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5945   assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5946   if (objAESCryptKey == NULL) return (Node *) NULL;
5947 
5948   // now have the array, need to get the start address of the K array
5949   Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5950   return k_start;
5951 }
5952 
5953 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5954 // Return node representing slow path of predicate check.
5955 // the pseudo code we want to emulate with this predicate is:
5956 // for encryption:
5957 //    if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5958 // for decryption:
5959 //    if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5960 //    note cipher==plain is more conservative than the original java code but that's OK
5961 //
5962 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5963   // First, check receiver for NULL since it is virtual method.
5964   Node* objCBC = argument(0);
5965   objCBC = null_check(objCBC);
5966 
5967   if (stopped()) return NULL; // Always NULL
5968 
5969   // Load embeddedCipher field of CipherBlockChaining object.
5970   Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5971 
5972   // get AESCrypt klass for instanceOf check
5973   // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5974   // will have same classloader as CipherBlockChaining object
5975   const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5976   assert(tinst != NULL, "CBCobj is null");
5977   assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5978 
5979   // we want to do an instanceof comparison against the AESCrypt class
5980   ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5981   if (!klass_AESCrypt->is_loaded()) {
5982     // if AESCrypt is not even loaded, we never take the intrinsic fast path
5983     Node* ctrl = control();
5984     set_control(top()); // no regular fast path
5985     return ctrl;
5986   }
5987   ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5988 
5989   Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5990   Node* cmp_instof  = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5991   Node* bool_instof  = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5992 
5993   Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5994 
5995   // for encryption, we are done
5996   if (!decrypting)
5997     return instof_false;  // even if it is NULL
5998 
5999   // for decryption, we need to add a further check to avoid
6000   // taking the intrinsic path when cipher and plain are the same
6001   // see the original java code for why.
6002   RegionNode* region = new(C) RegionNode(3);
6003   region->init_req(1, instof_false);
6004   Node* src = argument(1);
6005   Node* dest = argument(4);
6006   Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6007   Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6008   Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6009   region->init_req(2, src_dest_conjoint);
6010 
6011   record_for_igvn(region);
6012   return _gvn.transform(region);
6013 }