rev 7390 : 8028595: WhiteBox API for stress testing of TieredCompilation
Reviewed-by:

   1 /*
   2  * Copyright (c) 1999, 2014, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.hpp"
  27 #include "classfile/systemDictionary.hpp"
  28 #include "classfile/vmSymbols.hpp"
  29 #include "compiler/compileBroker.hpp"
  30 #include "compiler/compileLog.hpp"
  31 #include "oops/objArrayKlass.hpp"
  32 #include "opto/addnode.hpp"
  33 #include "opto/callGenerator.hpp"
  34 #include "opto/castnode.hpp"
  35 #include "opto/cfgnode.hpp"
  36 #include "opto/convertnode.hpp"
  37 #include "opto/countbitsnode.hpp"
  38 #include "opto/intrinsicnode.hpp"
  39 #include "opto/idealKit.hpp"
  40 #include "opto/mathexactnode.hpp"
  41 #include "opto/movenode.hpp"
  42 #include "opto/mulnode.hpp"
  43 #include "opto/narrowptrnode.hpp"
  44 #include "opto/parse.hpp"
  45 #include "opto/runtime.hpp"
  46 #include "opto/subnode.hpp"
  47 #include "prims/nativeLookup.hpp"
  48 #include "runtime/sharedRuntime.hpp"
  49 #include "trace/traceMacros.hpp"
  50 
  51 class LibraryIntrinsic : public InlineCallGenerator {
  52   // Extend the set of intrinsics known to the runtime:
  53  public:
  54  private:
  55   bool             _is_virtual;
  56   bool             _does_virtual_dispatch;
  57   int8_t           _predicates_count;  // Intrinsic is predicated by several conditions
  58   int8_t           _last_predicate; // Last generated predicate
  59   vmIntrinsics::ID _intrinsic_id;
  60 
  61  public:
  62   LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
  63     : InlineCallGenerator(m),
  64       _is_virtual(is_virtual),
  65       _does_virtual_dispatch(does_virtual_dispatch),
  66       _predicates_count((int8_t)predicates_count),
  67       _last_predicate((int8_t)-1),
  68       _intrinsic_id(id)
  69   {
  70   }
  71   virtual bool is_intrinsic() const { return true; }
  72   virtual bool is_virtual()   const { return _is_virtual; }
  73   virtual bool is_predicated() const { return _predicates_count > 0; }
  74   virtual int  predicates_count() const { return _predicates_count; }
  75   virtual bool does_virtual_dispatch()   const { return _does_virtual_dispatch; }
  76   virtual JVMState* generate(JVMState* jvms);
  77   virtual Node* generate_predicate(JVMState* jvms, int predicate);
  78   vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
  79 };
  80 
  81 
  82 // Local helper class for LibraryIntrinsic:
  83 class LibraryCallKit : public GraphKit {
  84  private:
  85   LibraryIntrinsic* _intrinsic;     // the library intrinsic being called
  86   Node*             _result;        // the result node, if any
  87   int               _reexecute_sp;  // the stack pointer when bytecode needs to be reexecuted
  88 
  89   const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
  90 
  91  public:
  92   LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
  93     : GraphKit(jvms),
  94       _intrinsic(intrinsic),
  95       _result(NULL)
  96   {
  97     // Check if this is a root compile.  In that case we don't have a caller.
  98     if (!jvms->has_method()) {
  99       _reexecute_sp = sp();
 100     } else {
 101       // Find out how many arguments the interpreter needs when deoptimizing
 102       // and save the stack pointer value so it can used by uncommon_trap.
 103       // We find the argument count by looking at the declared signature.
 104       bool ignored_will_link;
 105       ciSignature* declared_signature = NULL;
 106       ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
 107       const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
 108       _reexecute_sp = sp() + nargs;  // "push" arguments back on stack
 109     }
 110   }
 111 
 112   virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
 113 
 114   ciMethod*         caller()    const    { return jvms()->method(); }
 115   int               bci()       const    { return jvms()->bci(); }
 116   LibraryIntrinsic* intrinsic() const    { return _intrinsic; }
 117   vmIntrinsics::ID  intrinsic_id() const { return _intrinsic->intrinsic_id(); }
 118   ciMethod*         callee()    const    { return _intrinsic->method(); }
 119 
 120   bool  try_to_inline(int predicate);
 121   Node* try_to_predicate(int predicate);
 122 
 123   void push_result() {
 124     // Push the result onto the stack.
 125     if (!stopped() && result() != NULL) {
 126       BasicType bt = result()->bottom_type()->basic_type();
 127       push_node(bt, result());
 128     }
 129   }
 130 
 131  private:
 132   void fatal_unexpected_iid(vmIntrinsics::ID iid) {
 133     fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
 134   }
 135 
 136   void  set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
 137   void  set_result(RegionNode* region, PhiNode* value);
 138   Node*     result() { return _result; }
 139 
 140   virtual int reexecute_sp() { return _reexecute_sp; }
 141 
 142   // Helper functions to inline natives
 143   Node* generate_guard(Node* test, RegionNode* region, float true_prob);
 144   Node* generate_slow_guard(Node* test, RegionNode* region);
 145   Node* generate_fair_guard(Node* test, RegionNode* region);
 146   Node* generate_negative_guard(Node* index, RegionNode* region,
 147                                 // resulting CastII of index:
 148                                 Node* *pos_index = NULL);
 149   Node* generate_limit_guard(Node* offset, Node* subseq_length,
 150                              Node* array_length,
 151                              RegionNode* region);
 152   Node* generate_current_thread(Node* &tls_output);
 153   Node* load_mirror_from_klass(Node* klass);
 154   Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
 155                                       RegionNode* region, int null_path,
 156                                       int offset);
 157   Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
 158                                RegionNode* region, int null_path) {
 159     int offset = java_lang_Class::klass_offset_in_bytes();
 160     return load_klass_from_mirror_common(mirror, never_see_null,
 161                                          region, null_path,
 162                                          offset);
 163   }
 164   Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
 165                                      RegionNode* region, int null_path) {
 166     int offset = java_lang_Class::array_klass_offset_in_bytes();
 167     return load_klass_from_mirror_common(mirror, never_see_null,
 168                                          region, null_path,
 169                                          offset);
 170   }
 171   Node* generate_access_flags_guard(Node* kls,
 172                                     int modifier_mask, int modifier_bits,
 173                                     RegionNode* region);
 174   Node* generate_interface_guard(Node* kls, RegionNode* region);
 175   Node* generate_array_guard(Node* kls, RegionNode* region) {
 176     return generate_array_guard_common(kls, region, false, false);
 177   }
 178   Node* generate_non_array_guard(Node* kls, RegionNode* region) {
 179     return generate_array_guard_common(kls, region, false, true);
 180   }
 181   Node* generate_objArray_guard(Node* kls, RegionNode* region) {
 182     return generate_array_guard_common(kls, region, true, false);
 183   }
 184   Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
 185     return generate_array_guard_common(kls, region, true, true);
 186   }
 187   Node* generate_array_guard_common(Node* kls, RegionNode* region,
 188                                     bool obj_array, bool not_array);
 189   Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
 190   CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
 191                                      bool is_virtual = false, bool is_static = false);
 192   CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
 193     return generate_method_call(method_id, false, true);
 194   }
 195   CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
 196     return generate_method_call(method_id, true, false);
 197   }
 198   Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
 199 
 200   Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
 201   Node* make_string_method_node(int opcode, Node* str1, Node* str2);
 202   bool inline_string_compareTo();
 203   bool inline_string_indexOf();
 204   Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
 205   bool inline_string_equals();
 206   Node* round_double_node(Node* n);
 207   bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
 208   bool inline_math_native(vmIntrinsics::ID id);
 209   bool inline_trig(vmIntrinsics::ID id);
 210   bool inline_math(vmIntrinsics::ID id);
 211   template <typename OverflowOp>
 212   bool inline_math_overflow(Node* arg1, Node* arg2);
 213   void inline_math_mathExact(Node* math, Node* test);
 214   bool inline_math_addExactI(bool is_increment);
 215   bool inline_math_addExactL(bool is_increment);
 216   bool inline_math_multiplyExactI();
 217   bool inline_math_multiplyExactL();
 218   bool inline_math_negateExactI();
 219   bool inline_math_negateExactL();
 220   bool inline_math_subtractExactI(bool is_decrement);
 221   bool inline_math_subtractExactL(bool is_decrement);
 222   bool inline_exp();
 223   bool inline_pow();
 224   Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
 225   bool inline_min_max(vmIntrinsics::ID id);
 226   Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
 227   // This returns Type::AnyPtr, RawPtr, or OopPtr.
 228   int classify_unsafe_addr(Node* &base, Node* &offset);
 229   Node* make_unsafe_address(Node* base, Node* offset);
 230   // Helper for inline_unsafe_access.
 231   // Generates the guards that check whether the result of
 232   // Unsafe.getObject should be recorded in an SATB log buffer.
 233   void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
 234   bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
 235   bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
 236   static bool klass_needs_init_guard(Node* kls);
 237   bool inline_unsafe_allocate();
 238   bool inline_unsafe_copyMemory();
 239   bool inline_native_currentThread();
 240 #ifdef TRACE_HAVE_INTRINSICS
 241   bool inline_native_classID();
 242   bool inline_native_threadID();
 243 #endif
 244   bool inline_native_time_funcs(address method, const char* funcName);
 245   bool inline_native_isInterrupted();
 246   bool inline_native_Class_query(vmIntrinsics::ID id);
 247   bool inline_native_subtype_check();
 248 
 249   bool inline_native_newArray();
 250   bool inline_native_getLength();
 251   bool inline_array_copyOf(bool is_copyOfRange);
 252   bool inline_array_equals();
 253   void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
 254   bool inline_native_clone(bool is_virtual);
 255   bool inline_native_Reflection_getCallerClass();
 256   // Helper function for inlining native object hash method
 257   bool inline_native_hashcode(bool is_virtual, bool is_static);
 258   bool inline_native_getClass();
 259 
 260   // Helper functions for inlining arraycopy
 261   bool inline_arraycopy();
 262   AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
 263                                                 RegionNode* slow_region);
 264   typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
 265   bool inline_unsafe_load_store(BasicType type,  LoadStoreKind kind);
 266   bool inline_unsafe_ordered_store(BasicType type);
 267   bool inline_unsafe_fence(vmIntrinsics::ID id);
 268   bool inline_fp_conversions(vmIntrinsics::ID id);
 269   bool inline_number_methods(vmIntrinsics::ID id);
 270   bool inline_reference_get();
 271   bool inline_Class_cast();
 272   bool inline_aescrypt_Block(vmIntrinsics::ID id);
 273   bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
 274   Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
 275   Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
 276   Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
 277   bool inline_sha_implCompress(vmIntrinsics::ID id);
 278   bool inline_digestBase_implCompressMB(int predicate);
 279   bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
 280                                  bool long_state, address stubAddr, const char *stubName,
 281                                  Node* src_start, Node* ofs, Node* limit);
 282   Node* get_state_from_sha_object(Node *sha_object);
 283   Node* get_state_from_sha5_object(Node *sha_object);
 284   Node* inline_digestBase_implCompressMB_predicate(int predicate);
 285   bool inline_encodeISOArray();
 286   bool inline_updateCRC32();
 287   bool inline_updateBytesCRC32();
 288   bool inline_updateByteBufferCRC32();
 289   bool inline_multiplyToLen();


 290 };
 291 
 292 
 293 //---------------------------make_vm_intrinsic----------------------------
 294 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
 295   vmIntrinsics::ID id = m->intrinsic_id();
 296   assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
 297 
 298   ccstr disable_intr = NULL;
 299 
 300   if ((DisableIntrinsic[0] != '\0'
 301        && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) ||
 302       (method_has_option_value("DisableIntrinsic", disable_intr)
 303        && strstr(disable_intr, vmIntrinsics::name_at(id)) != NULL)) {
 304     // disabled by a user request on the command line:
 305     // example: -XX:DisableIntrinsic=_hashCode,_getClass
 306     return NULL;
 307   }
 308 
 309   if (!m->is_loaded()) {
 310     // do not attempt to inline unloaded methods
 311     return NULL;
 312   }
 313 
 314   // Only a few intrinsics implement a virtual dispatch.
 315   // They are expensive calls which are also frequently overridden.
 316   if (is_virtual) {
 317     switch (id) {
 318     case vmIntrinsics::_hashCode:
 319     case vmIntrinsics::_clone:
 320       // OK, Object.hashCode and Object.clone intrinsics come in both flavors
 321       break;
 322     default:
 323       return NULL;
 324     }
 325   }
 326 
 327   // -XX:-InlineNatives disables nearly all intrinsics:
 328   if (!InlineNatives) {
 329     switch (id) {
 330     case vmIntrinsics::_indexOf:
 331     case vmIntrinsics::_compareTo:
 332     case vmIntrinsics::_equals:
 333     case vmIntrinsics::_equalsC:
 334     case vmIntrinsics::_getAndAddInt:
 335     case vmIntrinsics::_getAndAddLong:
 336     case vmIntrinsics::_getAndSetInt:
 337     case vmIntrinsics::_getAndSetLong:
 338     case vmIntrinsics::_getAndSetObject:
 339     case vmIntrinsics::_loadFence:
 340     case vmIntrinsics::_storeFence:
 341     case vmIntrinsics::_fullFence:
 342       break;  // InlineNatives does not control String.compareTo
 343     case vmIntrinsics::_Reference_get:
 344       break;  // InlineNatives does not control Reference.get
 345     default:
 346       return NULL;
 347     }
 348   }
 349 
 350   int predicates = 0;
 351   bool does_virtual_dispatch = false;
 352 
 353   switch (id) {
 354   case vmIntrinsics::_compareTo:
 355     if (!SpecialStringCompareTo)  return NULL;
 356     if (!Matcher::match_rule_supported(Op_StrComp))  return NULL;
 357     break;
 358   case vmIntrinsics::_indexOf:
 359     if (!SpecialStringIndexOf)  return NULL;
 360     break;
 361   case vmIntrinsics::_equals:
 362     if (!SpecialStringEquals)  return NULL;
 363     if (!Matcher::match_rule_supported(Op_StrEquals))  return NULL;
 364     break;
 365   case vmIntrinsics::_equalsC:
 366     if (!SpecialArraysEquals)  return NULL;
 367     if (!Matcher::match_rule_supported(Op_AryEq))  return NULL;
 368     break;
 369   case vmIntrinsics::_arraycopy:
 370     if (!InlineArrayCopy)  return NULL;
 371     break;
 372   case vmIntrinsics::_copyMemory:
 373     if (StubRoutines::unsafe_arraycopy() == NULL)  return NULL;
 374     if (!InlineArrayCopy)  return NULL;
 375     break;
 376   case vmIntrinsics::_hashCode:
 377     if (!InlineObjectHash)  return NULL;
 378     does_virtual_dispatch = true;
 379     break;
 380   case vmIntrinsics::_clone:
 381     does_virtual_dispatch = true;
 382   case vmIntrinsics::_copyOf:
 383   case vmIntrinsics::_copyOfRange:
 384     if (!InlineObjectCopy)  return NULL;
 385     // These also use the arraycopy intrinsic mechanism:
 386     if (!InlineArrayCopy)  return NULL;
 387     break;
 388   case vmIntrinsics::_encodeISOArray:
 389     if (!SpecialEncodeISOArray)  return NULL;
 390     if (!Matcher::match_rule_supported(Op_EncodeISOArray))  return NULL;
 391     break;
 392   case vmIntrinsics::_checkIndex:
 393     // We do not intrinsify this.  The optimizer does fine with it.
 394     return NULL;
 395 
 396   case vmIntrinsics::_getCallerClass:
 397     if (!InlineReflectionGetCallerClass)  return NULL;
 398     if (SystemDictionary::reflect_CallerSensitive_klass() == NULL)  return NULL;
 399     break;
 400 
 401   case vmIntrinsics::_bitCount_i:
 402     if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
 403     break;
 404 
 405   case vmIntrinsics::_bitCount_l:
 406     if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
 407     break;
 408 
 409   case vmIntrinsics::_numberOfLeadingZeros_i:
 410     if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
 411     break;
 412 
 413   case vmIntrinsics::_numberOfLeadingZeros_l:
 414     if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
 415     break;
 416 
 417   case vmIntrinsics::_numberOfTrailingZeros_i:
 418     if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
 419     break;
 420 
 421   case vmIntrinsics::_numberOfTrailingZeros_l:
 422     if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
 423     break;
 424 
 425   case vmIntrinsics::_reverseBytes_c:
 426     if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
 427     break;
 428   case vmIntrinsics::_reverseBytes_s:
 429     if (!Matcher::match_rule_supported(Op_ReverseBytesS))  return NULL;
 430     break;
 431   case vmIntrinsics::_reverseBytes_i:
 432     if (!Matcher::match_rule_supported(Op_ReverseBytesI))  return NULL;
 433     break;
 434   case vmIntrinsics::_reverseBytes_l:
 435     if (!Matcher::match_rule_supported(Op_ReverseBytesL))  return NULL;
 436     break;
 437 
 438   case vmIntrinsics::_Reference_get:
 439     // Use the intrinsic version of Reference.get() so that the value in
 440     // the referent field can be registered by the G1 pre-barrier code.
 441     // Also add memory barrier to prevent commoning reads from this field
 442     // across safepoint since GC can change it value.
 443     break;
 444 
 445   case vmIntrinsics::_compareAndSwapObject:
 446 #ifdef _LP64
 447     if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
 448 #endif
 449     break;
 450 
 451   case vmIntrinsics::_compareAndSwapLong:
 452     if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
 453     break;
 454 
 455   case vmIntrinsics::_getAndAddInt:
 456     if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
 457     break;
 458 
 459   case vmIntrinsics::_getAndAddLong:
 460     if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
 461     break;
 462 
 463   case vmIntrinsics::_getAndSetInt:
 464     if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
 465     break;
 466 
 467   case vmIntrinsics::_getAndSetLong:
 468     if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
 469     break;
 470 
 471   case vmIntrinsics::_getAndSetObject:
 472 #ifdef _LP64
 473     if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
 474     if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
 475     break;
 476 #else
 477     if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
 478     break;
 479 #endif
 480 
 481   case vmIntrinsics::_aescrypt_encryptBlock:
 482   case vmIntrinsics::_aescrypt_decryptBlock:
 483     if (!UseAESIntrinsics) return NULL;
 484     break;
 485 
 486   case vmIntrinsics::_multiplyToLen:
 487     if (!UseMultiplyToLenIntrinsic) return NULL;
 488     break;
 489 
 490   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 491   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 492     if (!UseAESIntrinsics) return NULL;
 493     // these two require the predicated logic
 494     predicates = 1;
 495     break;
 496 
 497   case vmIntrinsics::_sha_implCompress:
 498     if (!UseSHA1Intrinsics) return NULL;
 499     break;
 500 
 501   case vmIntrinsics::_sha2_implCompress:
 502     if (!UseSHA256Intrinsics) return NULL;
 503     break;
 504 
 505   case vmIntrinsics::_sha5_implCompress:
 506     if (!UseSHA512Intrinsics) return NULL;
 507     break;
 508 
 509   case vmIntrinsics::_digestBase_implCompressMB:
 510     if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) return NULL;
 511     predicates = 3;
 512     break;
 513 
 514   case vmIntrinsics::_updateCRC32:
 515   case vmIntrinsics::_updateBytesCRC32:
 516   case vmIntrinsics::_updateByteBufferCRC32:
 517     if (!UseCRC32Intrinsics) return NULL;
 518     break;
 519 
 520   case vmIntrinsics::_incrementExactI:
 521   case vmIntrinsics::_addExactI:
 522     if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
 523     break;
 524   case vmIntrinsics::_incrementExactL:
 525   case vmIntrinsics::_addExactL:
 526     if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
 527     break;
 528   case vmIntrinsics::_decrementExactI:
 529   case vmIntrinsics::_subtractExactI:
 530     if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
 531     break;
 532   case vmIntrinsics::_decrementExactL:
 533   case vmIntrinsics::_subtractExactL:
 534     if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
 535     break;
 536   case vmIntrinsics::_negateExactI:
 537     if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
 538     break;
 539   case vmIntrinsics::_negateExactL:
 540     if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
 541     break;
 542   case vmIntrinsics::_multiplyExactI:
 543     if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
 544     break;
 545   case vmIntrinsics::_multiplyExactL:
 546     if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
 547     break;
 548 
 549  default:
 550     assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
 551     assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
 552     break;
 553   }
 554 
 555   // -XX:-InlineClassNatives disables natives from the Class class.
 556   // The flag applies to all reflective calls, notably Array.newArray
 557   // (visible to Java programmers as Array.newInstance).
 558   if (m->holder()->name() == ciSymbol::java_lang_Class() ||
 559       m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
 560     if (!InlineClassNatives)  return NULL;
 561   }
 562 
 563   // -XX:-InlineThreadNatives disables natives from the Thread class.
 564   if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
 565     if (!InlineThreadNatives)  return NULL;
 566   }
 567 
 568   // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
 569   if (m->holder()->name() == ciSymbol::java_lang_Math() ||
 570       m->holder()->name() == ciSymbol::java_lang_Float() ||
 571       m->holder()->name() == ciSymbol::java_lang_Double()) {
 572     if (!InlineMathNatives)  return NULL;
 573   }
 574 
 575   // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
 576   if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
 577     if (!InlineUnsafeOps)  return NULL;
 578   }
 579 
 580   return new LibraryIntrinsic(m, is_virtual, predicates, does_virtual_dispatch, (vmIntrinsics::ID) id);
 581 }
 582 
 583 //----------------------register_library_intrinsics-----------------------
 584 // Initialize this file's data structures, for each Compile instance.
 585 void Compile::register_library_intrinsics() {
 586   // Nothing to do here.
 587 }
 588 
 589 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
 590   LibraryCallKit kit(jvms, this);
 591   Compile* C = kit.C;
 592   int nodes = C->unique();
 593 #ifndef PRODUCT
 594   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
 595     char buf[1000];
 596     const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
 597     tty->print_cr("Intrinsic %s", str);
 598   }
 599 #endif
 600   ciMethod* callee = kit.callee();
 601   const int bci    = kit.bci();
 602 
 603   // Try to inline the intrinsic.
 604   if (kit.try_to_inline(_last_predicate)) {
 605     if (C->print_intrinsics() || C->print_inlining()) {
 606       C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
 607     }
 608     C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
 609     if (C->log()) {
 610       C->log()->elem("intrinsic id='%s'%s nodes='%d'",
 611                      vmIntrinsics::name_at(intrinsic_id()),
 612                      (is_virtual() ? " virtual='1'" : ""),
 613                      C->unique() - nodes);
 614     }
 615     // Push the result from the inlined method onto the stack.
 616     kit.push_result();
 617     C->print_inlining_update(this);
 618     return kit.transfer_exceptions_into_jvms();
 619   }
 620 
 621   // The intrinsic bailed out
 622   if (C->print_intrinsics() || C->print_inlining()) {
 623     if (jvms->has_method()) {
 624       // Not a root compile.
 625       const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
 626       C->print_inlining(callee, jvms->depth() - 1, bci, msg);
 627     } else {
 628       // Root compile
 629       tty->print("Did not generate intrinsic %s%s at bci:%d in",
 630                vmIntrinsics::name_at(intrinsic_id()),
 631                (is_virtual() ? " (virtual)" : ""), bci);
 632     }
 633   }
 634   C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
 635   C->print_inlining_update(this);
 636   return NULL;
 637 }
 638 
 639 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
 640   LibraryCallKit kit(jvms, this);
 641   Compile* C = kit.C;
 642   int nodes = C->unique();
 643   _last_predicate = predicate;
 644 #ifndef PRODUCT
 645   assert(is_predicated() && predicate < predicates_count(), "sanity");
 646   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
 647     char buf[1000];
 648     const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
 649     tty->print_cr("Predicate for intrinsic %s", str);
 650   }
 651 #endif
 652   ciMethod* callee = kit.callee();
 653   const int bci    = kit.bci();
 654 
 655   Node* slow_ctl = kit.try_to_predicate(predicate);
 656   if (!kit.failing()) {
 657     if (C->print_intrinsics() || C->print_inlining()) {
 658       C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
 659     }
 660     C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
 661     if (C->log()) {
 662       C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
 663                      vmIntrinsics::name_at(intrinsic_id()),
 664                      (is_virtual() ? " virtual='1'" : ""),
 665                      C->unique() - nodes);
 666     }
 667     return slow_ctl; // Could be NULL if the check folds.
 668   }
 669 
 670   // The intrinsic bailed out
 671   if (C->print_intrinsics() || C->print_inlining()) {
 672     if (jvms->has_method()) {
 673       // Not a root compile.
 674       const char* msg = "failed to generate predicate for intrinsic";
 675       C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
 676     } else {
 677       // Root compile
 678       C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
 679                                         vmIntrinsics::name_at(intrinsic_id()),
 680                                         (is_virtual() ? " (virtual)" : ""), bci);
 681     }
 682   }
 683   C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
 684   return NULL;
 685 }
 686 
 687 bool LibraryCallKit::try_to_inline(int predicate) {
 688   // Handle symbolic names for otherwise undistinguished boolean switches:
 689   const bool is_store       = true;
 690   const bool is_native_ptr  = true;
 691   const bool is_static      = true;
 692   const bool is_volatile    = true;
 693 
 694   if (!jvms()->has_method()) {
 695     // Root JVMState has a null method.
 696     assert(map()->memory()->Opcode() == Op_Parm, "");
 697     // Insert the memory aliasing node
 698     set_all_memory(reset_memory());
 699   }
 700   assert(merged_memory(), "");
 701 
 702 
 703   switch (intrinsic_id()) {
 704   case vmIntrinsics::_hashCode:                 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
 705   case vmIntrinsics::_identityHashCode:         return inline_native_hashcode(/*!virtual*/ false,         is_static);
 706   case vmIntrinsics::_getClass:                 return inline_native_getClass();
 707 
 708   case vmIntrinsics::_dsin:
 709   case vmIntrinsics::_dcos:
 710   case vmIntrinsics::_dtan:
 711   case vmIntrinsics::_dabs:
 712   case vmIntrinsics::_datan2:
 713   case vmIntrinsics::_dsqrt:
 714   case vmIntrinsics::_dexp:
 715   case vmIntrinsics::_dlog:
 716   case vmIntrinsics::_dlog10:
 717   case vmIntrinsics::_dpow:                     return inline_math_native(intrinsic_id());
 718 
 719   case vmIntrinsics::_min:
 720   case vmIntrinsics::_max:                      return inline_min_max(intrinsic_id());
 721 
 722   case vmIntrinsics::_addExactI:                return inline_math_addExactI(false /* add */);
 723   case vmIntrinsics::_addExactL:                return inline_math_addExactL(false /* add */);
 724   case vmIntrinsics::_decrementExactI:          return inline_math_subtractExactI(true /* decrement */);
 725   case vmIntrinsics::_decrementExactL:          return inline_math_subtractExactL(true /* decrement */);
 726   case vmIntrinsics::_incrementExactI:          return inline_math_addExactI(true /* increment */);
 727   case vmIntrinsics::_incrementExactL:          return inline_math_addExactL(true /* increment */);
 728   case vmIntrinsics::_multiplyExactI:           return inline_math_multiplyExactI();
 729   case vmIntrinsics::_multiplyExactL:           return inline_math_multiplyExactL();
 730   case vmIntrinsics::_negateExactI:             return inline_math_negateExactI();
 731   case vmIntrinsics::_negateExactL:             return inline_math_negateExactL();
 732   case vmIntrinsics::_subtractExactI:           return inline_math_subtractExactI(false /* subtract */);
 733   case vmIntrinsics::_subtractExactL:           return inline_math_subtractExactL(false /* subtract */);
 734 
 735   case vmIntrinsics::_arraycopy:                return inline_arraycopy();
 736 
 737   case vmIntrinsics::_compareTo:                return inline_string_compareTo();
 738   case vmIntrinsics::_indexOf:                  return inline_string_indexOf();
 739   case vmIntrinsics::_equals:                   return inline_string_equals();
 740 
 741   case vmIntrinsics::_getObject:                return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT,  !is_volatile);
 742   case vmIntrinsics::_getBoolean:               return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
 743   case vmIntrinsics::_getByte:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE,    !is_volatile);
 744   case vmIntrinsics::_getShort:                 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,   !is_volatile);
 745   case vmIntrinsics::_getChar:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,    !is_volatile);
 746   case vmIntrinsics::_getInt:                   return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,     !is_volatile);
 747   case vmIntrinsics::_getLong:                  return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,    !is_volatile);
 748   case vmIntrinsics::_getFloat:                 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT,   !is_volatile);
 749   case vmIntrinsics::_getDouble:                return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE,  !is_volatile);
 750 
 751   case vmIntrinsics::_putObject:                return inline_unsafe_access(!is_native_ptr,  is_store, T_OBJECT,  !is_volatile);
 752   case vmIntrinsics::_putBoolean:               return inline_unsafe_access(!is_native_ptr,  is_store, T_BOOLEAN, !is_volatile);
 753   case vmIntrinsics::_putByte:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_BYTE,    !is_volatile);
 754   case vmIntrinsics::_putShort:                 return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,   !is_volatile);
 755   case vmIntrinsics::_putChar:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,    !is_volatile);
 756   case vmIntrinsics::_putInt:                   return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,     !is_volatile);
 757   case vmIntrinsics::_putLong:                  return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,    !is_volatile);
 758   case vmIntrinsics::_putFloat:                 return inline_unsafe_access(!is_native_ptr,  is_store, T_FLOAT,   !is_volatile);
 759   case vmIntrinsics::_putDouble:                return inline_unsafe_access(!is_native_ptr,  is_store, T_DOUBLE,  !is_volatile);
 760 
 761   case vmIntrinsics::_getByte_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE,    !is_volatile);
 762   case vmIntrinsics::_getShort_raw:             return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT,   !is_volatile);
 763   case vmIntrinsics::_getChar_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR,    !is_volatile);
 764   case vmIntrinsics::_getInt_raw:               return inline_unsafe_access( is_native_ptr, !is_store, T_INT,     !is_volatile);
 765   case vmIntrinsics::_getLong_raw:              return inline_unsafe_access( is_native_ptr, !is_store, T_LONG,    !is_volatile);
 766   case vmIntrinsics::_getFloat_raw:             return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT,   !is_volatile);
 767   case vmIntrinsics::_getDouble_raw:            return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE,  !is_volatile);
 768   case vmIntrinsics::_getAddress_raw:           return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
 769 
 770   case vmIntrinsics::_putByte_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_BYTE,    !is_volatile);
 771   case vmIntrinsics::_putShort_raw:             return inline_unsafe_access( is_native_ptr,  is_store, T_SHORT,   !is_volatile);
 772   case vmIntrinsics::_putChar_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_CHAR,    !is_volatile);
 773   case vmIntrinsics::_putInt_raw:               return inline_unsafe_access( is_native_ptr,  is_store, T_INT,     !is_volatile);
 774   case vmIntrinsics::_putLong_raw:              return inline_unsafe_access( is_native_ptr,  is_store, T_LONG,    !is_volatile);
 775   case vmIntrinsics::_putFloat_raw:             return inline_unsafe_access( is_native_ptr,  is_store, T_FLOAT,   !is_volatile);
 776   case vmIntrinsics::_putDouble_raw:            return inline_unsafe_access( is_native_ptr,  is_store, T_DOUBLE,  !is_volatile);
 777   case vmIntrinsics::_putAddress_raw:           return inline_unsafe_access( is_native_ptr,  is_store, T_ADDRESS, !is_volatile);
 778 
 779   case vmIntrinsics::_getObjectVolatile:        return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT,   is_volatile);
 780   case vmIntrinsics::_getBooleanVolatile:       return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN,  is_volatile);
 781   case vmIntrinsics::_getByteVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE,     is_volatile);
 782   case vmIntrinsics::_getShortVolatile:         return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT,    is_volatile);
 783   case vmIntrinsics::_getCharVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR,     is_volatile);
 784   case vmIntrinsics::_getIntVolatile:           return inline_unsafe_access(!is_native_ptr, !is_store, T_INT,      is_volatile);
 785   case vmIntrinsics::_getLongVolatile:          return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG,     is_volatile);
 786   case vmIntrinsics::_getFloatVolatile:         return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT,    is_volatile);
 787   case vmIntrinsics::_getDoubleVolatile:        return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE,   is_volatile);
 788 
 789   case vmIntrinsics::_putObjectVolatile:        return inline_unsafe_access(!is_native_ptr,  is_store, T_OBJECT,   is_volatile);
 790   case vmIntrinsics::_putBooleanVolatile:       return inline_unsafe_access(!is_native_ptr,  is_store, T_BOOLEAN,  is_volatile);
 791   case vmIntrinsics::_putByteVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_BYTE,     is_volatile);
 792   case vmIntrinsics::_putShortVolatile:         return inline_unsafe_access(!is_native_ptr,  is_store, T_SHORT,    is_volatile);
 793   case vmIntrinsics::_putCharVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_CHAR,     is_volatile);
 794   case vmIntrinsics::_putIntVolatile:           return inline_unsafe_access(!is_native_ptr,  is_store, T_INT,      is_volatile);
 795   case vmIntrinsics::_putLongVolatile:          return inline_unsafe_access(!is_native_ptr,  is_store, T_LONG,     is_volatile);
 796   case vmIntrinsics::_putFloatVolatile:         return inline_unsafe_access(!is_native_ptr,  is_store, T_FLOAT,    is_volatile);
 797   case vmIntrinsics::_putDoubleVolatile:        return inline_unsafe_access(!is_native_ptr,  is_store, T_DOUBLE,   is_volatile);
 798 
 799   case vmIntrinsics::_prefetchRead:             return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
 800   case vmIntrinsics::_prefetchWrite:            return inline_unsafe_prefetch(!is_native_ptr,  is_store, !is_static);
 801   case vmIntrinsics::_prefetchReadStatic:       return inline_unsafe_prefetch(!is_native_ptr, !is_store,  is_static);
 802   case vmIntrinsics::_prefetchWriteStatic:      return inline_unsafe_prefetch(!is_native_ptr,  is_store,  is_static);
 803 
 804   case vmIntrinsics::_compareAndSwapObject:     return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
 805   case vmIntrinsics::_compareAndSwapInt:        return inline_unsafe_load_store(T_INT,    LS_cmpxchg);
 806   case vmIntrinsics::_compareAndSwapLong:       return inline_unsafe_load_store(T_LONG,   LS_cmpxchg);
 807 
 808   case vmIntrinsics::_putOrderedObject:         return inline_unsafe_ordered_store(T_OBJECT);
 809   case vmIntrinsics::_putOrderedInt:            return inline_unsafe_ordered_store(T_INT);
 810   case vmIntrinsics::_putOrderedLong:           return inline_unsafe_ordered_store(T_LONG);
 811 
 812   case vmIntrinsics::_getAndAddInt:             return inline_unsafe_load_store(T_INT,    LS_xadd);
 813   case vmIntrinsics::_getAndAddLong:            return inline_unsafe_load_store(T_LONG,   LS_xadd);
 814   case vmIntrinsics::_getAndSetInt:             return inline_unsafe_load_store(T_INT,    LS_xchg);
 815   case vmIntrinsics::_getAndSetLong:            return inline_unsafe_load_store(T_LONG,   LS_xchg);
 816   case vmIntrinsics::_getAndSetObject:          return inline_unsafe_load_store(T_OBJECT, LS_xchg);
 817 
 818   case vmIntrinsics::_loadFence:
 819   case vmIntrinsics::_storeFence:
 820   case vmIntrinsics::_fullFence:                return inline_unsafe_fence(intrinsic_id());
 821 


 822   case vmIntrinsics::_currentThread:            return inline_native_currentThread();
 823   case vmIntrinsics::_isInterrupted:            return inline_native_isInterrupted();
 824 
 825 #ifdef TRACE_HAVE_INTRINSICS
 826   case vmIntrinsics::_classID:                  return inline_native_classID();
 827   case vmIntrinsics::_threadID:                 return inline_native_threadID();
 828   case vmIntrinsics::_counterTime:              return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
 829 #endif
 830   case vmIntrinsics::_currentTimeMillis:        return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
 831   case vmIntrinsics::_nanoTime:                 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
 832   case vmIntrinsics::_allocateInstance:         return inline_unsafe_allocate();
 833   case vmIntrinsics::_copyMemory:               return inline_unsafe_copyMemory();
 834   case vmIntrinsics::_newArray:                 return inline_native_newArray();
 835   case vmIntrinsics::_getLength:                return inline_native_getLength();
 836   case vmIntrinsics::_copyOf:                   return inline_array_copyOf(false);
 837   case vmIntrinsics::_copyOfRange:              return inline_array_copyOf(true);
 838   case vmIntrinsics::_equalsC:                  return inline_array_equals();
 839   case vmIntrinsics::_clone:                    return inline_native_clone(intrinsic()->is_virtual());
 840 
 841   case vmIntrinsics::_isAssignableFrom:         return inline_native_subtype_check();
 842 
 843   case vmIntrinsics::_isInstance:
 844   case vmIntrinsics::_getModifiers:
 845   case vmIntrinsics::_isInterface:
 846   case vmIntrinsics::_isArray:
 847   case vmIntrinsics::_isPrimitive:
 848   case vmIntrinsics::_getSuperclass:
 849   case vmIntrinsics::_getClassAccessFlags:      return inline_native_Class_query(intrinsic_id());
 850 
 851   case vmIntrinsics::_floatToRawIntBits:
 852   case vmIntrinsics::_floatToIntBits:
 853   case vmIntrinsics::_intBitsToFloat:
 854   case vmIntrinsics::_doubleToRawLongBits:
 855   case vmIntrinsics::_doubleToLongBits:
 856   case vmIntrinsics::_longBitsToDouble:         return inline_fp_conversions(intrinsic_id());
 857 
 858   case vmIntrinsics::_numberOfLeadingZeros_i:
 859   case vmIntrinsics::_numberOfLeadingZeros_l:
 860   case vmIntrinsics::_numberOfTrailingZeros_i:
 861   case vmIntrinsics::_numberOfTrailingZeros_l:
 862   case vmIntrinsics::_bitCount_i:
 863   case vmIntrinsics::_bitCount_l:
 864   case vmIntrinsics::_reverseBytes_i:
 865   case vmIntrinsics::_reverseBytes_l:
 866   case vmIntrinsics::_reverseBytes_s:
 867   case vmIntrinsics::_reverseBytes_c:           return inline_number_methods(intrinsic_id());
 868 
 869   case vmIntrinsics::_getCallerClass:           return inline_native_Reflection_getCallerClass();
 870 
 871   case vmIntrinsics::_Reference_get:            return inline_reference_get();
 872 
 873   case vmIntrinsics::_Class_cast:               return inline_Class_cast();
 874 
 875   case vmIntrinsics::_aescrypt_encryptBlock:
 876   case vmIntrinsics::_aescrypt_decryptBlock:    return inline_aescrypt_Block(intrinsic_id());
 877 
 878   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 879   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 880     return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
 881 
 882   case vmIntrinsics::_sha_implCompress:
 883   case vmIntrinsics::_sha2_implCompress:
 884   case vmIntrinsics::_sha5_implCompress:
 885     return inline_sha_implCompress(intrinsic_id());
 886 
 887   case vmIntrinsics::_digestBase_implCompressMB:
 888     return inline_digestBase_implCompressMB(predicate);
 889 
 890   case vmIntrinsics::_multiplyToLen:
 891     return inline_multiplyToLen();
 892 
 893   case vmIntrinsics::_encodeISOArray:
 894     return inline_encodeISOArray();
 895 
 896   case vmIntrinsics::_updateCRC32:
 897     return inline_updateCRC32();
 898   case vmIntrinsics::_updateBytesCRC32:
 899     return inline_updateBytesCRC32();
 900   case vmIntrinsics::_updateByteBufferCRC32:
 901     return inline_updateByteBufferCRC32();
 902 
 903   default:
 904     // If you get here, it may be that someone has added a new intrinsic
 905     // to the list in vmSymbols.hpp without implementing it here.
 906 #ifndef PRODUCT
 907     if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
 908       tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
 909                     vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
 910     }
 911 #endif
 912     return false;
 913   }
 914 }
 915 
 916 Node* LibraryCallKit::try_to_predicate(int predicate) {
 917   if (!jvms()->has_method()) {
 918     // Root JVMState has a null method.
 919     assert(map()->memory()->Opcode() == Op_Parm, "");
 920     // Insert the memory aliasing node
 921     set_all_memory(reset_memory());
 922   }
 923   assert(merged_memory(), "");
 924 
 925   switch (intrinsic_id()) {
 926   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
 927     return inline_cipherBlockChaining_AESCrypt_predicate(false);
 928   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
 929     return inline_cipherBlockChaining_AESCrypt_predicate(true);
 930   case vmIntrinsics::_digestBase_implCompressMB:
 931     return inline_digestBase_implCompressMB_predicate(predicate);
 932 
 933   default:
 934     // If you get here, it may be that someone has added a new intrinsic
 935     // to the list in vmSymbols.hpp without implementing it here.
 936 #ifndef PRODUCT
 937     if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
 938       tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
 939                     vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
 940     }
 941 #endif
 942     Node* slow_ctl = control();
 943     set_control(top()); // No fast path instrinsic
 944     return slow_ctl;
 945   }
 946 }
 947 
 948 //------------------------------set_result-------------------------------
 949 // Helper function for finishing intrinsics.
 950 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
 951   record_for_igvn(region);
 952   set_control(_gvn.transform(region));
 953   set_result( _gvn.transform(value));
 954   assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
 955 }
 956 
 957 //------------------------------generate_guard---------------------------
 958 // Helper function for generating guarded fast-slow graph structures.
 959 // The given 'test', if true, guards a slow path.  If the test fails
 960 // then a fast path can be taken.  (We generally hope it fails.)
 961 // In all cases, GraphKit::control() is updated to the fast path.
 962 // The returned value represents the control for the slow path.
 963 // The return value is never 'top'; it is either a valid control
 964 // or NULL if it is obvious that the slow path can never be taken.
 965 // Also, if region and the slow control are not NULL, the slow edge
 966 // is appended to the region.
 967 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
 968   if (stopped()) {
 969     // Already short circuited.
 970     return NULL;
 971   }
 972 
 973   // Build an if node and its projections.
 974   // If test is true we take the slow path, which we assume is uncommon.
 975   if (_gvn.type(test) == TypeInt::ZERO) {
 976     // The slow branch is never taken.  No need to build this guard.
 977     return NULL;
 978   }
 979 
 980   IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
 981 
 982   Node* if_slow = _gvn.transform(new IfTrueNode(iff));
 983   if (if_slow == top()) {
 984     // The slow branch is never taken.  No need to build this guard.
 985     return NULL;
 986   }
 987 
 988   if (region != NULL)
 989     region->add_req(if_slow);
 990 
 991   Node* if_fast = _gvn.transform(new IfFalseNode(iff));
 992   set_control(if_fast);
 993 
 994   return if_slow;
 995 }
 996 
 997 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
 998   return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
 999 }
1000 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
1001   return generate_guard(test, region, PROB_FAIR);
1002 }
1003 
1004 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
1005                                                      Node* *pos_index) {
1006   if (stopped())
1007     return NULL;                // already stopped
1008   if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
1009     return NULL;                // index is already adequately typed
1010   Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
1011   Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1012   Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1013   if (is_neg != NULL && pos_index != NULL) {
1014     // Emulate effect of Parse::adjust_map_after_if.
1015     Node* ccast = new CastIINode(index, TypeInt::POS);
1016     ccast->set_req(0, control());
1017     (*pos_index) = _gvn.transform(ccast);
1018   }
1019   return is_neg;
1020 }
1021 
1022 // Make sure that 'position' is a valid limit index, in [0..length].
1023 // There are two equivalent plans for checking this:
1024 //   A. (offset + copyLength)  unsigned<=  arrayLength
1025 //   B. offset  <=  (arrayLength - copyLength)
1026 // We require that all of the values above, except for the sum and
1027 // difference, are already known to be non-negative.
1028 // Plan A is robust in the face of overflow, if offset and copyLength
1029 // are both hugely positive.
1030 //
1031 // Plan B is less direct and intuitive, but it does not overflow at
1032 // all, since the difference of two non-negatives is always
1033 // representable.  Whenever Java methods must perform the equivalent
1034 // check they generally use Plan B instead of Plan A.
1035 // For the moment we use Plan A.
1036 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1037                                                   Node* subseq_length,
1038                                                   Node* array_length,
1039                                                   RegionNode* region) {
1040   if (stopped())
1041     return NULL;                // already stopped
1042   bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1043   if (zero_offset && subseq_length->eqv_uncast(array_length))
1044     return NULL;                // common case of whole-array copy
1045   Node* last = subseq_length;
1046   if (!zero_offset)             // last += offset
1047     last = _gvn.transform(new AddINode(last, offset));
1048   Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1049   Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1050   Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1051   return is_over;
1052 }
1053 
1054 
1055 //--------------------------generate_current_thread--------------------
1056 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1057   ciKlass*    thread_klass = env()->Thread_klass();
1058   const Type* thread_type  = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1059   Node* thread = _gvn.transform(new ThreadLocalNode());
1060   Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1061   Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1062   tls_output = thread;
1063   return threadObj;
1064 }
1065 
1066 
1067 //------------------------------make_string_method_node------------------------
1068 // Helper method for String intrinsic functions. This version is called
1069 // with str1 and str2 pointing to String object nodes.
1070 //
1071 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1072   Node* no_ctrl = NULL;
1073 
1074   // Get start addr of string
1075   Node* str1_value   = load_String_value(no_ctrl, str1);
1076   Node* str1_offset  = load_String_offset(no_ctrl, str1);
1077   Node* str1_start   = array_element_address(str1_value, str1_offset, T_CHAR);
1078 
1079   // Get length of string 1
1080   Node* str1_len  = load_String_length(no_ctrl, str1);
1081 
1082   Node* str2_value   = load_String_value(no_ctrl, str2);
1083   Node* str2_offset  = load_String_offset(no_ctrl, str2);
1084   Node* str2_start   = array_element_address(str2_value, str2_offset, T_CHAR);
1085 
1086   Node* str2_len = NULL;
1087   Node* result = NULL;
1088 
1089   switch (opcode) {
1090   case Op_StrIndexOf:
1091     // Get length of string 2
1092     str2_len = load_String_length(no_ctrl, str2);
1093 
1094     result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1095                                 str1_start, str1_len, str2_start, str2_len);
1096     break;
1097   case Op_StrComp:
1098     // Get length of string 2
1099     str2_len = load_String_length(no_ctrl, str2);
1100 
1101     result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1102                              str1_start, str1_len, str2_start, str2_len);
1103     break;
1104   case Op_StrEquals:
1105     result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1106                                str1_start, str2_start, str1_len);
1107     break;
1108   default:
1109     ShouldNotReachHere();
1110     return NULL;
1111   }
1112 
1113   // All these intrinsics have checks.
1114   C->set_has_split_ifs(true); // Has chance for split-if optimization
1115 
1116   return _gvn.transform(result);
1117 }
1118 
1119 // Helper method for String intrinsic functions. This version is called
1120 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1121 // to Int nodes containing the lenghts of str1 and str2.
1122 //
1123 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1124   Node* result = NULL;
1125   switch (opcode) {
1126   case Op_StrIndexOf:
1127     result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1128                                 str1_start, cnt1, str2_start, cnt2);
1129     break;
1130   case Op_StrComp:
1131     result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
1132                              str1_start, cnt1, str2_start, cnt2);
1133     break;
1134   case Op_StrEquals:
1135     result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1136                                str1_start, str2_start, cnt1);
1137     break;
1138   default:
1139     ShouldNotReachHere();
1140     return NULL;
1141   }
1142 
1143   // All these intrinsics have checks.
1144   C->set_has_split_ifs(true); // Has chance for split-if optimization
1145 
1146   return _gvn.transform(result);
1147 }
1148 
1149 //------------------------------inline_string_compareTo------------------------
1150 // public int java.lang.String.compareTo(String anotherString);
1151 bool LibraryCallKit::inline_string_compareTo() {
1152   Node* receiver = null_check(argument(0));
1153   Node* arg      = null_check(argument(1));
1154   if (stopped()) {
1155     return true;
1156   }
1157   set_result(make_string_method_node(Op_StrComp, receiver, arg));
1158   return true;
1159 }
1160 
1161 //------------------------------inline_string_equals------------------------
1162 bool LibraryCallKit::inline_string_equals() {
1163   Node* receiver = null_check_receiver();
1164   // NOTE: Do not null check argument for String.equals() because spec
1165   // allows to specify NULL as argument.
1166   Node* argument = this->argument(1);
1167   if (stopped()) {
1168     return true;
1169   }
1170 
1171   // paths (plus control) merge
1172   RegionNode* region = new RegionNode(5);
1173   Node* phi = new PhiNode(region, TypeInt::BOOL);
1174 
1175   // does source == target string?
1176   Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1177   Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1178 
1179   Node* if_eq = generate_slow_guard(bol, NULL);
1180   if (if_eq != NULL) {
1181     // receiver == argument
1182     phi->init_req(2, intcon(1));
1183     region->init_req(2, if_eq);
1184   }
1185 
1186   // get String klass for instanceOf
1187   ciInstanceKlass* klass = env()->String_klass();
1188 
1189   if (!stopped()) {
1190     Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1191     Node* cmp  = _gvn.transform(new CmpINode(inst, intcon(1)));
1192     Node* bol  = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1193 
1194     Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1195     //instanceOf == true, fallthrough
1196 
1197     if (inst_false != NULL) {
1198       phi->init_req(3, intcon(0));
1199       region->init_req(3, inst_false);
1200     }
1201   }
1202 
1203   if (!stopped()) {
1204     const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1205 
1206     // Properly cast the argument to String
1207     argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1208     // This path is taken only when argument's type is String:NotNull.
1209     argument = cast_not_null(argument, false);
1210 
1211     Node* no_ctrl = NULL;
1212 
1213     // Get start addr of receiver
1214     Node* receiver_val    = load_String_value(no_ctrl, receiver);
1215     Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1216     Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1217 
1218     // Get length of receiver
1219     Node* receiver_cnt  = load_String_length(no_ctrl, receiver);
1220 
1221     // Get start addr of argument
1222     Node* argument_val    = load_String_value(no_ctrl, argument);
1223     Node* argument_offset = load_String_offset(no_ctrl, argument);
1224     Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1225 
1226     // Get length of argument
1227     Node* argument_cnt  = load_String_length(no_ctrl, argument);
1228 
1229     // Check for receiver count != argument count
1230     Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1231     Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1232     Node* if_ne = generate_slow_guard(bol, NULL);
1233     if (if_ne != NULL) {
1234       phi->init_req(4, intcon(0));
1235       region->init_req(4, if_ne);
1236     }
1237 
1238     // Check for count == 0 is done by assembler code for StrEquals.
1239 
1240     if (!stopped()) {
1241       Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1242       phi->init_req(1, equals);
1243       region->init_req(1, control());
1244     }
1245   }
1246 
1247   // post merge
1248   set_control(_gvn.transform(region));
1249   record_for_igvn(region);
1250 
1251   set_result(_gvn.transform(phi));
1252   return true;
1253 }
1254 
1255 //------------------------------inline_array_equals----------------------------
1256 bool LibraryCallKit::inline_array_equals() {
1257   Node* arg1 = argument(0);
1258   Node* arg2 = argument(1);
1259   set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1260   return true;
1261 }
1262 
1263 // Java version of String.indexOf(constant string)
1264 // class StringDecl {
1265 //   StringDecl(char[] ca) {
1266 //     offset = 0;
1267 //     count = ca.length;
1268 //     value = ca;
1269 //   }
1270 //   int offset;
1271 //   int count;
1272 //   char[] value;
1273 // }
1274 //
1275 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1276 //                             int targetOffset, int cache_i, int md2) {
1277 //   int cache = cache_i;
1278 //   int sourceOffset = string_object.offset;
1279 //   int sourceCount = string_object.count;
1280 //   int targetCount = target_object.length;
1281 //
1282 //   int targetCountLess1 = targetCount - 1;
1283 //   int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1284 //
1285 //   char[] source = string_object.value;
1286 //   char[] target = target_object;
1287 //   int lastChar = target[targetCountLess1];
1288 //
1289 //  outer_loop:
1290 //   for (int i = sourceOffset; i < sourceEnd; ) {
1291 //     int src = source[i + targetCountLess1];
1292 //     if (src == lastChar) {
1293 //       // With random strings and a 4-character alphabet,
1294 //       // reverse matching at this point sets up 0.8% fewer
1295 //       // frames, but (paradoxically) makes 0.3% more probes.
1296 //       // Since those probes are nearer the lastChar probe,
1297 //       // there is may be a net D$ win with reverse matching.
1298 //       // But, reversing loop inhibits unroll of inner loop
1299 //       // for unknown reason.  So, does running outer loop from
1300 //       // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1301 //       for (int j = 0; j < targetCountLess1; j++) {
1302 //         if (target[targetOffset + j] != source[i+j]) {
1303 //           if ((cache & (1 << source[i+j])) == 0) {
1304 //             if (md2 < j+1) {
1305 //               i += j+1;
1306 //               continue outer_loop;
1307 //             }
1308 //           }
1309 //           i += md2;
1310 //           continue outer_loop;
1311 //         }
1312 //       }
1313 //       return i - sourceOffset;
1314 //     }
1315 //     if ((cache & (1 << src)) == 0) {
1316 //       i += targetCountLess1;
1317 //     } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1318 //     i++;
1319 //   }
1320 //   return -1;
1321 // }
1322 
1323 //------------------------------string_indexOf------------------------
1324 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1325                                      jint cache_i, jint md2_i) {
1326 
1327   Node* no_ctrl  = NULL;
1328   float likely   = PROB_LIKELY(0.9);
1329   float unlikely = PROB_UNLIKELY(0.9);
1330 
1331   const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1332 
1333   Node* source        = load_String_value(no_ctrl, string_object);
1334   Node* sourceOffset  = load_String_offset(no_ctrl, string_object);
1335   Node* sourceCount   = load_String_length(no_ctrl, string_object);
1336 
1337   Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1338   jint target_length = target_array->length();
1339   const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1340   const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1341 
1342   // String.value field is known to be @Stable.
1343   if (UseImplicitStableValues) {
1344     target = cast_array_to_stable(target, target_type);
1345   }
1346 
1347   IdealKit kit(this, false, true);
1348 #define __ kit.
1349   Node* zero             = __ ConI(0);
1350   Node* one              = __ ConI(1);
1351   Node* cache            = __ ConI(cache_i);
1352   Node* md2              = __ ConI(md2_i);
1353   Node* lastChar         = __ ConI(target_array->char_at(target_length - 1));
1354   Node* targetCount      = __ ConI(target_length);
1355   Node* targetCountLess1 = __ ConI(target_length - 1);
1356   Node* targetOffset     = __ ConI(targetOffset_i);
1357   Node* sourceEnd        = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1358 
1359   IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1360   Node* outer_loop = __ make_label(2 /* goto */);
1361   Node* return_    = __ make_label(1);
1362 
1363   __ set(rtn,__ ConI(-1));
1364   __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1365        Node* i2  = __ AddI(__ value(i), targetCountLess1);
1366        // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1367        Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1368        __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1369          __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1370               Node* tpj = __ AddI(targetOffset, __ value(j));
1371               Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1372               Node* ipj  = __ AddI(__ value(i), __ value(j));
1373               Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1374               __ if_then(targ, BoolTest::ne, src2); {
1375                 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1376                   __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1377                     __ increment(i, __ AddI(__ value(j), one));
1378                     __ goto_(outer_loop);
1379                   } __ end_if(); __ dead(j);
1380                 }__ end_if(); __ dead(j);
1381                 __ increment(i, md2);
1382                 __ goto_(outer_loop);
1383               }__ end_if();
1384               __ increment(j, one);
1385          }__ end_loop(); __ dead(j);
1386          __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1387          __ goto_(return_);
1388        }__ end_if();
1389        __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1390          __ increment(i, targetCountLess1);
1391        }__ end_if();
1392        __ increment(i, one);
1393        __ bind(outer_loop);
1394   }__ end_loop(); __ dead(i);
1395   __ bind(return_);
1396 
1397   // Final sync IdealKit and GraphKit.
1398   final_sync(kit);
1399   Node* result = __ value(rtn);
1400 #undef __
1401   C->set_has_loops(true);
1402   return result;
1403 }
1404 
1405 //------------------------------inline_string_indexOf------------------------
1406 bool LibraryCallKit::inline_string_indexOf() {
1407   Node* receiver = argument(0);
1408   Node* arg      = argument(1);
1409 
1410   Node* result;
1411   // Disable the use of pcmpestri until it can be guaranteed that
1412   // the load doesn't cross into the uncommited space.
1413   if (Matcher::has_match_rule(Op_StrIndexOf) &&
1414       UseSSE42Intrinsics) {
1415     // Generate SSE4.2 version of indexOf
1416     // We currently only have match rules that use SSE4.2
1417 
1418     receiver = null_check(receiver);
1419     arg      = null_check(arg);
1420     if (stopped()) {
1421       return true;
1422     }
1423 
1424     ciInstanceKlass* str_klass = env()->String_klass();
1425     const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1426 
1427     // Make the merge point
1428     RegionNode* result_rgn = new RegionNode(4);
1429     Node*       result_phi = new PhiNode(result_rgn, TypeInt::INT);
1430     Node* no_ctrl  = NULL;
1431 
1432     // Get start addr of source string
1433     Node* source = load_String_value(no_ctrl, receiver);
1434     Node* source_offset = load_String_offset(no_ctrl, receiver);
1435     Node* source_start = array_element_address(source, source_offset, T_CHAR);
1436 
1437     // Get length of source string
1438     Node* source_cnt  = load_String_length(no_ctrl, receiver);
1439 
1440     // Get start addr of substring
1441     Node* substr = load_String_value(no_ctrl, arg);
1442     Node* substr_offset = load_String_offset(no_ctrl, arg);
1443     Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1444 
1445     // Get length of source string
1446     Node* substr_cnt  = load_String_length(no_ctrl, arg);
1447 
1448     // Check for substr count > string count
1449     Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1450     Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1451     Node* if_gt = generate_slow_guard(bol, NULL);
1452     if (if_gt != NULL) {
1453       result_phi->init_req(2, intcon(-1));
1454       result_rgn->init_req(2, if_gt);
1455     }
1456 
1457     if (!stopped()) {
1458       // Check for substr count == 0
1459       cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1460       bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1461       Node* if_zero = generate_slow_guard(bol, NULL);
1462       if (if_zero != NULL) {
1463         result_phi->init_req(3, intcon(0));
1464         result_rgn->init_req(3, if_zero);
1465       }
1466     }
1467 
1468     if (!stopped()) {
1469       result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1470       result_phi->init_req(1, result);
1471       result_rgn->init_req(1, control());
1472     }
1473     set_control(_gvn.transform(result_rgn));
1474     record_for_igvn(result_rgn);
1475     result = _gvn.transform(result_phi);
1476 
1477   } else { // Use LibraryCallKit::string_indexOf
1478     // don't intrinsify if argument isn't a constant string.
1479     if (!arg->is_Con()) {
1480      return false;
1481     }
1482     const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1483     if (str_type == NULL) {
1484       return false;
1485     }
1486     ciInstanceKlass* klass = env()->String_klass();
1487     ciObject* str_const = str_type->const_oop();
1488     if (str_const == NULL || str_const->klass() != klass) {
1489       return false;
1490     }
1491     ciInstance* str = str_const->as_instance();
1492     assert(str != NULL, "must be instance");
1493 
1494     ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1495     ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1496 
1497     int o;
1498     int c;
1499     if (java_lang_String::has_offset_field()) {
1500       o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1501       c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1502     } else {
1503       o = 0;
1504       c = pat->length();
1505     }
1506 
1507     // constant strings have no offset and count == length which
1508     // simplifies the resulting code somewhat so lets optimize for that.
1509     if (o != 0 || c != pat->length()) {
1510      return false;
1511     }
1512 
1513     receiver = null_check(receiver, T_OBJECT);
1514     // NOTE: No null check on the argument is needed since it's a constant String oop.
1515     if (stopped()) {
1516       return true;
1517     }
1518 
1519     // The null string as a pattern always returns 0 (match at beginning of string)
1520     if (c == 0) {
1521       set_result(intcon(0));
1522       return true;
1523     }
1524 
1525     // Generate default indexOf
1526     jchar lastChar = pat->char_at(o + (c - 1));
1527     int cache = 0;
1528     int i;
1529     for (i = 0; i < c - 1; i++) {
1530       assert(i < pat->length(), "out of range");
1531       cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1532     }
1533 
1534     int md2 = c;
1535     for (i = 0; i < c - 1; i++) {
1536       assert(i < pat->length(), "out of range");
1537       if (pat->char_at(o + i) == lastChar) {
1538         md2 = (c - 1) - i;
1539       }
1540     }
1541 
1542     result = string_indexOf(receiver, pat, o, cache, md2);
1543   }
1544   set_result(result);
1545   return true;
1546 }
1547 
1548 //--------------------------round_double_node--------------------------------
1549 // Round a double node if necessary.
1550 Node* LibraryCallKit::round_double_node(Node* n) {
1551   if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1552     n = _gvn.transform(new RoundDoubleNode(0, n));
1553   return n;
1554 }
1555 
1556 //------------------------------inline_math-----------------------------------
1557 // public static double Math.abs(double)
1558 // public static double Math.sqrt(double)
1559 // public static double Math.log(double)
1560 // public static double Math.log10(double)
1561 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1562   Node* arg = round_double_node(argument(0));
1563   Node* n;
1564   switch (id) {
1565   case vmIntrinsics::_dabs:   n = new AbsDNode(                arg);  break;
1566   case vmIntrinsics::_dsqrt:  n = new SqrtDNode(C, control(),  arg);  break;
1567   case vmIntrinsics::_dlog:   n = new LogDNode(C, control(),   arg);  break;
1568   case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg);  break;
1569   default:  fatal_unexpected_iid(id);  break;
1570   }
1571   set_result(_gvn.transform(n));
1572   return true;
1573 }
1574 
1575 //------------------------------inline_trig----------------------------------
1576 // Inline sin/cos/tan instructions, if possible.  If rounding is required, do
1577 // argument reduction which will turn into a fast/slow diamond.
1578 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1579   Node* arg = round_double_node(argument(0));
1580   Node* n = NULL;
1581 
1582   switch (id) {
1583   case vmIntrinsics::_dsin:  n = new SinDNode(C, control(), arg);  break;
1584   case vmIntrinsics::_dcos:  n = new CosDNode(C, control(), arg);  break;
1585   case vmIntrinsics::_dtan:  n = new TanDNode(C, control(), arg);  break;
1586   default:  fatal_unexpected_iid(id);  break;
1587   }
1588   n = _gvn.transform(n);
1589 
1590   // Rounding required?  Check for argument reduction!
1591   if (Matcher::strict_fp_requires_explicit_rounding) {
1592     static const double     pi_4 =  0.7853981633974483;
1593     static const double neg_pi_4 = -0.7853981633974483;
1594     // pi/2 in 80-bit extended precision
1595     // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1596     // -pi/2 in 80-bit extended precision
1597     // 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};
1598     // Cutoff value for using this argument reduction technique
1599     //static const double    pi_2_minus_epsilon =  1.564660403643354;
1600     //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1601 
1602     // Pseudocode for sin:
1603     // if (x <= Math.PI / 4.0) {
1604     //   if (x >= -Math.PI / 4.0) return  fsin(x);
1605     //   if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1606     // } else {
1607     //   if (x <=  Math.PI / 2.0) return  fcos(x - Math.PI / 2.0);
1608     // }
1609     // return StrictMath.sin(x);
1610 
1611     // Pseudocode for cos:
1612     // if (x <= Math.PI / 4.0) {
1613     //   if (x >= -Math.PI / 4.0) return  fcos(x);
1614     //   if (x >= -Math.PI / 2.0) return  fsin(x + Math.PI / 2.0);
1615     // } else {
1616     //   if (x <=  Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1617     // }
1618     // return StrictMath.cos(x);
1619 
1620     // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1621     // requires a special machine instruction to load it.  Instead we'll try
1622     // the 'easy' case.  If we really need the extra range +/- PI/2 we'll
1623     // probably do the math inside the SIN encoding.
1624 
1625     // Make the merge point
1626     RegionNode* r = new RegionNode(3);
1627     Node* phi = new PhiNode(r, Type::DOUBLE);
1628 
1629     // Flatten arg so we need only 1 test
1630     Node *abs = _gvn.transform(new AbsDNode(arg));
1631     // Node for PI/4 constant
1632     Node *pi4 = makecon(TypeD::make(pi_4));
1633     // Check PI/4 : abs(arg)
1634     Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1635     // Check: If PI/4 < abs(arg) then go slow
1636     Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1637     // Branch either way
1638     IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1639     set_control(opt_iff(r,iff));
1640 
1641     // Set fast path result
1642     phi->init_req(2, n);
1643 
1644     // Slow path - non-blocking leaf call
1645     Node* call = NULL;
1646     switch (id) {
1647     case vmIntrinsics::_dsin:
1648       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1649                                CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1650                                "Sin", NULL, arg, top());
1651       break;
1652     case vmIntrinsics::_dcos:
1653       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1654                                CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1655                                "Cos", NULL, arg, top());
1656       break;
1657     case vmIntrinsics::_dtan:
1658       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1659                                CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1660                                "Tan", NULL, arg, top());
1661       break;
1662     }
1663     assert(control()->in(0) == call, "");
1664     Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1665     r->init_req(1, control());
1666     phi->init_req(1, slow_result);
1667 
1668     // Post-merge
1669     set_control(_gvn.transform(r));
1670     record_for_igvn(r);
1671     n = _gvn.transform(phi);
1672 
1673     C->set_has_split_ifs(true); // Has chance for split-if optimization
1674   }
1675   set_result(n);
1676   return true;
1677 }
1678 
1679 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1680   //-------------------
1681   //result=(result.isNaN())? funcAddr():result;
1682   // Check: If isNaN() by checking result!=result? then either trap
1683   // or go to runtime
1684   Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1685   // Build the boolean node
1686   Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1687 
1688   if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1689     { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1690       // The pow or exp intrinsic returned a NaN, which requires a call
1691       // to the runtime.  Recompile with the runtime call.
1692       uncommon_trap(Deoptimization::Reason_intrinsic,
1693                     Deoptimization::Action_make_not_entrant);
1694     }
1695     return result;
1696   } else {
1697     // If this inlining ever returned NaN in the past, we compile a call
1698     // to the runtime to properly handle corner cases
1699 
1700     IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1701     Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1702     Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1703 
1704     if (!if_slow->is_top()) {
1705       RegionNode* result_region = new RegionNode(3);
1706       PhiNode*    result_val = new PhiNode(result_region, Type::DOUBLE);
1707 
1708       result_region->init_req(1, if_fast);
1709       result_val->init_req(1, result);
1710 
1711       set_control(if_slow);
1712 
1713       const TypePtr* no_memory_effects = NULL;
1714       Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1715                                    no_memory_effects,
1716                                    x, top(), y, y ? top() : NULL);
1717       Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1718 #ifdef ASSERT
1719       Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1720       assert(value_top == top(), "second value must be top");
1721 #endif
1722 
1723       result_region->init_req(2, control());
1724       result_val->init_req(2, value);
1725       set_control(_gvn.transform(result_region));
1726       return _gvn.transform(result_val);
1727     } else {
1728       return result;
1729     }
1730   }
1731 }
1732 
1733 //------------------------------inline_exp-------------------------------------
1734 // Inline exp instructions, if possible.  The Intel hardware only misses
1735 // really odd corner cases (+/- Infinity).  Just uncommon-trap them.
1736 bool LibraryCallKit::inline_exp() {
1737   Node* arg = round_double_node(argument(0));
1738   Node* n   = _gvn.transform(new ExpDNode(C, control(), arg));
1739 
1740   n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1741   set_result(n);
1742 
1743   C->set_has_split_ifs(true); // Has chance for split-if optimization
1744   return true;
1745 }
1746 
1747 //------------------------------inline_pow-------------------------------------
1748 // Inline power instructions, if possible.
1749 bool LibraryCallKit::inline_pow() {
1750   // Pseudocode for pow
1751   // if (y == 2) {
1752   //   return x * x;
1753   // } else {
1754   //   if (x <= 0.0) {
1755   //     long longy = (long)y;
1756   //     if ((double)longy == y) { // if y is long
1757   //       if (y + 1 == y) longy = 0; // huge number: even
1758   //       result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1759   //     } else {
1760   //       result = NaN;
1761   //     }
1762   //   } else {
1763   //     result = DPow(x,y);
1764   //   }
1765   //   if (result != result)?  {
1766   //     result = uncommon_trap() or runtime_call();
1767   //   }
1768   //   return result;
1769   // }
1770 
1771   Node* x = round_double_node(argument(0));
1772   Node* y = round_double_node(argument(2));
1773 
1774   Node* result = NULL;
1775 
1776   Node*   const_two_node = makecon(TypeD::make(2.0));
1777   Node*   cmp_node       = _gvn.transform(new CmpDNode(y, const_two_node));
1778   Node*   bool_node      = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1779   IfNode* if_node        = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1780   Node*   if_true        = _gvn.transform(new IfTrueNode(if_node));
1781   Node*   if_false       = _gvn.transform(new IfFalseNode(if_node));
1782 
1783   RegionNode* region_node = new RegionNode(3);
1784   region_node->init_req(1, if_true);
1785 
1786   Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1787   // special case for x^y where y == 2, we can convert it to x * x
1788   phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1789 
1790   // set control to if_false since we will now process the false branch
1791   set_control(if_false);
1792 
1793   if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1794     // Short form: skip the fancy tests and just check for NaN result.
1795     result = _gvn.transform(new PowDNode(C, control(), x, y));
1796   } else {
1797     // If this inlining ever returned NaN in the past, include all
1798     // checks + call to the runtime.
1799 
1800     // Set the merge point for If node with condition of (x <= 0.0)
1801     // There are four possible paths to region node and phi node
1802     RegionNode *r = new RegionNode(4);
1803     Node *phi = new PhiNode(r, Type::DOUBLE);
1804 
1805     // Build the first if node: if (x <= 0.0)
1806     // Node for 0 constant
1807     Node *zeronode = makecon(TypeD::ZERO);
1808     // Check x:0
1809     Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1810     // Check: If (x<=0) then go complex path
1811     Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1812     // Branch either way
1813     IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1814     // Fast path taken; set region slot 3
1815     Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1816     r->init_req(3,fast_taken); // Capture fast-control
1817 
1818     // Fast path not-taken, i.e. slow path
1819     Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1820 
1821     // Set fast path result
1822     Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1823     phi->init_req(3, fast_result);
1824 
1825     // Complex path
1826     // Build the second if node (if y is long)
1827     // Node for (long)y
1828     Node *longy = _gvn.transform(new ConvD2LNode(y));
1829     // Node for (double)((long) y)
1830     Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1831     // Check (double)((long) y) : y
1832     Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1833     // Check if (y isn't long) then go to slow path
1834 
1835     Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1836     // Branch either way
1837     IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1838     Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1839 
1840     Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1841 
1842     // Calculate DPow(abs(x), y)*(1 & (long)y)
1843     // Node for constant 1
1844     Node *conone = longcon(1);
1845     // 1& (long)y
1846     Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1847 
1848     // A huge number is always even. Detect a huge number by checking
1849     // if y + 1 == y and set integer to be tested for parity to 0.
1850     // Required for corner case:
1851     // (long)9.223372036854776E18 = max_jlong
1852     // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1853     // max_jlong is odd but 9.223372036854776E18 is even
1854     Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1855     Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1856     Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1857     Node* correctedsign = NULL;
1858     if (ConditionalMoveLimit != 0) {
1859       correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1860     } else {
1861       IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1862       RegionNode *r = new RegionNode(3);
1863       Node *phi = new PhiNode(r, TypeLong::LONG);
1864       r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1865       r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1866       phi->init_req(1, signnode);
1867       phi->init_req(2, longcon(0));
1868       correctedsign = _gvn.transform(phi);
1869       ylong_path = _gvn.transform(r);
1870       record_for_igvn(r);
1871     }
1872 
1873     // zero node
1874     Node *conzero = longcon(0);
1875     // Check (1&(long)y)==0?
1876     Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1877     // Check if (1&(long)y)!=0?, if so the result is negative
1878     Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1879     // abs(x)
1880     Node *absx=_gvn.transform(new AbsDNode(x));
1881     // abs(x)^y
1882     Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1883     // -abs(x)^y
1884     Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1885     // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1886     Node *signresult = NULL;
1887     if (ConditionalMoveLimit != 0) {
1888       signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1889     } else {
1890       IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1891       RegionNode *r = new RegionNode(3);
1892       Node *phi = new PhiNode(r, Type::DOUBLE);
1893       r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1894       r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1895       phi->init_req(1, absxpowy);
1896       phi->init_req(2, negabsxpowy);
1897       signresult = _gvn.transform(phi);
1898       ylong_path = _gvn.transform(r);
1899       record_for_igvn(r);
1900     }
1901     // Set complex path fast result
1902     r->init_req(2, ylong_path);
1903     phi->init_req(2, signresult);
1904 
1905     static const jlong nan_bits = CONST64(0x7ff8000000000000);
1906     Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1907     r->init_req(1,slow_path);
1908     phi->init_req(1,slow_result);
1909 
1910     // Post merge
1911     set_control(_gvn.transform(r));
1912     record_for_igvn(r);
1913     result = _gvn.transform(phi);
1914   }
1915 
1916   result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1917 
1918   // control from finish_pow_exp is now input to the region node
1919   region_node->set_req(2, control());
1920   // the result from finish_pow_exp is now input to the phi node
1921   phi_node->init_req(2, result);
1922   set_control(_gvn.transform(region_node));
1923   record_for_igvn(region_node);
1924   set_result(_gvn.transform(phi_node));
1925 
1926   C->set_has_split_ifs(true); // Has chance for split-if optimization
1927   return true;
1928 }
1929 
1930 //------------------------------runtime_math-----------------------------
1931 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1932   assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1933          "must be (DD)D or (D)D type");
1934 
1935   // Inputs
1936   Node* a = round_double_node(argument(0));
1937   Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1938 
1939   const TypePtr* no_memory_effects = NULL;
1940   Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1941                                  no_memory_effects,
1942                                  a, top(), b, b ? top() : NULL);
1943   Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1944 #ifdef ASSERT
1945   Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1946   assert(value_top == top(), "second value must be top");
1947 #endif
1948 
1949   set_result(value);
1950   return true;
1951 }
1952 
1953 //------------------------------inline_math_native-----------------------------
1954 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1955 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1956   switch (id) {
1957     // These intrinsics are not properly supported on all hardware
1958   case vmIntrinsics::_dcos:   return Matcher::has_match_rule(Op_CosD)   ? inline_trig(id) :
1959     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos),   "COS");
1960   case vmIntrinsics::_dsin:   return Matcher::has_match_rule(Op_SinD)   ? inline_trig(id) :
1961     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin),   "SIN");
1962   case vmIntrinsics::_dtan:   return Matcher::has_match_rule(Op_TanD)   ? inline_trig(id) :
1963     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan),   "TAN");
1964 
1965   case vmIntrinsics::_dlog:   return Matcher::has_match_rule(Op_LogD)   ? inline_math(id) :
1966     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog),   "LOG");
1967   case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1968     runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1969 
1970     // These intrinsics are supported on all hardware
1971   case vmIntrinsics::_dsqrt:  return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1972   case vmIntrinsics::_dabs:   return Matcher::has_match_rule(Op_AbsD)   ? inline_math(id) : false;
1973 
1974   case vmIntrinsics::_dexp:   return Matcher::has_match_rule(Op_ExpD)   ? inline_exp()    :
1975     runtime_math(OptoRuntime::Math_D_D_Type(),  FN_PTR(SharedRuntime::dexp),  "EXP");
1976   case vmIntrinsics::_dpow:   return Matcher::has_match_rule(Op_PowD)   ? inline_pow()    :
1977     runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow),  "POW");
1978 #undef FN_PTR
1979 
1980    // These intrinsics are not yet correctly implemented
1981   case vmIntrinsics::_datan2:
1982     return false;
1983 
1984   default:
1985     fatal_unexpected_iid(id);
1986     return false;
1987   }
1988 }
1989 
1990 static bool is_simple_name(Node* n) {
1991   return (n->req() == 1         // constant
1992           || (n->is_Type() && n->as_Type()->type()->singleton())
1993           || n->is_Proj()       // parameter or return value
1994           || n->is_Phi()        // local of some sort
1995           );
1996 }
1997 
1998 //----------------------------inline_min_max-----------------------------------
1999 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2000   set_result(generate_min_max(id, argument(0), argument(1)));
2001   return true;
2002 }
2003 
2004 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2005   Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2006   IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2007   Node* fast_path = _gvn.transform( new IfFalseNode(check));
2008   Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2009 
2010   {
2011     PreserveJVMState pjvms(this);
2012     PreserveReexecuteState preexecs(this);
2013     jvms()->set_should_reexecute(true);
2014 
2015     set_control(slow_path);
2016     set_i_o(i_o());
2017 
2018     uncommon_trap(Deoptimization::Reason_intrinsic,
2019                   Deoptimization::Action_none);
2020   }
2021 
2022   set_control(fast_path);
2023   set_result(math);
2024 }
2025 
2026 template <typename OverflowOp>
2027 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2028   typedef typename OverflowOp::MathOp MathOp;
2029 
2030   MathOp* mathOp = new MathOp(arg1, arg2);
2031   Node* operation = _gvn.transform( mathOp );
2032   Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2033   inline_math_mathExact(operation, ofcheck);
2034   return true;
2035 }
2036 
2037 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2038   return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2039 }
2040 
2041 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2042   return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2043 }
2044 
2045 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2046   return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2047 }
2048 
2049 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2050   return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2051 }
2052 
2053 bool LibraryCallKit::inline_math_negateExactI() {
2054   return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2055 }
2056 
2057 bool LibraryCallKit::inline_math_negateExactL() {
2058   return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2059 }
2060 
2061 bool LibraryCallKit::inline_math_multiplyExactI() {
2062   return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2063 }
2064 
2065 bool LibraryCallKit::inline_math_multiplyExactL() {
2066   return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2067 }
2068 
2069 Node*
2070 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2071   // These are the candidate return value:
2072   Node* xvalue = x0;
2073   Node* yvalue = y0;
2074 
2075   if (xvalue == yvalue) {
2076     return xvalue;
2077   }
2078 
2079   bool want_max = (id == vmIntrinsics::_max);
2080 
2081   const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2082   const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2083   if (txvalue == NULL || tyvalue == NULL)  return top();
2084   // This is not really necessary, but it is consistent with a
2085   // hypothetical MaxINode::Value method:
2086   int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2087 
2088   // %%% This folding logic should (ideally) be in a different place.
2089   // Some should be inside IfNode, and there to be a more reliable
2090   // transformation of ?: style patterns into cmoves.  We also want
2091   // more powerful optimizations around cmove and min/max.
2092 
2093   // Try to find a dominating comparison of these guys.
2094   // It can simplify the index computation for Arrays.copyOf
2095   // and similar uses of System.arraycopy.
2096   // First, compute the normalized version of CmpI(x, y).
2097   int   cmp_op = Op_CmpI;
2098   Node* xkey = xvalue;
2099   Node* ykey = yvalue;
2100   Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2101   if (ideal_cmpxy->is_Cmp()) {
2102     // E.g., if we have CmpI(length - offset, count),
2103     // it might idealize to CmpI(length, count + offset)
2104     cmp_op = ideal_cmpxy->Opcode();
2105     xkey = ideal_cmpxy->in(1);
2106     ykey = ideal_cmpxy->in(2);
2107   }
2108 
2109   // Start by locating any relevant comparisons.
2110   Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2111   Node* cmpxy = NULL;
2112   Node* cmpyx = NULL;
2113   for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2114     Node* cmp = start_from->fast_out(k);
2115     if (cmp->outcnt() > 0 &&            // must have prior uses
2116         cmp->in(0) == NULL &&           // must be context-independent
2117         cmp->Opcode() == cmp_op) {      // right kind of compare
2118       if (cmp->in(1) == xkey && cmp->in(2) == ykey)  cmpxy = cmp;
2119       if (cmp->in(1) == ykey && cmp->in(2) == xkey)  cmpyx = cmp;
2120     }
2121   }
2122 
2123   const int NCMPS = 2;
2124   Node* cmps[NCMPS] = { cmpxy, cmpyx };
2125   int cmpn;
2126   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2127     if (cmps[cmpn] != NULL)  break;     // find a result
2128   }
2129   if (cmpn < NCMPS) {
2130     // Look for a dominating test that tells us the min and max.
2131     int depth = 0;                // Limit search depth for speed
2132     Node* dom = control();
2133     for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2134       if (++depth >= 100)  break;
2135       Node* ifproj = dom;
2136       if (!ifproj->is_Proj())  continue;
2137       Node* iff = ifproj->in(0);
2138       if (!iff->is_If())  continue;
2139       Node* bol = iff->in(1);
2140       if (!bol->is_Bool())  continue;
2141       Node* cmp = bol->in(1);
2142       if (cmp == NULL)  continue;
2143       for (cmpn = 0; cmpn < NCMPS; cmpn++)
2144         if (cmps[cmpn] == cmp)  break;
2145       if (cmpn == NCMPS)  continue;
2146       BoolTest::mask btest = bol->as_Bool()->_test._test;
2147       if (ifproj->is_IfFalse())  btest = BoolTest(btest).negate();
2148       if (cmp->in(1) == ykey)    btest = BoolTest(btest).commute();
2149       // At this point, we know that 'x btest y' is true.
2150       switch (btest) {
2151       case BoolTest::eq:
2152         // They are proven equal, so we can collapse the min/max.
2153         // Either value is the answer.  Choose the simpler.
2154         if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2155           return yvalue;
2156         return xvalue;
2157       case BoolTest::lt:          // x < y
2158       case BoolTest::le:          // x <= y
2159         return (want_max ? yvalue : xvalue);
2160       case BoolTest::gt:          // x > y
2161       case BoolTest::ge:          // x >= y
2162         return (want_max ? xvalue : yvalue);
2163       }
2164     }
2165   }
2166 
2167   // We failed to find a dominating test.
2168   // Let's pick a test that might GVN with prior tests.
2169   Node*          best_bol   = NULL;
2170   BoolTest::mask best_btest = BoolTest::illegal;
2171   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2172     Node* cmp = cmps[cmpn];
2173     if (cmp == NULL)  continue;
2174     for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2175       Node* bol = cmp->fast_out(j);
2176       if (!bol->is_Bool())  continue;
2177       BoolTest::mask btest = bol->as_Bool()->_test._test;
2178       if (btest == BoolTest::eq || btest == BoolTest::ne)  continue;
2179       if (cmp->in(1) == ykey)   btest = BoolTest(btest).commute();
2180       if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2181         best_bol   = bol->as_Bool();
2182         best_btest = btest;
2183       }
2184     }
2185   }
2186 
2187   Node* answer_if_true  = NULL;
2188   Node* answer_if_false = NULL;
2189   switch (best_btest) {
2190   default:
2191     if (cmpxy == NULL)
2192       cmpxy = ideal_cmpxy;
2193     best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2194     // and fall through:
2195   case BoolTest::lt:          // x < y
2196   case BoolTest::le:          // x <= y
2197     answer_if_true  = (want_max ? yvalue : xvalue);
2198     answer_if_false = (want_max ? xvalue : yvalue);
2199     break;
2200   case BoolTest::gt:          // x > y
2201   case BoolTest::ge:          // x >= y
2202     answer_if_true  = (want_max ? xvalue : yvalue);
2203     answer_if_false = (want_max ? yvalue : xvalue);
2204     break;
2205   }
2206 
2207   jint hi, lo;
2208   if (want_max) {
2209     // We can sharpen the minimum.
2210     hi = MAX2(txvalue->_hi, tyvalue->_hi);
2211     lo = MAX2(txvalue->_lo, tyvalue->_lo);
2212   } else {
2213     // We can sharpen the maximum.
2214     hi = MIN2(txvalue->_hi, tyvalue->_hi);
2215     lo = MIN2(txvalue->_lo, tyvalue->_lo);
2216   }
2217 
2218   // Use a flow-free graph structure, to avoid creating excess control edges
2219   // which could hinder other optimizations.
2220   // Since Math.min/max is often used with arraycopy, we want
2221   // tightly_coupled_allocation to be able to see beyond min/max expressions.
2222   Node* cmov = CMoveNode::make(NULL, best_bol,
2223                                answer_if_false, answer_if_true,
2224                                TypeInt::make(lo, hi, widen));
2225 
2226   return _gvn.transform(cmov);
2227 
2228   /*
2229   // This is not as desirable as it may seem, since Min and Max
2230   // nodes do not have a full set of optimizations.
2231   // And they would interfere, anyway, with 'if' optimizations
2232   // and with CMoveI canonical forms.
2233   switch (id) {
2234   case vmIntrinsics::_min:
2235     result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2236   case vmIntrinsics::_max:
2237     result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2238   default:
2239     ShouldNotReachHere();
2240   }
2241   */
2242 }
2243 
2244 inline int
2245 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2246   const TypePtr* base_type = TypePtr::NULL_PTR;
2247   if (base != NULL)  base_type = _gvn.type(base)->isa_ptr();
2248   if (base_type == NULL) {
2249     // Unknown type.
2250     return Type::AnyPtr;
2251   } else if (base_type == TypePtr::NULL_PTR) {
2252     // Since this is a NULL+long form, we have to switch to a rawptr.
2253     base   = _gvn.transform(new CastX2PNode(offset));
2254     offset = MakeConX(0);
2255     return Type::RawPtr;
2256   } else if (base_type->base() == Type::RawPtr) {
2257     return Type::RawPtr;
2258   } else if (base_type->isa_oopptr()) {
2259     // Base is never null => always a heap address.
2260     if (base_type->ptr() == TypePtr::NotNull) {
2261       return Type::OopPtr;
2262     }
2263     // Offset is small => always a heap address.
2264     const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2265     if (offset_type != NULL &&
2266         base_type->offset() == 0 &&     // (should always be?)
2267         offset_type->_lo >= 0 &&
2268         !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2269       return Type::OopPtr;
2270     }
2271     // Otherwise, it might either be oop+off or NULL+addr.
2272     return Type::AnyPtr;
2273   } else {
2274     // No information:
2275     return Type::AnyPtr;
2276   }
2277 }
2278 
2279 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2280   int kind = classify_unsafe_addr(base, offset);
2281   if (kind == Type::RawPtr) {
2282     return basic_plus_adr(top(), base, offset);
2283   } else {
2284     return basic_plus_adr(base, offset);
2285   }
2286 }
2287 
2288 //--------------------------inline_number_methods-----------------------------
2289 // inline int     Integer.numberOfLeadingZeros(int)
2290 // inline int        Long.numberOfLeadingZeros(long)
2291 //
2292 // inline int     Integer.numberOfTrailingZeros(int)
2293 // inline int        Long.numberOfTrailingZeros(long)
2294 //
2295 // inline int     Integer.bitCount(int)
2296 // inline int        Long.bitCount(long)
2297 //
2298 // inline char  Character.reverseBytes(char)
2299 // inline short     Short.reverseBytes(short)
2300 // inline int     Integer.reverseBytes(int)
2301 // inline long       Long.reverseBytes(long)
2302 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2303   Node* arg = argument(0);
2304   Node* n;
2305   switch (id) {
2306   case vmIntrinsics::_numberOfLeadingZeros_i:   n = new CountLeadingZerosINode( arg);  break;
2307   case vmIntrinsics::_numberOfLeadingZeros_l:   n = new CountLeadingZerosLNode( arg);  break;
2308   case vmIntrinsics::_numberOfTrailingZeros_i:  n = new CountTrailingZerosINode(arg);  break;
2309   case vmIntrinsics::_numberOfTrailingZeros_l:  n = new CountTrailingZerosLNode(arg);  break;
2310   case vmIntrinsics::_bitCount_i:               n = new PopCountINode(          arg);  break;
2311   case vmIntrinsics::_bitCount_l:               n = new PopCountLNode(          arg);  break;
2312   case vmIntrinsics::_reverseBytes_c:           n = new ReverseBytesUSNode(0,   arg);  break;
2313   case vmIntrinsics::_reverseBytes_s:           n = new ReverseBytesSNode( 0,   arg);  break;
2314   case vmIntrinsics::_reverseBytes_i:           n = new ReverseBytesINode( 0,   arg);  break;
2315   case vmIntrinsics::_reverseBytes_l:           n = new ReverseBytesLNode( 0,   arg);  break;
2316   default:  fatal_unexpected_iid(id);  break;
2317   }
2318   set_result(_gvn.transform(n));
2319   return true;
2320 }
2321 
2322 //----------------------------inline_unsafe_access----------------------------
2323 
2324 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2325 
2326 // Helper that guards and inserts a pre-barrier.
2327 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2328                                         Node* pre_val, bool need_mem_bar) {
2329   // We could be accessing the referent field of a reference object. If so, when G1
2330   // is enabled, we need to log the value in the referent field in an SATB buffer.
2331   // This routine performs some compile time filters and generates suitable
2332   // runtime filters that guard the pre-barrier code.
2333   // Also add memory barrier for non volatile load from the referent field
2334   // to prevent commoning of loads across safepoint.
2335   if (!UseG1GC && !need_mem_bar)
2336     return;
2337 
2338   // Some compile time checks.
2339 
2340   // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2341   const TypeX* otype = offset->find_intptr_t_type();
2342   if (otype != NULL && otype->is_con() &&
2343       otype->get_con() != java_lang_ref_Reference::referent_offset) {
2344     // Constant offset but not the reference_offset so just return
2345     return;
2346   }
2347 
2348   // We only need to generate the runtime guards for instances.
2349   const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2350   if (btype != NULL) {
2351     if (btype->isa_aryptr()) {
2352       // Array type so nothing to do
2353       return;
2354     }
2355 
2356     const TypeInstPtr* itype = btype->isa_instptr();
2357     if (itype != NULL) {
2358       // Can the klass of base_oop be statically determined to be
2359       // _not_ a sub-class of Reference and _not_ Object?
2360       ciKlass* klass = itype->klass();
2361       if ( klass->is_loaded() &&
2362           !klass->is_subtype_of(env()->Reference_klass()) &&
2363           !env()->Object_klass()->is_subtype_of(klass)) {
2364         return;
2365       }
2366     }
2367   }
2368 
2369   // The compile time filters did not reject base_oop/offset so
2370   // we need to generate the following runtime filters
2371   //
2372   // if (offset == java_lang_ref_Reference::_reference_offset) {
2373   //   if (instance_of(base, java.lang.ref.Reference)) {
2374   //     pre_barrier(_, pre_val, ...);
2375   //   }
2376   // }
2377 
2378   float likely   = PROB_LIKELY(  0.999);
2379   float unlikely = PROB_UNLIKELY(0.999);
2380 
2381   IdealKit ideal(this);
2382 #define __ ideal.
2383 
2384   Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2385 
2386   __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2387       // Update graphKit memory and control from IdealKit.
2388       sync_kit(ideal);
2389 
2390       Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2391       Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2392 
2393       // Update IdealKit memory and control from graphKit.
2394       __ sync_kit(this);
2395 
2396       Node* one = __ ConI(1);
2397       // is_instof == 0 if base_oop == NULL
2398       __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2399 
2400         // Update graphKit from IdeakKit.
2401         sync_kit(ideal);
2402 
2403         // Use the pre-barrier to record the value in the referent field
2404         pre_barrier(false /* do_load */,
2405                     __ ctrl(),
2406                     NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2407                     pre_val /* pre_val */,
2408                     T_OBJECT);
2409         if (need_mem_bar) {
2410           // Add memory barrier to prevent commoning reads from this field
2411           // across safepoint since GC can change its value.
2412           insert_mem_bar(Op_MemBarCPUOrder);
2413         }
2414         // Update IdealKit from graphKit.
2415         __ sync_kit(this);
2416 
2417       } __ end_if(); // _ref_type != ref_none
2418   } __ end_if(); // offset == referent_offset
2419 
2420   // Final sync IdealKit and GraphKit.
2421   final_sync(ideal);
2422 #undef __
2423 }
2424 
2425 
2426 // Interpret Unsafe.fieldOffset cookies correctly:
2427 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2428 
2429 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2430   // Attempt to infer a sharper value type from the offset and base type.
2431   ciKlass* sharpened_klass = NULL;
2432 
2433   // See if it is an instance field, with an object type.
2434   if (alias_type->field() != NULL) {
2435     assert(!is_native_ptr, "native pointer op cannot use a java address");
2436     if (alias_type->field()->type()->is_klass()) {
2437       sharpened_klass = alias_type->field()->type()->as_klass();
2438     }
2439   }
2440 
2441   // See if it is a narrow oop array.
2442   if (adr_type->isa_aryptr()) {
2443     if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2444       const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2445       if (elem_type != NULL) {
2446         sharpened_klass = elem_type->klass();
2447       }
2448     }
2449   }
2450 
2451   // The sharpened class might be unloaded if there is no class loader
2452   // contraint in place.
2453   if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2454     const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2455 
2456 #ifndef PRODUCT
2457     if (C->print_intrinsics() || C->print_inlining()) {
2458       tty->print("  from base type: ");  adr_type->dump();
2459       tty->print("  sharpened value: ");  tjp->dump();
2460     }
2461 #endif
2462     // Sharpen the value type.
2463     return tjp;
2464   }
2465   return NULL;
2466 }
2467 
2468 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2469   if (callee()->is_static())  return false;  // caller must have the capability!
2470 
2471 #ifndef PRODUCT
2472   {
2473     ResourceMark rm;
2474     // Check the signatures.
2475     ciSignature* sig = callee()->signature();
2476 #ifdef ASSERT
2477     if (!is_store) {
2478       // Object getObject(Object base, int/long offset), etc.
2479       BasicType rtype = sig->return_type()->basic_type();
2480       if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2481           rtype = T_ADDRESS;  // it is really a C void*
2482       assert(rtype == type, "getter must return the expected value");
2483       if (!is_native_ptr) {
2484         assert(sig->count() == 2, "oop getter has 2 arguments");
2485         assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2486         assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2487       } else {
2488         assert(sig->count() == 1, "native getter has 1 argument");
2489         assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2490       }
2491     } else {
2492       // void putObject(Object base, int/long offset, Object x), etc.
2493       assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2494       if (!is_native_ptr) {
2495         assert(sig->count() == 3, "oop putter has 3 arguments");
2496         assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2497         assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2498       } else {
2499         assert(sig->count() == 2, "native putter has 2 arguments");
2500         assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2501       }
2502       BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2503       if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2504         vtype = T_ADDRESS;  // it is really a C void*
2505       assert(vtype == type, "putter must accept the expected value");
2506     }
2507 #endif // ASSERT
2508  }
2509 #endif //PRODUCT
2510 
2511   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2512 
2513   Node* receiver = argument(0);  // type: oop
2514 
2515   // Build address expression.  See the code in inline_unsafe_prefetch.
2516   Node* adr;
2517   Node* heap_base_oop = top();
2518   Node* offset = top();
2519   Node* val;
2520 
2521   if (!is_native_ptr) {
2522     // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2523     Node* base = argument(1);  // type: oop
2524     // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2525     offset = argument(2);  // type: long
2526     // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2527     // to be plain byte offsets, which are also the same as those accepted
2528     // by oopDesc::field_base.
2529     assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2530            "fieldOffset must be byte-scaled");
2531     // 32-bit machines ignore the high half!
2532     offset = ConvL2X(offset);
2533     adr = make_unsafe_address(base, offset);
2534     heap_base_oop = base;
2535     val = is_store ? argument(4) : NULL;
2536   } else {
2537     Node* ptr = argument(1);  // type: long
2538     ptr = ConvL2X(ptr);  // adjust Java long to machine word
2539     adr = make_unsafe_address(NULL, ptr);
2540     val = is_store ? argument(3) : NULL;
2541   }
2542 
2543   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2544 
2545   // First guess at the value type.
2546   const Type *value_type = Type::get_const_basic_type(type);
2547 
2548   // Try to categorize the address.  If it comes up as TypeJavaPtr::BOTTOM,
2549   // there was not enough information to nail it down.
2550   Compile::AliasType* alias_type = C->alias_type(adr_type);
2551   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2552 
2553   // We will need memory barriers unless we can determine a unique
2554   // alias category for this reference.  (Note:  If for some reason
2555   // the barriers get omitted and the unsafe reference begins to "pollute"
2556   // the alias analysis of the rest of the graph, either Compile::can_alias
2557   // or Compile::must_alias will throw a diagnostic assert.)
2558   bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2559 
2560   // If we are reading the value of the referent field of a Reference
2561   // object (either by using Unsafe directly or through reflection)
2562   // then, if G1 is enabled, we need to record the referent in an
2563   // SATB log buffer using the pre-barrier mechanism.
2564   // Also we need to add memory barrier to prevent commoning reads
2565   // from this field across safepoint since GC can change its value.
2566   bool need_read_barrier = !is_native_ptr && !is_store &&
2567                            offset != top() && heap_base_oop != top();
2568 
2569   if (!is_store && type == T_OBJECT) {
2570     const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2571     if (tjp != NULL) {
2572       value_type = tjp;
2573     }
2574   }
2575 
2576   receiver = null_check(receiver);
2577   if (stopped()) {
2578     return true;
2579   }
2580   // Heap pointers get a null-check from the interpreter,
2581   // as a courtesy.  However, this is not guaranteed by Unsafe,
2582   // and it is not possible to fully distinguish unintended nulls
2583   // from intended ones in this API.
2584 
2585   if (is_volatile) {
2586     // We need to emit leading and trailing CPU membars (see below) in
2587     // addition to memory membars when is_volatile. This is a little
2588     // too strong, but avoids the need to insert per-alias-type
2589     // volatile membars (for stores; compare Parse::do_put_xxx), which
2590     // we cannot do effectively here because we probably only have a
2591     // rough approximation of type.
2592     need_mem_bar = true;
2593     // For Stores, place a memory ordering barrier now.
2594     if (is_store) {
2595       insert_mem_bar(Op_MemBarRelease);
2596     } else {
2597       if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2598         insert_mem_bar(Op_MemBarVolatile);
2599       }
2600     }
2601   }
2602 
2603   // Memory barrier to prevent normal and 'unsafe' accesses from
2604   // bypassing each other.  Happens after null checks, so the
2605   // exception paths do not take memory state from the memory barrier,
2606   // so there's no problems making a strong assert about mixing users
2607   // of safe & unsafe memory.  Otherwise fails in a CTW of rt.jar
2608   // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2609   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2610 
2611   if (!is_store) {
2612     MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2613     Node* p = make_load(control(), adr, value_type, type, adr_type, mo, is_volatile);
2614     // load value
2615     switch (type) {
2616     case T_BOOLEAN:
2617     case T_CHAR:
2618     case T_BYTE:
2619     case T_SHORT:
2620     case T_INT:
2621     case T_LONG:
2622     case T_FLOAT:
2623     case T_DOUBLE:
2624       break;
2625     case T_OBJECT:
2626       if (need_read_barrier) {
2627         insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2628       }
2629       break;
2630     case T_ADDRESS:
2631       // Cast to an int type.
2632       p = _gvn.transform(new CastP2XNode(NULL, p));
2633       p = ConvX2UL(p);
2634       break;
2635     default:
2636       fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2637       break;
2638     }
2639     // The load node has the control of the preceding MemBarCPUOrder.  All
2640     // following nodes will have the control of the MemBarCPUOrder inserted at
2641     // the end of this method.  So, pushing the load onto the stack at a later
2642     // point is fine.
2643     set_result(p);
2644   } else {
2645     // place effect of store into memory
2646     switch (type) {
2647     case T_DOUBLE:
2648       val = dstore_rounding(val);
2649       break;
2650     case T_ADDRESS:
2651       // Repackage the long as a pointer.
2652       val = ConvL2X(val);
2653       val = _gvn.transform(new CastX2PNode(val));
2654       break;
2655     }
2656 
2657     MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2658     if (type != T_OBJECT ) {
2659       (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2660     } else {
2661       // Possibly an oop being stored to Java heap or native memory
2662       if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2663         // oop to Java heap.
2664         (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2665       } else {
2666         // We can't tell at compile time if we are storing in the Java heap or outside
2667         // of it. So we need to emit code to conditionally do the proper type of
2668         // store.
2669 
2670         IdealKit ideal(this);
2671 #define __ ideal.
2672         // QQQ who knows what probability is here??
2673         __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2674           // Sync IdealKit and graphKit.
2675           sync_kit(ideal);
2676           Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2677           // Update IdealKit memory.
2678           __ sync_kit(this);
2679         } __ else_(); {
2680           __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2681         } __ end_if();
2682         // Final sync IdealKit and GraphKit.
2683         final_sync(ideal);
2684 #undef __
2685       }
2686     }
2687   }
2688 
2689   if (is_volatile) {
2690     if (!is_store) {
2691       insert_mem_bar(Op_MemBarAcquire);
2692     } else {
2693       if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2694         insert_mem_bar(Op_MemBarVolatile);
2695       }
2696     }
2697   }
2698 
2699   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2700 
2701   return true;
2702 }
2703 
2704 //----------------------------inline_unsafe_prefetch----------------------------
2705 
2706 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2707 #ifndef PRODUCT
2708   {
2709     ResourceMark rm;
2710     // Check the signatures.
2711     ciSignature* sig = callee()->signature();
2712 #ifdef ASSERT
2713     // Object getObject(Object base, int/long offset), etc.
2714     BasicType rtype = sig->return_type()->basic_type();
2715     if (!is_native_ptr) {
2716       assert(sig->count() == 2, "oop prefetch has 2 arguments");
2717       assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2718       assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2719     } else {
2720       assert(sig->count() == 1, "native prefetch has 1 argument");
2721       assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2722     }
2723 #endif // ASSERT
2724   }
2725 #endif // !PRODUCT
2726 
2727   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2728 
2729   const int idx = is_static ? 0 : 1;
2730   if (!is_static) {
2731     null_check_receiver();
2732     if (stopped()) {
2733       return true;
2734     }
2735   }
2736 
2737   // Build address expression.  See the code in inline_unsafe_access.
2738   Node *adr;
2739   if (!is_native_ptr) {
2740     // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2741     Node* base   = argument(idx + 0);  // type: oop
2742     // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2743     Node* offset = argument(idx + 1);  // type: long
2744     // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2745     // to be plain byte offsets, which are also the same as those accepted
2746     // by oopDesc::field_base.
2747     assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2748            "fieldOffset must be byte-scaled");
2749     // 32-bit machines ignore the high half!
2750     offset = ConvL2X(offset);
2751     adr = make_unsafe_address(base, offset);
2752   } else {
2753     Node* ptr = argument(idx + 0);  // type: long
2754     ptr = ConvL2X(ptr);  // adjust Java long to machine word
2755     adr = make_unsafe_address(NULL, ptr);
2756   }
2757 
2758   // Generate the read or write prefetch
2759   Node *prefetch;
2760   if (is_store) {
2761     prefetch = new PrefetchWriteNode(i_o(), adr);
2762   } else {
2763     prefetch = new PrefetchReadNode(i_o(), adr);
2764   }
2765   prefetch->init_req(0, control());
2766   set_i_o(_gvn.transform(prefetch));
2767 
2768   return true;
2769 }
2770 
2771 //----------------------------inline_unsafe_load_store----------------------------
2772 // This method serves a couple of different customers (depending on LoadStoreKind):
2773 //
2774 // LS_cmpxchg:
2775 //   public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2776 //   public final native boolean compareAndSwapInt(   Object o, long offset, int    expected, int    x);
2777 //   public final native boolean compareAndSwapLong(  Object o, long offset, long   expected, long   x);
2778 //
2779 // LS_xadd:
2780 //   public int  getAndAddInt( Object o, long offset, int  delta)
2781 //   public long getAndAddLong(Object o, long offset, long delta)
2782 //
2783 // LS_xchg:
2784 //   int    getAndSet(Object o, long offset, int    newValue)
2785 //   long   getAndSet(Object o, long offset, long   newValue)
2786 //   Object getAndSet(Object o, long offset, Object newValue)
2787 //
2788 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2789   // This basic scheme here is the same as inline_unsafe_access, but
2790   // differs in enough details that combining them would make the code
2791   // overly confusing.  (This is a true fact! I originally combined
2792   // them, but even I was confused by it!) As much code/comments as
2793   // possible are retained from inline_unsafe_access though to make
2794   // the correspondences clearer. - dl
2795 
2796   if (callee()->is_static())  return false;  // caller must have the capability!
2797 
2798 #ifndef PRODUCT
2799   BasicType rtype;
2800   {
2801     ResourceMark rm;
2802     // Check the signatures.
2803     ciSignature* sig = callee()->signature();
2804     rtype = sig->return_type()->basic_type();
2805     if (kind == LS_xadd || kind == LS_xchg) {
2806       // Check the signatures.
2807 #ifdef ASSERT
2808       assert(rtype == type, "get and set must return the expected type");
2809       assert(sig->count() == 3, "get and set has 3 arguments");
2810       assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2811       assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2812       assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2813 #endif // ASSERT
2814     } else if (kind == LS_cmpxchg) {
2815       // Check the signatures.
2816 #ifdef ASSERT
2817       assert(rtype == T_BOOLEAN, "CAS must return boolean");
2818       assert(sig->count() == 4, "CAS has 4 arguments");
2819       assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2820       assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2821 #endif // ASSERT
2822     } else {
2823       ShouldNotReachHere();
2824     }
2825   }
2826 #endif //PRODUCT
2827 
2828   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2829 
2830   // Get arguments:
2831   Node* receiver = NULL;
2832   Node* base     = NULL;
2833   Node* offset   = NULL;
2834   Node* oldval   = NULL;
2835   Node* newval   = NULL;
2836   if (kind == LS_cmpxchg) {
2837     const bool two_slot_type = type2size[type] == 2;
2838     receiver = argument(0);  // type: oop
2839     base     = argument(1);  // type: oop
2840     offset   = argument(2);  // type: long
2841     oldval   = argument(4);  // type: oop, int, or long
2842     newval   = argument(two_slot_type ? 6 : 5);  // type: oop, int, or long
2843   } else if (kind == LS_xadd || kind == LS_xchg){
2844     receiver = argument(0);  // type: oop
2845     base     = argument(1);  // type: oop
2846     offset   = argument(2);  // type: long
2847     oldval   = NULL;
2848     newval   = argument(4);  // type: oop, int, or long
2849   }
2850 
2851   // Null check receiver.
2852   receiver = null_check(receiver);
2853   if (stopped()) {
2854     return true;
2855   }
2856 
2857   // Build field offset expression.
2858   // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2859   // to be plain byte offsets, which are also the same as those accepted
2860   // by oopDesc::field_base.
2861   assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2862   // 32-bit machines ignore the high half of long offsets
2863   offset = ConvL2X(offset);
2864   Node* adr = make_unsafe_address(base, offset);
2865   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2866 
2867   // For CAS, unlike inline_unsafe_access, there seems no point in
2868   // trying to refine types. Just use the coarse types here.
2869   const Type *value_type = Type::get_const_basic_type(type);
2870   Compile::AliasType* alias_type = C->alias_type(adr_type);
2871   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2872 
2873   if (kind == LS_xchg && type == T_OBJECT) {
2874     const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2875     if (tjp != NULL) {
2876       value_type = tjp;
2877     }
2878   }
2879 
2880   int alias_idx = C->get_alias_index(adr_type);
2881 
2882   // Memory-model-wise, a LoadStore acts like a little synchronized
2883   // block, so needs barriers on each side.  These don't translate
2884   // into actual barriers on most machines, but we still need rest of
2885   // compiler to respect ordering.
2886 
2887   insert_mem_bar(Op_MemBarRelease);
2888   insert_mem_bar(Op_MemBarCPUOrder);
2889 
2890   // 4984716: MemBars must be inserted before this
2891   //          memory node in order to avoid a false
2892   //          dependency which will confuse the scheduler.
2893   Node *mem = memory(alias_idx);
2894 
2895   // For now, we handle only those cases that actually exist: ints,
2896   // longs, and Object. Adding others should be straightforward.
2897   Node* load_store;
2898   switch(type) {
2899   case T_INT:
2900     if (kind == LS_xadd) {
2901       load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2902     } else if (kind == LS_xchg) {
2903       load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2904     } else if (kind == LS_cmpxchg) {
2905       load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2906     } else {
2907       ShouldNotReachHere();
2908     }
2909     break;
2910   case T_LONG:
2911     if (kind == LS_xadd) {
2912       load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2913     } else if (kind == LS_xchg) {
2914       load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2915     } else if (kind == LS_cmpxchg) {
2916       load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2917     } else {
2918       ShouldNotReachHere();
2919     }
2920     break;
2921   case T_OBJECT:
2922     // Transformation of a value which could be NULL pointer (CastPP #NULL)
2923     // could be delayed during Parse (for example, in adjust_map_after_if()).
2924     // Execute transformation here to avoid barrier generation in such case.
2925     if (_gvn.type(newval) == TypePtr::NULL_PTR)
2926       newval = _gvn.makecon(TypePtr::NULL_PTR);
2927 
2928     // Reference stores need a store barrier.
2929     if (kind == LS_xchg) {
2930       // If pre-barrier must execute before the oop store, old value will require do_load here.
2931       if (!can_move_pre_barrier()) {
2932         pre_barrier(true /* do_load*/,
2933                     control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2934                     NULL /* pre_val*/,
2935                     T_OBJECT);
2936       } // Else move pre_barrier to use load_store value, see below.
2937     } else if (kind == LS_cmpxchg) {
2938       // Same as for newval above:
2939       if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2940         oldval = _gvn.makecon(TypePtr::NULL_PTR);
2941       }
2942       // The only known value which might get overwritten is oldval.
2943       pre_barrier(false /* do_load */,
2944                   control(), NULL, NULL, max_juint, NULL, NULL,
2945                   oldval /* pre_val */,
2946                   T_OBJECT);
2947     } else {
2948       ShouldNotReachHere();
2949     }
2950 
2951 #ifdef _LP64
2952     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2953       Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2954       if (kind == LS_xchg) {
2955         load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2956                                                        newval_enc, adr_type, value_type->make_narrowoop()));
2957       } else {
2958         assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2959         Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2960         load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2961                                                                 newval_enc, oldval_enc));
2962       }
2963     } else
2964 #endif
2965     {
2966       if (kind == LS_xchg) {
2967         load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2968       } else {
2969         assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2970         load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2971       }
2972     }
2973     post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2974     break;
2975   default:
2976     fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2977     break;
2978   }
2979 
2980   // SCMemProjNodes represent the memory state of a LoadStore. Their
2981   // main role is to prevent LoadStore nodes from being optimized away
2982   // when their results aren't used.
2983   Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2984   set_memory(proj, alias_idx);
2985 
2986   if (type == T_OBJECT && kind == LS_xchg) {
2987 #ifdef _LP64
2988     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2989       load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2990     }
2991 #endif
2992     if (can_move_pre_barrier()) {
2993       // Don't need to load pre_val. The old value is returned by load_store.
2994       // The pre_barrier can execute after the xchg as long as no safepoint
2995       // gets inserted between them.
2996       pre_barrier(false /* do_load */,
2997                   control(), NULL, NULL, max_juint, NULL, NULL,
2998                   load_store /* pre_val */,
2999                   T_OBJECT);
3000     }
3001   }
3002 
3003   // Add the trailing membar surrounding the access
3004   insert_mem_bar(Op_MemBarCPUOrder);
3005   insert_mem_bar(Op_MemBarAcquire);
3006 
3007   assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3008   set_result(load_store);
3009   return true;
3010 }
3011 
3012 //----------------------------inline_unsafe_ordered_store----------------------
3013 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3014 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3015 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3016 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3017   // This is another variant of inline_unsafe_access, differing in
3018   // that it always issues store-store ("release") barrier and ensures
3019   // store-atomicity (which only matters for "long").
3020 
3021   if (callee()->is_static())  return false;  // caller must have the capability!
3022 
3023 #ifndef PRODUCT
3024   {
3025     ResourceMark rm;
3026     // Check the signatures.
3027     ciSignature* sig = callee()->signature();
3028 #ifdef ASSERT
3029     BasicType rtype = sig->return_type()->basic_type();
3030     assert(rtype == T_VOID, "must return void");
3031     assert(sig->count() == 3, "has 3 arguments");
3032     assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3033     assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3034 #endif // ASSERT
3035   }
3036 #endif //PRODUCT
3037 
3038   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
3039 
3040   // Get arguments:
3041   Node* receiver = argument(0);  // type: oop
3042   Node* base     = argument(1);  // type: oop
3043   Node* offset   = argument(2);  // type: long
3044   Node* val      = argument(4);  // type: oop, int, or long
3045 
3046   // Null check receiver.
3047   receiver = null_check(receiver);
3048   if (stopped()) {
3049     return true;
3050   }
3051 
3052   // Build field offset expression.
3053   assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3054   // 32-bit machines ignore the high half of long offsets
3055   offset = ConvL2X(offset);
3056   Node* adr = make_unsafe_address(base, offset);
3057   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3058   const Type *value_type = Type::get_const_basic_type(type);
3059   Compile::AliasType* alias_type = C->alias_type(adr_type);
3060 
3061   insert_mem_bar(Op_MemBarRelease);
3062   insert_mem_bar(Op_MemBarCPUOrder);
3063   // Ensure that the store is atomic for longs:
3064   const bool require_atomic_access = true;
3065   Node* store;
3066   if (type == T_OBJECT) // reference stores need a store barrier.
3067     store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3068   else {
3069     store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3070   }
3071   insert_mem_bar(Op_MemBarCPUOrder);
3072   return true;
3073 }
3074 
3075 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3076   // Regardless of form, don't allow previous ld/st to move down,
3077   // then issue acquire, release, or volatile mem_bar.
3078   insert_mem_bar(Op_MemBarCPUOrder);
3079   switch(id) {
3080     case vmIntrinsics::_loadFence:
3081       insert_mem_bar(Op_LoadFence);
3082       return true;
3083     case vmIntrinsics::_storeFence:
3084       insert_mem_bar(Op_StoreFence);
3085       return true;
3086     case vmIntrinsics::_fullFence:
3087       insert_mem_bar(Op_MemBarVolatile);
3088       return true;
3089     default:
3090       fatal_unexpected_iid(id);
3091       return false;
3092   }
3093 }
3094 
3095 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3096   if (!kls->is_Con()) {
3097     return true;
3098   }
3099   const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3100   if (klsptr == NULL) {
3101     return true;
3102   }
3103   ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3104   // don't need a guard for a klass that is already initialized
3105   return !ik->is_initialized();
3106 }
3107 
3108 //----------------------------inline_unsafe_allocate---------------------------
3109 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3110 bool LibraryCallKit::inline_unsafe_allocate() {
3111   if (callee()->is_static())  return false;  // caller must have the capability!
3112 
3113   null_check_receiver();  // null-check, then ignore
3114   Node* cls = null_check(argument(1));
3115   if (stopped())  return true;
3116 
3117   Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3118   kls = null_check(kls);
3119   if (stopped())  return true;  // argument was like int.class
3120 
3121   Node* test = NULL;
3122   if (LibraryCallKit::klass_needs_init_guard(kls)) {
3123     // Note:  The argument might still be an illegal value like
3124     // Serializable.class or Object[].class.   The runtime will handle it.
3125     // But we must make an explicit check for initialization.
3126     Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3127     // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3128     // can generate code to load it as unsigned byte.
3129     Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3130     Node* bits = intcon(InstanceKlass::fully_initialized);
3131     test = _gvn.transform(new SubINode(inst, bits));
3132     // The 'test' is non-zero if we need to take a slow path.
3133   }
3134 
3135   Node* obj = new_instance(kls, test);
3136   set_result(obj);
3137   return true;
3138 }
3139 
3140 #ifdef TRACE_HAVE_INTRINSICS
3141 /*
3142  * oop -> myklass
3143  * myklass->trace_id |= USED
3144  * return myklass->trace_id & ~0x3
3145  */
3146 bool LibraryCallKit::inline_native_classID() {
3147   null_check_receiver();  // null-check, then ignore
3148   Node* cls = null_check(argument(1), T_OBJECT);
3149   Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3150   kls = null_check(kls, T_OBJECT);
3151   ByteSize offset = TRACE_ID_OFFSET;
3152   Node* insp = basic_plus_adr(kls, in_bytes(offset));
3153   Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3154   Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3155   Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3156   Node* clsused = longcon(0x01l); // set the class bit
3157   Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3158 
3159   const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3160   store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3161   set_result(andl);
3162   return true;
3163 }
3164 
3165 bool LibraryCallKit::inline_native_threadID() {
3166   Node* tls_ptr = NULL;
3167   Node* cur_thr = generate_current_thread(tls_ptr);
3168   Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3169   Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3170   p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3171 
3172   Node* threadid = NULL;
3173   size_t thread_id_size = OSThread::thread_id_size();
3174   if (thread_id_size == (size_t) BytesPerLong) {
3175     threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3176   } else if (thread_id_size == (size_t) BytesPerInt) {
3177     threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3178   } else {
3179     ShouldNotReachHere();
3180   }
3181   set_result(threadid);
3182   return true;
3183 }
3184 #endif
3185 
3186 //------------------------inline_native_time_funcs--------------
3187 // inline code for System.currentTimeMillis() and System.nanoTime()
3188 // these have the same type and signature
3189 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3190   const TypeFunc* tf = OptoRuntime::void_long_Type();
3191   const TypePtr* no_memory_effects = NULL;
3192   Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3193   Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3194 #ifdef ASSERT
3195   Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3196   assert(value_top == top(), "second value must be top");
3197 #endif
3198   set_result(value);





3199   return true;
3200 }
3201 
3202 //------------------------inline_native_currentThread------------------
3203 bool LibraryCallKit::inline_native_currentThread() {
3204   Node* junk = NULL;
3205   set_result(generate_current_thread(junk));
3206   return true;
3207 }
3208 
3209 //------------------------inline_native_isInterrupted------------------
3210 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3211 bool LibraryCallKit::inline_native_isInterrupted() {
3212   // Add a fast path to t.isInterrupted(clear_int):
3213   //   (t == Thread.current() &&
3214   //    (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3215   //   ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3216   // So, in the common case that the interrupt bit is false,
3217   // we avoid making a call into the VM.  Even if the interrupt bit
3218   // is true, if the clear_int argument is false, we avoid the VM call.
3219   // However, if the receiver is not currentThread, we must call the VM,
3220   // because there must be some locking done around the operation.
3221 
3222   // We only go to the fast case code if we pass two guards.
3223   // Paths which do not pass are accumulated in the slow_region.
3224 
3225   enum {
3226     no_int_result_path   = 1, // t == Thread.current() && !TLS._osthread._interrupted
3227     no_clear_result_path = 2, // t == Thread.current() &&  TLS._osthread._interrupted && !clear_int
3228     slow_result_path     = 3, // slow path: t.isInterrupted(clear_int)
3229     PATH_LIMIT
3230   };
3231 
3232   // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3233   // out of the function.
3234   insert_mem_bar(Op_MemBarCPUOrder);
3235 
3236   RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3237   PhiNode*    result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3238 
3239   RegionNode* slow_region = new RegionNode(1);
3240   record_for_igvn(slow_region);
3241 
3242   // (a) Receiving thread must be the current thread.
3243   Node* rec_thr = argument(0);
3244   Node* tls_ptr = NULL;
3245   Node* cur_thr = generate_current_thread(tls_ptr);
3246   Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3247   Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3248 
3249   generate_slow_guard(bol_thr, slow_region);
3250 
3251   // (b) Interrupt bit on TLS must be false.
3252   Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3253   Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3254   p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3255 
3256   // Set the control input on the field _interrupted read to prevent it floating up.
3257   Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3258   Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3259   Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3260 
3261   IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3262 
3263   // First fast path:  if (!TLS._interrupted) return false;
3264   Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3265   result_rgn->init_req(no_int_result_path, false_bit);
3266   result_val->init_req(no_int_result_path, intcon(0));
3267 
3268   // drop through to next case
3269   set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3270 
3271 #ifndef TARGET_OS_FAMILY_windows
3272   // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3273   Node* clr_arg = argument(1);
3274   Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3275   Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3276   IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3277 
3278   // Second fast path:  ... else if (!clear_int) return true;
3279   Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3280   result_rgn->init_req(no_clear_result_path, false_arg);
3281   result_val->init_req(no_clear_result_path, intcon(1));
3282 
3283   // drop through to next case
3284   set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3285 #else
3286   // To return true on Windows you must read the _interrupted field
3287   // and check the the event state i.e. take the slow path.
3288 #endif // TARGET_OS_FAMILY_windows
3289 
3290   // (d) Otherwise, go to the slow path.
3291   slow_region->add_req(control());
3292   set_control( _gvn.transform(slow_region));
3293 
3294   if (stopped()) {
3295     // There is no slow path.
3296     result_rgn->init_req(slow_result_path, top());
3297     result_val->init_req(slow_result_path, top());
3298   } else {
3299     // non-virtual because it is a private non-static
3300     CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3301 
3302     Node* slow_val = set_results_for_java_call(slow_call);
3303     // this->control() comes from set_results_for_java_call
3304 
3305     Node* fast_io  = slow_call->in(TypeFunc::I_O);
3306     Node* fast_mem = slow_call->in(TypeFunc::Memory);
3307 
3308     // These two phis are pre-filled with copies of of the fast IO and Memory
3309     PhiNode* result_mem  = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3310     PhiNode* result_io   = PhiNode::make(result_rgn, fast_io,  Type::ABIO);
3311 
3312     result_rgn->init_req(slow_result_path, control());
3313     result_io ->init_req(slow_result_path, i_o());
3314     result_mem->init_req(slow_result_path, reset_memory());
3315     result_val->init_req(slow_result_path, slow_val);
3316 
3317     set_all_memory(_gvn.transform(result_mem));
3318     set_i_o(       _gvn.transform(result_io));
3319   }
3320 
3321   C->set_has_split_ifs(true); // Has chance for split-if optimization
3322   set_result(result_rgn, result_val);
3323   return true;
3324 }
3325 
3326 //---------------------------load_mirror_from_klass----------------------------
3327 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3328 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3329   Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3330   return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3331 }
3332 
3333 //-----------------------load_klass_from_mirror_common-------------------------
3334 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3335 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3336 // and branch to the given path on the region.
3337 // If never_see_null, take an uncommon trap on null, so we can optimistically
3338 // compile for the non-null case.
3339 // If the region is NULL, force never_see_null = true.
3340 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3341                                                     bool never_see_null,
3342                                                     RegionNode* region,
3343                                                     int null_path,
3344                                                     int offset) {
3345   if (region == NULL)  never_see_null = true;
3346   Node* p = basic_plus_adr(mirror, offset);
3347   const TypeKlassPtr*  kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3348   Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3349   Node* null_ctl = top();
3350   kls = null_check_oop(kls, &null_ctl, never_see_null);
3351   if (region != NULL) {
3352     // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3353     region->init_req(null_path, null_ctl);
3354   } else {
3355     assert(null_ctl == top(), "no loose ends");
3356   }
3357   return kls;
3358 }
3359 
3360 //--------------------(inline_native_Class_query helpers)---------------------
3361 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3362 // Fall through if (mods & mask) == bits, take the guard otherwise.
3363 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3364   // Branch around if the given klass has the given modifier bit set.
3365   // Like generate_guard, adds a new path onto the region.
3366   Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3367   Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3368   Node* mask = intcon(modifier_mask);
3369   Node* bits = intcon(modifier_bits);
3370   Node* mbit = _gvn.transform(new AndINode(mods, mask));
3371   Node* cmp  = _gvn.transform(new CmpINode(mbit, bits));
3372   Node* bol  = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3373   return generate_fair_guard(bol, region);
3374 }
3375 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3376   return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3377 }
3378 
3379 //-------------------------inline_native_Class_query-------------------
3380 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3381   const Type* return_type = TypeInt::BOOL;
3382   Node* prim_return_value = top();  // what happens if it's a primitive class?
3383   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3384   bool expect_prim = false;     // most of these guys expect to work on refs
3385 
3386   enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3387 
3388   Node* mirror = argument(0);
3389   Node* obj    = top();
3390 
3391   switch (id) {
3392   case vmIntrinsics::_isInstance:
3393     // nothing is an instance of a primitive type
3394     prim_return_value = intcon(0);
3395     obj = argument(1);
3396     break;
3397   case vmIntrinsics::_getModifiers:
3398     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3399     assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3400     return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3401     break;
3402   case vmIntrinsics::_isInterface:
3403     prim_return_value = intcon(0);
3404     break;
3405   case vmIntrinsics::_isArray:
3406     prim_return_value = intcon(0);
3407     expect_prim = true;  // cf. ObjectStreamClass.getClassSignature
3408     break;
3409   case vmIntrinsics::_isPrimitive:
3410     prim_return_value = intcon(1);
3411     expect_prim = true;  // obviously
3412     break;
3413   case vmIntrinsics::_getSuperclass:
3414     prim_return_value = null();
3415     return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3416     break;
3417   case vmIntrinsics::_getClassAccessFlags:
3418     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3419     return_type = TypeInt::INT;  // not bool!  6297094
3420     break;
3421   default:
3422     fatal_unexpected_iid(id);
3423     break;
3424   }
3425 
3426   const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3427   if (mirror_con == NULL)  return false;  // cannot happen?
3428 
3429 #ifndef PRODUCT
3430   if (C->print_intrinsics() || C->print_inlining()) {
3431     ciType* k = mirror_con->java_mirror_type();
3432     if (k) {
3433       tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3434       k->print_name();
3435       tty->cr();
3436     }
3437   }
3438 #endif
3439 
3440   // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3441   RegionNode* region = new RegionNode(PATH_LIMIT);
3442   record_for_igvn(region);
3443   PhiNode* phi = new PhiNode(region, return_type);
3444 
3445   // The mirror will never be null of Reflection.getClassAccessFlags, however
3446   // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3447   // if it is. See bug 4774291.
3448 
3449   // For Reflection.getClassAccessFlags(), the null check occurs in
3450   // the wrong place; see inline_unsafe_access(), above, for a similar
3451   // situation.
3452   mirror = null_check(mirror);
3453   // If mirror or obj is dead, only null-path is taken.
3454   if (stopped())  return true;
3455 
3456   if (expect_prim)  never_see_null = false;  // expect nulls (meaning prims)
3457 
3458   // Now load the mirror's klass metaobject, and null-check it.
3459   // Side-effects region with the control path if the klass is null.
3460   Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3461   // If kls is null, we have a primitive mirror.
3462   phi->init_req(_prim_path, prim_return_value);
3463   if (stopped()) { set_result(region, phi); return true; }
3464   bool safe_for_replace = (region->in(_prim_path) == top());
3465 
3466   Node* p;  // handy temp
3467   Node* null_ctl;
3468 
3469   // Now that we have the non-null klass, we can perform the real query.
3470   // For constant classes, the query will constant-fold in LoadNode::Value.
3471   Node* query_value = top();
3472   switch (id) {
3473   case vmIntrinsics::_isInstance:
3474     // nothing is an instance of a primitive type
3475     query_value = gen_instanceof(obj, kls, safe_for_replace);
3476     break;
3477 
3478   case vmIntrinsics::_getModifiers:
3479     p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3480     query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3481     break;
3482 
3483   case vmIntrinsics::_isInterface:
3484     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3485     if (generate_interface_guard(kls, region) != NULL)
3486       // A guard was added.  If the guard is taken, it was an interface.
3487       phi->add_req(intcon(1));
3488     // If we fall through, it's a plain class.
3489     query_value = intcon(0);
3490     break;
3491 
3492   case vmIntrinsics::_isArray:
3493     // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3494     if (generate_array_guard(kls, region) != NULL)
3495       // A guard was added.  If the guard is taken, it was an array.
3496       phi->add_req(intcon(1));
3497     // If we fall through, it's a plain class.
3498     query_value = intcon(0);
3499     break;
3500 
3501   case vmIntrinsics::_isPrimitive:
3502     query_value = intcon(0); // "normal" path produces false
3503     break;
3504 
3505   case vmIntrinsics::_getSuperclass:
3506     // The rules here are somewhat unfortunate, but we can still do better
3507     // with random logic than with a JNI call.
3508     // Interfaces store null or Object as _super, but must report null.
3509     // Arrays store an intermediate super as _super, but must report Object.
3510     // Other types can report the actual _super.
3511     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3512     if (generate_interface_guard(kls, region) != NULL)
3513       // A guard was added.  If the guard is taken, it was an interface.
3514       phi->add_req(null());
3515     if (generate_array_guard(kls, region) != NULL)
3516       // A guard was added.  If the guard is taken, it was an array.
3517       phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3518     // If we fall through, it's a plain class.  Get its _super.
3519     p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3520     kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3521     null_ctl = top();
3522     kls = null_check_oop(kls, &null_ctl);
3523     if (null_ctl != top()) {
3524       // If the guard is taken, Object.superClass is null (both klass and mirror).
3525       region->add_req(null_ctl);
3526       phi   ->add_req(null());
3527     }
3528     if (!stopped()) {
3529       query_value = load_mirror_from_klass(kls);
3530     }
3531     break;
3532 
3533   case vmIntrinsics::_getClassAccessFlags:
3534     p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3535     query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3536     break;
3537 
3538   default:
3539     fatal_unexpected_iid(id);
3540     break;
3541   }
3542 
3543   // Fall-through is the normal case of a query to a real class.
3544   phi->init_req(1, query_value);
3545   region->init_req(1, control());
3546 
3547   C->set_has_split_ifs(true); // Has chance for split-if optimization
3548   set_result(region, phi);
3549   return true;
3550 }
3551 
3552 //-------------------------inline_Class_cast-------------------
3553 bool LibraryCallKit::inline_Class_cast() {
3554   Node* mirror = argument(0); // Class
3555   Node* obj    = argument(1);
3556   const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3557   if (mirror_con == NULL) {
3558     return false;  // dead path (mirror->is_top()).
3559   }
3560   if (obj == NULL || obj->is_top()) {
3561     return false;  // dead path
3562   }
3563   const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3564 
3565   // First, see if Class.cast() can be folded statically.
3566   // java_mirror_type() returns non-null for compile-time Class constants.
3567   ciType* tm = mirror_con->java_mirror_type();
3568   if (tm != NULL && tm->is_klass() &&
3569       tp != NULL && tp->klass() != NULL) {
3570     if (!tp->klass()->is_loaded()) {
3571       // Don't use intrinsic when class is not loaded.
3572       return false;
3573     } else {
3574       int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3575       if (static_res == Compile::SSC_always_true) {
3576         // isInstance() is true - fold the code.
3577         set_result(obj);
3578         return true;
3579       } else if (static_res == Compile::SSC_always_false) {
3580         // Don't use intrinsic, have to throw ClassCastException.
3581         // If the reference is null, the non-intrinsic bytecode will
3582         // be optimized appropriately.
3583         return false;
3584       }
3585     }
3586   }
3587 
3588   // Bailout intrinsic and do normal inlining if exception path is frequent.
3589   if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3590     return false;
3591   }
3592 
3593   // Generate dynamic checks.
3594   // Class.cast() is java implementation of _checkcast bytecode.
3595   // Do checkcast (Parse::do_checkcast()) optimizations here.
3596 
3597   mirror = null_check(mirror);
3598   // If mirror is dead, only null-path is taken.
3599   if (stopped()) {
3600     return true;
3601   }
3602 
3603   // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3604   enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3605   RegionNode* region = new RegionNode(PATH_LIMIT);
3606   record_for_igvn(region);
3607 
3608   // Now load the mirror's klass metaobject, and null-check it.
3609   // If kls is null, we have a primitive mirror and
3610   // nothing is an instance of a primitive type.
3611   Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3612 
3613   Node* res = top();
3614   if (!stopped()) {
3615     Node* bad_type_ctrl = top();
3616     // Do checkcast optimizations.
3617     res = gen_checkcast(obj, kls, &bad_type_ctrl);
3618     region->init_req(_bad_type_path, bad_type_ctrl);
3619   }
3620   if (region->in(_prim_path) != top() ||
3621       region->in(_bad_type_path) != top()) {
3622     // Let Interpreter throw ClassCastException.
3623     PreserveJVMState pjvms(this);
3624     set_control(_gvn.transform(region));
3625     uncommon_trap(Deoptimization::Reason_intrinsic,
3626                   Deoptimization::Action_maybe_recompile);
3627   }
3628   if (!stopped()) {
3629     set_result(res);
3630   }
3631   return true;
3632 }
3633 
3634 
3635 //--------------------------inline_native_subtype_check------------------------
3636 // This intrinsic takes the JNI calls out of the heart of
3637 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3638 bool LibraryCallKit::inline_native_subtype_check() {
3639   // Pull both arguments off the stack.
3640   Node* args[2];                // two java.lang.Class mirrors: superc, subc
3641   args[0] = argument(0);
3642   args[1] = argument(1);
3643   Node* klasses[2];             // corresponding Klasses: superk, subk
3644   klasses[0] = klasses[1] = top();
3645 
3646   enum {
3647     // A full decision tree on {superc is prim, subc is prim}:
3648     _prim_0_path = 1,           // {P,N} => false
3649                                 // {P,P} & superc!=subc => false
3650     _prim_same_path,            // {P,P} & superc==subc => true
3651     _prim_1_path,               // {N,P} => false
3652     _ref_subtype_path,          // {N,N} & subtype check wins => true
3653     _both_ref_path,             // {N,N} & subtype check loses => false
3654     PATH_LIMIT
3655   };
3656 
3657   RegionNode* region = new RegionNode(PATH_LIMIT);
3658   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
3659   record_for_igvn(region);
3660 
3661   const TypePtr* adr_type = TypeRawPtr::BOTTOM;   // memory type of loads
3662   const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3663   int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3664 
3665   // First null-check both mirrors and load each mirror's klass metaobject.
3666   int which_arg;
3667   for (which_arg = 0; which_arg <= 1; which_arg++) {
3668     Node* arg = args[which_arg];
3669     arg = null_check(arg);
3670     if (stopped())  break;
3671     args[which_arg] = arg;
3672 
3673     Node* p = basic_plus_adr(arg, class_klass_offset);
3674     Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3675     klasses[which_arg] = _gvn.transform(kls);
3676   }
3677 
3678   // Having loaded both klasses, test each for null.
3679   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3680   for (which_arg = 0; which_arg <= 1; which_arg++) {
3681     Node* kls = klasses[which_arg];
3682     Node* null_ctl = top();
3683     kls = null_check_oop(kls, &null_ctl, never_see_null);
3684     int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3685     region->init_req(prim_path, null_ctl);
3686     if (stopped())  break;
3687     klasses[which_arg] = kls;
3688   }
3689 
3690   if (!stopped()) {
3691     // now we have two reference types, in klasses[0..1]
3692     Node* subk   = klasses[1];  // the argument to isAssignableFrom
3693     Node* superk = klasses[0];  // the receiver
3694     region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3695     // now we have a successful reference subtype check
3696     region->set_req(_ref_subtype_path, control());
3697   }
3698 
3699   // If both operands are primitive (both klasses null), then
3700   // we must return true when they are identical primitives.
3701   // It is convenient to test this after the first null klass check.
3702   set_control(region->in(_prim_0_path)); // go back to first null check
3703   if (!stopped()) {
3704     // Since superc is primitive, make a guard for the superc==subc case.
3705     Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3706     Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3707     generate_guard(bol_eq, region, PROB_FAIR);
3708     if (region->req() == PATH_LIMIT+1) {
3709       // A guard was added.  If the added guard is taken, superc==subc.
3710       region->swap_edges(PATH_LIMIT, _prim_same_path);
3711       region->del_req(PATH_LIMIT);
3712     }
3713     region->set_req(_prim_0_path, control()); // Not equal after all.
3714   }
3715 
3716   // these are the only paths that produce 'true':
3717   phi->set_req(_prim_same_path,   intcon(1));
3718   phi->set_req(_ref_subtype_path, intcon(1));
3719 
3720   // pull together the cases:
3721   assert(region->req() == PATH_LIMIT, "sane region");
3722   for (uint i = 1; i < region->req(); i++) {
3723     Node* ctl = region->in(i);
3724     if (ctl == NULL || ctl == top()) {
3725       region->set_req(i, top());
3726       phi   ->set_req(i, top());
3727     } else if (phi->in(i) == NULL) {
3728       phi->set_req(i, intcon(0)); // all other paths produce 'false'
3729     }
3730   }
3731 
3732   set_control(_gvn.transform(region));
3733   set_result(_gvn.transform(phi));
3734   return true;
3735 }
3736 
3737 //---------------------generate_array_guard_common------------------------
3738 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3739                                                   bool obj_array, bool not_array) {
3740   // If obj_array/non_array==false/false:
3741   // Branch around if the given klass is in fact an array (either obj or prim).
3742   // If obj_array/non_array==false/true:
3743   // Branch around if the given klass is not an array klass of any kind.
3744   // If obj_array/non_array==true/true:
3745   // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3746   // If obj_array/non_array==true/false:
3747   // Branch around if the kls is an oop array (Object[] or subtype)
3748   //
3749   // Like generate_guard, adds a new path onto the region.
3750   jint  layout_con = 0;
3751   Node* layout_val = get_layout_helper(kls, layout_con);
3752   if (layout_val == NULL) {
3753     bool query = (obj_array
3754                   ? Klass::layout_helper_is_objArray(layout_con)
3755                   : Klass::layout_helper_is_array(layout_con));
3756     if (query == not_array) {
3757       return NULL;                       // never a branch
3758     } else {                             // always a branch
3759       Node* always_branch = control();
3760       if (region != NULL)
3761         region->add_req(always_branch);
3762       set_control(top());
3763       return always_branch;
3764     }
3765   }
3766   // Now test the correct condition.
3767   jint  nval = (obj_array
3768                 ? ((jint)Klass::_lh_array_tag_type_value
3769                    <<    Klass::_lh_array_tag_shift)
3770                 : Klass::_lh_neutral_value);
3771   Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3772   BoolTest::mask btest = BoolTest::lt;  // correct for testing is_[obj]array
3773   // invert the test if we are looking for a non-array
3774   if (not_array)  btest = BoolTest(btest).negate();
3775   Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3776   return generate_fair_guard(bol, region);
3777 }
3778 
3779 
3780 //-----------------------inline_native_newArray--------------------------
3781 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3782 bool LibraryCallKit::inline_native_newArray() {
3783   Node* mirror    = argument(0);
3784   Node* count_val = argument(1);
3785 
3786   mirror = null_check(mirror);
3787   // If mirror or obj is dead, only null-path is taken.
3788   if (stopped())  return true;
3789 
3790   enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3791   RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3792   PhiNode*    result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3793   PhiNode*    result_io  = new PhiNode(result_reg, Type::ABIO);
3794   PhiNode*    result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3795 
3796   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3797   Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3798                                                   result_reg, _slow_path);
3799   Node* normal_ctl   = control();
3800   Node* no_array_ctl = result_reg->in(_slow_path);
3801 
3802   // Generate code for the slow case.  We make a call to newArray().
3803   set_control(no_array_ctl);
3804   if (!stopped()) {
3805     // Either the input type is void.class, or else the
3806     // array klass has not yet been cached.  Either the
3807     // ensuing call will throw an exception, or else it
3808     // will cache the array klass for next time.
3809     PreserveJVMState pjvms(this);
3810     CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3811     Node* slow_result = set_results_for_java_call(slow_call);
3812     // this->control() comes from set_results_for_java_call
3813     result_reg->set_req(_slow_path, control());
3814     result_val->set_req(_slow_path, slow_result);
3815     result_io ->set_req(_slow_path, i_o());
3816     result_mem->set_req(_slow_path, reset_memory());
3817   }
3818 
3819   set_control(normal_ctl);
3820   if (!stopped()) {
3821     // Normal case:  The array type has been cached in the java.lang.Class.
3822     // The following call works fine even if the array type is polymorphic.
3823     // It could be a dynamic mix of int[], boolean[], Object[], etc.
3824     Node* obj = new_array(klass_node, count_val, 0);  // no arguments to push
3825     result_reg->init_req(_normal_path, control());
3826     result_val->init_req(_normal_path, obj);
3827     result_io ->init_req(_normal_path, i_o());
3828     result_mem->init_req(_normal_path, reset_memory());
3829   }
3830 
3831   // Return the combined state.
3832   set_i_o(        _gvn.transform(result_io)  );
3833   set_all_memory( _gvn.transform(result_mem));
3834 
3835   C->set_has_split_ifs(true); // Has chance for split-if optimization
3836   set_result(result_reg, result_val);
3837   return true;
3838 }
3839 
3840 //----------------------inline_native_getLength--------------------------
3841 // public static native int java.lang.reflect.Array.getLength(Object array);
3842 bool LibraryCallKit::inline_native_getLength() {
3843   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3844 
3845   Node* array = null_check(argument(0));
3846   // If array is dead, only null-path is taken.
3847   if (stopped())  return true;
3848 
3849   // Deoptimize if it is a non-array.
3850   Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3851 
3852   if (non_array != NULL) {
3853     PreserveJVMState pjvms(this);
3854     set_control(non_array);
3855     uncommon_trap(Deoptimization::Reason_intrinsic,
3856                   Deoptimization::Action_maybe_recompile);
3857   }
3858 
3859   // If control is dead, only non-array-path is taken.
3860   if (stopped())  return true;
3861 
3862   // The works fine even if the array type is polymorphic.
3863   // It could be a dynamic mix of int[], boolean[], Object[], etc.
3864   Node* result = load_array_length(array);
3865 
3866   C->set_has_split_ifs(true);  // Has chance for split-if optimization
3867   set_result(result);
3868   return true;
3869 }
3870 
3871 //------------------------inline_array_copyOf----------------------------
3872 // public static <T,U> T[] java.util.Arrays.copyOf(     U[] original, int newLength,         Class<? extends T[]> newType);
3873 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from,      int to, Class<? extends T[]> newType);
3874 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3875   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3876 
3877   // Get the arguments.
3878   Node* original          = argument(0);
3879   Node* start             = is_copyOfRange? argument(1): intcon(0);
3880   Node* end               = is_copyOfRange? argument(2): argument(1);
3881   Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3882 
3883   Node* newcopy;
3884 
3885   // Set the original stack and the reexecute bit for the interpreter to reexecute
3886   // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3887   { PreserveReexecuteState preexecs(this);
3888     jvms()->set_should_reexecute(true);
3889 
3890     array_type_mirror = null_check(array_type_mirror);
3891     original          = null_check(original);
3892 
3893     // Check if a null path was taken unconditionally.
3894     if (stopped())  return true;
3895 
3896     Node* orig_length = load_array_length(original);
3897 
3898     Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3899     klass_node = null_check(klass_node);
3900 
3901     RegionNode* bailout = new RegionNode(1);
3902     record_for_igvn(bailout);
3903 
3904     // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3905     // Bail out if that is so.
3906     Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3907     if (not_objArray != NULL) {
3908       // Improve the klass node's type from the new optimistic assumption:
3909       ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3910       const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3911       Node* cast = new CastPPNode(klass_node, akls);
3912       cast->init_req(0, control());
3913       klass_node = _gvn.transform(cast);
3914     }
3915 
3916     // Bail out if either start or end is negative.
3917     generate_negative_guard(start, bailout, &start);
3918     generate_negative_guard(end,   bailout, &end);
3919 
3920     Node* length = end;
3921     if (_gvn.type(start) != TypeInt::ZERO) {
3922       length = _gvn.transform(new SubINode(end, start));
3923     }
3924 
3925     // Bail out if length is negative.
3926     // Without this the new_array would throw
3927     // NegativeArraySizeException but IllegalArgumentException is what
3928     // should be thrown
3929     generate_negative_guard(length, bailout, &length);
3930 
3931     if (bailout->req() > 1) {
3932       PreserveJVMState pjvms(this);
3933       set_control(_gvn.transform(bailout));
3934       uncommon_trap(Deoptimization::Reason_intrinsic,
3935                     Deoptimization::Action_maybe_recompile);
3936     }
3937 
3938     if (!stopped()) {
3939       // How many elements will we copy from the original?
3940       // The answer is MinI(orig_length - start, length).
3941       Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3942       Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3943 
3944       newcopy = new_array(klass_node, length, 0);  // no arguments to push
3945 
3946       // Generate a direct call to the right arraycopy function(s).
3947       // We know the copy is disjoint but we might not know if the
3948       // oop stores need checking.
3949       // Extreme case:  Arrays.copyOf((Integer[])x, 10, String[].class).
3950       // This will fail a store-check if x contains any non-nulls.
3951 
3952       Node* alloc = tightly_coupled_allocation(newcopy, NULL);
3953 
3954       ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, alloc != NULL,
3955                                               load_object_klass(original), klass_node);
3956       if (!is_copyOfRange) {
3957         ac->set_copyof();
3958       } else {
3959         ac->set_copyofrange();
3960       }
3961       Node* n = _gvn.transform(ac);
3962       assert(n == ac, "cannot disappear");
3963       ac->connect_outputs(this);
3964     }
3965   } // original reexecute is set back here
3966 
3967   C->set_has_split_ifs(true); // Has chance for split-if optimization
3968   if (!stopped()) {
3969     set_result(newcopy);
3970   }
3971   return true;
3972 }
3973 
3974 
3975 //----------------------generate_virtual_guard---------------------------
3976 // Helper for hashCode and clone.  Peeks inside the vtable to avoid a call.
3977 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3978                                              RegionNode* slow_region) {
3979   ciMethod* method = callee();
3980   int vtable_index = method->vtable_index();
3981   assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3982          err_msg_res("bad index %d", vtable_index));
3983   // Get the Method* out of the appropriate vtable entry.
3984   int entry_offset  = (InstanceKlass::vtable_start_offset() +
3985                      vtable_index*vtableEntry::size()) * wordSize +
3986                      vtableEntry::method_offset_in_bytes();
3987   Node* entry_addr  = basic_plus_adr(obj_klass, entry_offset);
3988   Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3989 
3990   // Compare the target method with the expected method (e.g., Object.hashCode).
3991   const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3992 
3993   Node* native_call = makecon(native_call_addr);
3994   Node* chk_native  = _gvn.transform(new CmpPNode(target_call, native_call));
3995   Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3996 
3997   return generate_slow_guard(test_native, slow_region);
3998 }
3999 
4000 //-----------------------generate_method_call----------------------------
4001 // Use generate_method_call to make a slow-call to the real
4002 // method if the fast path fails.  An alternative would be to
4003 // use a stub like OptoRuntime::slow_arraycopy_Java.
4004 // This only works for expanding the current library call,
4005 // not another intrinsic.  (E.g., don't use this for making an
4006 // arraycopy call inside of the copyOf intrinsic.)
4007 CallJavaNode*
4008 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4009   // When compiling the intrinsic method itself, do not use this technique.
4010   guarantee(callee() != C->method(), "cannot make slow-call to self");
4011 
4012   ciMethod* method = callee();
4013   // ensure the JVMS we have will be correct for this call
4014   guarantee(method_id == method->intrinsic_id(), "must match");
4015 
4016   const TypeFunc* tf = TypeFunc::make(method);
4017   CallJavaNode* slow_call;
4018   if (is_static) {
4019     assert(!is_virtual, "");
4020     slow_call = new CallStaticJavaNode(C, tf,
4021                            SharedRuntime::get_resolve_static_call_stub(),
4022                            method, bci());
4023   } else if (is_virtual) {
4024     null_check_receiver();
4025     int vtable_index = Method::invalid_vtable_index;
4026     if (UseInlineCaches) {
4027       // Suppress the vtable call
4028     } else {
4029       // hashCode and clone are not a miranda methods,
4030       // so the vtable index is fixed.
4031       // No need to use the linkResolver to get it.
4032        vtable_index = method->vtable_index();
4033        assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4034               err_msg_res("bad index %d", vtable_index));
4035     }
4036     slow_call = new CallDynamicJavaNode(tf,
4037                           SharedRuntime::get_resolve_virtual_call_stub(),
4038                           method, vtable_index, bci());
4039   } else {  // neither virtual nor static:  opt_virtual
4040     null_check_receiver();
4041     slow_call = new CallStaticJavaNode(C, tf,
4042                                 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4043                                 method, bci());
4044     slow_call->set_optimized_virtual(true);
4045   }
4046   set_arguments_for_java_call(slow_call);
4047   set_edges_for_java_call(slow_call);
4048   return slow_call;
4049 }
4050 
4051 
4052 /**
4053  * Build special case code for calls to hashCode on an object. This call may
4054  * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4055  * slightly different code.
4056  */
4057 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4058   assert(is_static == callee()->is_static(), "correct intrinsic selection");
4059   assert(!(is_virtual && is_static), "either virtual, special, or static");
4060 
4061   enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4062 
4063   RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4064   PhiNode*    result_val = new PhiNode(result_reg, TypeInt::INT);
4065   PhiNode*    result_io  = new PhiNode(result_reg, Type::ABIO);
4066   PhiNode*    result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4067   Node* obj = NULL;
4068   if (!is_static) {
4069     // Check for hashing null object
4070     obj = null_check_receiver();
4071     if (stopped())  return true;        // unconditionally null
4072     result_reg->init_req(_null_path, top());
4073     result_val->init_req(_null_path, top());
4074   } else {
4075     // Do a null check, and return zero if null.
4076     // System.identityHashCode(null) == 0
4077     obj = argument(0);
4078     Node* null_ctl = top();
4079     obj = null_check_oop(obj, &null_ctl);
4080     result_reg->init_req(_null_path, null_ctl);
4081     result_val->init_req(_null_path, _gvn.intcon(0));
4082   }
4083 
4084   // Unconditionally null?  Then return right away.
4085   if (stopped()) {
4086     set_control( result_reg->in(_null_path));
4087     if (!stopped())
4088       set_result(result_val->in(_null_path));
4089     return true;
4090   }
4091 
4092   // We only go to the fast case code if we pass a number of guards.  The
4093   // paths which do not pass are accumulated in the slow_region.
4094   RegionNode* slow_region = new RegionNode(1);
4095   record_for_igvn(slow_region);
4096 
4097   // If this is a virtual call, we generate a funny guard.  We pull out
4098   // the vtable entry corresponding to hashCode() from the target object.
4099   // If the target method which we are calling happens to be the native
4100   // Object hashCode() method, we pass the guard.  We do not need this
4101   // guard for non-virtual calls -- the caller is known to be the native
4102   // Object hashCode().
4103   if (is_virtual) {
4104     // After null check, get the object's klass.
4105     Node* obj_klass = load_object_klass(obj);
4106     generate_virtual_guard(obj_klass, slow_region);
4107   }
4108 
4109   // Get the header out of the object, use LoadMarkNode when available
4110   Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4111   // The control of the load must be NULL. Otherwise, the load can move before
4112   // the null check after castPP removal.
4113   Node* no_ctrl = NULL;
4114   Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4115 
4116   // Test the header to see if it is unlocked.
4117   Node *lock_mask      = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4118   Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4119   Node *unlocked_val   = _gvn.MakeConX(markOopDesc::unlocked_value);
4120   Node *chk_unlocked   = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4121   Node *test_unlocked  = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4122 
4123   generate_slow_guard(test_unlocked, slow_region);
4124 
4125   // Get the hash value and check to see that it has been properly assigned.
4126   // We depend on hash_mask being at most 32 bits and avoid the use of
4127   // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4128   // vm: see markOop.hpp.
4129   Node *hash_mask      = _gvn.intcon(markOopDesc::hash_mask);
4130   Node *hash_shift     = _gvn.intcon(markOopDesc::hash_shift);
4131   Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4132   // This hack lets the hash bits live anywhere in the mark object now, as long
4133   // as the shift drops the relevant bits into the low 32 bits.  Note that
4134   // Java spec says that HashCode is an int so there's no point in capturing
4135   // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4136   hshifted_header      = ConvX2I(hshifted_header);
4137   Node *hash_val       = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4138 
4139   Node *no_hash_val    = _gvn.intcon(markOopDesc::no_hash);
4140   Node *chk_assigned   = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4141   Node *test_assigned  = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4142 
4143   generate_slow_guard(test_assigned, slow_region);
4144 
4145   Node* init_mem = reset_memory();
4146   // fill in the rest of the null path:
4147   result_io ->init_req(_null_path, i_o());
4148   result_mem->init_req(_null_path, init_mem);
4149 
4150   result_val->init_req(_fast_path, hash_val);
4151   result_reg->init_req(_fast_path, control());
4152   result_io ->init_req(_fast_path, i_o());
4153   result_mem->init_req(_fast_path, init_mem);
4154 
4155   // Generate code for the slow case.  We make a call to hashCode().
4156   set_control(_gvn.transform(slow_region));
4157   if (!stopped()) {
4158     // No need for PreserveJVMState, because we're using up the present state.
4159     set_all_memory(init_mem);
4160     vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4161     CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4162     Node* slow_result = set_results_for_java_call(slow_call);
4163     // this->control() comes from set_results_for_java_call
4164     result_reg->init_req(_slow_path, control());
4165     result_val->init_req(_slow_path, slow_result);
4166     result_io  ->set_req(_slow_path, i_o());
4167     result_mem ->set_req(_slow_path, reset_memory());
4168   }
4169 
4170   // Return the combined state.
4171   set_i_o(        _gvn.transform(result_io)  );
4172   set_all_memory( _gvn.transform(result_mem));
4173 
4174   set_result(result_reg, result_val);
4175   return true;
4176 }
4177 
4178 //---------------------------inline_native_getClass----------------------------
4179 // public final native Class<?> java.lang.Object.getClass();
4180 //
4181 // Build special case code for calls to getClass on an object.
4182 bool LibraryCallKit::inline_native_getClass() {
4183   Node* obj = null_check_receiver();
4184   if (stopped())  return true;
4185   set_result(load_mirror_from_klass(load_object_klass(obj)));
4186   return true;
4187 }
4188 
4189 //-----------------inline_native_Reflection_getCallerClass---------------------
4190 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4191 //
4192 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4193 //
4194 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4195 // in that it must skip particular security frames and checks for
4196 // caller sensitive methods.
4197 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4198 #ifndef PRODUCT
4199   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4200     tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4201   }
4202 #endif
4203 
4204   if (!jvms()->has_method()) {
4205 #ifndef PRODUCT
4206     if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4207       tty->print_cr("  Bailing out because intrinsic was inlined at top level");
4208     }
4209 #endif
4210     return false;
4211   }
4212 
4213   // Walk back up the JVM state to find the caller at the required
4214   // depth.
4215   JVMState* caller_jvms = jvms();
4216 
4217   // Cf. JVM_GetCallerClass
4218   // NOTE: Start the loop at depth 1 because the current JVM state does
4219   // not include the Reflection.getCallerClass() frame.
4220   for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4221     ciMethod* m = caller_jvms->method();
4222     switch (n) {
4223     case 0:
4224       fatal("current JVM state does not include the Reflection.getCallerClass frame");
4225       break;
4226     case 1:
4227       // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4228       if (!m->caller_sensitive()) {
4229 #ifndef PRODUCT
4230         if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4231           tty->print_cr("  Bailing out: CallerSensitive annotation expected at frame %d", n);
4232         }
4233 #endif
4234         return false;  // bail-out; let JVM_GetCallerClass do the work
4235       }
4236       break;
4237     default:
4238       if (!m->is_ignored_by_security_stack_walk()) {
4239         // We have reached the desired frame; return the holder class.
4240         // Acquire method holder as java.lang.Class and push as constant.
4241         ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4242         ciInstance* caller_mirror = caller_klass->java_mirror();
4243         set_result(makecon(TypeInstPtr::make(caller_mirror)));
4244 
4245 #ifndef PRODUCT
4246         if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4247           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());
4248           tty->print_cr("  JVM state at this point:");
4249           for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4250             ciMethod* m = jvms()->of_depth(i)->method();
4251             tty->print_cr("   %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4252           }
4253         }
4254 #endif
4255         return true;
4256       }
4257       break;
4258     }
4259   }
4260 
4261 #ifndef PRODUCT
4262   if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4263     tty->print_cr("  Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4264     tty->print_cr("  JVM state at this point:");
4265     for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4266       ciMethod* m = jvms()->of_depth(i)->method();
4267       tty->print_cr("   %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4268     }
4269   }
4270 #endif
4271 
4272   return false;  // bail-out; let JVM_GetCallerClass do the work
4273 }
4274 
4275 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4276   Node* arg = argument(0);
4277   Node* result;
4278 
4279   switch (id) {
4280   case vmIntrinsics::_floatToRawIntBits:    result = new MoveF2INode(arg);  break;
4281   case vmIntrinsics::_intBitsToFloat:       result = new MoveI2FNode(arg);  break;
4282   case vmIntrinsics::_doubleToRawLongBits:  result = new MoveD2LNode(arg);  break;
4283   case vmIntrinsics::_longBitsToDouble:     result = new MoveL2DNode(arg);  break;
4284 
4285   case vmIntrinsics::_doubleToLongBits: {
4286     // two paths (plus control) merge in a wood
4287     RegionNode *r = new RegionNode(3);
4288     Node *phi = new PhiNode(r, TypeLong::LONG);
4289 
4290     Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4291     // Build the boolean node
4292     Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4293 
4294     // Branch either way.
4295     // NaN case is less traveled, which makes all the difference.
4296     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4297     Node *opt_isnan = _gvn.transform(ifisnan);
4298     assert( opt_isnan->is_If(), "Expect an IfNode");
4299     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4300     Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4301 
4302     set_control(iftrue);
4303 
4304     static const jlong nan_bits = CONST64(0x7ff8000000000000);
4305     Node *slow_result = longcon(nan_bits); // return NaN
4306     phi->init_req(1, _gvn.transform( slow_result ));
4307     r->init_req(1, iftrue);
4308 
4309     // Else fall through
4310     Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4311     set_control(iffalse);
4312 
4313     phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4314     r->init_req(2, iffalse);
4315 
4316     // Post merge
4317     set_control(_gvn.transform(r));
4318     record_for_igvn(r);
4319 
4320     C->set_has_split_ifs(true); // Has chance for split-if optimization
4321     result = phi;
4322     assert(result->bottom_type()->isa_long(), "must be");
4323     break;
4324   }
4325 
4326   case vmIntrinsics::_floatToIntBits: {
4327     // two paths (plus control) merge in a wood
4328     RegionNode *r = new RegionNode(3);
4329     Node *phi = new PhiNode(r, TypeInt::INT);
4330 
4331     Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4332     // Build the boolean node
4333     Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4334 
4335     // Branch either way.
4336     // NaN case is less traveled, which makes all the difference.
4337     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4338     Node *opt_isnan = _gvn.transform(ifisnan);
4339     assert( opt_isnan->is_If(), "Expect an IfNode");
4340     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4341     Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4342 
4343     set_control(iftrue);
4344 
4345     static const jint nan_bits = 0x7fc00000;
4346     Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4347     phi->init_req(1, _gvn.transform( slow_result ));
4348     r->init_req(1, iftrue);
4349 
4350     // Else fall through
4351     Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4352     set_control(iffalse);
4353 
4354     phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4355     r->init_req(2, iffalse);
4356 
4357     // Post merge
4358     set_control(_gvn.transform(r));
4359     record_for_igvn(r);
4360 
4361     C->set_has_split_ifs(true); // Has chance for split-if optimization
4362     result = phi;
4363     assert(result->bottom_type()->isa_int(), "must be");
4364     break;
4365   }
4366 
4367   default:
4368     fatal_unexpected_iid(id);
4369     break;
4370   }
4371   set_result(_gvn.transform(result));
4372   return true;
4373 }
4374 
4375 #ifdef _LP64
4376 #define XTOP ,top() /*additional argument*/
4377 #else  //_LP64
4378 #define XTOP        /*no additional argument*/
4379 #endif //_LP64
4380 
4381 //----------------------inline_unsafe_copyMemory-------------------------
4382 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4383 bool LibraryCallKit::inline_unsafe_copyMemory() {
4384   if (callee()->is_static())  return false;  // caller must have the capability!
4385   null_check_receiver();  // null-check receiver
4386   if (stopped())  return true;
4387 
4388   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
4389 
4390   Node* src_ptr =         argument(1);   // type: oop
4391   Node* src_off = ConvL2X(argument(2));  // type: long
4392   Node* dst_ptr =         argument(4);   // type: oop
4393   Node* dst_off = ConvL2X(argument(5));  // type: long
4394   Node* size    = ConvL2X(argument(7));  // type: long
4395 
4396   assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4397          "fieldOffset must be byte-scaled");
4398 
4399   Node* src = make_unsafe_address(src_ptr, src_off);
4400   Node* dst = make_unsafe_address(dst_ptr, dst_off);
4401 
4402   // Conservatively insert a memory barrier on all memory slices.
4403   // Do not let writes of the copy source or destination float below the copy.
4404   insert_mem_bar(Op_MemBarCPUOrder);
4405 
4406   // Call it.  Note that the length argument is not scaled.
4407   make_runtime_call(RC_LEAF|RC_NO_FP,
4408                     OptoRuntime::fast_arraycopy_Type(),
4409                     StubRoutines::unsafe_arraycopy(),
4410                     "unsafe_arraycopy",
4411                     TypeRawPtr::BOTTOM,
4412                     src, dst, size XTOP);
4413 
4414   // Do not let reads of the copy destination float above the copy.
4415   insert_mem_bar(Op_MemBarCPUOrder);
4416 
4417   return true;
4418 }
4419 
4420 //------------------------clone_coping-----------------------------------
4421 // Helper function for inline_native_clone.
4422 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4423   assert(obj_size != NULL, "");
4424   Node* raw_obj = alloc_obj->in(1);
4425   assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4426 
4427   AllocateNode* alloc = NULL;
4428   if (ReduceBulkZeroing) {
4429     // We will be completely responsible for initializing this object -
4430     // mark Initialize node as complete.
4431     alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4432     // The object was just allocated - there should be no any stores!
4433     guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4434     // Mark as complete_with_arraycopy so that on AllocateNode
4435     // expansion, we know this AllocateNode is initialized by an array
4436     // copy and a StoreStore barrier exists after the array copy.
4437     alloc->initialization()->set_complete_with_arraycopy();
4438   }
4439 
4440   // Copy the fastest available way.
4441   // TODO: generate fields copies for small objects instead.
4442   Node* src  = obj;
4443   Node* dest = alloc_obj;
4444   Node* size = _gvn.transform(obj_size);
4445 
4446   // Exclude the header but include array length to copy by 8 bytes words.
4447   // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4448   int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4449                             instanceOopDesc::base_offset_in_bytes();
4450   // base_off:
4451   // 8  - 32-bit VM
4452   // 12 - 64-bit VM, compressed klass
4453   // 16 - 64-bit VM, normal klass
4454   if (base_off % BytesPerLong != 0) {
4455     assert(UseCompressedClassPointers, "");
4456     if (is_array) {
4457       // Exclude length to copy by 8 bytes words.
4458       base_off += sizeof(int);
4459     } else {
4460       // Include klass to copy by 8 bytes words.
4461       base_off = instanceOopDesc::klass_offset_in_bytes();
4462     }
4463     assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4464   }
4465   src  = basic_plus_adr(src,  base_off);
4466   dest = basic_plus_adr(dest, base_off);
4467 
4468   // Compute the length also, if needed:
4469   Node* countx = size;
4470   countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4471   countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4472 
4473   const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4474 
4475   ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4476   ac->set_clonebasic();
4477   Node* n = _gvn.transform(ac);
4478   assert(n == ac, "cannot disappear");
4479   set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4480 
4481   // If necessary, emit some card marks afterwards.  (Non-arrays only.)
4482   if (card_mark) {
4483     assert(!is_array, "");
4484     // Put in store barrier for any and all oops we are sticking
4485     // into this object.  (We could avoid this if we could prove
4486     // that the object type contains no oop fields at all.)
4487     Node* no_particular_value = NULL;
4488     Node* no_particular_field = NULL;
4489     int raw_adr_idx = Compile::AliasIdxRaw;
4490     post_barrier(control(),
4491                  memory(raw_adr_type),
4492                  alloc_obj,
4493                  no_particular_field,
4494                  raw_adr_idx,
4495                  no_particular_value,
4496                  T_OBJECT,
4497                  false);
4498   }
4499 
4500   // Do not let reads from the cloned object float above the arraycopy.
4501   if (alloc != NULL) {
4502     // Do not let stores that initialize this object be reordered with
4503     // a subsequent store that would make this object accessible by
4504     // other threads.
4505     // Record what AllocateNode this StoreStore protects so that
4506     // escape analysis can go from the MemBarStoreStoreNode to the
4507     // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4508     // based on the escape status of the AllocateNode.
4509     insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4510   } else {
4511     insert_mem_bar(Op_MemBarCPUOrder);
4512   }
4513 }
4514 
4515 //------------------------inline_native_clone----------------------------
4516 // protected native Object java.lang.Object.clone();
4517 //
4518 // Here are the simple edge cases:
4519 //  null receiver => normal trap
4520 //  virtual and clone was overridden => slow path to out-of-line clone
4521 //  not cloneable or finalizer => slow path to out-of-line Object.clone
4522 //
4523 // The general case has two steps, allocation and copying.
4524 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4525 //
4526 // Copying also has two cases, oop arrays and everything else.
4527 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4528 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4529 //
4530 // These steps fold up nicely if and when the cloned object's klass
4531 // can be sharply typed as an object array, a type array, or an instance.
4532 //
4533 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4534   PhiNode* result_val;
4535 
4536   // Set the reexecute bit for the interpreter to reexecute
4537   // the bytecode that invokes Object.clone if deoptimization happens.
4538   { PreserveReexecuteState preexecs(this);
4539     jvms()->set_should_reexecute(true);
4540 
4541     Node* obj = null_check_receiver();
4542     if (stopped())  return true;
4543 
4544     Node* obj_klass = load_object_klass(obj);
4545     const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4546     const TypeOopPtr*   toop   = ((tklass != NULL)
4547                                 ? tklass->as_instance_type()
4548                                 : TypeInstPtr::NOTNULL);
4549 
4550     // Conservatively insert a memory barrier on all memory slices.
4551     // Do not let writes into the original float below the clone.
4552     insert_mem_bar(Op_MemBarCPUOrder);
4553 
4554     // paths into result_reg:
4555     enum {
4556       _slow_path = 1,     // out-of-line call to clone method (virtual or not)
4557       _objArray_path,     // plain array allocation, plus arrayof_oop_arraycopy
4558       _array_path,        // plain array allocation, plus arrayof_long_arraycopy
4559       _instance_path,     // plain instance allocation, plus arrayof_long_arraycopy
4560       PATH_LIMIT
4561     };
4562     RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4563     result_val             = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4564     PhiNode*    result_i_o = new PhiNode(result_reg, Type::ABIO);
4565     PhiNode*    result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4566     record_for_igvn(result_reg);
4567 
4568     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4569     int raw_adr_idx = Compile::AliasIdxRaw;
4570 
4571     Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4572     if (array_ctl != NULL) {
4573       // It's an array.
4574       PreserveJVMState pjvms(this);
4575       set_control(array_ctl);
4576       Node* obj_length = load_array_length(obj);
4577       Node* obj_size  = NULL;
4578       Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);  // no arguments to push
4579 
4580       if (!use_ReduceInitialCardMarks()) {
4581         // If it is an oop array, it requires very special treatment,
4582         // because card marking is required on each card of the array.
4583         Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4584         if (is_obja != NULL) {
4585           PreserveJVMState pjvms2(this);
4586           set_control(is_obja);
4587           // Generate a direct call to the right arraycopy function(s).
4588           Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4589           ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4590           ac->set_cloneoop();
4591           Node* n = _gvn.transform(ac);
4592           assert(n == ac, "cannot disappear");
4593           ac->connect_outputs(this);
4594 
4595           result_reg->init_req(_objArray_path, control());
4596           result_val->init_req(_objArray_path, alloc_obj);
4597           result_i_o ->set_req(_objArray_path, i_o());
4598           result_mem ->set_req(_objArray_path, reset_memory());
4599         }
4600       }
4601       // Otherwise, there are no card marks to worry about.
4602       // (We can dispense with card marks if we know the allocation
4603       //  comes out of eden (TLAB)...  In fact, ReduceInitialCardMarks
4604       //  causes the non-eden paths to take compensating steps to
4605       //  simulate a fresh allocation, so that no further
4606       //  card marks are required in compiled code to initialize
4607       //  the object.)
4608 
4609       if (!stopped()) {
4610         copy_to_clone(obj, alloc_obj, obj_size, true, false);
4611 
4612         // Present the results of the copy.
4613         result_reg->init_req(_array_path, control());
4614         result_val->init_req(_array_path, alloc_obj);
4615         result_i_o ->set_req(_array_path, i_o());
4616         result_mem ->set_req(_array_path, reset_memory());
4617       }
4618     }
4619 
4620     // We only go to the instance fast case code if we pass a number of guards.
4621     // The paths which do not pass are accumulated in the slow_region.
4622     RegionNode* slow_region = new RegionNode(1);
4623     record_for_igvn(slow_region);
4624     if (!stopped()) {
4625       // It's an instance (we did array above).  Make the slow-path tests.
4626       // If this is a virtual call, we generate a funny guard.  We grab
4627       // the vtable entry corresponding to clone() from the target object.
4628       // If the target method which we are calling happens to be the
4629       // Object clone() method, we pass the guard.  We do not need this
4630       // guard for non-virtual calls; the caller is known to be the native
4631       // Object clone().
4632       if (is_virtual) {
4633         generate_virtual_guard(obj_klass, slow_region);
4634       }
4635 
4636       // The object must be cloneable and must not have a finalizer.
4637       // Both of these conditions may be checked in a single test.
4638       // We could optimize the cloneable test further, but we don't care.
4639       generate_access_flags_guard(obj_klass,
4640                                   // Test both conditions:
4641                                   JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4642                                   // Must be cloneable but not finalizer:
4643                                   JVM_ACC_IS_CLONEABLE,
4644                                   slow_region);
4645     }
4646 
4647     if (!stopped()) {
4648       // It's an instance, and it passed the slow-path tests.
4649       PreserveJVMState pjvms(this);
4650       Node* obj_size  = NULL;
4651       // Need to deoptimize on exception from allocation since Object.clone intrinsic
4652       // is reexecuted if deoptimization occurs and there could be problems when merging
4653       // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4654       Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4655 
4656       copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4657 
4658       // Present the results of the slow call.
4659       result_reg->init_req(_instance_path, control());
4660       result_val->init_req(_instance_path, alloc_obj);
4661       result_i_o ->set_req(_instance_path, i_o());
4662       result_mem ->set_req(_instance_path, reset_memory());
4663     }
4664 
4665     // Generate code for the slow case.  We make a call to clone().
4666     set_control(_gvn.transform(slow_region));
4667     if (!stopped()) {
4668       PreserveJVMState pjvms(this);
4669       CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4670       Node* slow_result = set_results_for_java_call(slow_call);
4671       // this->control() comes from set_results_for_java_call
4672       result_reg->init_req(_slow_path, control());
4673       result_val->init_req(_slow_path, slow_result);
4674       result_i_o ->set_req(_slow_path, i_o());
4675       result_mem ->set_req(_slow_path, reset_memory());
4676     }
4677 
4678     // Return the combined state.
4679     set_control(    _gvn.transform(result_reg));
4680     set_i_o(        _gvn.transform(result_i_o));
4681     set_all_memory( _gvn.transform(result_mem));
4682   } // original reexecute is set back here
4683 
4684   set_result(_gvn.transform(result_val));
4685   return true;
4686 }
4687 
4688 //------------------------------inline_arraycopy-----------------------
4689 // public static native void java.lang.System.arraycopy(Object src,  int  srcPos,
4690 //                                                      Object dest, int destPos,
4691 //                                                      int length);
4692 bool LibraryCallKit::inline_arraycopy() {
4693   // Get the arguments.
4694   Node* src         = argument(0);  // type: oop
4695   Node* src_offset  = argument(1);  // type: int
4696   Node* dest        = argument(2);  // type: oop
4697   Node* dest_offset = argument(3);  // type: int
4698   Node* length      = argument(4);  // type: int
4699 
4700   // Check for allocation before we add nodes that would confuse
4701   // tightly_coupled_allocation()
4702   AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4703 
4704   // The following tests must be performed
4705   // (1) src and dest are arrays.
4706   // (2) src and dest arrays must have elements of the same BasicType
4707   // (3) src and dest must not be null.
4708   // (4) src_offset must not be negative.
4709   // (5) dest_offset must not be negative.
4710   // (6) length must not be negative.
4711   // (7) src_offset + length must not exceed length of src.
4712   // (8) dest_offset + length must not exceed length of dest.
4713   // (9) each element of an oop array must be assignable
4714 
4715   // (3) src and dest must not be null.
4716   // always do this here because we need the JVM state for uncommon traps
4717   src  = null_check(src,  T_ARRAY);
4718   dest = null_check(dest, T_ARRAY);
4719 
4720   bool notest = false;
4721 
4722   const Type* src_type  = _gvn.type(src);
4723   const Type* dest_type = _gvn.type(dest);
4724   const TypeAryPtr* top_src  = src_type->isa_aryptr();
4725   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4726 
4727   // Do we have the type of src?
4728   bool has_src = (top_src != NULL && top_src->klass() != NULL);
4729   // Do we have the type of dest?
4730   bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4731   // Is the type for src from speculation?
4732   bool src_spec = false;
4733   // Is the type for dest from speculation?
4734   bool dest_spec = false;
4735 
4736   if (!has_src || !has_dest) {
4737     // We don't have sufficient type information, let's see if
4738     // speculative types can help. We need to have types for both src
4739     // and dest so that it pays off.
4740 
4741     // Do we already have or could we have type information for src
4742     bool could_have_src = has_src;
4743     // Do we already have or could we have type information for dest
4744     bool could_have_dest = has_dest;
4745 
4746     ciKlass* src_k = NULL;
4747     if (!has_src) {
4748       src_k = src_type->speculative_type_not_null();
4749       if (src_k != NULL && src_k->is_array_klass()) {
4750         could_have_src = true;
4751       }
4752     }
4753 
4754     ciKlass* dest_k = NULL;
4755     if (!has_dest) {
4756       dest_k = dest_type->speculative_type_not_null();
4757       if (dest_k != NULL && dest_k->is_array_klass()) {
4758         could_have_dest = true;
4759       }
4760     }
4761 
4762     if (could_have_src && could_have_dest) {
4763       // This is going to pay off so emit the required guards
4764       if (!has_src) {
4765         src = maybe_cast_profiled_obj(src, src_k);
4766         src_type  = _gvn.type(src);
4767         top_src  = src_type->isa_aryptr();
4768         has_src = (top_src != NULL && top_src->klass() != NULL);
4769         src_spec = true;
4770       }
4771       if (!has_dest) {
4772         dest = maybe_cast_profiled_obj(dest, dest_k);
4773         dest_type  = _gvn.type(dest);
4774         top_dest  = dest_type->isa_aryptr();
4775         has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4776         dest_spec = true;
4777       }
4778     }
4779   }
4780 
4781   if (has_src && has_dest) {
4782     BasicType src_elem  = top_src->klass()->as_array_klass()->element_type()->basic_type();
4783     BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4784     if (src_elem  == T_ARRAY)  src_elem  = T_OBJECT;
4785     if (dest_elem == T_ARRAY)  dest_elem = T_OBJECT;
4786 
4787     if (src_elem == dest_elem && src_elem == T_OBJECT) {
4788       // If both arrays are object arrays then having the exact types
4789       // for both will remove the need for a subtype check at runtime
4790       // before the call and may make it possible to pick a faster copy
4791       // routine (without a subtype check on every element)
4792       // Do we have the exact type of src?
4793       bool could_have_src = src_spec;
4794       // Do we have the exact type of dest?
4795       bool could_have_dest = dest_spec;
4796       ciKlass* src_k = top_src->klass();
4797       ciKlass* dest_k = top_dest->klass();
4798       if (!src_spec) {
4799         src_k = src_type->speculative_type_not_null();
4800         if (src_k != NULL && src_k->is_array_klass()) {
4801           could_have_src = true;
4802         }
4803       }
4804       if (!dest_spec) {
4805         dest_k = dest_type->speculative_type_not_null();
4806         if (dest_k != NULL && dest_k->is_array_klass()) {
4807           could_have_dest = true;
4808         }
4809       }
4810       if (could_have_src && could_have_dest) {
4811         // If we can have both exact types, emit the missing guards
4812         if (could_have_src && !src_spec) {
4813           src = maybe_cast_profiled_obj(src, src_k);
4814         }
4815         if (could_have_dest && !dest_spec) {
4816           dest = maybe_cast_profiled_obj(dest, dest_k);
4817         }
4818       }
4819     }
4820   }
4821 
4822   if (!too_many_traps(Deoptimization::Reason_intrinsic) && !src->is_top() && !dest->is_top()) {
4823     // validate arguments: enables transformation the ArrayCopyNode
4824     notest = true;
4825 
4826     RegionNode* slow_region = new RegionNode(1);
4827     record_for_igvn(slow_region);
4828 
4829     // (1) src and dest are arrays.
4830     generate_non_array_guard(load_object_klass(src), slow_region);
4831     generate_non_array_guard(load_object_klass(dest), slow_region);
4832 
4833     // (2) src and dest arrays must have elements of the same BasicType
4834     // done at macro expansion or at Ideal transformation time
4835 
4836     // (4) src_offset must not be negative.
4837     generate_negative_guard(src_offset, slow_region);
4838 
4839     // (5) dest_offset must not be negative.
4840     generate_negative_guard(dest_offset, slow_region);
4841 
4842     // (7) src_offset + length must not exceed length of src.
4843     generate_limit_guard(src_offset, length,
4844                          load_array_length(src),
4845                          slow_region);
4846 
4847     // (8) dest_offset + length must not exceed length of dest.
4848     generate_limit_guard(dest_offset, length,
4849                          load_array_length(dest),
4850                          slow_region);
4851 
4852     // (9) each element of an oop array must be assignable
4853     Node* src_klass  = load_object_klass(src);
4854     Node* dest_klass = load_object_klass(dest);
4855     Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4856 
4857     if (not_subtype_ctrl != top()) {
4858       PreserveJVMState pjvms(this);
4859       set_control(not_subtype_ctrl);
4860       uncommon_trap(Deoptimization::Reason_intrinsic,
4861                     Deoptimization::Action_make_not_entrant);
4862       assert(stopped(), "Should be stopped");
4863     }
4864     {
4865       PreserveJVMState pjvms(this);
4866       set_control(_gvn.transform(slow_region));
4867       uncommon_trap(Deoptimization::Reason_intrinsic,
4868                     Deoptimization::Action_make_not_entrant);
4869       assert(stopped(), "Should be stopped");
4870     }
4871   }
4872 
4873   if (stopped()) {
4874     return true;
4875   }
4876 
4877   ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
4878                                           // Create LoadRange and LoadKlass nodes for use during macro expansion here
4879                                           // so the compiler has a chance to eliminate them: during macro expansion,
4880                                           // we have to set their control (CastPP nodes are eliminated).
4881                                           load_object_klass(src), load_object_klass(dest),
4882                                           load_array_length(src), load_array_length(dest));
4883 
4884   if (notest) {
4885     ac->set_arraycopy_notest();
4886   }
4887 
4888   Node* n = _gvn.transform(ac);
4889   assert(n == ac, "cannot disappear");
4890   ac->connect_outputs(this);
4891 
4892   return true;
4893 }
4894 
4895 
4896 // Helper function which determines if an arraycopy immediately follows
4897 // an allocation, with no intervening tests or other escapes for the object.
4898 AllocateArrayNode*
4899 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4900                                            RegionNode* slow_region) {
4901   if (stopped())             return NULL;  // no fast path
4902   if (C->AliasLevel() == 0)  return NULL;  // no MergeMems around
4903 
4904   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4905   if (alloc == NULL)  return NULL;
4906 
4907   Node* rawmem = memory(Compile::AliasIdxRaw);
4908   // Is the allocation's memory state untouched?
4909   if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4910     // Bail out if there have been raw-memory effects since the allocation.
4911     // (Example:  There might have been a call or safepoint.)
4912     return NULL;
4913   }
4914   rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4915   if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4916     return NULL;
4917   }
4918 
4919   // There must be no unexpected observers of this allocation.
4920   for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4921     Node* obs = ptr->fast_out(i);
4922     if (obs != this->map()) {
4923       return NULL;
4924     }
4925   }
4926 
4927   // This arraycopy must unconditionally follow the allocation of the ptr.
4928   Node* alloc_ctl = ptr->in(0);
4929   assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4930 
4931   Node* ctl = control();
4932   while (ctl != alloc_ctl) {
4933     // There may be guards which feed into the slow_region.
4934     // Any other control flow means that we might not get a chance
4935     // to finish initializing the allocated object.
4936     if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4937       IfNode* iff = ctl->in(0)->as_If();
4938       Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4939       assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4940       if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4941         ctl = iff->in(0);       // This test feeds the known slow_region.
4942         continue;
4943       }
4944       // One more try:  Various low-level checks bottom out in
4945       // uncommon traps.  If the debug-info of the trap omits
4946       // any reference to the allocation, as we've already
4947       // observed, then there can be no objection to the trap.
4948       bool found_trap = false;
4949       for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4950         Node* obs = not_ctl->fast_out(j);
4951         if (obs->in(0) == not_ctl && obs->is_Call() &&
4952             (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4953           found_trap = true; break;
4954         }
4955       }
4956       if (found_trap) {
4957         ctl = iff->in(0);       // This test feeds a harmless uncommon trap.
4958         continue;
4959       }
4960     }
4961     return NULL;
4962   }
4963 
4964   // If we get this far, we have an allocation which immediately
4965   // precedes the arraycopy, and we can take over zeroing the new object.
4966   // The arraycopy will finish the initialization, and provide
4967   // a new control state to which we will anchor the destination pointer.
4968 
4969   return alloc;
4970 }
4971 
4972 //-------------inline_encodeISOArray-----------------------------------
4973 // encode char[] to byte[] in ISO_8859_1
4974 bool LibraryCallKit::inline_encodeISOArray() {
4975   assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4976   // no receiver since it is static method
4977   Node *src         = argument(0);
4978   Node *src_offset  = argument(1);
4979   Node *dst         = argument(2);
4980   Node *dst_offset  = argument(3);
4981   Node *length      = argument(4);
4982 
4983   const Type* src_type = src->Value(&_gvn);
4984   const Type* dst_type = dst->Value(&_gvn);
4985   const TypeAryPtr* top_src = src_type->isa_aryptr();
4986   const TypeAryPtr* top_dest = dst_type->isa_aryptr();
4987   if (top_src  == NULL || top_src->klass()  == NULL ||
4988       top_dest == NULL || top_dest->klass() == NULL) {
4989     // failed array check
4990     return false;
4991   }
4992 
4993   // Figure out the size and type of the elements we will be copying.
4994   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4995   BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
4996   if (src_elem != T_CHAR || dst_elem != T_BYTE) {
4997     return false;
4998   }
4999   Node* src_start = array_element_address(src, src_offset, src_elem);
5000   Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5001   // 'src_start' points to src array + scaled offset
5002   // 'dst_start' points to dst array + scaled offset
5003 
5004   const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5005   Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5006   enc = _gvn.transform(enc);
5007   Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5008   set_memory(res_mem, mtype);
5009   set_result(enc);
5010   return true;
5011 }
5012 
5013 //-------------inline_multiplyToLen-----------------------------------
5014 bool LibraryCallKit::inline_multiplyToLen() {
5015   assert(UseMultiplyToLenIntrinsic, "not implementated on this platform");
5016 
5017   address stubAddr = StubRoutines::multiplyToLen();
5018   if (stubAddr == NULL) {
5019     return false; // Intrinsic's stub is not implemented on this platform
5020   }
5021   const char* stubName = "multiplyToLen";
5022 
5023   assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5024 
5025   Node* x    = argument(1);
5026   Node* xlen = argument(2);
5027   Node* y    = argument(3);
5028   Node* ylen = argument(4);
5029   Node* z    = argument(5);
5030 
5031   const Type* x_type = x->Value(&_gvn);
5032   const Type* y_type = y->Value(&_gvn);
5033   const TypeAryPtr* top_x = x_type->isa_aryptr();
5034   const TypeAryPtr* top_y = y_type->isa_aryptr();
5035   if (top_x  == NULL || top_x->klass()  == NULL ||
5036       top_y == NULL || top_y->klass() == NULL) {
5037     // failed array check
5038     return false;
5039   }
5040 
5041   BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5042   BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5043   if (x_elem != T_INT || y_elem != T_INT) {
5044     return false;
5045   }
5046 
5047   // Set the original stack and the reexecute bit for the interpreter to reexecute
5048   // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5049   // on the return from z array allocation in runtime.
5050   { PreserveReexecuteState preexecs(this);
5051     jvms()->set_should_reexecute(true);
5052 
5053     Node* x_start = array_element_address(x, intcon(0), x_elem);
5054     Node* y_start = array_element_address(y, intcon(0), y_elem);
5055     // 'x_start' points to x array + scaled xlen
5056     // 'y_start' points to y array + scaled ylen
5057 
5058     // Allocate the result array
5059     Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5060     ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5061     Node* klass_node = makecon(TypeKlassPtr::make(klass));
5062 
5063     IdealKit ideal(this);
5064 
5065 #define __ ideal.
5066      Node* one = __ ConI(1);
5067      Node* zero = __ ConI(0);
5068      IdealVariable need_alloc(ideal), z_alloc(ideal);  __ declarations_done();
5069      __ set(need_alloc, zero);
5070      __ set(z_alloc, z);
5071      __ if_then(z, BoolTest::eq, null()); {
5072        __ increment (need_alloc, one);
5073      } __ else_(); {
5074        // Update graphKit memory and control from IdealKit.
5075        sync_kit(ideal);
5076        Node* zlen_arg = load_array_length(z);
5077        // Update IdealKit memory and control from graphKit.
5078        __ sync_kit(this);
5079        __ if_then(zlen_arg, BoolTest::lt, zlen); {
5080          __ increment (need_alloc, one);
5081        } __ end_if();
5082      } __ end_if();
5083 
5084      __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5085        // Update graphKit memory and control from IdealKit.
5086        sync_kit(ideal);
5087        Node * narr = new_array(klass_node, zlen, 1);
5088        // Update IdealKit memory and control from graphKit.
5089        __ sync_kit(this);
5090        __ set(z_alloc, narr);
5091      } __ end_if();
5092 
5093      sync_kit(ideal);
5094      z = __ value(z_alloc);
5095      // Can't use TypeAryPtr::INTS which uses Bottom offset.
5096      _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5097      // Final sync IdealKit and GraphKit.
5098      final_sync(ideal);
5099 #undef __
5100 
5101     Node* z_start = array_element_address(z, intcon(0), T_INT);
5102 
5103     Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5104                                    OptoRuntime::multiplyToLen_Type(),
5105                                    stubAddr, stubName, TypePtr::BOTTOM,
5106                                    x_start, xlen, y_start, ylen, z_start, zlen);
5107   } // original reexecute is set back here
5108 
5109   C->set_has_split_ifs(true); // Has chance for split-if optimization
5110   set_result(z);
5111   return true;
5112 }
5113 
5114 
5115 /**
5116  * Calculate CRC32 for byte.
5117  * int java.util.zip.CRC32.update(int crc, int b)
5118  */
5119 bool LibraryCallKit::inline_updateCRC32() {
5120   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5121   assert(callee()->signature()->size() == 2, "update has 2 parameters");
5122   // no receiver since it is static method
5123   Node* crc  = argument(0); // type: int
5124   Node* b    = argument(1); // type: int
5125 
5126   /*
5127    *    int c = ~ crc;
5128    *    b = timesXtoThe32[(b ^ c) & 0xFF];
5129    *    b = b ^ (c >>> 8);
5130    *    crc = ~b;
5131    */
5132 
5133   Node* M1 = intcon(-1);
5134   crc = _gvn.transform(new XorINode(crc, M1));
5135   Node* result = _gvn.transform(new XorINode(crc, b));
5136   result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5137 
5138   Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5139   Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5140   Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5141   result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5142 
5143   crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5144   result = _gvn.transform(new XorINode(crc, result));
5145   result = _gvn.transform(new XorINode(result, M1));
5146   set_result(result);
5147   return true;
5148 }
5149 
5150 /**
5151  * Calculate CRC32 for byte[] array.
5152  * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5153  */
5154 bool LibraryCallKit::inline_updateBytesCRC32() {
5155   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5156   assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5157   // no receiver since it is static method
5158   Node* crc     = argument(0); // type: int
5159   Node* src     = argument(1); // type: oop
5160   Node* offset  = argument(2); // type: int
5161   Node* length  = argument(3); // type: int
5162 
5163   const Type* src_type = src->Value(&_gvn);
5164   const TypeAryPtr* top_src = src_type->isa_aryptr();
5165   if (top_src  == NULL || top_src->klass()  == NULL) {
5166     // failed array check
5167     return false;
5168   }
5169 
5170   // Figure out the size and type of the elements we will be copying.
5171   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5172   if (src_elem != T_BYTE) {
5173     return false;
5174   }
5175 
5176   // 'src_start' points to src array + scaled offset
5177   Node* src_start = array_element_address(src, offset, src_elem);
5178 
5179   // We assume that range check is done by caller.
5180   // TODO: generate range check (offset+length < src.length) in debug VM.
5181 
5182   // Call the stub.
5183   address stubAddr = StubRoutines::updateBytesCRC32();
5184   const char *stubName = "updateBytesCRC32";
5185 
5186   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5187                                  stubAddr, stubName, TypePtr::BOTTOM,
5188                                  crc, src_start, length);
5189   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5190   set_result(result);
5191   return true;
5192 }
5193 
5194 /**
5195  * Calculate CRC32 for ByteBuffer.
5196  * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5197  */
5198 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5199   assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5200   assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5201   // no receiver since it is static method
5202   Node* crc     = argument(0); // type: int
5203   Node* src     = argument(1); // type: long
5204   Node* offset  = argument(3); // type: int
5205   Node* length  = argument(4); // type: int
5206 
5207   src = ConvL2X(src);  // adjust Java long to machine word
5208   Node* base = _gvn.transform(new CastX2PNode(src));
5209   offset = ConvI2X(offset);
5210 
5211   // 'src_start' points to src array + scaled offset
5212   Node* src_start = basic_plus_adr(top(), base, offset);
5213 
5214   // Call the stub.
5215   address stubAddr = StubRoutines::updateBytesCRC32();
5216   const char *stubName = "updateBytesCRC32";
5217 
5218   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5219                                  stubAddr, stubName, TypePtr::BOTTOM,
5220                                  crc, src_start, length);
5221   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5222   set_result(result);
5223   return true;
5224 }
5225 
5226 //----------------------------inline_reference_get----------------------------
5227 // public T java.lang.ref.Reference.get();
5228 bool LibraryCallKit::inline_reference_get() {
5229   const int referent_offset = java_lang_ref_Reference::referent_offset;
5230   guarantee(referent_offset > 0, "should have already been set");
5231 
5232   // Get the argument:
5233   Node* reference_obj = null_check_receiver();
5234   if (stopped()) return true;
5235 
5236   Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5237 
5238   ciInstanceKlass* klass = env()->Object_klass();
5239   const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5240 
5241   Node* no_ctrl = NULL;
5242   Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5243 
5244   // Use the pre-barrier to record the value in the referent field
5245   pre_barrier(false /* do_load */,
5246               control(),
5247               NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5248               result /* pre_val */,
5249               T_OBJECT);
5250 
5251   // Add memory barrier to prevent commoning reads from this field
5252   // across safepoint since GC can change its value.
5253   insert_mem_bar(Op_MemBarCPUOrder);
5254 
5255   set_result(result);
5256   return true;
5257 }
5258 
5259 
5260 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5261                                               bool is_exact=true, bool is_static=false) {
5262 
5263   const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5264   assert(tinst != NULL, "obj is null");
5265   assert(tinst->klass()->is_loaded(), "obj is not loaded");
5266   assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5267 
5268   ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5269                                                                           ciSymbol::make(fieldTypeString),
5270                                                                           is_static);
5271   if (field == NULL) return (Node *) NULL;
5272   assert (field != NULL, "undefined field");
5273 
5274   // Next code  copied from Parse::do_get_xxx():
5275 
5276   // Compute address and memory type.
5277   int offset  = field->offset_in_bytes();
5278   bool is_vol = field->is_volatile();
5279   ciType* field_klass = field->type();
5280   assert(field_klass->is_loaded(), "should be loaded");
5281   const TypePtr* adr_type = C->alias_type(field)->adr_type();
5282   Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5283   BasicType bt = field->layout_type();
5284 
5285   // Build the resultant type of the load
5286   const Type *type;
5287   if (bt == T_OBJECT) {
5288     type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5289   } else {
5290     type = Type::get_const_basic_type(bt);
5291   }
5292 
5293   if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5294     insert_mem_bar(Op_MemBarVolatile);   // StoreLoad barrier
5295   }
5296   // Build the load.
5297   MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5298   Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, is_vol);
5299   // If reference is volatile, prevent following memory ops from
5300   // floating up past the volatile read.  Also prevents commoning
5301   // another volatile read.
5302   if (is_vol) {
5303     // Memory barrier includes bogus read of value to force load BEFORE membar
5304     insert_mem_bar(Op_MemBarAcquire, loadedField);
5305   }
5306   return loadedField;
5307 }
5308 
5309 
5310 //------------------------------inline_aescrypt_Block-----------------------
5311 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5312   address stubAddr;
5313   const char *stubName;
5314   assert(UseAES, "need AES instruction support");
5315 
5316   switch(id) {
5317   case vmIntrinsics::_aescrypt_encryptBlock:
5318     stubAddr = StubRoutines::aescrypt_encryptBlock();
5319     stubName = "aescrypt_encryptBlock";
5320     break;
5321   case vmIntrinsics::_aescrypt_decryptBlock:
5322     stubAddr = StubRoutines::aescrypt_decryptBlock();
5323     stubName = "aescrypt_decryptBlock";
5324     break;
5325   }
5326   if (stubAddr == NULL) return false;
5327 
5328   Node* aescrypt_object = argument(0);
5329   Node* src             = argument(1);
5330   Node* src_offset      = argument(2);
5331   Node* dest            = argument(3);
5332   Node* dest_offset     = argument(4);
5333 
5334   // (1) src and dest are arrays.
5335   const Type* src_type = src->Value(&_gvn);
5336   const Type* dest_type = dest->Value(&_gvn);
5337   const TypeAryPtr* top_src = src_type->isa_aryptr();
5338   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5339   assert (top_src  != NULL && top_src->klass()  != NULL &&  top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5340 
5341   // for the quick and dirty code we will skip all the checks.
5342   // we are just trying to get the call to be generated.
5343   Node* src_start  = src;
5344   Node* dest_start = dest;
5345   if (src_offset != NULL || dest_offset != NULL) {
5346     assert(src_offset != NULL && dest_offset != NULL, "");
5347     src_start  = array_element_address(src,  src_offset,  T_BYTE);
5348     dest_start = array_element_address(dest, dest_offset, T_BYTE);
5349   }
5350 
5351   // now need to get the start of its expanded key array
5352   // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5353   Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5354   if (k_start == NULL) return false;
5355 
5356   if (Matcher::pass_original_key_for_aes()) {
5357     // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5358     // compatibility issues between Java key expansion and SPARC crypto instructions
5359     Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5360     if (original_k_start == NULL) return false;
5361 
5362     // Call the stub.
5363     make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5364                       stubAddr, stubName, TypePtr::BOTTOM,
5365                       src_start, dest_start, k_start, original_k_start);
5366   } else {
5367     // Call the stub.
5368     make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5369                       stubAddr, stubName, TypePtr::BOTTOM,
5370                       src_start, dest_start, k_start);
5371   }
5372 
5373   return true;
5374 }
5375 
5376 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5377 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5378   address stubAddr;
5379   const char *stubName;
5380 
5381   assert(UseAES, "need AES instruction support");
5382 
5383   switch(id) {
5384   case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5385     stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5386     stubName = "cipherBlockChaining_encryptAESCrypt";
5387     break;
5388   case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5389     stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5390     stubName = "cipherBlockChaining_decryptAESCrypt";
5391     break;
5392   }
5393   if (stubAddr == NULL) return false;
5394 
5395   Node* cipherBlockChaining_object = argument(0);
5396   Node* src                        = argument(1);
5397   Node* src_offset                 = argument(2);
5398   Node* len                        = argument(3);
5399   Node* dest                       = argument(4);
5400   Node* dest_offset                = argument(5);
5401 
5402   // (1) src and dest are arrays.
5403   const Type* src_type = src->Value(&_gvn);
5404   const Type* dest_type = dest->Value(&_gvn);
5405   const TypeAryPtr* top_src = src_type->isa_aryptr();
5406   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5407   assert (top_src  != NULL && top_src->klass()  != NULL
5408           &&  top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5409 
5410   // checks are the responsibility of the caller
5411   Node* src_start  = src;
5412   Node* dest_start = dest;
5413   if (src_offset != NULL || dest_offset != NULL) {
5414     assert(src_offset != NULL && dest_offset != NULL, "");
5415     src_start  = array_element_address(src,  src_offset,  T_BYTE);
5416     dest_start = array_element_address(dest, dest_offset, T_BYTE);
5417   }
5418 
5419   // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5420   // (because of the predicated logic executed earlier).
5421   // so we cast it here safely.
5422   // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5423 
5424   Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5425   if (embeddedCipherObj == NULL) return false;
5426 
5427   // cast it to what we know it will be at runtime
5428   const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5429   assert(tinst != NULL, "CBC obj is null");
5430   assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5431   ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5432   assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5433 
5434   ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5435   const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5436   const TypeOopPtr* xtype = aklass->as_instance_type();
5437   Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5438   aescrypt_object = _gvn.transform(aescrypt_object);
5439 
5440   // we need to get the start of the aescrypt_object's expanded key array
5441   Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5442   if (k_start == NULL) return false;
5443 
5444   // similarly, get the start address of the r vector
5445   Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5446   if (objRvec == NULL) return false;
5447   Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5448 
5449   Node* cbcCrypt;
5450   if (Matcher::pass_original_key_for_aes()) {
5451     // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5452     // compatibility issues between Java key expansion and SPARC crypto instructions
5453     Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5454     if (original_k_start == NULL) return false;
5455 
5456     // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5457     cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5458                                  OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5459                                  stubAddr, stubName, TypePtr::BOTTOM,
5460                                  src_start, dest_start, k_start, r_start, len, original_k_start);
5461   } else {
5462     // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5463     cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5464                                  OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5465                                  stubAddr, stubName, TypePtr::BOTTOM,
5466                                  src_start, dest_start, k_start, r_start, len);
5467   }
5468 
5469   // return cipher length (int)
5470   Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5471   set_result(retvalue);
5472   return true;
5473 }
5474 
5475 //------------------------------get_key_start_from_aescrypt_object-----------------------
5476 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5477   Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5478   assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5479   if (objAESCryptKey == NULL) return (Node *) NULL;
5480 
5481   // now have the array, need to get the start address of the K array
5482   Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5483   return k_start;
5484 }
5485 
5486 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5487 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5488   Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5489   assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5490   if (objAESCryptKey == NULL) return (Node *) NULL;
5491 
5492   // now have the array, need to get the start address of the lastKey array
5493   Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5494   return original_k_start;
5495 }
5496 
5497 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5498 // Return node representing slow path of predicate check.
5499 // the pseudo code we want to emulate with this predicate is:
5500 // for encryption:
5501 //    if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5502 // for decryption:
5503 //    if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5504 //    note cipher==plain is more conservative than the original java code but that's OK
5505 //
5506 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5507   // The receiver was checked for NULL already.
5508   Node* objCBC = argument(0);
5509 
5510   // Load embeddedCipher field of CipherBlockChaining object.
5511   Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5512 
5513   // get AESCrypt klass for instanceOf check
5514   // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5515   // will have same classloader as CipherBlockChaining object
5516   const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5517   assert(tinst != NULL, "CBCobj is null");
5518   assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5519 
5520   // we want to do an instanceof comparison against the AESCrypt class
5521   ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5522   if (!klass_AESCrypt->is_loaded()) {
5523     // if AESCrypt is not even loaded, we never take the intrinsic fast path
5524     Node* ctrl = control();
5525     set_control(top()); // no regular fast path
5526     return ctrl;
5527   }
5528   ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5529 
5530   Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5531   Node* cmp_instof  = _gvn.transform(new CmpINode(instof, intcon(1)));
5532   Node* bool_instof  = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5533 
5534   Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5535 
5536   // for encryption, we are done
5537   if (!decrypting)
5538     return instof_false;  // even if it is NULL
5539 
5540   // for decryption, we need to add a further check to avoid
5541   // taking the intrinsic path when cipher and plain are the same
5542   // see the original java code for why.
5543   RegionNode* region = new RegionNode(3);
5544   region->init_req(1, instof_false);
5545   Node* src = argument(1);
5546   Node* dest = argument(4);
5547   Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5548   Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5549   Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5550   region->init_req(2, src_dest_conjoint);
5551 
5552   record_for_igvn(region);
5553   return _gvn.transform(region);
5554 }
5555 
5556 //------------------------------inline_sha_implCompress-----------------------
5557 //
5558 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
5559 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
5560 //
5561 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
5562 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
5563 //
5564 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
5565 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
5566 //
5567 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
5568   assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
5569 
5570   Node* sha_obj = argument(0);
5571   Node* src     = argument(1); // type oop
5572   Node* ofs     = argument(2); // type int
5573 
5574   const Type* src_type = src->Value(&_gvn);
5575   const TypeAryPtr* top_src = src_type->isa_aryptr();
5576   if (top_src  == NULL || top_src->klass()  == NULL) {
5577     // failed array check
5578     return false;
5579   }
5580   // Figure out the size and type of the elements we will be copying.
5581   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5582   if (src_elem != T_BYTE) {
5583     return false;
5584   }
5585   // 'src_start' points to src array + offset
5586   Node* src_start = array_element_address(src, ofs, src_elem);
5587   Node* state = NULL;
5588   address stubAddr;
5589   const char *stubName;
5590 
5591   switch(id) {
5592   case vmIntrinsics::_sha_implCompress:
5593     assert(UseSHA1Intrinsics, "need SHA1 instruction support");
5594     state = get_state_from_sha_object(sha_obj);
5595     stubAddr = StubRoutines::sha1_implCompress();
5596     stubName = "sha1_implCompress";
5597     break;
5598   case vmIntrinsics::_sha2_implCompress:
5599     assert(UseSHA256Intrinsics, "need SHA256 instruction support");
5600     state = get_state_from_sha_object(sha_obj);
5601     stubAddr = StubRoutines::sha256_implCompress();
5602     stubName = "sha256_implCompress";
5603     break;
5604   case vmIntrinsics::_sha5_implCompress:
5605     assert(UseSHA512Intrinsics, "need SHA512 instruction support");
5606     state = get_state_from_sha5_object(sha_obj);
5607     stubAddr = StubRoutines::sha512_implCompress();
5608     stubName = "sha512_implCompress";
5609     break;
5610   default:
5611     fatal_unexpected_iid(id);
5612     return false;
5613   }
5614   if (state == NULL) return false;
5615 
5616   // Call the stub.
5617   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
5618                                  stubAddr, stubName, TypePtr::BOTTOM,
5619                                  src_start, state);
5620 
5621   return true;
5622 }
5623 
5624 //------------------------------inline_digestBase_implCompressMB-----------------------
5625 //
5626 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
5627 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
5628 //
5629 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
5630   assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
5631          "need SHA1/SHA256/SHA512 instruction support");
5632   assert((uint)predicate < 3, "sanity");
5633   assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
5634 
5635   Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
5636   Node* src            = argument(1); // byte[] array
5637   Node* ofs            = argument(2); // type int
5638   Node* limit          = argument(3); // type int
5639 
5640   const Type* src_type = src->Value(&_gvn);
5641   const TypeAryPtr* top_src = src_type->isa_aryptr();
5642   if (top_src  == NULL || top_src->klass()  == NULL) {
5643     // failed array check
5644     return false;
5645   }
5646   // Figure out the size and type of the elements we will be copying.
5647   BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5648   if (src_elem != T_BYTE) {
5649     return false;
5650   }
5651   // 'src_start' points to src array + offset
5652   Node* src_start = array_element_address(src, ofs, src_elem);
5653 
5654   const char* klass_SHA_name = NULL;
5655   const char* stub_name = NULL;
5656   address     stub_addr = NULL;
5657   bool        long_state = false;
5658 
5659   switch (predicate) {
5660   case 0:
5661     if (UseSHA1Intrinsics) {
5662       klass_SHA_name = "sun/security/provider/SHA";
5663       stub_name = "sha1_implCompressMB";
5664       stub_addr = StubRoutines::sha1_implCompressMB();
5665     }
5666     break;
5667   case 1:
5668     if (UseSHA256Intrinsics) {
5669       klass_SHA_name = "sun/security/provider/SHA2";
5670       stub_name = "sha256_implCompressMB";
5671       stub_addr = StubRoutines::sha256_implCompressMB();
5672     }
5673     break;
5674   case 2:
5675     if (UseSHA512Intrinsics) {
5676       klass_SHA_name = "sun/security/provider/SHA5";
5677       stub_name = "sha512_implCompressMB";
5678       stub_addr = StubRoutines::sha512_implCompressMB();
5679       long_state = true;
5680     }
5681     break;
5682   default:
5683     fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
5684   }
5685   if (klass_SHA_name != NULL) {
5686     // get DigestBase klass to lookup for SHA klass
5687     const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
5688     assert(tinst != NULL, "digestBase_obj is not instance???");
5689     assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
5690 
5691     ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
5692     assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
5693     ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
5694     return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
5695   }
5696   return false;
5697 }
5698 //------------------------------inline_sha_implCompressMB-----------------------
5699 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
5700                                                bool long_state, address stubAddr, const char *stubName,
5701                                                Node* src_start, Node* ofs, Node* limit) {
5702   const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
5703   const TypeOopPtr* xtype = aklass->as_instance_type();
5704   Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
5705   sha_obj = _gvn.transform(sha_obj);
5706 
5707   Node* state;
5708   if (long_state) {
5709     state = get_state_from_sha5_object(sha_obj);
5710   } else {
5711     state = get_state_from_sha_object(sha_obj);
5712   }
5713   if (state == NULL) return false;
5714 
5715   // Call the stub.
5716   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5717                                  OptoRuntime::digestBase_implCompressMB_Type(),
5718                                  stubAddr, stubName, TypePtr::BOTTOM,
5719                                  src_start, state, ofs, limit);
5720   // return ofs (int)
5721   Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5722   set_result(result);
5723 
5724   return true;
5725 }
5726 
5727 //------------------------------get_state_from_sha_object-----------------------
5728 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
5729   Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
5730   assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
5731   if (sha_state == NULL) return (Node *) NULL;
5732 
5733   // now have the array, need to get the start address of the state array
5734   Node* state = array_element_address(sha_state, intcon(0), T_INT);
5735   return state;
5736 }
5737 
5738 //------------------------------get_state_from_sha5_object-----------------------
5739 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
5740   Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
5741   assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
5742   if (sha_state == NULL) return (Node *) NULL;
5743 
5744   // now have the array, need to get the start address of the state array
5745   Node* state = array_element_address(sha_state, intcon(0), T_LONG);
5746   return state;
5747 }
5748 
5749 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
5750 // Return node representing slow path of predicate check.
5751 // the pseudo code we want to emulate with this predicate is:
5752 //    if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
5753 //
5754 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
5755   assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
5756          "need SHA1/SHA256/SHA512 instruction support");
5757   assert((uint)predicate < 3, "sanity");
5758 
5759   // The receiver was checked for NULL already.
5760   Node* digestBaseObj = argument(0);
5761 
5762   // get DigestBase klass for instanceOf check
5763   const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
5764   assert(tinst != NULL, "digestBaseObj is null");
5765   assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
5766 
5767   const char* klass_SHA_name = NULL;
5768   switch (predicate) {
5769   case 0:
5770     if (UseSHA1Intrinsics) {
5771       // we want to do an instanceof comparison against the SHA class
5772       klass_SHA_name = "sun/security/provider/SHA";
5773     }
5774     break;
5775   case 1:
5776     if (UseSHA256Intrinsics) {
5777       // we want to do an instanceof comparison against the SHA2 class
5778       klass_SHA_name = "sun/security/provider/SHA2";
5779     }
5780     break;
5781   case 2:
5782     if (UseSHA512Intrinsics) {
5783       // we want to do an instanceof comparison against the SHA5 class
5784       klass_SHA_name = "sun/security/provider/SHA5";
5785     }
5786     break;
5787   default:
5788     fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
5789   }
5790 
5791   ciKlass* klass_SHA = NULL;
5792   if (klass_SHA_name != NULL) {
5793     klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
5794   }
5795   if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
5796     // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
5797     Node* ctrl = control();
5798     set_control(top()); // no intrinsic path
5799     return ctrl;
5800   }
5801   ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
5802 
5803   Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
5804   Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
5805   Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5806   Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5807 
5808   return instof_false;  // even if it is NULL
5809 }
--- EOF ---