1 /*
   2  * Copyright (c) 1999, 2011, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/systemDictionary.hpp"
  27 #include "classfile/vmSymbols.hpp"
  28 #include "compiler/compileBroker.hpp"
  29 #include "compiler/compileLog.hpp"
  30 #include "oops/objArrayKlass.hpp"
  31 #include "opto/addnode.hpp"
  32 #include "opto/callGenerator.hpp"
  33 #include "opto/cfgnode.hpp"
  34 #include "opto/idealKit.hpp"
  35 #include "opto/mulnode.hpp"
  36 #include "opto/parse.hpp"
  37 #include "opto/runtime.hpp"
  38 #include "opto/subnode.hpp"
  39 #include "prims/nativeLookup.hpp"
  40 #include "runtime/sharedRuntime.hpp"
  41 
  42 class LibraryIntrinsic : public InlineCallGenerator {
  43   // Extend the set of intrinsics known to the runtime:
  44  public:
  45  private:
  46   bool             _is_virtual;
  47   vmIntrinsics::ID _intrinsic_id;
  48 
  49  public:
  50   LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
  51     : InlineCallGenerator(m),
  52       _is_virtual(is_virtual),
  53       _intrinsic_id(id)
  54   {
  55   }
  56   virtual bool is_intrinsic() const { return true; }
  57   virtual bool is_virtual()   const { return _is_virtual; }
  58   virtual JVMState* generate(JVMState* jvms);
  59   vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
  60 };
  61 
  62 
  63 // Local helper class for LibraryIntrinsic:
  64 class LibraryCallKit : public GraphKit {
  65  private:
  66   LibraryIntrinsic* _intrinsic;   // the library intrinsic being called
  67 
  68  public:
  69   LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
  70     : GraphKit(caller),
  71       _intrinsic(intrinsic)
  72   {
  73   }
  74 
  75   ciMethod*         caller()    const    { return jvms()->method(); }
  76   int               bci()       const    { return jvms()->bci(); }
  77   LibraryIntrinsic* intrinsic() const    { return _intrinsic; }
  78   vmIntrinsics::ID  intrinsic_id() const { return _intrinsic->intrinsic_id(); }
  79   ciMethod*         callee()    const    { return _intrinsic->method(); }
  80   ciSignature*      signature() const    { return callee()->signature(); }
  81   int               arg_size()  const    { return callee()->arg_size(); }
  82 
  83   bool try_to_inline();
  84 
  85   // Helper functions to inline natives
  86   void push_result(RegionNode* region, PhiNode* value);
  87   Node* generate_guard(Node* test, RegionNode* region, float true_prob);
  88   Node* generate_slow_guard(Node* test, RegionNode* region);
  89   Node* generate_fair_guard(Node* test, RegionNode* region);
  90   Node* generate_negative_guard(Node* index, RegionNode* region,
  91                                 // resulting CastII of index:
  92                                 Node* *pos_index = NULL);
  93   Node* generate_nonpositive_guard(Node* index, bool never_negative,
  94                                    // resulting CastII of index:
  95                                    Node* *pos_index = NULL);
  96   Node* generate_limit_guard(Node* offset, Node* subseq_length,
  97                              Node* array_length,
  98                              RegionNode* region);
  99   Node* generate_current_thread(Node* &tls_output);
 100   address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
 101                               bool disjoint_bases, const char* &name, bool dest_uninitialized);
 102   Node* load_mirror_from_klass(Node* klass);
 103   Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
 104                                       int nargs,
 105                                       RegionNode* region, int null_path,
 106                                       int offset);
 107   Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
 108                                RegionNode* region, int null_path) {
 109     int offset = java_lang_Class::klass_offset_in_bytes();
 110     return load_klass_from_mirror_common(mirror, never_see_null, nargs,
 111                                          region, null_path,
 112                                          offset);
 113   }
 114   Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
 115                                      int nargs,
 116                                      RegionNode* region, int null_path) {
 117     int offset = java_lang_Class::array_klass_offset_in_bytes();
 118     return load_klass_from_mirror_common(mirror, never_see_null, nargs,
 119                                          region, null_path,
 120                                          offset);
 121   }
 122   Node* generate_access_flags_guard(Node* kls,
 123                                     int modifier_mask, int modifier_bits,
 124                                     RegionNode* region);
 125   Node* generate_interface_guard(Node* kls, RegionNode* region);
 126   Node* generate_array_guard(Node* kls, RegionNode* region) {
 127     return generate_array_guard_common(kls, region, false, false);
 128   }
 129   Node* generate_non_array_guard(Node* kls, RegionNode* region) {
 130     return generate_array_guard_common(kls, region, false, true);
 131   }
 132   Node* generate_objArray_guard(Node* kls, RegionNode* region) {
 133     return generate_array_guard_common(kls, region, true, false);
 134   }
 135   Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
 136     return generate_array_guard_common(kls, region, true, true);
 137   }
 138   Node* generate_array_guard_common(Node* kls, RegionNode* region,
 139                                     bool obj_array, bool not_array);
 140   Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
 141   CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
 142                                      bool is_virtual = false, bool is_static = false);
 143   CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
 144     return generate_method_call(method_id, false, true);
 145   }
 146   CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
 147     return generate_method_call(method_id, true, false);
 148   }
 149 
 150   Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2);
 151   bool inline_string_compareTo();
 152   bool inline_string_indexOf();
 153   Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
 154   bool inline_string_equals();
 155   Node* pop_math_arg();
 156   bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
 157   bool inline_math_native(vmIntrinsics::ID id);
 158   bool inline_trig(vmIntrinsics::ID id);
 159   bool inline_trans(vmIntrinsics::ID id);
 160   bool inline_abs(vmIntrinsics::ID id);
 161   bool inline_sqrt(vmIntrinsics::ID id);
 162   bool inline_pow(vmIntrinsics::ID id);
 163   bool inline_exp(vmIntrinsics::ID id);
 164   bool inline_min_max(vmIntrinsics::ID id);
 165   Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
 166   // This returns Type::AnyPtr, RawPtr, or OopPtr.
 167   int classify_unsafe_addr(Node* &base, Node* &offset);
 168   Node* make_unsafe_address(Node* base, Node* offset);
 169   bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
 170   bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
 171   bool inline_unsafe_allocate();
 172   bool inline_unsafe_copyMemory();
 173   bool inline_native_currentThread();
 174   bool inline_native_time_funcs(bool isNano);
 175   bool inline_native_isInterrupted();
 176   bool inline_native_Class_query(vmIntrinsics::ID id);
 177   bool inline_native_subtype_check();
 178 
 179   bool inline_native_newArray();
 180   bool inline_native_getLength();
 181   bool inline_array_copyOf(bool is_copyOfRange);
 182   bool inline_array_equals();
 183   void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
 184   bool inline_native_clone(bool is_virtual);
 185   bool inline_native_Reflection_getCallerClass();
 186   bool inline_native_AtomicLong_get();
 187   bool inline_native_AtomicLong_attemptUpdate();
 188   bool is_method_invoke_or_aux_frame(JVMState* jvms);
 189   // Helper function for inlining native object hash method
 190   bool inline_native_hashcode(bool is_virtual, bool is_static);
 191   bool inline_native_getClass();
 192 
 193   // Helper functions for inlining arraycopy
 194   bool inline_arraycopy();
 195   void generate_arraycopy(const TypePtr* adr_type,
 196                           BasicType basic_elem_type,
 197                           Node* src,  Node* src_offset,
 198                           Node* dest, Node* dest_offset,
 199                           Node* copy_length,
 200                           bool disjoint_bases = false,
 201                           bool length_never_negative = false,
 202                           RegionNode* slow_region = NULL);
 203   AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
 204                                                 RegionNode* slow_region);
 205   void generate_clear_array(const TypePtr* adr_type,
 206                             Node* dest,
 207                             BasicType basic_elem_type,
 208                             Node* slice_off,
 209                             Node* slice_len,
 210                             Node* slice_end);
 211   bool generate_block_arraycopy(const TypePtr* adr_type,
 212                                 BasicType basic_elem_type,
 213                                 AllocateNode* alloc,
 214                                 Node* src,  Node* src_offset,
 215                                 Node* dest, Node* dest_offset,
 216                                 Node* dest_size, bool dest_uninitialized);
 217   void generate_slow_arraycopy(const TypePtr* adr_type,
 218                                Node* src,  Node* src_offset,
 219                                Node* dest, Node* dest_offset,
 220                                Node* copy_length, bool dest_uninitialized);
 221   Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
 222                                      Node* dest_elem_klass,
 223                                      Node* src,  Node* src_offset,
 224                                      Node* dest, Node* dest_offset,
 225                                      Node* copy_length, bool dest_uninitialized);
 226   Node* generate_generic_arraycopy(const TypePtr* adr_type,
 227                                    Node* src,  Node* src_offset,
 228                                    Node* dest, Node* dest_offset,
 229                                    Node* copy_length, bool dest_uninitialized);
 230   void generate_unchecked_arraycopy(const TypePtr* adr_type,
 231                                     BasicType basic_elem_type,
 232                                     bool disjoint_bases,
 233                                     Node* src,  Node* src_offset,
 234                                     Node* dest, Node* dest_offset,
 235                                     Node* copy_length, bool dest_uninitialized);
 236   bool inline_unsafe_CAS(BasicType type);
 237   bool inline_unsafe_ordered_store(BasicType type);
 238   bool inline_fp_conversions(vmIntrinsics::ID id);
 239   bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
 240   bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
 241   bool inline_bitCount(vmIntrinsics::ID id);
 242   bool inline_reverseBytes(vmIntrinsics::ID id);
 243 };
 244 
 245 
 246 //---------------------------make_vm_intrinsic----------------------------
 247 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
 248   vmIntrinsics::ID id = m->intrinsic_id();
 249   assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
 250 
 251   if (DisableIntrinsic[0] != '\0'
 252       && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
 253     // disabled by a user request on the command line:
 254     // example: -XX:DisableIntrinsic=_hashCode,_getClass
 255     return NULL;
 256   }
 257 
 258   if (!m->is_loaded()) {
 259     // do not attempt to inline unloaded methods
 260     return NULL;
 261   }
 262 
 263   // Only a few intrinsics implement a virtual dispatch.
 264   // They are expensive calls which are also frequently overridden.
 265   if (is_virtual) {
 266     switch (id) {
 267     case vmIntrinsics::_hashCode:
 268     case vmIntrinsics::_clone:
 269       // OK, Object.hashCode and Object.clone intrinsics come in both flavors
 270       break;
 271     default:
 272       return NULL;
 273     }
 274   }
 275 
 276   // -XX:-InlineNatives disables nearly all intrinsics:
 277   if (!InlineNatives) {
 278     switch (id) {
 279     case vmIntrinsics::_indexOf:
 280     case vmIntrinsics::_compareTo:
 281     case vmIntrinsics::_equals:
 282     case vmIntrinsics::_equalsC:
 283       break;  // InlineNatives does not control String.compareTo
 284     default:
 285       return NULL;
 286     }
 287   }
 288 
 289   switch (id) {
 290   case vmIntrinsics::_compareTo:
 291     if (!SpecialStringCompareTo)  return NULL;
 292     break;
 293   case vmIntrinsics::_indexOf:
 294     if (!SpecialStringIndexOf)  return NULL;
 295     break;
 296   case vmIntrinsics::_equals:
 297     if (!SpecialStringEquals)  return NULL;
 298     break;
 299   case vmIntrinsics::_equalsC:
 300     if (!SpecialArraysEquals)  return NULL;
 301     break;
 302   case vmIntrinsics::_arraycopy:
 303     if (!InlineArrayCopy)  return NULL;
 304     break;
 305   case vmIntrinsics::_copyMemory:
 306     if (StubRoutines::unsafe_arraycopy() == NULL)  return NULL;
 307     if (!InlineArrayCopy)  return NULL;
 308     break;
 309   case vmIntrinsics::_hashCode:
 310     if (!InlineObjectHash)  return NULL;
 311     break;
 312   case vmIntrinsics::_clone:
 313   case vmIntrinsics::_copyOf:
 314   case vmIntrinsics::_copyOfRange:
 315     if (!InlineObjectCopy)  return NULL;
 316     // These also use the arraycopy intrinsic mechanism:
 317     if (!InlineArrayCopy)  return NULL;
 318     break;
 319   case vmIntrinsics::_checkIndex:
 320     // We do not intrinsify this.  The optimizer does fine with it.
 321     return NULL;
 322 
 323   case vmIntrinsics::_get_AtomicLong:
 324   case vmIntrinsics::_attemptUpdate:
 325     if (!InlineAtomicLong)  return NULL;
 326     break;
 327 
 328   case vmIntrinsics::_getCallerClass:
 329     if (!UseNewReflection)  return NULL;
 330     if (!InlineReflectionGetCallerClass)  return NULL;
 331     if (!JDK_Version::is_gte_jdk14x_version())  return NULL;
 332     break;
 333 
 334   case vmIntrinsics::_bitCount_i:
 335   case vmIntrinsics::_bitCount_l:
 336     if (!UsePopCountInstruction)  return NULL;
 337     break;
 338 
 339  default:
 340     assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
 341     assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
 342     break;
 343   }
 344 
 345   // -XX:-InlineClassNatives disables natives from the Class class.
 346   // The flag applies to all reflective calls, notably Array.newArray
 347   // (visible to Java programmers as Array.newInstance).
 348   if (m->holder()->name() == ciSymbol::java_lang_Class() ||
 349       m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
 350     if (!InlineClassNatives)  return NULL;
 351   }
 352 
 353   // -XX:-InlineThreadNatives disables natives from the Thread class.
 354   if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
 355     if (!InlineThreadNatives)  return NULL;
 356   }
 357 
 358   // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
 359   if (m->holder()->name() == ciSymbol::java_lang_Math() ||
 360       m->holder()->name() == ciSymbol::java_lang_Float() ||
 361       m->holder()->name() == ciSymbol::java_lang_Double()) {
 362     if (!InlineMathNatives)  return NULL;
 363   }
 364 
 365   // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
 366   if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
 367     if (!InlineUnsafeOps)  return NULL;
 368   }
 369 
 370   return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
 371 }
 372 
 373 //----------------------register_library_intrinsics-----------------------
 374 // Initialize this file's data structures, for each Compile instance.
 375 void Compile::register_library_intrinsics() {
 376   // Nothing to do here.
 377 }
 378 
 379 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
 380   LibraryCallKit kit(jvms, this);
 381   Compile* C = kit.C;
 382   int nodes = C->unique();
 383 #ifndef PRODUCT
 384   if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
 385     char buf[1000];
 386     const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
 387     tty->print_cr("Intrinsic %s", str);
 388   }
 389 #endif
 390   if (kit.try_to_inline()) {
 391     if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
 392       CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
 393     }
 394     C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
 395     if (C->log()) {
 396       C->log()->elem("intrinsic id='%s'%s nodes='%d'",
 397                      vmIntrinsics::name_at(intrinsic_id()),
 398                      (is_virtual() ? " virtual='1'" : ""),
 399                      C->unique() - nodes);
 400     }
 401     return kit.transfer_exceptions_into_jvms();
 402   }
 403 
 404   if (PrintIntrinsics) {
 405     tty->print("Did not inline intrinsic %s%s at bci:%d in",
 406                vmIntrinsics::name_at(intrinsic_id()),
 407                (is_virtual() ? " (virtual)" : ""), kit.bci());
 408     kit.caller()->print_short_name(tty);
 409     tty->print_cr(" (%d bytes)", kit.caller()->code_size());
 410   }
 411   C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
 412   return NULL;
 413 }
 414 
 415 bool LibraryCallKit::try_to_inline() {
 416   // Handle symbolic names for otherwise undistinguished boolean switches:
 417   const bool is_store       = true;
 418   const bool is_native_ptr  = true;
 419   const bool is_static      = true;
 420 
 421   switch (intrinsic_id()) {
 422   case vmIntrinsics::_hashCode:
 423     return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
 424   case vmIntrinsics::_identityHashCode:
 425     return inline_native_hashcode(/*!virtual*/ false, is_static);
 426   case vmIntrinsics::_getClass:
 427     return inline_native_getClass();
 428 
 429   case vmIntrinsics::_dsin:
 430   case vmIntrinsics::_dcos:
 431   case vmIntrinsics::_dtan:
 432   case vmIntrinsics::_dabs:
 433   case vmIntrinsics::_datan2:
 434   case vmIntrinsics::_dsqrt:
 435   case vmIntrinsics::_dexp:
 436   case vmIntrinsics::_dlog:
 437   case vmIntrinsics::_dlog10:
 438   case vmIntrinsics::_dpow:
 439     return inline_math_native(intrinsic_id());
 440 
 441   case vmIntrinsics::_min:
 442   case vmIntrinsics::_max:
 443     return inline_min_max(intrinsic_id());
 444 
 445   case vmIntrinsics::_arraycopy:
 446     return inline_arraycopy();
 447 
 448   case vmIntrinsics::_compareTo:
 449     return inline_string_compareTo();
 450   case vmIntrinsics::_indexOf:
 451     return inline_string_indexOf();
 452   case vmIntrinsics::_equals:
 453     return inline_string_equals();
 454 
 455   case vmIntrinsics::_getObject:
 456     return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
 457   case vmIntrinsics::_getBoolean:
 458     return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
 459   case vmIntrinsics::_getByte:
 460     return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
 461   case vmIntrinsics::_getShort:
 462     return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
 463   case vmIntrinsics::_getChar:
 464     return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
 465   case vmIntrinsics::_getInt:
 466     return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
 467   case vmIntrinsics::_getLong:
 468     return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
 469   case vmIntrinsics::_getFloat:
 470     return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
 471   case vmIntrinsics::_getDouble:
 472     return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
 473 
 474   case vmIntrinsics::_putObject:
 475     return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
 476   case vmIntrinsics::_putBoolean:
 477     return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
 478   case vmIntrinsics::_putByte:
 479     return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
 480   case vmIntrinsics::_putShort:
 481     return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
 482   case vmIntrinsics::_putChar:
 483     return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
 484   case vmIntrinsics::_putInt:
 485     return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
 486   case vmIntrinsics::_putLong:
 487     return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
 488   case vmIntrinsics::_putFloat:
 489     return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
 490   case vmIntrinsics::_putDouble:
 491     return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
 492 
 493   case vmIntrinsics::_getByte_raw:
 494     return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
 495   case vmIntrinsics::_getShort_raw:
 496     return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
 497   case vmIntrinsics::_getChar_raw:
 498     return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
 499   case vmIntrinsics::_getInt_raw:
 500     return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
 501   case vmIntrinsics::_getLong_raw:
 502     return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
 503   case vmIntrinsics::_getFloat_raw:
 504     return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
 505   case vmIntrinsics::_getDouble_raw:
 506     return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
 507   case vmIntrinsics::_getAddress_raw:
 508     return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
 509 
 510   case vmIntrinsics::_putByte_raw:
 511     return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
 512   case vmIntrinsics::_putShort_raw:
 513     return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
 514   case vmIntrinsics::_putChar_raw:
 515     return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
 516   case vmIntrinsics::_putInt_raw:
 517     return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
 518   case vmIntrinsics::_putLong_raw:
 519     return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
 520   case vmIntrinsics::_putFloat_raw:
 521     return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
 522   case vmIntrinsics::_putDouble_raw:
 523     return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
 524   case vmIntrinsics::_putAddress_raw:
 525     return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
 526 
 527   case vmIntrinsics::_getObjectVolatile:
 528     return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
 529   case vmIntrinsics::_getBooleanVolatile:
 530     return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
 531   case vmIntrinsics::_getByteVolatile:
 532     return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
 533   case vmIntrinsics::_getShortVolatile:
 534     return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
 535   case vmIntrinsics::_getCharVolatile:
 536     return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
 537   case vmIntrinsics::_getIntVolatile:
 538     return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
 539   case vmIntrinsics::_getLongVolatile:
 540     return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
 541   case vmIntrinsics::_getFloatVolatile:
 542     return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
 543   case vmIntrinsics::_getDoubleVolatile:
 544     return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
 545 
 546   case vmIntrinsics::_putObjectVolatile:
 547     return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
 548   case vmIntrinsics::_putBooleanVolatile:
 549     return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
 550   case vmIntrinsics::_putByteVolatile:
 551     return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
 552   case vmIntrinsics::_putShortVolatile:
 553     return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
 554   case vmIntrinsics::_putCharVolatile:
 555     return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
 556   case vmIntrinsics::_putIntVolatile:
 557     return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
 558   case vmIntrinsics::_putLongVolatile:
 559     return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
 560   case vmIntrinsics::_putFloatVolatile:
 561     return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
 562   case vmIntrinsics::_putDoubleVolatile:
 563     return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
 564 
 565   case vmIntrinsics::_prefetchRead:
 566     return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
 567   case vmIntrinsics::_prefetchWrite:
 568     return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
 569   case vmIntrinsics::_prefetchReadStatic:
 570     return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
 571   case vmIntrinsics::_prefetchWriteStatic:
 572     return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
 573 
 574   case vmIntrinsics::_compareAndSwapObject:
 575     return inline_unsafe_CAS(T_OBJECT);
 576   case vmIntrinsics::_compareAndSwapInt:
 577     return inline_unsafe_CAS(T_INT);
 578   case vmIntrinsics::_compareAndSwapLong:
 579     return inline_unsafe_CAS(T_LONG);
 580 
 581   case vmIntrinsics::_putOrderedObject:
 582     return inline_unsafe_ordered_store(T_OBJECT);
 583   case vmIntrinsics::_putOrderedInt:
 584     return inline_unsafe_ordered_store(T_INT);
 585   case vmIntrinsics::_putOrderedLong:
 586     return inline_unsafe_ordered_store(T_LONG);
 587 
 588   case vmIntrinsics::_currentThread:
 589     return inline_native_currentThread();
 590   case vmIntrinsics::_isInterrupted:
 591     return inline_native_isInterrupted();
 592 
 593   case vmIntrinsics::_currentTimeMillis:
 594     return inline_native_time_funcs(false);
 595   case vmIntrinsics::_nanoTime:
 596     return inline_native_time_funcs(true);
 597   case vmIntrinsics::_allocateInstance:
 598     return inline_unsafe_allocate();
 599   case vmIntrinsics::_copyMemory:
 600     return inline_unsafe_copyMemory();
 601   case vmIntrinsics::_newArray:
 602     return inline_native_newArray();
 603   case vmIntrinsics::_getLength:
 604     return inline_native_getLength();
 605   case vmIntrinsics::_copyOf:
 606     return inline_array_copyOf(false);
 607   case vmIntrinsics::_copyOfRange:
 608     return inline_array_copyOf(true);
 609   case vmIntrinsics::_equalsC:
 610     return inline_array_equals();
 611   case vmIntrinsics::_clone:
 612     return inline_native_clone(intrinsic()->is_virtual());
 613 
 614   case vmIntrinsics::_isAssignableFrom:
 615     return inline_native_subtype_check();
 616 
 617   case vmIntrinsics::_isInstance:
 618   case vmIntrinsics::_getModifiers:
 619   case vmIntrinsics::_isInterface:
 620   case vmIntrinsics::_isArray:
 621   case vmIntrinsics::_isPrimitive:
 622   case vmIntrinsics::_getSuperclass:
 623   case vmIntrinsics::_getComponentType:
 624   case vmIntrinsics::_getClassAccessFlags:
 625     return inline_native_Class_query(intrinsic_id());
 626 
 627   case vmIntrinsics::_floatToRawIntBits:
 628   case vmIntrinsics::_floatToIntBits:
 629   case vmIntrinsics::_intBitsToFloat:
 630   case vmIntrinsics::_doubleToRawLongBits:
 631   case vmIntrinsics::_doubleToLongBits:
 632   case vmIntrinsics::_longBitsToDouble:
 633     return inline_fp_conversions(intrinsic_id());
 634 
 635   case vmIntrinsics::_numberOfLeadingZeros_i:
 636   case vmIntrinsics::_numberOfLeadingZeros_l:
 637     return inline_numberOfLeadingZeros(intrinsic_id());
 638 
 639   case vmIntrinsics::_numberOfTrailingZeros_i:
 640   case vmIntrinsics::_numberOfTrailingZeros_l:
 641     return inline_numberOfTrailingZeros(intrinsic_id());
 642 
 643   case vmIntrinsics::_bitCount_i:
 644   case vmIntrinsics::_bitCount_l:
 645     return inline_bitCount(intrinsic_id());
 646 
 647   case vmIntrinsics::_reverseBytes_i:
 648   case vmIntrinsics::_reverseBytes_l:
 649   case vmIntrinsics::_reverseBytes_s:
 650   case vmIntrinsics::_reverseBytes_c:
 651     return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
 652 
 653   case vmIntrinsics::_get_AtomicLong:
 654     return inline_native_AtomicLong_get();
 655   case vmIntrinsics::_attemptUpdate:
 656     return inline_native_AtomicLong_attemptUpdate();
 657 
 658   case vmIntrinsics::_getCallerClass:
 659     return inline_native_Reflection_getCallerClass();
 660 
 661   default:
 662     // If you get here, it may be that someone has added a new intrinsic
 663     // to the list in vmSymbols.hpp without implementing it here.
 664 #ifndef PRODUCT
 665     if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
 666       tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
 667                     vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
 668     }
 669 #endif
 670     return false;
 671   }
 672 }
 673 
 674 //------------------------------push_result------------------------------
 675 // Helper function for finishing intrinsics.
 676 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
 677   record_for_igvn(region);
 678   set_control(_gvn.transform(region));
 679   BasicType value_type = value->type()->basic_type();
 680   push_node(value_type, _gvn.transform(value));
 681 }
 682 
 683 //------------------------------generate_guard---------------------------
 684 // Helper function for generating guarded fast-slow graph structures.
 685 // The given 'test', if true, guards a slow path.  If the test fails
 686 // then a fast path can be taken.  (We generally hope it fails.)
 687 // In all cases, GraphKit::control() is updated to the fast path.
 688 // The returned value represents the control for the slow path.
 689 // The return value is never 'top'; it is either a valid control
 690 // or NULL if it is obvious that the slow path can never be taken.
 691 // Also, if region and the slow control are not NULL, the slow edge
 692 // is appended to the region.
 693 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
 694   if (stopped()) {
 695     // Already short circuited.
 696     return NULL;
 697   }
 698 
 699   // Build an if node and its projections.
 700   // If test is true we take the slow path, which we assume is uncommon.
 701   if (_gvn.type(test) == TypeInt::ZERO) {
 702     // The slow branch is never taken.  No need to build this guard.
 703     return NULL;
 704   }
 705 
 706   IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
 707 
 708   Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
 709   if (if_slow == top()) {
 710     // The slow branch is never taken.  No need to build this guard.
 711     return NULL;
 712   }
 713 
 714   if (region != NULL)
 715     region->add_req(if_slow);
 716 
 717   Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
 718   set_control(if_fast);
 719 
 720   return if_slow;
 721 }
 722 
 723 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
 724   return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
 725 }
 726 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
 727   return generate_guard(test, region, PROB_FAIR);
 728 }
 729 
 730 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
 731                                                      Node* *pos_index) {
 732   if (stopped())
 733     return NULL;                // already stopped
 734   if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
 735     return NULL;                // index is already adequately typed
 736   Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
 737   Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
 738   Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
 739   if (is_neg != NULL && pos_index != NULL) {
 740     // Emulate effect of Parse::adjust_map_after_if.
 741     Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
 742     ccast->set_req(0, control());
 743     (*pos_index) = _gvn.transform(ccast);
 744   }
 745   return is_neg;
 746 }
 747 
 748 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
 749                                                         Node* *pos_index) {
 750   if (stopped())
 751     return NULL;                // already stopped
 752   if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
 753     return NULL;                // index is already adequately typed
 754   Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
 755   BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
 756   Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
 757   Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
 758   if (is_notp != NULL && pos_index != NULL) {
 759     // Emulate effect of Parse::adjust_map_after_if.
 760     Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
 761     ccast->set_req(0, control());
 762     (*pos_index) = _gvn.transform(ccast);
 763   }
 764   return is_notp;
 765 }
 766 
 767 // Make sure that 'position' is a valid limit index, in [0..length].
 768 // There are two equivalent plans for checking this:
 769 //   A. (offset + copyLength)  unsigned<=  arrayLength
 770 //   B. offset  <=  (arrayLength - copyLength)
 771 // We require that all of the values above, except for the sum and
 772 // difference, are already known to be non-negative.
 773 // Plan A is robust in the face of overflow, if offset and copyLength
 774 // are both hugely positive.
 775 //
 776 // Plan B is less direct and intuitive, but it does not overflow at
 777 // all, since the difference of two non-negatives is always
 778 // representable.  Whenever Java methods must perform the equivalent
 779 // check they generally use Plan B instead of Plan A.
 780 // For the moment we use Plan A.
 781 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
 782                                                   Node* subseq_length,
 783                                                   Node* array_length,
 784                                                   RegionNode* region) {
 785   if (stopped())
 786     return NULL;                // already stopped
 787   bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
 788   if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
 789     return NULL;                // common case of whole-array copy
 790   Node* last = subseq_length;
 791   if (!zero_offset)             // last += offset
 792     last = _gvn.transform( new (C, 3) AddINode(last, offset));
 793   Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
 794   Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
 795   Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
 796   return is_over;
 797 }
 798 
 799 
 800 //--------------------------generate_current_thread--------------------
 801 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
 802   ciKlass*    thread_klass = env()->Thread_klass();
 803   const Type* thread_type  = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
 804   Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
 805   Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
 806   Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
 807   tls_output = thread;
 808   return threadObj;
 809 }
 810 
 811 
 812 //------------------------------make_string_method_node------------------------
 813 // Helper method for String intrinsic finctions.
 814 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
 815   const int value_offset  = java_lang_String::value_offset_in_bytes();
 816   const int count_offset  = java_lang_String::count_offset_in_bytes();
 817   const int offset_offset = java_lang_String::offset_offset_in_bytes();
 818 
 819   Node* no_ctrl = NULL;
 820 
 821   ciInstanceKlass* klass = env()->String_klass();
 822   const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
 823 
 824   const TypeAryPtr* value_type =
 825         TypeAryPtr::make(TypePtr::NotNull,
 826                          TypeAry::make(TypeInt::CHAR,TypeInt::POS),
 827                          ciTypeArrayKlass::make(T_CHAR), true, 0);
 828 
 829   // Get start addr of string and substring
 830   Node* str1_valuea  = basic_plus_adr(str1, str1, value_offset);
 831   Node* str1_value   = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
 832   Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
 833   Node* str1_offset  = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
 834   Node* str1_start   = array_element_address(str1_value, str1_offset, T_CHAR);
 835 
 836   // Pin loads from String::equals() argument since it could be NULL.
 837   Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl;
 838   Node* str2_valuea  = basic_plus_adr(str2, str2, value_offset);
 839   Node* str2_value   = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
 840   Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
 841   Node* str2_offset  = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
 842   Node* str2_start   = array_element_address(str2_value, str2_offset, T_CHAR);
 843 
 844   Node* result = NULL;
 845   switch (opcode) {
 846   case Op_StrIndexOf:
 847     result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
 848                                        str1_start, cnt1, str2_start, cnt2);
 849     break;
 850   case Op_StrComp:
 851     result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
 852                                     str1_start, cnt1, str2_start, cnt2);
 853     break;
 854   case Op_StrEquals:
 855     result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
 856                                       str1_start, str2_start, cnt1);
 857     break;
 858   default:
 859     ShouldNotReachHere();
 860     return NULL;
 861   }
 862 
 863   // All these intrinsics have checks.
 864   C->set_has_split_ifs(true); // Has chance for split-if optimization
 865 
 866   return _gvn.transform(result);
 867 }
 868 
 869 //------------------------------inline_string_compareTo------------------------
 870 bool LibraryCallKit::inline_string_compareTo() {
 871 
 872   if (!Matcher::has_match_rule(Op_StrComp)) return false;
 873 
 874   const int value_offset = java_lang_String::value_offset_in_bytes();
 875   const int count_offset = java_lang_String::count_offset_in_bytes();
 876   const int offset_offset = java_lang_String::offset_offset_in_bytes();
 877 
 878   _sp += 2;
 879   Node *argument = pop();  // pop non-receiver first:  it was pushed second
 880   Node *receiver = pop();
 881 
 882   // Null check on self without removing any arguments.  The argument
 883   // null check technically happens in the wrong place, which can lead to
 884   // invalid stack traces when string compare is inlined into a method
 885   // which handles NullPointerExceptions.
 886   _sp += 2;
 887   receiver = do_null_check(receiver, T_OBJECT);
 888   argument = do_null_check(argument, T_OBJECT);
 889   _sp -= 2;
 890   if (stopped()) {
 891     return true;
 892   }
 893 
 894   ciInstanceKlass* klass = env()->String_klass();
 895   const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
 896   Node* no_ctrl = NULL;
 897 
 898   // Get counts for string and argument
 899   Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
 900   Node* receiver_cnt  = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
 901 
 902   Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
 903   Node* argument_cnt  = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
 904 
 905   Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
 906   push(compare);
 907   return true;
 908 }
 909 
 910 //------------------------------inline_string_equals------------------------
 911 bool LibraryCallKit::inline_string_equals() {
 912 
 913   if (!Matcher::has_match_rule(Op_StrEquals)) return false;
 914 
 915   const int value_offset = java_lang_String::value_offset_in_bytes();
 916   const int count_offset = java_lang_String::count_offset_in_bytes();
 917   const int offset_offset = java_lang_String::offset_offset_in_bytes();
 918 
 919   int nargs = 2;
 920   _sp += nargs;
 921   Node* argument = pop();  // pop non-receiver first:  it was pushed second
 922   Node* receiver = pop();
 923 
 924   // Null check on self without removing any arguments.  The argument
 925   // null check technically happens in the wrong place, which can lead to
 926   // invalid stack traces when string compare is inlined into a method
 927   // which handles NullPointerExceptions.
 928   _sp += nargs;
 929   receiver = do_null_check(receiver, T_OBJECT);
 930   //should not do null check for argument for String.equals(), because spec
 931   //allows to specify NULL as argument.
 932   _sp -= nargs;
 933 
 934   if (stopped()) {
 935     return true;
 936   }
 937 
 938   // paths (plus control) merge
 939   RegionNode* region = new (C, 5) RegionNode(5);
 940   Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
 941 
 942   // does source == target string?
 943   Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
 944   Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
 945 
 946   Node* if_eq = generate_slow_guard(bol, NULL);
 947   if (if_eq != NULL) {
 948     // receiver == argument
 949     phi->init_req(2, intcon(1));
 950     region->init_req(2, if_eq);
 951   }
 952 
 953   // get String klass for instanceOf
 954   ciInstanceKlass* klass = env()->String_klass();
 955 
 956   if (!stopped()) {
 957     _sp += nargs;          // gen_instanceof might do an uncommon trap
 958     Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
 959     _sp -= nargs;
 960     Node* cmp  = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
 961     Node* bol  = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
 962 
 963     Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
 964     //instanceOf == true, fallthrough
 965 
 966     if (inst_false != NULL) {
 967       phi->init_req(3, intcon(0));
 968       region->init_req(3, inst_false);
 969     }
 970   }
 971 
 972   const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
 973 
 974   Node* no_ctrl = NULL;
 975   Node* receiver_cnt;
 976   Node* argument_cnt;
 977 
 978   if (!stopped()) {
 979     // Properly cast the argument to String
 980     argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
 981 
 982     // Get counts for string and argument
 983     Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
 984     receiver_cnt  = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
 985 
 986     // Pin load from argument string since it could be NULL.
 987     Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
 988     argument_cnt  = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
 989 
 990     // Check for receiver count != argument count
 991     Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
 992     Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
 993     Node* if_ne = generate_slow_guard(bol, NULL);
 994     if (if_ne != NULL) {
 995       phi->init_req(4, intcon(0));
 996       region->init_req(4, if_ne);
 997     }
 998   }
 999 
1000   // Check for count == 0 is done by mach node StrEquals.
1001 
1002   if (!stopped()) {
1003     Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
1004     phi->init_req(1, equals);
1005     region->init_req(1, control());
1006   }
1007 
1008   // post merge
1009   set_control(_gvn.transform(region));
1010   record_for_igvn(region);
1011 
1012   push(_gvn.transform(phi));
1013 
1014   return true;
1015 }
1016 
1017 //------------------------------inline_array_equals----------------------------
1018 bool LibraryCallKit::inline_array_equals() {
1019 
1020   if (!Matcher::has_match_rule(Op_AryEq)) return false;
1021 
1022   _sp += 2;
1023   Node *argument2 = pop();
1024   Node *argument1 = pop();
1025 
1026   Node* equals =
1027     _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1028                                         argument1, argument2) );
1029   push(equals);
1030   return true;
1031 }
1032 
1033 // Java version of String.indexOf(constant string)
1034 // class StringDecl {
1035 //   StringDecl(char[] ca) {
1036 //     offset = 0;
1037 //     count = ca.length;
1038 //     value = ca;
1039 //   }
1040 //   int offset;
1041 //   int count;
1042 //   char[] value;
1043 // }
1044 //
1045 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1046 //                             int targetOffset, int cache_i, int md2) {
1047 //   int cache = cache_i;
1048 //   int sourceOffset = string_object.offset;
1049 //   int sourceCount = string_object.count;
1050 //   int targetCount = target_object.length;
1051 //
1052 //   int targetCountLess1 = targetCount - 1;
1053 //   int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1054 //
1055 //   char[] source = string_object.value;
1056 //   char[] target = target_object;
1057 //   int lastChar = target[targetCountLess1];
1058 //
1059 //  outer_loop:
1060 //   for (int i = sourceOffset; i < sourceEnd; ) {
1061 //     int src = source[i + targetCountLess1];
1062 //     if (src == lastChar) {
1063 //       // With random strings and a 4-character alphabet,
1064 //       // reverse matching at this point sets up 0.8% fewer
1065 //       // frames, but (paradoxically) makes 0.3% more probes.
1066 //       // Since those probes are nearer the lastChar probe,
1067 //       // there is may be a net D$ win with reverse matching.
1068 //       // But, reversing loop inhibits unroll of inner loop
1069 //       // for unknown reason.  So, does running outer loop from
1070 //       // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1071 //       for (int j = 0; j < targetCountLess1; j++) {
1072 //         if (target[targetOffset + j] != source[i+j]) {
1073 //           if ((cache & (1 << source[i+j])) == 0) {
1074 //             if (md2 < j+1) {
1075 //               i += j+1;
1076 //               continue outer_loop;
1077 //             }
1078 //           }
1079 //           i += md2;
1080 //           continue outer_loop;
1081 //         }
1082 //       }
1083 //       return i - sourceOffset;
1084 //     }
1085 //     if ((cache & (1 << src)) == 0) {
1086 //       i += targetCountLess1;
1087 //     } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1088 //     i++;
1089 //   }
1090 //   return -1;
1091 // }
1092 
1093 //------------------------------string_indexOf------------------------
1094 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1095                                      jint cache_i, jint md2_i) {
1096 
1097   Node* no_ctrl  = NULL;
1098   float likely   = PROB_LIKELY(0.9);
1099   float unlikely = PROB_UNLIKELY(0.9);
1100 
1101   const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1102 
1103   const int value_offset  = java_lang_String::value_offset_in_bytes();
1104   const int count_offset  = java_lang_String::count_offset_in_bytes();
1105   const int offset_offset = java_lang_String::offset_offset_in_bytes();
1106 
1107   ciInstanceKlass* klass = env()->String_klass();
1108   const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1109   const TypeAryPtr*  source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1110 
1111   Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1112   Node* sourceOffset  = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1113   Node* sourceCounta  = basic_plus_adr(string_object, string_object, count_offset);
1114   Node* sourceCount   = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1115   Node* sourcea       = basic_plus_adr(string_object, string_object, value_offset);
1116   Node* source        = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1117 
1118   Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1119   jint target_length = target_array->length();
1120   const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1121   const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1122 
1123   IdealKit kit(gvn(), control(), merged_memory(), false, true);
1124 #define __ kit.
1125   Node* zero             = __ ConI(0);
1126   Node* one              = __ ConI(1);
1127   Node* cache            = __ ConI(cache_i);
1128   Node* md2              = __ ConI(md2_i);
1129   Node* lastChar         = __ ConI(target_array->char_at(target_length - 1));
1130   Node* targetCount      = __ ConI(target_length);
1131   Node* targetCountLess1 = __ ConI(target_length - 1);
1132   Node* targetOffset     = __ ConI(targetOffset_i);
1133   Node* sourceEnd        = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1134 
1135   IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1136   Node* outer_loop = __ make_label(2 /* goto */);
1137   Node* return_    = __ make_label(1);
1138 
1139   __ set(rtn,__ ConI(-1));
1140   __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1141        Node* i2  = __ AddI(__ value(i), targetCountLess1);
1142        // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1143        Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1144        __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1145          __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1146               Node* tpj = __ AddI(targetOffset, __ value(j));
1147               Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1148               Node* ipj  = __ AddI(__ value(i), __ value(j));
1149               Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1150               __ if_then(targ, BoolTest::ne, src2); {
1151                 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1152                   __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1153                     __ increment(i, __ AddI(__ value(j), one));
1154                     __ goto_(outer_loop);
1155                   } __ end_if(); __ dead(j);
1156                 }__ end_if(); __ dead(j);
1157                 __ increment(i, md2);
1158                 __ goto_(outer_loop);
1159               }__ end_if();
1160               __ increment(j, one);
1161          }__ end_loop(); __ dead(j);
1162          __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1163          __ goto_(return_);
1164        }__ end_if();
1165        __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1166          __ increment(i, targetCountLess1);
1167        }__ end_if();
1168        __ increment(i, one);
1169        __ bind(outer_loop);
1170   }__ end_loop(); __ dead(i);
1171   __ bind(return_);
1172 
1173   // Final sync IdealKit and GraphKit.
1174   sync_kit(kit);
1175   Node* result = __ value(rtn);
1176 #undef __
1177   C->set_has_loops(true);
1178   return result;
1179 }
1180 
1181 //------------------------------inline_string_indexOf------------------------
1182 bool LibraryCallKit::inline_string_indexOf() {
1183 
1184   const int value_offset  = java_lang_String::value_offset_in_bytes();
1185   const int count_offset  = java_lang_String::count_offset_in_bytes();
1186   const int offset_offset = java_lang_String::offset_offset_in_bytes();
1187 
1188   _sp += 2;
1189   Node *argument = pop();  // pop non-receiver first:  it was pushed second
1190   Node *receiver = pop();
1191 
1192   Node* result;
1193   // Disable the use of pcmpestri until it can be guaranteed that
1194   // the load doesn't cross into the uncommited space.
1195   if (Matcher::has_match_rule(Op_StrIndexOf) &&
1196       UseSSE42Intrinsics) {
1197     // Generate SSE4.2 version of indexOf
1198     // We currently only have match rules that use SSE4.2
1199 
1200     // Null check on self without removing any arguments.  The argument
1201     // null check technically happens in the wrong place, which can lead to
1202     // invalid stack traces when string compare is inlined into a method
1203     // which handles NullPointerExceptions.
1204     _sp += 2;
1205     receiver = do_null_check(receiver, T_OBJECT);
1206     argument = do_null_check(argument, T_OBJECT);
1207     _sp -= 2;
1208 
1209     if (stopped()) {
1210       return true;
1211     }
1212 
1213     ciInstanceKlass* str_klass = env()->String_klass();
1214     const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1215 
1216     // Make the merge point
1217     RegionNode* result_rgn = new (C, 4) RegionNode(4);
1218     Node*       result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1219     Node* no_ctrl  = NULL;
1220 
1221     // Get counts for string and substr
1222     Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1223     Node* source_cnt  = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1224 
1225     Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1226     Node* substr_cnt  = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1227 
1228     // Check for substr count > string count
1229     Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1230     Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1231     Node* if_gt = generate_slow_guard(bol, NULL);
1232     if (if_gt != NULL) {
1233       result_phi->init_req(2, intcon(-1));
1234       result_rgn->init_req(2, if_gt);
1235     }
1236 
1237     if (!stopped()) {
1238       // Check for substr count == 0
1239       cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1240       bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1241       Node* if_zero = generate_slow_guard(bol, NULL);
1242       if (if_zero != NULL) {
1243         result_phi->init_req(3, intcon(0));
1244         result_rgn->init_req(3, if_zero);
1245       }
1246     }
1247 
1248     if (!stopped()) {
1249       result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1250       result_phi->init_req(1, result);
1251       result_rgn->init_req(1, control());
1252     }
1253     set_control(_gvn.transform(result_rgn));
1254     record_for_igvn(result_rgn);
1255     result = _gvn.transform(result_phi);
1256 
1257   } else { // Use LibraryCallKit::string_indexOf
1258     // don't intrinsify if argument isn't a constant string.
1259     if (!argument->is_Con()) {
1260      return false;
1261     }
1262     const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1263     if (str_type == NULL) {
1264       return false;
1265     }
1266     ciInstanceKlass* klass = env()->String_klass();
1267     ciObject* str_const = str_type->const_oop();
1268     if (str_const == NULL || str_const->klass() != klass) {
1269       return false;
1270     }
1271     ciInstance* str = str_const->as_instance();
1272     assert(str != NULL, "must be instance");
1273 
1274     ciObject* v = str->field_value_by_offset(value_offset).as_object();
1275     int       o = str->field_value_by_offset(offset_offset).as_int();
1276     int       c = str->field_value_by_offset(count_offset).as_int();
1277     ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1278 
1279     // constant strings have no offset and count == length which
1280     // simplifies the resulting code somewhat so lets optimize for that.
1281     if (o != 0 || c != pat->length()) {
1282      return false;
1283     }
1284 
1285     // Null check on self without removing any arguments.  The argument
1286     // null check technically happens in the wrong place, which can lead to
1287     // invalid stack traces when string compare is inlined into a method
1288     // which handles NullPointerExceptions.
1289     _sp += 2;
1290     receiver = do_null_check(receiver, T_OBJECT);
1291     // No null check on the argument is needed since it's a constant String oop.
1292     _sp -= 2;
1293     if (stopped()) {
1294       return true;
1295     }
1296 
1297     // The null string as a pattern always returns 0 (match at beginning of string)
1298     if (c == 0) {
1299       push(intcon(0));
1300       return true;
1301     }
1302 
1303     // Generate default indexOf
1304     jchar lastChar = pat->char_at(o + (c - 1));
1305     int cache = 0;
1306     int i;
1307     for (i = 0; i < c - 1; i++) {
1308       assert(i < pat->length(), "out of range");
1309       cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1310     }
1311 
1312     int md2 = c;
1313     for (i = 0; i < c - 1; i++) {
1314       assert(i < pat->length(), "out of range");
1315       if (pat->char_at(o + i) == lastChar) {
1316         md2 = (c - 1) - i;
1317       }
1318     }
1319 
1320     result = string_indexOf(receiver, pat, o, cache, md2);
1321   }
1322 
1323   push(result);
1324   return true;
1325 }
1326 
1327 //--------------------------pop_math_arg--------------------------------
1328 // Pop a double argument to a math function from the stack
1329 // rounding it if necessary.
1330 Node * LibraryCallKit::pop_math_arg() {
1331   Node *arg = pop_pair();
1332   if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1333     arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1334   return arg;
1335 }
1336 
1337 //------------------------------inline_trig----------------------------------
1338 // Inline sin/cos/tan instructions, if possible.  If rounding is required, do
1339 // argument reduction which will turn into a fast/slow diamond.
1340 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1341   _sp += arg_size();            // restore stack pointer
1342   Node* arg = pop_math_arg();
1343   Node* trig = NULL;
1344 
1345   switch (id) {
1346   case vmIntrinsics::_dsin:
1347     trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1348     break;
1349   case vmIntrinsics::_dcos:
1350     trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1351     break;
1352   case vmIntrinsics::_dtan:
1353     trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1354     break;
1355   default:
1356     assert(false, "bad intrinsic was passed in");
1357     return false;
1358   }
1359 
1360   // Rounding required?  Check for argument reduction!
1361   if( Matcher::strict_fp_requires_explicit_rounding ) {
1362 
1363     static const double     pi_4 =  0.7853981633974483;
1364     static const double neg_pi_4 = -0.7853981633974483;
1365     // pi/2 in 80-bit extended precision
1366     // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1367     // -pi/2 in 80-bit extended precision
1368     // 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};
1369     // Cutoff value for using this argument reduction technique
1370     //static const double    pi_2_minus_epsilon =  1.564660403643354;
1371     //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1372 
1373     // Pseudocode for sin:
1374     // if (x <= Math.PI / 4.0) {
1375     //   if (x >= -Math.PI / 4.0) return  fsin(x);
1376     //   if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1377     // } else {
1378     //   if (x <=  Math.PI / 2.0) return  fcos(x - Math.PI / 2.0);
1379     // }
1380     // return StrictMath.sin(x);
1381 
1382     // Pseudocode for cos:
1383     // if (x <= Math.PI / 4.0) {
1384     //   if (x >= -Math.PI / 4.0) return  fcos(x);
1385     //   if (x >= -Math.PI / 2.0) return  fsin(x + Math.PI / 2.0);
1386     // } else {
1387     //   if (x <=  Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1388     // }
1389     // return StrictMath.cos(x);
1390 
1391     // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1392     // requires a special machine instruction to load it.  Instead we'll try
1393     // the 'easy' case.  If we really need the extra range +/- PI/2 we'll
1394     // probably do the math inside the SIN encoding.
1395 
1396     // Make the merge point
1397     RegionNode *r = new (C, 3) RegionNode(3);
1398     Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1399 
1400     // Flatten arg so we need only 1 test
1401     Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1402     // Node for PI/4 constant
1403     Node *pi4 = makecon(TypeD::make(pi_4));
1404     // Check PI/4 : abs(arg)
1405     Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1406     // Check: If PI/4 < abs(arg) then go slow
1407     Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1408     // Branch either way
1409     IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1410     set_control(opt_iff(r,iff));
1411 
1412     // Set fast path result
1413     phi->init_req(2,trig);
1414 
1415     // Slow path - non-blocking leaf call
1416     Node* call = NULL;
1417     switch (id) {
1418     case vmIntrinsics::_dsin:
1419       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1420                                CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1421                                "Sin", NULL, arg, top());
1422       break;
1423     case vmIntrinsics::_dcos:
1424       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1425                                CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1426                                "Cos", NULL, arg, top());
1427       break;
1428     case vmIntrinsics::_dtan:
1429       call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1430                                CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1431                                "Tan", NULL, arg, top());
1432       break;
1433     }
1434     assert(control()->in(0) == call, "");
1435     Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1436     r->init_req(1,control());
1437     phi->init_req(1,slow_result);
1438 
1439     // Post-merge
1440     set_control(_gvn.transform(r));
1441     record_for_igvn(r);
1442     trig = _gvn.transform(phi);
1443 
1444     C->set_has_split_ifs(true); // Has chance for split-if optimization
1445   }
1446   // Push result back on JVM stack
1447   push_pair(trig);
1448   return true;
1449 }
1450 
1451 //------------------------------inline_sqrt-------------------------------------
1452 // Inline square root instruction, if possible.
1453 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1454   assert(id == vmIntrinsics::_dsqrt, "Not square root");
1455   _sp += arg_size();        // restore stack pointer
1456   push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1457   return true;
1458 }
1459 
1460 //------------------------------inline_abs-------------------------------------
1461 // Inline absolute value instruction, if possible.
1462 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1463   assert(id == vmIntrinsics::_dabs, "Not absolute value");
1464   _sp += arg_size();        // restore stack pointer
1465   push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1466   return true;
1467 }
1468 
1469 //------------------------------inline_exp-------------------------------------
1470 // Inline exp instructions, if possible.  The Intel hardware only misses
1471 // really odd corner cases (+/- Infinity).  Just uncommon-trap them.
1472 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1473   assert(id == vmIntrinsics::_dexp, "Not exp");
1474 
1475   // If this inlining ever returned NaN in the past, we do not intrinsify it
1476   // every again.  NaN results requires StrictMath.exp handling.
1477   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
1478 
1479   // Do not intrinsify on older platforms which lack cmove.
1480   if (ConditionalMoveLimit == 0)  return false;
1481 
1482   _sp += arg_size();        // restore stack pointer
1483   Node *x = pop_math_arg();
1484   Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1485 
1486   //-------------------
1487   //result=(result.isNaN())? StrictMath::exp():result;
1488   // Check: If isNaN() by checking result!=result? then go to Strict Math
1489   Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1490   // Build the boolean node
1491   Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1492 
1493   { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1494     // End the current control-flow path
1495     push_pair(x);
1496     // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1497     // to handle.  Recompile without intrinsifying Math.exp
1498     uncommon_trap(Deoptimization::Reason_intrinsic,
1499                   Deoptimization::Action_make_not_entrant);
1500   }
1501 
1502   C->set_has_split_ifs(true); // Has chance for split-if optimization
1503 
1504   push_pair(result);
1505 
1506   return true;
1507 }
1508 
1509 //------------------------------inline_pow-------------------------------------
1510 // Inline power instructions, if possible.
1511 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1512   assert(id == vmIntrinsics::_dpow, "Not pow");
1513 
1514   // If this inlining ever returned NaN in the past, we do not intrinsify it
1515   // every again.  NaN results requires StrictMath.pow handling.
1516   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
1517 
1518   // Do not intrinsify on older platforms which lack cmove.
1519   if (ConditionalMoveLimit == 0)  return false;
1520 
1521   // Pseudocode for pow
1522   // if (x <= 0.0) {
1523   //   if ((double)((int)y)==y) { // if y is int
1524   //     result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1525   //   } else {
1526   //     result = NaN;
1527   //   }
1528   // } else {
1529   //   result = DPow(x,y);
1530   // }
1531   // if (result != result)?  {
1532   //   uncommon_trap();
1533   // }
1534   // return result;
1535 
1536   _sp += arg_size();        // restore stack pointer
1537   Node* y = pop_math_arg();
1538   Node* x = pop_math_arg();
1539 
1540   Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1541 
1542   // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1543   // inside of something) then skip the fancy tests and just check for
1544   // NaN result.
1545   Node *result = NULL;
1546   if( jvms()->depth() >= 1 ) {
1547     result = fast_result;
1548   } else {
1549 
1550     // Set the merge point for If node with condition of (x <= 0.0)
1551     // There are four possible paths to region node and phi node
1552     RegionNode *r = new (C, 4) RegionNode(4);
1553     Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1554 
1555     // Build the first if node: if (x <= 0.0)
1556     // Node for 0 constant
1557     Node *zeronode = makecon(TypeD::ZERO);
1558     // Check x:0
1559     Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1560     // Check: If (x<=0) then go complex path
1561     Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1562     // Branch either way
1563     IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1564     Node *opt_test = _gvn.transform(if1);
1565     //assert( opt_test->is_If(), "Expect an IfNode");
1566     IfNode *opt_if1 = (IfNode*)opt_test;
1567     // Fast path taken; set region slot 3
1568     Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1569     r->init_req(3,fast_taken); // Capture fast-control
1570 
1571     // Fast path not-taken, i.e. slow path
1572     Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1573 
1574     // Set fast path result
1575     Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1576     phi->init_req(3, fast_result);
1577 
1578     // Complex path
1579     // Build the second if node (if y is int)
1580     // Node for (int)y
1581     Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1582     // Node for (double)((int) y)
1583     Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1584     // Check (double)((int) y) : y
1585     Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1586     // Check if (y isn't int) then go to slow path
1587 
1588     Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1589     // Branch either way
1590     IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1591     Node *slow_path = opt_iff(r,if2); // Set region path 2
1592 
1593     // Calculate DPow(abs(x), y)*(1 & (int)y)
1594     // Node for constant 1
1595     Node *conone = intcon(1);
1596     // 1& (int)y
1597     Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1598     // zero node
1599     Node *conzero = intcon(0);
1600     // Check (1&(int)y)==0?
1601     Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1602     // Check if (1&(int)y)!=0?, if so the result is negative
1603     Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1604     // abs(x)
1605     Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1606     // abs(x)^y
1607     Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1608     // -abs(x)^y
1609     Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1610     // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1611     Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1612     // Set complex path fast result
1613     phi->init_req(2, signresult);
1614 
1615     static const jlong nan_bits = CONST64(0x7ff8000000000000);
1616     Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1617     r->init_req(1,slow_path);
1618     phi->init_req(1,slow_result);
1619 
1620     // Post merge
1621     set_control(_gvn.transform(r));
1622     record_for_igvn(r);
1623     result=_gvn.transform(phi);
1624   }
1625 
1626   //-------------------
1627   //result=(result.isNaN())? uncommon_trap():result;
1628   // Check: If isNaN() by checking result!=result? then go to Strict Math
1629   Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1630   // Build the boolean node
1631   Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1632 
1633   { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1634     // End the current control-flow path
1635     push_pair(x);
1636     push_pair(y);
1637     // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1638     // to handle.  Recompile without intrinsifying Math.pow.
1639     uncommon_trap(Deoptimization::Reason_intrinsic,
1640                   Deoptimization::Action_make_not_entrant);
1641   }
1642 
1643   C->set_has_split_ifs(true); // Has chance for split-if optimization
1644 
1645   push_pair(result);
1646 
1647   return true;
1648 }
1649 
1650 //------------------------------inline_trans-------------------------------------
1651 // Inline transcendental instructions, if possible.  The Intel hardware gets
1652 // these right, no funny corner cases missed.
1653 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1654   _sp += arg_size();        // restore stack pointer
1655   Node* arg = pop_math_arg();
1656   Node* trans = NULL;
1657 
1658   switch (id) {
1659   case vmIntrinsics::_dlog:
1660     trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1661     break;
1662   case vmIntrinsics::_dlog10:
1663     trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1664     break;
1665   default:
1666     assert(false, "bad intrinsic was passed in");
1667     return false;
1668   }
1669 
1670   // Push result back on JVM stack
1671   push_pair(trans);
1672   return true;
1673 }
1674 
1675 //------------------------------runtime_math-----------------------------
1676 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1677   Node* a = NULL;
1678   Node* b = NULL;
1679 
1680   assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1681          "must be (DD)D or (D)D type");
1682 
1683   // Inputs
1684   _sp += arg_size();        // restore stack pointer
1685   if (call_type == OptoRuntime::Math_DD_D_Type()) {
1686     b = pop_math_arg();
1687   }
1688   a = pop_math_arg();
1689 
1690   const TypePtr* no_memory_effects = NULL;
1691   Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1692                                  no_memory_effects,
1693                                  a, top(), b, b ? top() : NULL);
1694   Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1695 #ifdef ASSERT
1696   Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1697   assert(value_top == top(), "second value must be top");
1698 #endif
1699 
1700   push_pair(value);
1701   return true;
1702 }
1703 
1704 //------------------------------inline_math_native-----------------------------
1705 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1706   switch (id) {
1707     // These intrinsics are not properly supported on all hardware
1708   case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1709     runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1710   case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1711     runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1712   case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1713     runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1714 
1715   case vmIntrinsics::_dlog:   return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1716     runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1717   case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1718     runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1719 
1720     // These intrinsics are supported on all hardware
1721   case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1722   case vmIntrinsics::_dabs:  return Matcher::has_match_rule(Op_AbsD)  ? inline_abs(id)  : false;
1723 
1724     // These intrinsics don't work on X86.  The ad implementation doesn't
1725     // handle NaN's properly.  Instead of returning infinity, the ad
1726     // implementation returns a NaN on overflow. See bug: 6304089
1727     // Once the ad implementations are fixed, change the code below
1728     // to match the intrinsics above
1729 
1730   case vmIntrinsics::_dexp:  return
1731     runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1732   case vmIntrinsics::_dpow:  return
1733     runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1734 
1735    // These intrinsics are not yet correctly implemented
1736   case vmIntrinsics::_datan2:
1737     return false;
1738 
1739   default:
1740     ShouldNotReachHere();
1741     return false;
1742   }
1743 }
1744 
1745 static bool is_simple_name(Node* n) {
1746   return (n->req() == 1         // constant
1747           || (n->is_Type() && n->as_Type()->type()->singleton())
1748           || n->is_Proj()       // parameter or return value
1749           || n->is_Phi()        // local of some sort
1750           );
1751 }
1752 
1753 //----------------------------inline_min_max-----------------------------------
1754 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1755   push(generate_min_max(id, argument(0), argument(1)));
1756 
1757   return true;
1758 }
1759 
1760 Node*
1761 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1762   // These are the candidate return value:
1763   Node* xvalue = x0;
1764   Node* yvalue = y0;
1765 
1766   if (xvalue == yvalue) {
1767     return xvalue;
1768   }
1769 
1770   bool want_max = (id == vmIntrinsics::_max);
1771 
1772   const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1773   const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1774   if (txvalue == NULL || tyvalue == NULL)  return top();
1775   // This is not really necessary, but it is consistent with a
1776   // hypothetical MaxINode::Value method:
1777   int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1778 
1779   // %%% This folding logic should (ideally) be in a different place.
1780   // Some should be inside IfNode, and there to be a more reliable
1781   // transformation of ?: style patterns into cmoves.  We also want
1782   // more powerful optimizations around cmove and min/max.
1783 
1784   // Try to find a dominating comparison of these guys.
1785   // It can simplify the index computation for Arrays.copyOf
1786   // and similar uses of System.arraycopy.
1787   // First, compute the normalized version of CmpI(x, y).
1788   int   cmp_op = Op_CmpI;
1789   Node* xkey = xvalue;
1790   Node* ykey = yvalue;
1791   Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1792   if (ideal_cmpxy->is_Cmp()) {
1793     // E.g., if we have CmpI(length - offset, count),
1794     // it might idealize to CmpI(length, count + offset)
1795     cmp_op = ideal_cmpxy->Opcode();
1796     xkey = ideal_cmpxy->in(1);
1797     ykey = ideal_cmpxy->in(2);
1798   }
1799 
1800   // Start by locating any relevant comparisons.
1801   Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1802   Node* cmpxy = NULL;
1803   Node* cmpyx = NULL;
1804   for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1805     Node* cmp = start_from->fast_out(k);
1806     if (cmp->outcnt() > 0 &&            // must have prior uses
1807         cmp->in(0) == NULL &&           // must be context-independent
1808         cmp->Opcode() == cmp_op) {      // right kind of compare
1809       if (cmp->in(1) == xkey && cmp->in(2) == ykey)  cmpxy = cmp;
1810       if (cmp->in(1) == ykey && cmp->in(2) == xkey)  cmpyx = cmp;
1811     }
1812   }
1813 
1814   const int NCMPS = 2;
1815   Node* cmps[NCMPS] = { cmpxy, cmpyx };
1816   int cmpn;
1817   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1818     if (cmps[cmpn] != NULL)  break;     // find a result
1819   }
1820   if (cmpn < NCMPS) {
1821     // Look for a dominating test that tells us the min and max.
1822     int depth = 0;                // Limit search depth for speed
1823     Node* dom = control();
1824     for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1825       if (++depth >= 100)  break;
1826       Node* ifproj = dom;
1827       if (!ifproj->is_Proj())  continue;
1828       Node* iff = ifproj->in(0);
1829       if (!iff->is_If())  continue;
1830       Node* bol = iff->in(1);
1831       if (!bol->is_Bool())  continue;
1832       Node* cmp = bol->in(1);
1833       if (cmp == NULL)  continue;
1834       for (cmpn = 0; cmpn < NCMPS; cmpn++)
1835         if (cmps[cmpn] == cmp)  break;
1836       if (cmpn == NCMPS)  continue;
1837       BoolTest::mask btest = bol->as_Bool()->_test._test;
1838       if (ifproj->is_IfFalse())  btest = BoolTest(btest).negate();
1839       if (cmp->in(1) == ykey)    btest = BoolTest(btest).commute();
1840       // At this point, we know that 'x btest y' is true.
1841       switch (btest) {
1842       case BoolTest::eq:
1843         // They are proven equal, so we can collapse the min/max.
1844         // Either value is the answer.  Choose the simpler.
1845         if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1846           return yvalue;
1847         return xvalue;
1848       case BoolTest::lt:          // x < y
1849       case BoolTest::le:          // x <= y
1850         return (want_max ? yvalue : xvalue);
1851       case BoolTest::gt:          // x > y
1852       case BoolTest::ge:          // x >= y
1853         return (want_max ? xvalue : yvalue);
1854       }
1855     }
1856   }
1857 
1858   // We failed to find a dominating test.
1859   // Let's pick a test that might GVN with prior tests.
1860   Node*          best_bol   = NULL;
1861   BoolTest::mask best_btest = BoolTest::illegal;
1862   for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1863     Node* cmp = cmps[cmpn];
1864     if (cmp == NULL)  continue;
1865     for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1866       Node* bol = cmp->fast_out(j);
1867       if (!bol->is_Bool())  continue;
1868       BoolTest::mask btest = bol->as_Bool()->_test._test;
1869       if (btest == BoolTest::eq || btest == BoolTest::ne)  continue;
1870       if (cmp->in(1) == ykey)   btest = BoolTest(btest).commute();
1871       if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1872         best_bol   = bol->as_Bool();
1873         best_btest = btest;
1874       }
1875     }
1876   }
1877 
1878   Node* answer_if_true  = NULL;
1879   Node* answer_if_false = NULL;
1880   switch (best_btest) {
1881   default:
1882     if (cmpxy == NULL)
1883       cmpxy = ideal_cmpxy;
1884     best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1885     // and fall through:
1886   case BoolTest::lt:          // x < y
1887   case BoolTest::le:          // x <= y
1888     answer_if_true  = (want_max ? yvalue : xvalue);
1889     answer_if_false = (want_max ? xvalue : yvalue);
1890     break;
1891   case BoolTest::gt:          // x > y
1892   case BoolTest::ge:          // x >= y
1893     answer_if_true  = (want_max ? xvalue : yvalue);
1894     answer_if_false = (want_max ? yvalue : xvalue);
1895     break;
1896   }
1897 
1898   jint hi, lo;
1899   if (want_max) {
1900     // We can sharpen the minimum.
1901     hi = MAX2(txvalue->_hi, tyvalue->_hi);
1902     lo = MAX2(txvalue->_lo, tyvalue->_lo);
1903   } else {
1904     // We can sharpen the maximum.
1905     hi = MIN2(txvalue->_hi, tyvalue->_hi);
1906     lo = MIN2(txvalue->_lo, tyvalue->_lo);
1907   }
1908 
1909   // Use a flow-free graph structure, to avoid creating excess control edges
1910   // which could hinder other optimizations.
1911   // Since Math.min/max is often used with arraycopy, we want
1912   // tightly_coupled_allocation to be able to see beyond min/max expressions.
1913   Node* cmov = CMoveNode::make(C, NULL, best_bol,
1914                                answer_if_false, answer_if_true,
1915                                TypeInt::make(lo, hi, widen));
1916 
1917   return _gvn.transform(cmov);
1918 
1919   /*
1920   // This is not as desirable as it may seem, since Min and Max
1921   // nodes do not have a full set of optimizations.
1922   // And they would interfere, anyway, with 'if' optimizations
1923   // and with CMoveI canonical forms.
1924   switch (id) {
1925   case vmIntrinsics::_min:
1926     result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1927   case vmIntrinsics::_max:
1928     result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1929   default:
1930     ShouldNotReachHere();
1931   }
1932   */
1933 }
1934 
1935 inline int
1936 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1937   const TypePtr* base_type = TypePtr::NULL_PTR;
1938   if (base != NULL)  base_type = _gvn.type(base)->isa_ptr();
1939   if (base_type == NULL) {
1940     // Unknown type.
1941     return Type::AnyPtr;
1942   } else if (base_type == TypePtr::NULL_PTR) {
1943     // Since this is a NULL+long form, we have to switch to a rawptr.
1944     base   = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1945     offset = MakeConX(0);
1946     return Type::RawPtr;
1947   } else if (base_type->base() == Type::RawPtr) {
1948     return Type::RawPtr;
1949   } else if (base_type->isa_oopptr()) {
1950     // Base is never null => always a heap address.
1951     if (base_type->ptr() == TypePtr::NotNull) {
1952       return Type::OopPtr;
1953     }
1954     // Offset is small => always a heap address.
1955     const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1956     if (offset_type != NULL &&
1957         base_type->offset() == 0 &&     // (should always be?)
1958         offset_type->_lo >= 0 &&
1959         !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1960       return Type::OopPtr;
1961     }
1962     // Otherwise, it might either be oop+off or NULL+addr.
1963     return Type::AnyPtr;
1964   } else {
1965     // No information:
1966     return Type::AnyPtr;
1967   }
1968 }
1969 
1970 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1971   int kind = classify_unsafe_addr(base, offset);
1972   if (kind == Type::RawPtr) {
1973     return basic_plus_adr(top(), base, offset);
1974   } else {
1975     return basic_plus_adr(base, offset);
1976   }
1977 }
1978 
1979 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1980 // inline int Integer.numberOfLeadingZeros(int)
1981 // inline int Long.numberOfLeadingZeros(long)
1982 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1983   assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1984   if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1985   if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1986   _sp += arg_size();  // restore stack pointer
1987   switch (id) {
1988   case vmIntrinsics::_numberOfLeadingZeros_i:
1989     push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1990     break;
1991   case vmIntrinsics::_numberOfLeadingZeros_l:
1992     push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1993     break;
1994   default:
1995     ShouldNotReachHere();
1996   }
1997   return true;
1998 }
1999 
2000 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2001 // inline int Integer.numberOfTrailingZeros(int)
2002 // inline int Long.numberOfTrailingZeros(long)
2003 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2004   assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2005   if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2006   if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2007   _sp += arg_size();  // restore stack pointer
2008   switch (id) {
2009   case vmIntrinsics::_numberOfTrailingZeros_i:
2010     push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2011     break;
2012   case vmIntrinsics::_numberOfTrailingZeros_l:
2013     push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2014     break;
2015   default:
2016     ShouldNotReachHere();
2017   }
2018   return true;
2019 }
2020 
2021 //----------------------------inline_bitCount_int/long-----------------------
2022 // inline int Integer.bitCount(int)
2023 // inline int Long.bitCount(long)
2024 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2025   assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2026   if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2027   if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2028   _sp += arg_size();  // restore stack pointer
2029   switch (id) {
2030   case vmIntrinsics::_bitCount_i:
2031     push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2032     break;
2033   case vmIntrinsics::_bitCount_l:
2034     push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2035     break;
2036   default:
2037     ShouldNotReachHere();
2038   }
2039   return true;
2040 }
2041 
2042 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2043 // inline Integer.reverseBytes(int)
2044 // inline Long.reverseBytes(long)
2045 // inline Character.reverseBytes(char)
2046 // inline Short.reverseBytes(short)
2047 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2048   assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2049          id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2050          "not reverse Bytes");
2051   if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI))  return false;
2052   if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL))  return false;
2053   if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2054   if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS))  return false;
2055   _sp += arg_size();        // restore stack pointer
2056   switch (id) {
2057   case vmIntrinsics::_reverseBytes_i:
2058     push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2059     break;
2060   case vmIntrinsics::_reverseBytes_l:
2061     push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2062     break;
2063   case vmIntrinsics::_reverseBytes_c:
2064     push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2065     break;
2066   case vmIntrinsics::_reverseBytes_s:
2067     push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2068     break;
2069   default:
2070     ;
2071   }
2072   return true;
2073 }
2074 
2075 //----------------------------inline_unsafe_access----------------------------
2076 
2077 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2078 
2079 // Interpret Unsafe.fieldOffset cookies correctly:
2080 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2081 
2082 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2083   if (callee()->is_static())  return false;  // caller must have the capability!
2084 
2085 #ifndef PRODUCT
2086   {
2087     ResourceMark rm;
2088     // Check the signatures.
2089     ciSignature* sig = signature();
2090 #ifdef ASSERT
2091     if (!is_store) {
2092       // Object getObject(Object base, int/long offset), etc.
2093       BasicType rtype = sig->return_type()->basic_type();
2094       if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2095           rtype = T_ADDRESS;  // it is really a C void*
2096       assert(rtype == type, "getter must return the expected value");
2097       if (!is_native_ptr) {
2098         assert(sig->count() == 2, "oop getter has 2 arguments");
2099         assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2100         assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2101       } else {
2102         assert(sig->count() == 1, "native getter has 1 argument");
2103         assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2104       }
2105     } else {
2106       // void putObject(Object base, int/long offset, Object x), etc.
2107       assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2108       if (!is_native_ptr) {
2109         assert(sig->count() == 3, "oop putter has 3 arguments");
2110         assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2111         assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2112       } else {
2113         assert(sig->count() == 2, "native putter has 2 arguments");
2114         assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2115       }
2116       BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2117       if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2118         vtype = T_ADDRESS;  // it is really a C void*
2119       assert(vtype == type, "putter must accept the expected value");
2120     }
2121 #endif // ASSERT
2122  }
2123 #endif //PRODUCT
2124 
2125   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2126 
2127   int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2128 
2129   // Argument words:  "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2130   int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2131 
2132   debug_only(int saved_sp = _sp);
2133   _sp += nargs;
2134 
2135   Node* val;
2136   debug_only(val = (Node*)(uintptr_t)-1);
2137 
2138 
2139   if (is_store) {
2140     // Get the value being stored.  (Pop it first; it was pushed last.)
2141     switch (type) {
2142     case T_DOUBLE:
2143     case T_LONG:
2144     case T_ADDRESS:
2145       val = pop_pair();
2146       break;
2147     default:
2148       val = pop();
2149     }
2150   }
2151 
2152   // Build address expression.  See the code in inline_unsafe_prefetch.
2153   Node *adr;
2154   Node *heap_base_oop = top();
2155   if (!is_native_ptr) {
2156     // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2157     Node* offset = pop_pair();
2158     // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2159     Node* base   = pop();
2160     // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2161     // to be plain byte offsets, which are also the same as those accepted
2162     // by oopDesc::field_base.
2163     assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2164            "fieldOffset must be byte-scaled");
2165     // 32-bit machines ignore the high half!
2166     offset = ConvL2X(offset);
2167     adr = make_unsafe_address(base, offset);
2168     heap_base_oop = base;
2169   } else {
2170     Node* ptr = pop_pair();
2171     // Adjust Java long to machine word:
2172     ptr = ConvL2X(ptr);
2173     adr = make_unsafe_address(NULL, ptr);
2174   }
2175 
2176   // Pop receiver last:  it was pushed first.
2177   Node *receiver = pop();
2178 
2179   assert(saved_sp == _sp, "must have correct argument count");
2180 
2181   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2182 
2183   // First guess at the value type.
2184   const Type *value_type = Type::get_const_basic_type(type);
2185 
2186   // Try to categorize the address.  If it comes up as TypeJavaPtr::BOTTOM,
2187   // there was not enough information to nail it down.
2188   Compile::AliasType* alias_type = C->alias_type(adr_type);
2189   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2190 
2191   // We will need memory barriers unless we can determine a unique
2192   // alias category for this reference.  (Note:  If for some reason
2193   // the barriers get omitted and the unsafe reference begins to "pollute"
2194   // the alias analysis of the rest of the graph, either Compile::can_alias
2195   // or Compile::must_alias will throw a diagnostic assert.)
2196   bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2197 
2198   if (!is_store && type == T_OBJECT) {
2199     // Attempt to infer a sharper value type from the offset and base type.
2200     ciKlass* sharpened_klass = NULL;
2201 
2202     // See if it is an instance field, with an object type.
2203     if (alias_type->field() != NULL) {
2204       assert(!is_native_ptr, "native pointer op cannot use a java address");
2205       if (alias_type->field()->type()->is_klass()) {
2206         sharpened_klass = alias_type->field()->type()->as_klass();
2207       }
2208     }
2209 
2210     // See if it is a narrow oop array.
2211     if (adr_type->isa_aryptr()) {
2212       if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2213         const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2214         if (elem_type != NULL) {
2215           sharpened_klass = elem_type->klass();
2216         }
2217       }
2218     }
2219 
2220     if (sharpened_klass != NULL) {
2221       const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2222 
2223       // Sharpen the value type.
2224       value_type = tjp;
2225 
2226 #ifndef PRODUCT
2227       if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2228         tty->print("  from base type:  ");   adr_type->dump();
2229         tty->print("  sharpened value: "); value_type->dump();
2230       }
2231 #endif
2232     }
2233   }
2234 
2235   // Null check on self without removing any arguments.  The argument
2236   // null check technically happens in the wrong place, which can lead to
2237   // invalid stack traces when the primitive is inlined into a method
2238   // which handles NullPointerExceptions.
2239   _sp += nargs;
2240   do_null_check(receiver, T_OBJECT);
2241   _sp -= nargs;
2242   if (stopped()) {
2243     return true;
2244   }
2245   // Heap pointers get a null-check from the interpreter,
2246   // as a courtesy.  However, this is not guaranteed by Unsafe,
2247   // and it is not possible to fully distinguish unintended nulls
2248   // from intended ones in this API.
2249 
2250   if (is_volatile) {
2251     // We need to emit leading and trailing CPU membars (see below) in
2252     // addition to memory membars when is_volatile. This is a little
2253     // too strong, but avoids the need to insert per-alias-type
2254     // volatile membars (for stores; compare Parse::do_put_xxx), which
2255     // we cannot do effectively here because we probably only have a
2256     // rough approximation of type.
2257     need_mem_bar = true;
2258     // For Stores, place a memory ordering barrier now.
2259     if (is_store)
2260       insert_mem_bar(Op_MemBarRelease);
2261   }
2262 
2263   // Memory barrier to prevent normal and 'unsafe' accesses from
2264   // bypassing each other.  Happens after null checks, so the
2265   // exception paths do not take memory state from the memory barrier,
2266   // so there's no problems making a strong assert about mixing users
2267   // of safe & unsafe memory.  Otherwise fails in a CTW of rt.jar
2268   // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2269   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2270 
2271   if (!is_store) {
2272     Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2273     // load value and push onto stack
2274     switch (type) {
2275     case T_BOOLEAN:
2276     case T_CHAR:
2277     case T_BYTE:
2278     case T_SHORT:
2279     case T_INT:
2280     case T_FLOAT:
2281     case T_OBJECT:
2282       push( p );
2283       break;
2284     case T_ADDRESS:
2285       // Cast to an int type.
2286       p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2287       p = ConvX2L(p);
2288       push_pair(p);
2289       break;
2290     case T_DOUBLE:
2291     case T_LONG:
2292       push_pair( p );
2293       break;
2294     default: ShouldNotReachHere();
2295     }
2296   } else {
2297     // place effect of store into memory
2298     switch (type) {
2299     case T_DOUBLE:
2300       val = dstore_rounding(val);
2301       break;
2302     case T_ADDRESS:
2303       // Repackage the long as a pointer.
2304       val = ConvL2X(val);
2305       val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2306       break;
2307     }
2308 
2309     if (type != T_OBJECT ) {
2310       (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2311     } else {
2312       // Possibly an oop being stored to Java heap or native memory
2313       if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2314         // oop to Java heap.
2315         (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2316       } else {
2317         // We can't tell at compile time if we are storing in the Java heap or outside
2318         // of it. So we need to emit code to conditionally do the proper type of
2319         // store.
2320 
2321         IdealKit ideal(gvn(), control(),  merged_memory());
2322 #define __ ideal.
2323         // QQQ who knows what probability is here??
2324         __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2325           // Sync IdealKit and graphKit.
2326           set_all_memory( __ merged_memory());
2327           set_control(__ ctrl());
2328           Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2329           // Update IdealKit memory.
2330           __ set_all_memory(merged_memory());
2331           __ set_ctrl(control());
2332         } __ else_(); {
2333           __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2334         } __ end_if();
2335         // Final sync IdealKit and GraphKit.
2336         sync_kit(ideal);
2337 #undef __
2338       }
2339     }
2340   }
2341 
2342   if (is_volatile) {
2343     if (!is_store)
2344       insert_mem_bar(Op_MemBarAcquire);
2345     else
2346       insert_mem_bar(Op_MemBarVolatile);
2347   }
2348 
2349   if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2350 
2351   return true;
2352 }
2353 
2354 //----------------------------inline_unsafe_prefetch----------------------------
2355 
2356 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2357 #ifndef PRODUCT
2358   {
2359     ResourceMark rm;
2360     // Check the signatures.
2361     ciSignature* sig = signature();
2362 #ifdef ASSERT
2363     // Object getObject(Object base, int/long offset), etc.
2364     BasicType rtype = sig->return_type()->basic_type();
2365     if (!is_native_ptr) {
2366       assert(sig->count() == 2, "oop prefetch has 2 arguments");
2367       assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2368       assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2369     } else {
2370       assert(sig->count() == 1, "native prefetch has 1 argument");
2371       assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2372     }
2373 #endif // ASSERT
2374   }
2375 #endif // !PRODUCT
2376 
2377   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2378 
2379   // Argument words:  "this" if not static, plus (oop/offset) or (lo/hi) args
2380   int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2381 
2382   debug_only(int saved_sp = _sp);
2383   _sp += nargs;
2384 
2385   // Build address expression.  See the code in inline_unsafe_access.
2386   Node *adr;
2387   if (!is_native_ptr) {
2388     // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2389     Node* offset = pop_pair();
2390     // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2391     Node* base   = pop();
2392     // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2393     // to be plain byte offsets, which are also the same as those accepted
2394     // by oopDesc::field_base.
2395     assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2396            "fieldOffset must be byte-scaled");
2397     // 32-bit machines ignore the high half!
2398     offset = ConvL2X(offset);
2399     adr = make_unsafe_address(base, offset);
2400   } else {
2401     Node* ptr = pop_pair();
2402     // Adjust Java long to machine word:
2403     ptr = ConvL2X(ptr);
2404     adr = make_unsafe_address(NULL, ptr);
2405   }
2406 
2407   if (is_static) {
2408     assert(saved_sp == _sp, "must have correct argument count");
2409   } else {
2410     // Pop receiver last:  it was pushed first.
2411     Node *receiver = pop();
2412     assert(saved_sp == _sp, "must have correct argument count");
2413 
2414     // Null check on self without removing any arguments.  The argument
2415     // null check technically happens in the wrong place, which can lead to
2416     // invalid stack traces when the primitive is inlined into a method
2417     // which handles NullPointerExceptions.
2418     _sp += nargs;
2419     do_null_check(receiver, T_OBJECT);
2420     _sp -= nargs;
2421     if (stopped()) {
2422       return true;
2423     }
2424   }
2425 
2426   // Generate the read or write prefetch
2427   Node *prefetch;
2428   if (is_store) {
2429     prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2430   } else {
2431     prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2432   }
2433   prefetch->init_req(0, control());
2434   set_i_o(_gvn.transform(prefetch));
2435 
2436   return true;
2437 }
2438 
2439 //----------------------------inline_unsafe_CAS----------------------------
2440 
2441 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2442   // This basic scheme here is the same as inline_unsafe_access, but
2443   // differs in enough details that combining them would make the code
2444   // overly confusing.  (This is a true fact! I originally combined
2445   // them, but even I was confused by it!) As much code/comments as
2446   // possible are retained from inline_unsafe_access though to make
2447   // the correspondences clearer. - dl
2448 
2449   if (callee()->is_static())  return false;  // caller must have the capability!
2450 
2451 #ifndef PRODUCT
2452   {
2453     ResourceMark rm;
2454     // Check the signatures.
2455     ciSignature* sig = signature();
2456 #ifdef ASSERT
2457     BasicType rtype = sig->return_type()->basic_type();
2458     assert(rtype == T_BOOLEAN, "CAS must return boolean");
2459     assert(sig->count() == 4, "CAS has 4 arguments");
2460     assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2461     assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2462 #endif // ASSERT
2463   }
2464 #endif //PRODUCT
2465 
2466   // number of stack slots per value argument (1 or 2)
2467   int type_words = type2size[type];
2468 
2469   // Cannot inline wide CAS on machines that don't support it natively
2470   if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2471     return false;
2472 
2473   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2474 
2475   // Argument words:  "this" plus oop plus offset plus oldvalue plus newvalue;
2476   int nargs = 1 + 1 + 2  + type_words + type_words;
2477 
2478   // pop arguments: newval, oldval, offset, base, and receiver
2479   debug_only(int saved_sp = _sp);
2480   _sp += nargs;
2481   Node* newval   = (type_words == 1) ? pop() : pop_pair();
2482   Node* oldval   = (type_words == 1) ? pop() : pop_pair();
2483   Node *offset   = pop_pair();
2484   Node *base     = pop();
2485   Node *receiver = pop();
2486   assert(saved_sp == _sp, "must have correct argument count");
2487 
2488   //  Null check receiver.
2489   _sp += nargs;
2490   do_null_check(receiver, T_OBJECT);
2491   _sp -= nargs;
2492   if (stopped()) {
2493     return true;
2494   }
2495 
2496   // Build field offset expression.
2497   // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2498   // to be plain byte offsets, which are also the same as those accepted
2499   // by oopDesc::field_base.
2500   assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2501   // 32-bit machines ignore the high half of long offsets
2502   offset = ConvL2X(offset);
2503   Node* adr = make_unsafe_address(base, offset);
2504   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2505 
2506   // (Unlike inline_unsafe_access, there seems no point in trying
2507   // to refine types. Just use the coarse types here.
2508   const Type *value_type = Type::get_const_basic_type(type);
2509   Compile::AliasType* alias_type = C->alias_type(adr_type);
2510   assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2511   int alias_idx = C->get_alias_index(adr_type);
2512 
2513   // Memory-model-wise, a CAS acts like a little synchronized block,
2514   // so needs barriers on each side.  These don't translate into
2515   // actual barriers on most machines, but we still need rest of
2516   // compiler to respect ordering.
2517 
2518   insert_mem_bar(Op_MemBarRelease);
2519   insert_mem_bar(Op_MemBarCPUOrder);
2520 
2521   // 4984716: MemBars must be inserted before this
2522   //          memory node in order to avoid a false
2523   //          dependency which will confuse the scheduler.
2524   Node *mem = memory(alias_idx);
2525 
2526   // For now, we handle only those cases that actually exist: ints,
2527   // longs, and Object. Adding others should be straightforward.
2528   Node* cas;
2529   switch(type) {
2530   case T_INT:
2531     cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2532     break;
2533   case T_LONG:
2534     cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2535     break;
2536   case T_OBJECT:
2537      // reference stores need a store barrier.
2538     // (They don't if CAS fails, but it isn't worth checking.)
2539     pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2540 #ifdef _LP64
2541     if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2542       Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2543       Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2544       cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2545                                                           newval_enc, oldval_enc));
2546     } else
2547 #endif
2548     {
2549       cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2550     }
2551     post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2552     break;
2553   default:
2554     ShouldNotReachHere();
2555     break;
2556   }
2557 
2558   // SCMemProjNodes represent the memory state of CAS. Their main
2559   // role is to prevent CAS nodes from being optimized away when their
2560   // results aren't used.
2561   Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2562   set_memory(proj, alias_idx);
2563 
2564   // Add the trailing membar surrounding the access
2565   insert_mem_bar(Op_MemBarCPUOrder);
2566   insert_mem_bar(Op_MemBarAcquire);
2567 
2568   push(cas);
2569   return true;
2570 }
2571 
2572 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2573   // This is another variant of inline_unsafe_access, differing in
2574   // that it always issues store-store ("release") barrier and ensures
2575   // store-atomicity (which only matters for "long").
2576 
2577   if (callee()->is_static())  return false;  // caller must have the capability!
2578 
2579 #ifndef PRODUCT
2580   {
2581     ResourceMark rm;
2582     // Check the signatures.
2583     ciSignature* sig = signature();
2584 #ifdef ASSERT
2585     BasicType rtype = sig->return_type()->basic_type();
2586     assert(rtype == T_VOID, "must return void");
2587     assert(sig->count() == 3, "has 3 arguments");
2588     assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2589     assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2590 #endif // ASSERT
2591   }
2592 #endif //PRODUCT
2593 
2594   // number of stack slots per value argument (1 or 2)
2595   int type_words = type2size[type];
2596 
2597   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
2598 
2599   // Argument words:  "this" plus oop plus offset plus value;
2600   int nargs = 1 + 1 + 2 + type_words;
2601 
2602   // pop arguments: val, offset, base, and receiver
2603   debug_only(int saved_sp = _sp);
2604   _sp += nargs;
2605   Node* val      = (type_words == 1) ? pop() : pop_pair();
2606   Node *offset   = pop_pair();
2607   Node *base     = pop();
2608   Node *receiver = pop();
2609   assert(saved_sp == _sp, "must have correct argument count");
2610 
2611   //  Null check receiver.
2612   _sp += nargs;
2613   do_null_check(receiver, T_OBJECT);
2614   _sp -= nargs;
2615   if (stopped()) {
2616     return true;
2617   }
2618 
2619   // Build field offset expression.
2620   assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2621   // 32-bit machines ignore the high half of long offsets
2622   offset = ConvL2X(offset);
2623   Node* adr = make_unsafe_address(base, offset);
2624   const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2625   const Type *value_type = Type::get_const_basic_type(type);
2626   Compile::AliasType* alias_type = C->alias_type(adr_type);
2627 
2628   insert_mem_bar(Op_MemBarRelease);
2629   insert_mem_bar(Op_MemBarCPUOrder);
2630   // Ensure that the store is atomic for longs:
2631   bool require_atomic_access = true;
2632   Node* store;
2633   if (type == T_OBJECT) // reference stores need a store barrier.
2634     store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2635   else {
2636     store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2637   }
2638   insert_mem_bar(Op_MemBarCPUOrder);
2639   return true;
2640 }
2641 
2642 bool LibraryCallKit::inline_unsafe_allocate() {
2643   if (callee()->is_static())  return false;  // caller must have the capability!
2644   int nargs = 1 + 1;
2645   assert(signature()->size() == nargs-1, "alloc has 1 argument");
2646   null_check_receiver(callee());  // check then ignore argument(0)
2647   _sp += nargs;  // set original stack for use by uncommon_trap
2648   Node* cls = do_null_check(argument(1), T_OBJECT);
2649   _sp -= nargs;
2650   if (stopped())  return true;
2651 
2652   Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2653   _sp += nargs;  // set original stack for use by uncommon_trap
2654   kls = do_null_check(kls, T_OBJECT);
2655   _sp -= nargs;
2656   if (stopped())  return true;  // argument was like int.class
2657 
2658   // Note:  The argument might still be an illegal value like
2659   // Serializable.class or Object[].class.   The runtime will handle it.
2660   // But we must make an explicit check for initialization.
2661   Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2662   Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2663   Node* bits = intcon(instanceKlass::fully_initialized);
2664   Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2665   // The 'test' is non-zero if we need to take a slow path.
2666 
2667   Node* obj = new_instance(kls, test);
2668   push(obj);
2669 
2670   return true;
2671 }
2672 
2673 //------------------------inline_native_time_funcs--------------
2674 // inline code for System.currentTimeMillis() and System.nanoTime()
2675 // these have the same type and signature
2676 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2677   address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2678                               CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2679   const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2680   const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2681   const TypePtr* no_memory_effects = NULL;
2682   Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2683   Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2684 #ifdef ASSERT
2685   Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2686   assert(value_top == top(), "second value must be top");
2687 #endif
2688   push_pair(value);
2689   return true;
2690 }
2691 
2692 //------------------------inline_native_currentThread------------------
2693 bool LibraryCallKit::inline_native_currentThread() {
2694   Node* junk = NULL;
2695   push(generate_current_thread(junk));
2696   return true;
2697 }
2698 
2699 //------------------------inline_native_isInterrupted------------------
2700 bool LibraryCallKit::inline_native_isInterrupted() {
2701   const int nargs = 1+1;  // receiver + boolean
2702   assert(nargs == arg_size(), "sanity");
2703   // Add a fast path to t.isInterrupted(clear_int):
2704   //   (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2705   //   ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2706   // So, in the common case that the interrupt bit is false,
2707   // we avoid making a call into the VM.  Even if the interrupt bit
2708   // is true, if the clear_int argument is false, we avoid the VM call.
2709   // However, if the receiver is not currentThread, we must call the VM,
2710   // because there must be some locking done around the operation.
2711 
2712   // We only go to the fast case code if we pass two guards.
2713   // Paths which do not pass are accumulated in the slow_region.
2714   RegionNode* slow_region = new (C, 1) RegionNode(1);
2715   record_for_igvn(slow_region);
2716   RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2717   PhiNode*    result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2718   enum { no_int_result_path   = 1,
2719          no_clear_result_path = 2,
2720          slow_result_path     = 3
2721   };
2722 
2723   // (a) Receiving thread must be the current thread.
2724   Node* rec_thr = argument(0);
2725   Node* tls_ptr = NULL;
2726   Node* cur_thr = generate_current_thread(tls_ptr);
2727   Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2728   Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2729 
2730   bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2731   if (!known_current_thread)
2732     generate_slow_guard(bol_thr, slow_region);
2733 
2734   // (b) Interrupt bit on TLS must be false.
2735   Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2736   Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2737   p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2738   // Set the control input on the field _interrupted read to prevent it floating up.
2739   Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2740   Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2741   Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2742 
2743   IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2744 
2745   // First fast path:  if (!TLS._interrupted) return false;
2746   Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2747   result_rgn->init_req(no_int_result_path, false_bit);
2748   result_val->init_req(no_int_result_path, intcon(0));
2749 
2750   // drop through to next case
2751   set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2752 
2753   // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2754   Node* clr_arg = argument(1);
2755   Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2756   Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2757   IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2758 
2759   // Second fast path:  ... else if (!clear_int) return true;
2760   Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2761   result_rgn->init_req(no_clear_result_path, false_arg);
2762   result_val->init_req(no_clear_result_path, intcon(1));
2763 
2764   // drop through to next case
2765   set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2766 
2767   // (d) Otherwise, go to the slow path.
2768   slow_region->add_req(control());
2769   set_control( _gvn.transform(slow_region) );
2770 
2771   if (stopped()) {
2772     // There is no slow path.
2773     result_rgn->init_req(slow_result_path, top());
2774     result_val->init_req(slow_result_path, top());
2775   } else {
2776     // non-virtual because it is a private non-static
2777     CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2778 
2779     Node* slow_val = set_results_for_java_call(slow_call);
2780     // this->control() comes from set_results_for_java_call
2781 
2782     // If we know that the result of the slow call will be true, tell the optimizer!
2783     if (known_current_thread)  slow_val = intcon(1);
2784 
2785     Node* fast_io  = slow_call->in(TypeFunc::I_O);
2786     Node* fast_mem = slow_call->in(TypeFunc::Memory);
2787     // These two phis are pre-filled with copies of of the fast IO and Memory
2788     Node* io_phi   = PhiNode::make(result_rgn, fast_io,  Type::ABIO);
2789     Node* mem_phi  = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2790 
2791     result_rgn->init_req(slow_result_path, control());
2792     io_phi    ->init_req(slow_result_path, i_o());
2793     mem_phi   ->init_req(slow_result_path, reset_memory());
2794     result_val->init_req(slow_result_path, slow_val);
2795 
2796     set_all_memory( _gvn.transform(mem_phi) );
2797     set_i_o(        _gvn.transform(io_phi) );
2798   }
2799 
2800   push_result(result_rgn, result_val);
2801   C->set_has_split_ifs(true); // Has chance for split-if optimization
2802 
2803   return true;
2804 }
2805 
2806 //---------------------------load_mirror_from_klass----------------------------
2807 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2808 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2809   Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2810   return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2811 }
2812 
2813 //-----------------------load_klass_from_mirror_common-------------------------
2814 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2815 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2816 // and branch to the given path on the region.
2817 // If never_see_null, take an uncommon trap on null, so we can optimistically
2818 // compile for the non-null case.
2819 // If the region is NULL, force never_see_null = true.
2820 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2821                                                     bool never_see_null,
2822                                                     int nargs,
2823                                                     RegionNode* region,
2824                                                     int null_path,
2825                                                     int offset) {
2826   if (region == NULL)  never_see_null = true;
2827   Node* p = basic_plus_adr(mirror, offset);
2828   const TypeKlassPtr*  kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2829   Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2830   _sp += nargs; // any deopt will start just before call to enclosing method
2831   Node* null_ctl = top();
2832   kls = null_check_oop(kls, &null_ctl, never_see_null);
2833   if (region != NULL) {
2834     // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2835     region->init_req(null_path, null_ctl);
2836   } else {
2837     assert(null_ctl == top(), "no loose ends");
2838   }
2839   _sp -= nargs;
2840   return kls;
2841 }
2842 
2843 //--------------------(inline_native_Class_query helpers)---------------------
2844 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2845 // Fall through if (mods & mask) == bits, take the guard otherwise.
2846 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2847   // Branch around if the given klass has the given modifier bit set.
2848   // Like generate_guard, adds a new path onto the region.
2849   Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2850   Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2851   Node* mask = intcon(modifier_mask);
2852   Node* bits = intcon(modifier_bits);
2853   Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2854   Node* cmp  = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2855   Node* bol  = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2856   return generate_fair_guard(bol, region);
2857 }
2858 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2859   return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2860 }
2861 
2862 //-------------------------inline_native_Class_query-------------------
2863 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2864   int nargs = 1+0;  // just the Class mirror, in most cases
2865   const Type* return_type = TypeInt::BOOL;
2866   Node* prim_return_value = top();  // what happens if it's a primitive class?
2867   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2868   bool expect_prim = false;     // most of these guys expect to work on refs
2869 
2870   enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2871 
2872   switch (id) {
2873   case vmIntrinsics::_isInstance:
2874     nargs = 1+1;  // the Class mirror, plus the object getting queried about
2875     // nothing is an instance of a primitive type
2876     prim_return_value = intcon(0);
2877     break;
2878   case vmIntrinsics::_getModifiers:
2879     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2880     assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2881     return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2882     break;
2883   case vmIntrinsics::_isInterface:
2884     prim_return_value = intcon(0);
2885     break;
2886   case vmIntrinsics::_isArray:
2887     prim_return_value = intcon(0);
2888     expect_prim = true;  // cf. ObjectStreamClass.getClassSignature
2889     break;
2890   case vmIntrinsics::_isPrimitive:
2891     prim_return_value = intcon(1);
2892     expect_prim = true;  // obviously
2893     break;
2894   case vmIntrinsics::_getSuperclass:
2895     prim_return_value = null();
2896     return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2897     break;
2898   case vmIntrinsics::_getComponentType:
2899     prim_return_value = null();
2900     return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2901     break;
2902   case vmIntrinsics::_getClassAccessFlags:
2903     prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2904     return_type = TypeInt::INT;  // not bool!  6297094
2905     break;
2906   default:
2907     ShouldNotReachHere();
2908   }
2909 
2910   Node* mirror =                      argument(0);
2911   Node* obj    = (nargs <= 1)? top(): argument(1);
2912 
2913   const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2914   if (mirror_con == NULL)  return false;  // cannot happen?
2915 
2916 #ifndef PRODUCT
2917   if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2918     ciType* k = mirror_con->java_mirror_type();
2919     if (k) {
2920       tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2921       k->print_name();
2922       tty->cr();
2923     }
2924   }
2925 #endif
2926 
2927   // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2928   RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2929   record_for_igvn(region);
2930   PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2931 
2932   // The mirror will never be null of Reflection.getClassAccessFlags, however
2933   // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2934   // if it is. See bug 4774291.
2935 
2936   // For Reflection.getClassAccessFlags(), the null check occurs in
2937   // the wrong place; see inline_unsafe_access(), above, for a similar
2938   // situation.
2939   _sp += nargs;  // set original stack for use by uncommon_trap
2940   mirror = do_null_check(mirror, T_OBJECT);
2941   _sp -= nargs;
2942   // If mirror or obj is dead, only null-path is taken.
2943   if (stopped())  return true;
2944 
2945   if (expect_prim)  never_see_null = false;  // expect nulls (meaning prims)
2946 
2947   // Now load the mirror's klass metaobject, and null-check it.
2948   // Side-effects region with the control path if the klass is null.
2949   Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2950                                      region, _prim_path);
2951   // If kls is null, we have a primitive mirror.
2952   phi->init_req(_prim_path, prim_return_value);
2953   if (stopped()) { push_result(region, phi); return true; }
2954 
2955   Node* p;  // handy temp
2956   Node* null_ctl;
2957 
2958   // Now that we have the non-null klass, we can perform the real query.
2959   // For constant classes, the query will constant-fold in LoadNode::Value.
2960   Node* query_value = top();
2961   switch (id) {
2962   case vmIntrinsics::_isInstance:
2963     // nothing is an instance of a primitive type
2964     _sp += nargs;          // gen_instanceof might do an uncommon trap
2965     query_value = gen_instanceof(obj, kls);
2966     _sp -= nargs;
2967     break;
2968 
2969   case vmIntrinsics::_getModifiers:
2970     p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2971     query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2972     break;
2973 
2974   case vmIntrinsics::_isInterface:
2975     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2976     if (generate_interface_guard(kls, region) != NULL)
2977       // A guard was added.  If the guard is taken, it was an interface.
2978       phi->add_req(intcon(1));
2979     // If we fall through, it's a plain class.
2980     query_value = intcon(0);
2981     break;
2982 
2983   case vmIntrinsics::_isArray:
2984     // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2985     if (generate_array_guard(kls, region) != NULL)
2986       // A guard was added.  If the guard is taken, it was an array.
2987       phi->add_req(intcon(1));
2988     // If we fall through, it's a plain class.
2989     query_value = intcon(0);
2990     break;
2991 
2992   case vmIntrinsics::_isPrimitive:
2993     query_value = intcon(0); // "normal" path produces false
2994     break;
2995 
2996   case vmIntrinsics::_getSuperclass:
2997     // The rules here are somewhat unfortunate, but we can still do better
2998     // with random logic than with a JNI call.
2999     // Interfaces store null or Object as _super, but must report null.
3000     // Arrays store an intermediate super as _super, but must report Object.
3001     // Other types can report the actual _super.
3002     // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3003     if (generate_interface_guard(kls, region) != NULL)
3004       // A guard was added.  If the guard is taken, it was an interface.
3005       phi->add_req(null());
3006     if (generate_array_guard(kls, region) != NULL)
3007       // A guard was added.  If the guard is taken, it was an array.
3008       phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3009     // If we fall through, it's a plain class.  Get its _super.
3010     p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
3011     kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3012     null_ctl = top();
3013     kls = null_check_oop(kls, &null_ctl);
3014     if (null_ctl != top()) {
3015       // If the guard is taken, Object.superClass is null (both klass and mirror).
3016       region->add_req(null_ctl);
3017       phi   ->add_req(null());
3018     }
3019     if (!stopped()) {
3020       query_value = load_mirror_from_klass(kls);
3021     }
3022     break;
3023 
3024   case vmIntrinsics::_getComponentType:
3025     if (generate_array_guard(kls, region) != NULL) {
3026       // Be sure to pin the oop load to the guard edge just created:
3027       Node* is_array_ctrl = region->in(region->req()-1);
3028       Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
3029       Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3030       phi->add_req(cmo);
3031     }
3032     query_value = null();  // non-array case is null
3033     break;
3034 
3035   case vmIntrinsics::_getClassAccessFlags:
3036     p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
3037     query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3038     break;
3039 
3040   default:
3041     ShouldNotReachHere();
3042   }
3043 
3044   // Fall-through is the normal case of a query to a real class.
3045   phi->init_req(1, query_value);
3046   region->init_req(1, control());
3047 
3048   push_result(region, phi);
3049   C->set_has_split_ifs(true); // Has chance for split-if optimization
3050 
3051   return true;
3052 }
3053 
3054 //--------------------------inline_native_subtype_check------------------------
3055 // This intrinsic takes the JNI calls out of the heart of
3056 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3057 bool LibraryCallKit::inline_native_subtype_check() {
3058   int nargs = 1+1;  // the Class mirror, plus the other class getting examined
3059 
3060   // Pull both arguments off the stack.
3061   Node* args[2];                // two java.lang.Class mirrors: superc, subc
3062   args[0] = argument(0);
3063   args[1] = argument(1);
3064   Node* klasses[2];             // corresponding Klasses: superk, subk
3065   klasses[0] = klasses[1] = top();
3066 
3067   enum {
3068     // A full decision tree on {superc is prim, subc is prim}:
3069     _prim_0_path = 1,           // {P,N} => false
3070                                 // {P,P} & superc!=subc => false
3071     _prim_same_path,            // {P,P} & superc==subc => true
3072     _prim_1_path,               // {N,P} => false
3073     _ref_subtype_path,          // {N,N} & subtype check wins => true
3074     _both_ref_path,             // {N,N} & subtype check loses => false
3075     PATH_LIMIT
3076   };
3077 
3078   RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3079   Node*       phi    = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3080   record_for_igvn(region);
3081 
3082   const TypePtr* adr_type = TypeRawPtr::BOTTOM;   // memory type of loads
3083   const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3084   int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3085 
3086   // First null-check both mirrors and load each mirror's klass metaobject.
3087   int which_arg;
3088   for (which_arg = 0; which_arg <= 1; which_arg++) {
3089     Node* arg = args[which_arg];
3090     _sp += nargs;  // set original stack for use by uncommon_trap
3091     arg = do_null_check(arg, T_OBJECT);
3092     _sp -= nargs;
3093     if (stopped())  break;
3094     args[which_arg] = _gvn.transform(arg);
3095 
3096     Node* p = basic_plus_adr(arg, class_klass_offset);
3097     Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3098     klasses[which_arg] = _gvn.transform(kls);
3099   }
3100 
3101   // Having loaded both klasses, test each for null.
3102   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3103   for (which_arg = 0; which_arg <= 1; which_arg++) {
3104     Node* kls = klasses[which_arg];
3105     Node* null_ctl = top();
3106     _sp += nargs;  // set original stack for use by uncommon_trap
3107     kls = null_check_oop(kls, &null_ctl, never_see_null);
3108     _sp -= nargs;
3109     int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3110     region->init_req(prim_path, null_ctl);
3111     if (stopped())  break;
3112     klasses[which_arg] = kls;
3113   }
3114 
3115   if (!stopped()) {
3116     // now we have two reference types, in klasses[0..1]
3117     Node* subk   = klasses[1];  // the argument to isAssignableFrom
3118     Node* superk = klasses[0];  // the receiver
3119     region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3120     // now we have a successful reference subtype check
3121     region->set_req(_ref_subtype_path, control());
3122   }
3123 
3124   // If both operands are primitive (both klasses null), then
3125   // we must return true when they are identical primitives.
3126   // It is convenient to test this after the first null klass check.
3127   set_control(region->in(_prim_0_path)); // go back to first null check
3128   if (!stopped()) {
3129     // Since superc is primitive, make a guard for the superc==subc case.
3130     Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3131     Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3132     generate_guard(bol_eq, region, PROB_FAIR);
3133     if (region->req() == PATH_LIMIT+1) {
3134       // A guard was added.  If the added guard is taken, superc==subc.
3135       region->swap_edges(PATH_LIMIT, _prim_same_path);
3136       region->del_req(PATH_LIMIT);
3137     }
3138     region->set_req(_prim_0_path, control()); // Not equal after all.
3139   }
3140 
3141   // these are the only paths that produce 'true':
3142   phi->set_req(_prim_same_path,   intcon(1));
3143   phi->set_req(_ref_subtype_path, intcon(1));
3144 
3145   // pull together the cases:
3146   assert(region->req() == PATH_LIMIT, "sane region");
3147   for (uint i = 1; i < region->req(); i++) {
3148     Node* ctl = region->in(i);
3149     if (ctl == NULL || ctl == top()) {
3150       region->set_req(i, top());
3151       phi   ->set_req(i, top());
3152     } else if (phi->in(i) == NULL) {
3153       phi->set_req(i, intcon(0)); // all other paths produce 'false'
3154     }
3155   }
3156 
3157   set_control(_gvn.transform(region));
3158   push(_gvn.transform(phi));
3159 
3160   return true;
3161 }
3162 
3163 //---------------------generate_array_guard_common------------------------
3164 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3165                                                   bool obj_array, bool not_array) {
3166   // If obj_array/non_array==false/false:
3167   // Branch around if the given klass is in fact an array (either obj or prim).
3168   // If obj_array/non_array==false/true:
3169   // Branch around if the given klass is not an array klass of any kind.
3170   // If obj_array/non_array==true/true:
3171   // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3172   // If obj_array/non_array==true/false:
3173   // Branch around if the kls is an oop array (Object[] or subtype)
3174   //
3175   // Like generate_guard, adds a new path onto the region.
3176   jint  layout_con = 0;
3177   Node* layout_val = get_layout_helper(kls, layout_con);
3178   if (layout_val == NULL) {
3179     bool query = (obj_array
3180                   ? Klass::layout_helper_is_objArray(layout_con)
3181                   : Klass::layout_helper_is_javaArray(layout_con));
3182     if (query == not_array) {
3183       return NULL;                       // never a branch
3184     } else {                             // always a branch
3185       Node* always_branch = control();
3186       if (region != NULL)
3187         region->add_req(always_branch);
3188       set_control(top());
3189       return always_branch;
3190     }
3191   }
3192   // Now test the correct condition.
3193   jint  nval = (obj_array
3194                 ? ((jint)Klass::_lh_array_tag_type_value
3195                    <<    Klass::_lh_array_tag_shift)
3196                 : Klass::_lh_neutral_value);
3197   Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3198   BoolTest::mask btest = BoolTest::lt;  // correct for testing is_[obj]array
3199   // invert the test if we are looking for a non-array
3200   if (not_array)  btest = BoolTest(btest).negate();
3201   Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3202   return generate_fair_guard(bol, region);
3203 }
3204 
3205 
3206 //-----------------------inline_native_newArray--------------------------
3207 bool LibraryCallKit::inline_native_newArray() {
3208   int nargs = 2;
3209   Node* mirror    = argument(0);
3210   Node* count_val = argument(1);
3211 
3212   _sp += nargs;  // set original stack for use by uncommon_trap
3213   mirror = do_null_check(mirror, T_OBJECT);
3214   _sp -= nargs;
3215   // If mirror or obj is dead, only null-path is taken.
3216   if (stopped())  return true;
3217 
3218   enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3219   RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3220   PhiNode*    result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3221                                                       TypeInstPtr::NOTNULL);
3222   PhiNode*    result_io  = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3223   PhiNode*    result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3224                                                       TypePtr::BOTTOM);
3225 
3226   bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3227   Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3228                                                   nargs,
3229                                                   result_reg, _slow_path);
3230   Node* normal_ctl   = control();
3231   Node* no_array_ctl = result_reg->in(_slow_path);
3232 
3233   // Generate code for the slow case.  We make a call to newArray().
3234   set_control(no_array_ctl);
3235   if (!stopped()) {
3236     // Either the input type is void.class, or else the
3237     // array klass has not yet been cached.  Either the
3238     // ensuing call will throw an exception, or else it
3239     // will cache the array klass for next time.
3240     PreserveJVMState pjvms(this);
3241     CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3242     Node* slow_result = set_results_for_java_call(slow_call);
3243     // this->control() comes from set_results_for_java_call
3244     result_reg->set_req(_slow_path, control());
3245     result_val->set_req(_slow_path, slow_result);
3246     result_io ->set_req(_slow_path, i_o());
3247     result_mem->set_req(_slow_path, reset_memory());
3248   }
3249 
3250   set_control(normal_ctl);
3251   if (!stopped()) {
3252     // Normal case:  The array type has been cached in the java.lang.Class.
3253     // The following call works fine even if the array type is polymorphic.
3254     // It could be a dynamic mix of int[], boolean[], Object[], etc.
3255     Node* obj = new_array(klass_node, count_val, nargs);
3256     result_reg->init_req(_normal_path, control());
3257     result_val->init_req(_normal_path, obj);
3258     result_io ->init_req(_normal_path, i_o());
3259     result_mem->init_req(_normal_path, reset_memory());
3260   }
3261 
3262   // Return the combined state.
3263   set_i_o(        _gvn.transform(result_io)  );
3264   set_all_memory( _gvn.transform(result_mem) );
3265   push_result(result_reg, result_val);
3266   C->set_has_split_ifs(true); // Has chance for split-if optimization
3267 
3268   return true;
3269 }
3270 
3271 //----------------------inline_native_getLength--------------------------
3272 bool LibraryCallKit::inline_native_getLength() {
3273   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3274 
3275   int nargs = 1;
3276   Node* array = argument(0);
3277 
3278   _sp += nargs;  // set original stack for use by uncommon_trap
3279   array = do_null_check(array, T_OBJECT);
3280   _sp -= nargs;
3281 
3282   // If array is dead, only null-path is taken.
3283   if (stopped())  return true;
3284 
3285   // Deoptimize if it is a non-array.
3286   Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3287 
3288   if (non_array != NULL) {
3289     PreserveJVMState pjvms(this);
3290     set_control(non_array);
3291     _sp += nargs;  // push the arguments back on the stack
3292     uncommon_trap(Deoptimization::Reason_intrinsic,
3293                   Deoptimization::Action_maybe_recompile);
3294   }
3295 
3296   // If control is dead, only non-array-path is taken.
3297   if (stopped())  return true;
3298 
3299   // The works fine even if the array type is polymorphic.
3300   // It could be a dynamic mix of int[], boolean[], Object[], etc.
3301   push( load_array_length(array) );
3302 
3303   C->set_has_split_ifs(true); // Has chance for split-if optimization
3304 
3305   return true;
3306 }
3307 
3308 //------------------------inline_array_copyOf----------------------------
3309 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3310   if (too_many_traps(Deoptimization::Reason_intrinsic))  return false;
3311 
3312   // Restore the stack and pop off the arguments.
3313   int nargs = 3 + (is_copyOfRange? 1: 0);
3314   Node* original          = argument(0);
3315   Node* start             = is_copyOfRange? argument(1): intcon(0);
3316   Node* end               = is_copyOfRange? argument(2): argument(1);
3317   Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3318 
3319   Node* newcopy;
3320 
3321   //set the original stack and the reexecute bit for the interpreter to reexecute
3322   //the bytecode that invokes Arrays.copyOf if deoptimization happens
3323   { PreserveReexecuteState preexecs(this);
3324     _sp += nargs;
3325     jvms()->set_should_reexecute(true);
3326 
3327     array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3328     original          = do_null_check(original, T_OBJECT);
3329 
3330     // Check if a null path was taken unconditionally.
3331     if (stopped())  return true;
3332 
3333     Node* orig_length = load_array_length(original);
3334 
3335     Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3336                                               NULL, 0);
3337     klass_node = do_null_check(klass_node, T_OBJECT);
3338 
3339     RegionNode* bailout = new (C, 1) RegionNode(1);
3340     record_for_igvn(bailout);
3341 
3342     // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3343     // Bail out if that is so.
3344     Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3345     if (not_objArray != NULL) {
3346       // Improve the klass node's type from the new optimistic assumption:
3347       ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3348       const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3349       Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3350       cast->init_req(0, control());
3351       klass_node = _gvn.transform(cast);
3352     }
3353 
3354     // Bail out if either start or end is negative.
3355     generate_negative_guard(start, bailout, &start);
3356     generate_negative_guard(end,   bailout, &end);
3357 
3358     Node* length = end;
3359     if (_gvn.type(start) != TypeInt::ZERO) {
3360       length = _gvn.transform( new (C, 3) SubINode(end, start) );
3361     }
3362 
3363     // Bail out if length is negative.
3364     // ...Not needed, since the new_array will throw the right exception.
3365     //generate_negative_guard(length, bailout, &length);
3366 
3367     if (bailout->req() > 1) {
3368       PreserveJVMState pjvms(this);
3369       set_control( _gvn.transform(bailout) );
3370       uncommon_trap(Deoptimization::Reason_intrinsic,
3371                     Deoptimization::Action_maybe_recompile);
3372     }
3373 
3374     if (!stopped()) {
3375 
3376       // How many elements will we copy from the original?
3377       // The answer is MinI(orig_length - start, length).
3378       Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3379       Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3380 
3381       const bool raw_mem_only = true;
3382       newcopy = new_array(klass_node, length, 0, raw_mem_only);
3383 
3384       // Generate a direct call to the right arraycopy function(s).
3385       // We know the copy is disjoint but we might not know if the
3386       // oop stores need checking.
3387       // Extreme case:  Arrays.copyOf((Integer[])x, 10, String[].class).
3388       // This will fail a store-check if x contains any non-nulls.
3389       bool disjoint_bases = true;
3390       bool length_never_negative = true;
3391       generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3392                          original, start, newcopy, intcon(0), moved,
3393                          disjoint_bases, length_never_negative);
3394     }
3395   } //original reexecute and sp are set back here
3396 
3397   if(!stopped()) {
3398     push(newcopy);
3399   }
3400 
3401   C->set_has_split_ifs(true); // Has chance for split-if optimization
3402 
3403   return true;
3404 }
3405 
3406 
3407 //----------------------generate_virtual_guard---------------------------
3408 // Helper for hashCode and clone.  Peeks inside the vtable to avoid a call.
3409 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3410                                              RegionNode* slow_region) {
3411   ciMethod* method = callee();
3412   int vtable_index = method->vtable_index();
3413   // Get the methodOop out of the appropriate vtable entry.
3414   int entry_offset  = (instanceKlass::vtable_start_offset() +
3415                      vtable_index*vtableEntry::size()) * wordSize +
3416                      vtableEntry::method_offset_in_bytes();
3417   Node* entry_addr  = basic_plus_adr(obj_klass, entry_offset);
3418   Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3419 
3420   // Compare the target method with the expected method (e.g., Object.hashCode).
3421   const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3422 
3423   Node* native_call = makecon(native_call_addr);
3424   Node* chk_native  = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3425   Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3426 
3427   return generate_slow_guard(test_native, slow_region);
3428 }
3429 
3430 //-----------------------generate_method_call----------------------------
3431 // Use generate_method_call to make a slow-call to the real
3432 // method if the fast path fails.  An alternative would be to
3433 // use a stub like OptoRuntime::slow_arraycopy_Java.
3434 // This only works for expanding the current library call,
3435 // not another intrinsic.  (E.g., don't use this for making an
3436 // arraycopy call inside of the copyOf intrinsic.)
3437 CallJavaNode*
3438 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3439   // When compiling the intrinsic method itself, do not use this technique.
3440   guarantee(callee() != C->method(), "cannot make slow-call to self");
3441 
3442   ciMethod* method = callee();
3443   // ensure the JVMS we have will be correct for this call
3444   guarantee(method_id == method->intrinsic_id(), "must match");
3445 
3446   const TypeFunc* tf = TypeFunc::make(method);
3447   int tfdc = tf->domain()->cnt();
3448   CallJavaNode* slow_call;
3449   if (is_static) {
3450     assert(!is_virtual, "");
3451     slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3452                                 SharedRuntime::get_resolve_static_call_stub(),
3453                                 method, bci());
3454   } else if (is_virtual) {
3455     null_check_receiver(method);
3456     int vtable_index = methodOopDesc::invalid_vtable_index;
3457     if (UseInlineCaches) {
3458       // Suppress the vtable call
3459     } else {
3460       // hashCode and clone are not a miranda methods,
3461       // so the vtable index is fixed.
3462       // No need to use the linkResolver to get it.
3463        vtable_index = method->vtable_index();
3464     }
3465     slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3466                                 SharedRuntime::get_resolve_virtual_call_stub(),
3467                                 method, vtable_index, bci());
3468   } else {  // neither virtual nor static:  opt_virtual
3469     null_check_receiver(method);
3470     slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3471                                 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3472                                 method, bci());
3473     slow_call->set_optimized_virtual(true);
3474   }
3475   set_arguments_for_java_call(slow_call);
3476   set_edges_for_java_call(slow_call);
3477   return slow_call;
3478 }
3479 
3480 
3481 //------------------------------inline_native_hashcode--------------------
3482 // Build special case code for calls to hashCode on an object.
3483 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3484   assert(is_static == callee()->is_static(), "correct intrinsic selection");
3485   assert(!(is_virtual && is_static), "either virtual, special, or static");
3486 
3487   enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3488 
3489   RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3490   PhiNode*    result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3491                                                       TypeInt::INT);
3492   PhiNode*    result_io  = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3493   PhiNode*    result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3494                                                       TypePtr::BOTTOM);
3495   Node* obj = NULL;
3496   if (!is_static) {
3497     // Check for hashing null object
3498     obj = null_check_receiver(callee());
3499     if (stopped())  return true;        // unconditionally null
3500     result_reg->init_req(_null_path, top());
3501     result_val->init_req(_null_path, top());
3502   } else {
3503     // Do a null check, and return zero if null.
3504     // System.identityHashCode(null) == 0
3505     obj = argument(0);
3506     Node* null_ctl = top();
3507     obj = null_check_oop(obj, &null_ctl);
3508     result_reg->init_req(_null_path, null_ctl);
3509     result_val->init_req(_null_path, _gvn.intcon(0));
3510   }
3511 
3512   // Unconditionally null?  Then return right away.
3513   if (stopped()) {
3514     set_control( result_reg->in(_null_path) );
3515     if (!stopped())
3516       push(      result_val ->in(_null_path) );
3517     return true;
3518   }
3519 
3520   // After null check, get the object's klass.
3521   Node* obj_klass = load_object_klass(obj);
3522 
3523   // This call may be virtual (invokevirtual) or bound (invokespecial).
3524   // For each case we generate slightly different code.
3525 
3526   // We only go to the fast case code if we pass a number of guards.  The
3527   // paths which do not pass are accumulated in the slow_region.
3528   RegionNode* slow_region = new (C, 1) RegionNode(1);
3529   record_for_igvn(slow_region);
3530 
3531   // If this is a virtual call, we generate a funny guard.  We pull out
3532   // the vtable entry corresponding to hashCode() from the target object.
3533   // If the target method which we are calling happens to be the native
3534   // Object hashCode() method, we pass the guard.  We do not need this
3535   // guard for non-virtual calls -- the caller is known to be the native
3536   // Object hashCode().
3537   if (is_virtual) {
3538     generate_virtual_guard(obj_klass, slow_region);
3539   }
3540 
3541   // Get the header out of the object, use LoadMarkNode when available
3542   Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3543   Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3544 
3545   // Test the header to see if it is unlocked.
3546   Node *lock_mask      = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3547   Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3548   Node *unlocked_val   = _gvn.MakeConX(markOopDesc::unlocked_value);
3549   Node *chk_unlocked   = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3550   Node *test_unlocked  = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3551 
3552   generate_slow_guard(test_unlocked, slow_region);
3553 
3554   // Get the hash value and check to see that it has been properly assigned.
3555   // We depend on hash_mask being at most 32 bits and avoid the use of
3556   // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3557   // vm: see markOop.hpp.
3558   Node *hash_mask      = _gvn.intcon(markOopDesc::hash_mask);
3559   Node *hash_shift     = _gvn.intcon(markOopDesc::hash_shift);
3560   Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3561   // This hack lets the hash bits live anywhere in the mark object now, as long
3562   // as the shift drops the relevant bits into the low 32 bits.  Note that
3563   // Java spec says that HashCode is an int so there's no point in capturing
3564   // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3565   hshifted_header      = ConvX2I(hshifted_header);
3566   Node *hash_val       = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3567 
3568   Node *no_hash_val    = _gvn.intcon(markOopDesc::no_hash);
3569   Node *chk_assigned   = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3570   Node *test_assigned  = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3571 
3572   generate_slow_guard(test_assigned, slow_region);
3573 
3574   Node* init_mem = reset_memory();
3575   // fill in the rest of the null path:
3576   result_io ->init_req(_null_path, i_o());
3577   result_mem->init_req(_null_path, init_mem);
3578 
3579   result_val->init_req(_fast_path, hash_val);
3580   result_reg->init_req(_fast_path, control());
3581   result_io ->init_req(_fast_path, i_o());
3582   result_mem->init_req(_fast_path, init_mem);
3583 
3584   // Generate code for the slow case.  We make a call to hashCode().
3585   set_control(_gvn.transform(slow_region));
3586   if (!stopped()) {
3587     // No need for PreserveJVMState, because we're using up the present state.
3588     set_all_memory(init_mem);
3589     vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3590     if (is_static)   hashCode_id = vmIntrinsics::_identityHashCode;
3591     CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3592     Node* slow_result = set_results_for_java_call(slow_call);
3593     // this->control() comes from set_results_for_java_call
3594     result_reg->init_req(_slow_path, control());
3595     result_val->init_req(_slow_path, slow_result);
3596     result_io  ->set_req(_slow_path, i_o());
3597     result_mem ->set_req(_slow_path, reset_memory());
3598   }
3599 
3600   // Return the combined state.
3601   set_i_o(        _gvn.transform(result_io)  );
3602   set_all_memory( _gvn.transform(result_mem) );
3603   push_result(result_reg, result_val);
3604 
3605   return true;
3606 }
3607 
3608 //---------------------------inline_native_getClass----------------------------
3609 // Build special case code for calls to getClass on an object.
3610 bool LibraryCallKit::inline_native_getClass() {
3611   Node* obj = null_check_receiver(callee());
3612   if (stopped())  return true;
3613   push( load_mirror_from_klass(load_object_klass(obj)) );
3614   return true;
3615 }
3616 
3617 //-----------------inline_native_Reflection_getCallerClass---------------------
3618 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3619 //
3620 // NOTE that this code must perform the same logic as
3621 // vframeStream::security_get_caller_frame in that it must skip
3622 // Method.invoke() and auxiliary frames.
3623 
3624 
3625 
3626 
3627 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3628   ciMethod*       method = callee();
3629 
3630 #ifndef PRODUCT
3631   if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3632     tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3633   }
3634 #endif
3635 
3636   debug_only(int saved_sp = _sp);
3637 
3638   // Argument words:  (int depth)
3639   int nargs = 1;
3640 
3641   _sp += nargs;
3642   Node* caller_depth_node = pop();
3643 
3644   assert(saved_sp == _sp, "must have correct argument count");
3645 
3646   // The depth value must be a constant in order for the runtime call
3647   // to be eliminated.
3648   const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3649   if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3650 #ifndef PRODUCT
3651     if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3652       tty->print_cr("  Bailing out because caller depth was not a constant");
3653     }
3654 #endif
3655     return false;
3656   }
3657   // Note that the JVM state at this point does not include the
3658   // getCallerClass() frame which we are trying to inline. The
3659   // semantics of getCallerClass(), however, are that the "first"
3660   // frame is the getCallerClass() frame, so we subtract one from the
3661   // requested depth before continuing. We don't inline requests of
3662   // getCallerClass(0).
3663   int caller_depth = caller_depth_type->get_con() - 1;
3664   if (caller_depth < 0) {
3665 #ifndef PRODUCT
3666     if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3667       tty->print_cr("  Bailing out because caller depth was %d", caller_depth);
3668     }
3669 #endif
3670     return false;
3671   }
3672 
3673   if (!jvms()->has_method()) {
3674 #ifndef PRODUCT
3675     if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3676       tty->print_cr("  Bailing out because intrinsic was inlined at top level");
3677     }
3678 #endif
3679     return false;
3680   }
3681   int _depth = jvms()->depth();  // cache call chain depth
3682 
3683   // Walk back up the JVM state to find the caller at the required
3684   // depth. NOTE that this code must perform the same logic as
3685   // vframeStream::security_get_caller_frame in that it must skip
3686   // Method.invoke() and auxiliary frames. Note also that depth is
3687   // 1-based (1 is the bottom of the inlining).
3688   int inlining_depth = _depth;
3689   JVMState* caller_jvms = NULL;
3690 
3691   if (inlining_depth > 0) {
3692     caller_jvms = jvms();
3693     assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3694     do {
3695       // The following if-tests should be performed in this order
3696       if (is_method_invoke_or_aux_frame(caller_jvms)) {
3697         // Skip a Method.invoke() or auxiliary frame
3698       } else if (caller_depth > 0) {
3699         // Skip real frame
3700         --caller_depth;
3701       } else {
3702         // We're done: reached desired caller after skipping.
3703         break;
3704       }
3705       caller_jvms = caller_jvms->caller();
3706       --inlining_depth;
3707     } while (inlining_depth > 0);
3708   }
3709 
3710   if (inlining_depth == 0) {
3711 #ifndef PRODUCT
3712     if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3713       tty->print_cr("  Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3714       tty->print_cr("  JVM state at this point:");
3715       for (int i = _depth; i >= 1; i--) {
3716         tty->print_cr("   %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3717       }
3718     }
3719 #endif
3720     return false; // Reached end of inlining
3721   }
3722 
3723   // Acquire method holder as java.lang.Class
3724   ciInstanceKlass* caller_klass  = caller_jvms->method()->holder();
3725   ciInstance*      caller_mirror = caller_klass->java_mirror();
3726   // Push this as a constant
3727   push(makecon(TypeInstPtr::make(caller_mirror)));
3728 #ifndef PRODUCT
3729   if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3730     tty->print_cr("  Succeeded: caller = %s.%s, caller depth = %d, depth = %d", caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), caller_depth_type->get_con(), _depth);
3731     tty->print_cr("  JVM state at this point:");
3732     for (int i = _depth; i >= 1; i--) {
3733       tty->print_cr("   %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3734     }
3735   }
3736 #endif
3737   return true;
3738 }
3739 
3740 // Helper routine for above
3741 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3742   ciMethod* method = jvms->method();
3743 
3744   // Is this the Method.invoke method itself?
3745   if (method->intrinsic_id() == vmIntrinsics::_invoke)
3746     return true;
3747 
3748   // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3749   ciKlass* k = method->holder();
3750   if (k->is_instance_klass()) {
3751     ciInstanceKlass* ik = k->as_instance_klass();
3752     for (; ik != NULL; ik = ik->super()) {
3753       if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3754           ik == env()->find_system_klass(ik->name())) {
3755         return true;
3756       }
3757     }
3758   }
3759   else if (method->is_method_handle_adapter()) {
3760     // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3761     return true;
3762   }
3763 
3764   return false;
3765 }
3766 
3767 static int value_field_offset = -1;  // offset of the "value" field of AtomicLongCSImpl.  This is needed by
3768                                      // inline_native_AtomicLong_attemptUpdate() but it has no way of
3769                                      // computing it since there is no lookup field by name function in the
3770                                      // CI interface.  This is computed and set by inline_native_AtomicLong_get().
3771                                      // Using a static variable here is safe even if we have multiple compilation
3772                                      // threads because the offset is constant.  At worst the same offset will be
3773                                      // computed and  stored multiple
3774 
3775 bool LibraryCallKit::inline_native_AtomicLong_get() {
3776   // Restore the stack and pop off the argument
3777   _sp+=1;
3778   Node *obj = pop();
3779 
3780   // get the offset of the "value" field. Since the CI interfaces
3781   // does not provide a way to look up a field by name, we scan the bytecodes
3782   // to get the field index.  We expect the first 2 instructions of the method
3783   // to be:
3784   //    0 aload_0
3785   //    1 getfield "value"
3786   ciMethod* method = callee();
3787   if (value_field_offset == -1)
3788   {
3789     ciField* value_field;
3790     ciBytecodeStream iter(method);
3791     Bytecodes::Code bc = iter.next();
3792 
3793     if ((bc != Bytecodes::_aload_0) &&
3794               ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3795       return false;
3796     bc = iter.next();
3797     if (bc != Bytecodes::_getfield)
3798       return false;
3799     bool ignore;
3800     value_field = iter.get_field(ignore);
3801     value_field_offset = value_field->offset_in_bytes();
3802   }
3803 
3804   // Null check without removing any arguments.
3805   _sp++;
3806   obj = do_null_check(obj, T_OBJECT);
3807   _sp--;
3808   // Check for locking null object
3809   if (stopped()) return true;
3810 
3811   Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3812   const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3813   int alias_idx = C->get_alias_index(adr_type);
3814 
3815   Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3816 
3817   push_pair(result);
3818 
3819   return true;
3820 }
3821 
3822 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3823   // Restore the stack and pop off the arguments
3824   _sp+=5;
3825   Node *newVal = pop_pair();
3826   Node *oldVal = pop_pair();
3827   Node *obj = pop();
3828 
3829   // we need the offset of the "value" field which was computed when
3830   // inlining the get() method.  Give up if we don't have it.
3831   if (value_field_offset == -1)
3832     return false;
3833 
3834   // Null check without removing any arguments.
3835   _sp+=5;
3836   obj = do_null_check(obj, T_OBJECT);
3837   _sp-=5;
3838   // Check for locking null object
3839   if (stopped()) return true;
3840 
3841   Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3842   const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3843   int alias_idx = C->get_alias_index(adr_type);
3844 
3845   Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3846   Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3847   set_memory(store_proj, alias_idx);
3848   Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3849 
3850   Node *result;
3851   // CMove node is not used to be able fold a possible check code
3852   // after attemptUpdate() call. This code could be transformed
3853   // into CMove node by loop optimizations.
3854   {
3855     RegionNode *r = new (C, 3) RegionNode(3);
3856     result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3857 
3858     Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3859     Node *iftrue = opt_iff(r, iff);
3860     r->init_req(1, iftrue);
3861     result->init_req(1, intcon(1));
3862     result->init_req(2, intcon(0));
3863 
3864     set_control(_gvn.transform(r));
3865     record_for_igvn(r);
3866 
3867     C->set_has_split_ifs(true); // Has chance for split-if optimization
3868   }
3869 
3870   push(_gvn.transform(result));
3871   return true;
3872 }
3873 
3874 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3875   // restore the arguments
3876   _sp += arg_size();
3877 
3878   switch (id) {
3879   case vmIntrinsics::_floatToRawIntBits:
3880     push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3881     break;
3882 
3883   case vmIntrinsics::_intBitsToFloat:
3884     push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3885     break;
3886 
3887   case vmIntrinsics::_doubleToRawLongBits:
3888     push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3889     break;
3890 
3891   case vmIntrinsics::_longBitsToDouble:
3892     push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3893     break;
3894 
3895   case vmIntrinsics::_doubleToLongBits: {
3896     Node* value = pop_pair();
3897 
3898     // two paths (plus control) merge in a wood
3899     RegionNode *r = new (C, 3) RegionNode(3);
3900     Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3901 
3902     Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3903     // Build the boolean node
3904     Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3905 
3906     // Branch either way.
3907     // NaN case is less traveled, which makes all the difference.
3908     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3909     Node *opt_isnan = _gvn.transform(ifisnan);
3910     assert( opt_isnan->is_If(), "Expect an IfNode");
3911     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3912     Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3913 
3914     set_control(iftrue);
3915 
3916     static const jlong nan_bits = CONST64(0x7ff8000000000000);
3917     Node *slow_result = longcon(nan_bits); // return NaN
3918     phi->init_req(1, _gvn.transform( slow_result ));
3919     r->init_req(1, iftrue);
3920 
3921     // Else fall through
3922     Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3923     set_control(iffalse);
3924 
3925     phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3926     r->init_req(2, iffalse);
3927 
3928     // Post merge
3929     set_control(_gvn.transform(r));
3930     record_for_igvn(r);
3931 
3932     Node* result = _gvn.transform(phi);
3933     assert(result->bottom_type()->isa_long(), "must be");
3934     push_pair(result);
3935 
3936     C->set_has_split_ifs(true); // Has chance for split-if optimization
3937 
3938     break;
3939   }
3940 
3941   case vmIntrinsics::_floatToIntBits: {
3942     Node* value = pop();
3943 
3944     // two paths (plus control) merge in a wood
3945     RegionNode *r = new (C, 3) RegionNode(3);
3946     Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3947 
3948     Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3949     // Build the boolean node
3950     Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3951 
3952     // Branch either way.
3953     // NaN case is less traveled, which makes all the difference.
3954     IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3955     Node *opt_isnan = _gvn.transform(ifisnan);
3956     assert( opt_isnan->is_If(), "Expect an IfNode");
3957     IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3958     Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3959 
3960     set_control(iftrue);
3961 
3962     static const jint nan_bits = 0x7fc00000;
3963     Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3964     phi->init_req(1, _gvn.transform( slow_result ));
3965     r->init_req(1, iftrue);
3966 
3967     // Else fall through
3968     Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3969     set_control(iffalse);
3970 
3971     phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3972     r->init_req(2, iffalse);
3973 
3974     // Post merge
3975     set_control(_gvn.transform(r));
3976     record_for_igvn(r);
3977 
3978     Node* result = _gvn.transform(phi);
3979     assert(result->bottom_type()->isa_int(), "must be");
3980     push(result);
3981 
3982     C->set_has_split_ifs(true); // Has chance for split-if optimization
3983 
3984     break;
3985   }
3986 
3987   default:
3988     ShouldNotReachHere();
3989   }
3990 
3991   return true;
3992 }
3993 
3994 #ifdef _LP64
3995 #define XTOP ,top() /*additional argument*/
3996 #else  //_LP64
3997 #define XTOP        /*no additional argument*/
3998 #endif //_LP64
3999 
4000 //----------------------inline_unsafe_copyMemory-------------------------
4001 bool LibraryCallKit::inline_unsafe_copyMemory() {
4002   if (callee()->is_static())  return false;  // caller must have the capability!
4003   int nargs = 1 + 5 + 3;  // 5 args:  (src: ptr,off, dst: ptr,off, size)
4004   assert(signature()->size() == nargs-1, "copy has 5 arguments");
4005   null_check_receiver(callee());  // check then ignore argument(0)
4006   if (stopped())  return true;
4007 
4008   C->set_has_unsafe_access(true);  // Mark eventual nmethod as "unsafe".
4009 
4010   Node* src_ptr = argument(1);
4011   Node* src_off = ConvL2X(argument(2));
4012   assert(argument(3)->is_top(), "2nd half of long");
4013   Node* dst_ptr = argument(4);
4014   Node* dst_off = ConvL2X(argument(5));
4015   assert(argument(6)->is_top(), "2nd half of long");
4016   Node* size    = ConvL2X(argument(7));
4017   assert(argument(8)->is_top(), "2nd half of long");
4018 
4019   assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4020          "fieldOffset must be byte-scaled");
4021 
4022   Node* src = make_unsafe_address(src_ptr, src_off);
4023   Node* dst = make_unsafe_address(dst_ptr, dst_off);
4024 
4025   // Conservatively insert a memory barrier on all memory slices.
4026   // Do not let writes of the copy source or destination float below the copy.
4027   insert_mem_bar(Op_MemBarCPUOrder);
4028 
4029   // Call it.  Note that the length argument is not scaled.
4030   make_runtime_call(RC_LEAF|RC_NO_FP,
4031                     OptoRuntime::fast_arraycopy_Type(),
4032                     StubRoutines::unsafe_arraycopy(),
4033                     "unsafe_arraycopy",
4034                     TypeRawPtr::BOTTOM,
4035                     src, dst, size XTOP);
4036 
4037   // Do not let reads of the copy destination float above the copy.
4038   insert_mem_bar(Op_MemBarCPUOrder);
4039 
4040   return true;
4041 }
4042 
4043 //------------------------clone_coping-----------------------------------
4044 // Helper function for inline_native_clone.
4045 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4046   assert(obj_size != NULL, "");
4047   Node* raw_obj = alloc_obj->in(1);
4048   assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4049 
4050   if (ReduceBulkZeroing) {
4051     // We will be completely responsible for initializing this object -
4052     // mark Initialize node as complete.
4053     AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4054     // The object was just allocated - there should be no any stores!
4055     guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4056   }
4057 
4058   // Copy the fastest available way.
4059   // TODO: generate fields copies for small objects instead.
4060   Node* src  = obj;
4061   Node* dest = alloc_obj;
4062   Node* size = _gvn.transform(obj_size);
4063 
4064   // Exclude the header but include array length to copy by 8 bytes words.
4065   // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4066   int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4067                             instanceOopDesc::base_offset_in_bytes();
4068   // base_off:
4069   // 8  - 32-bit VM
4070   // 12 - 64-bit VM, compressed oops
4071   // 16 - 64-bit VM, normal oops
4072   if (base_off % BytesPerLong != 0) {
4073     assert(UseCompressedOops, "");
4074     if (is_array) {
4075       // Exclude length to copy by 8 bytes words.
4076       base_off += sizeof(int);
4077     } else {
4078       // Include klass to copy by 8 bytes words.
4079       base_off = instanceOopDesc::klass_offset_in_bytes();
4080     }
4081     assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4082   }
4083   src  = basic_plus_adr(src,  base_off);
4084   dest = basic_plus_adr(dest, base_off);
4085 
4086   // Compute the length also, if needed:
4087   Node* countx = size;
4088   countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4089   countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4090 
4091   const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4092   bool disjoint_bases = true;
4093   generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4094                                src, NULL, dest, NULL, countx,
4095                                /*dest_uninitialized*/true);
4096 
4097   // If necessary, emit some card marks afterwards.  (Non-arrays only.)
4098   if (card_mark) {
4099     assert(!is_array, "");
4100     // Put in store barrier for any and all oops we are sticking
4101     // into this object.  (We could avoid this if we could prove
4102     // that the object type contains no oop fields at all.)
4103     Node* no_particular_value = NULL;
4104     Node* no_particular_field = NULL;
4105     int raw_adr_idx = Compile::AliasIdxRaw;
4106     post_barrier(control(),
4107                  memory(raw_adr_type),
4108                  alloc_obj,
4109                  no_particular_field,
4110                  raw_adr_idx,
4111                  no_particular_value,
4112                  T_OBJECT,
4113                  false);
4114   }
4115 
4116   // Do not let reads from the cloned object float above the arraycopy.
4117   insert_mem_bar(Op_MemBarCPUOrder);
4118 }
4119 
4120 //------------------------inline_native_clone----------------------------
4121 // Here are the simple edge cases:
4122 //  null receiver => normal trap
4123 //  virtual and clone was overridden => slow path to out-of-line clone
4124 //  not cloneable or finalizer => slow path to out-of-line Object.clone
4125 //
4126 // The general case has two steps, allocation and copying.
4127 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4128 //
4129 // Copying also has two cases, oop arrays and everything else.
4130 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4131 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4132 //
4133 // These steps fold up nicely if and when the cloned object's klass
4134 // can be sharply typed as an object array, a type array, or an instance.
4135 //
4136 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4137   int nargs = 1;
4138   PhiNode* result_val;
4139 
4140   //set the original stack and the reexecute bit for the interpreter to reexecute
4141   //the bytecode that invokes Object.clone if deoptimization happens
4142   { PreserveReexecuteState preexecs(this);
4143     jvms()->set_should_reexecute(true);
4144 
4145     //null_check_receiver will adjust _sp (push and pop)
4146     Node* obj = null_check_receiver(callee());
4147     if (stopped())  return true;
4148 
4149     _sp += nargs;
4150 
4151     Node* obj_klass = load_object_klass(obj);
4152     const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4153     const TypeOopPtr*   toop   = ((tklass != NULL)
4154                                 ? tklass->as_instance_type()
4155                                 : TypeInstPtr::NOTNULL);
4156 
4157     // Conservatively insert a memory barrier on all memory slices.
4158     // Do not let writes into the original float below the clone.
4159     insert_mem_bar(Op_MemBarCPUOrder);
4160 
4161     // paths into result_reg:
4162     enum {
4163       _slow_path = 1,     // out-of-line call to clone method (virtual or not)
4164       _objArray_path,     // plain array allocation, plus arrayof_oop_arraycopy
4165       _array_path,        // plain array allocation, plus arrayof_long_arraycopy
4166       _instance_path,     // plain instance allocation, plus arrayof_long_arraycopy
4167       PATH_LIMIT
4168     };
4169     RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4170     result_val             = new(C, PATH_LIMIT) PhiNode(result_reg,
4171                                                         TypeInstPtr::NOTNULL);
4172     PhiNode*    result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4173     PhiNode*    result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4174                                                         TypePtr::BOTTOM);
4175     record_for_igvn(result_reg);
4176 
4177     const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4178     int raw_adr_idx = Compile::AliasIdxRaw;
4179     const bool raw_mem_only = true;
4180 
4181 
4182     Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4183     if (array_ctl != NULL) {
4184       // It's an array.
4185       PreserveJVMState pjvms(this);
4186       set_control(array_ctl);
4187       Node* obj_length = load_array_length(obj);
4188       Node* obj_size  = NULL;
4189       Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4190                                   raw_mem_only, &obj_size);
4191 
4192       if (!use_ReduceInitialCardMarks()) {
4193         // If it is an oop array, it requires very special treatment,
4194         // because card marking is required on each card of the array.
4195         Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4196         if (is_obja != NULL) {
4197           PreserveJVMState pjvms2(this);
4198           set_control(is_obja);
4199           // Generate a direct call to the right arraycopy function(s).
4200           bool disjoint_bases = true;
4201           bool length_never_negative = true;
4202           generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4203                              obj, intcon(0), alloc_obj, intcon(0),
4204                              obj_length,
4205                              disjoint_bases, length_never_negative);
4206           result_reg->init_req(_objArray_path, control());
4207           result_val->init_req(_objArray_path, alloc_obj);
4208           result_i_o ->set_req(_objArray_path, i_o());
4209           result_mem ->set_req(_objArray_path, reset_memory());
4210         }
4211       }
4212       // Otherwise, there are no card marks to worry about.
4213       // (We can dispense with card marks if we know the allocation
4214       //  comes out of eden (TLAB)...  In fact, ReduceInitialCardMarks
4215       //  causes the non-eden paths to take compensating steps to
4216       //  simulate a fresh allocation, so that no further
4217       //  card marks are required in compiled code to initialize
4218       //  the object.)
4219 
4220       if (!stopped()) {
4221         copy_to_clone(obj, alloc_obj, obj_size, true, false);
4222 
4223         // Present the results of the copy.
4224         result_reg->init_req(_array_path, control());
4225         result_val->init_req(_array_path, alloc_obj);
4226         result_i_o ->set_req(_array_path, i_o());
4227         result_mem ->set_req(_array_path, reset_memory());
4228       }
4229     }
4230 
4231     // We only go to the instance fast case code if we pass a number of guards.
4232     // The paths which do not pass are accumulated in the slow_region.
4233     RegionNode* slow_region = new (C, 1) RegionNode(1);
4234     record_for_igvn(slow_region);
4235     if (!stopped()) {
4236       // It's an instance (we did array above).  Make the slow-path tests.
4237       // If this is a virtual call, we generate a funny guard.  We grab
4238       // the vtable entry corresponding to clone() from the target object.
4239       // If the target method which we are calling happens to be the
4240       // Object clone() method, we pass the guard.  We do not need this
4241       // guard for non-virtual calls; the caller is known to be the native
4242       // Object clone().
4243       if (is_virtual) {
4244         generate_virtual_guard(obj_klass, slow_region);
4245       }
4246 
4247       // The object must be cloneable and must not have a finalizer.
4248       // Both of these conditions may be checked in a single test.
4249       // We could optimize the cloneable test further, but we don't care.
4250       generate_access_flags_guard(obj_klass,
4251                                   // Test both conditions:
4252                                   JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4253                                   // Must be cloneable but not finalizer:
4254                                   JVM_ACC_IS_CLONEABLE,
4255                                   slow_region);
4256     }
4257 
4258     if (!stopped()) {
4259       // It's an instance, and it passed the slow-path tests.
4260       PreserveJVMState pjvms(this);
4261       Node* obj_size  = NULL;
4262       Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4263 
4264       copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4265 
4266       // Present the results of the slow call.
4267       result_reg->init_req(_instance_path, control());
4268       result_val->init_req(_instance_path, alloc_obj);
4269       result_i_o ->set_req(_instance_path, i_o());
4270       result_mem ->set_req(_instance_path, reset_memory());
4271     }
4272 
4273     // Generate code for the slow case.  We make a call to clone().
4274     set_control(_gvn.transform(slow_region));
4275     if (!stopped()) {
4276       PreserveJVMState pjvms(this);
4277       CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4278       Node* slow_result = set_results_for_java_call(slow_call);
4279       // this->control() comes from set_results_for_java_call
4280       result_reg->init_req(_slow_path, control());
4281       result_val->init_req(_slow_path, slow_result);
4282       result_i_o ->set_req(_slow_path, i_o());
4283       result_mem ->set_req(_slow_path, reset_memory());
4284     }
4285 
4286     // Return the combined state.
4287     set_control(    _gvn.transform(result_reg) );
4288     set_i_o(        _gvn.transform(result_i_o) );
4289     set_all_memory( _gvn.transform(result_mem) );
4290   } //original reexecute and sp are set back here
4291 
4292   push(_gvn.transform(result_val));
4293 
4294   return true;
4295 }
4296 
4297 
4298 // constants for computing the copy function
4299 enum {
4300   COPYFUNC_UNALIGNED = 0,
4301   COPYFUNC_ALIGNED = 1,                 // src, dest aligned to HeapWordSize
4302   COPYFUNC_CONJOINT = 0,
4303   COPYFUNC_DISJOINT = 2                 // src != dest, or transfer can descend
4304 };
4305 
4306 // Note:  The condition "disjoint" applies also for overlapping copies
4307 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4308 static address
4309 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name, bool dest_uninitialized) {
4310   int selector =
4311     (aligned  ? COPYFUNC_ALIGNED  : COPYFUNC_UNALIGNED) +
4312     (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4313 
4314 #define RETURN_STUB(xxx_arraycopy) { \
4315   name = #xxx_arraycopy; \
4316   return StubRoutines::xxx_arraycopy(); }
4317 
4318 #define RETURN_STUB_PARM(xxx_arraycopy, parm) {           \
4319   name = #xxx_arraycopy; \
4320   return StubRoutines::xxx_arraycopy(parm); }
4321 
4322   switch (t) {
4323   case T_BYTE:
4324   case T_BOOLEAN:
4325     switch (selector) {
4326     case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jbyte_arraycopy);
4327     case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jbyte_arraycopy);
4328     case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jbyte_disjoint_arraycopy);
4329     case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4330     }
4331   case T_CHAR:
4332   case T_SHORT:
4333     switch (selector) {
4334     case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jshort_arraycopy);
4335     case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jshort_arraycopy);
4336     case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jshort_disjoint_arraycopy);
4337     case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4338     }
4339   case T_INT:
4340   case T_FLOAT:
4341     switch (selector) {
4342     case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jint_arraycopy);
4343     case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jint_arraycopy);
4344     case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jint_disjoint_arraycopy);
4345     case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4346     }
4347   case T_DOUBLE:
4348   case T_LONG:
4349     switch (selector) {
4350     case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jlong_arraycopy);
4351     case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jlong_arraycopy);
4352     case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB(jlong_disjoint_arraycopy);
4353     case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4354     }
4355   case T_ARRAY:
4356   case T_OBJECT:
4357     switch (selector) {
4358     case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB_PARM(oop_arraycopy, dest_uninitialized);
4359     case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED:    RETURN_STUB_PARM(arrayof_oop_arraycopy, dest_uninitialized);
4360     case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED:  RETURN_STUB_PARM(oop_disjoint_arraycopy, dest_uninitialized);
4361     case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED:    RETURN_STUB_PARM(arrayof_oop_disjoint_arraycopy, dest_uninitialized);
4362     }
4363   default:
4364     ShouldNotReachHere();
4365     return NULL;
4366   }
4367 
4368 #undef RETURN_STUB
4369 #undef RETURN_STUB_PARM
4370 }
4371 
4372 //------------------------------basictype2arraycopy----------------------------
4373 address LibraryCallKit::basictype2arraycopy(BasicType t,
4374                                             Node* src_offset,
4375                                             Node* dest_offset,
4376                                             bool disjoint_bases,
4377                                             const char* &name,
4378                                             bool dest_uninitialized) {
4379   const TypeInt* src_offset_inttype  = gvn().find_int_type(src_offset);;
4380   const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4381 
4382   bool aligned = false;
4383   bool disjoint = disjoint_bases;
4384 
4385   // if the offsets are the same, we can treat the memory regions as
4386   // disjoint, because either the memory regions are in different arrays,
4387   // or they are identical (which we can treat as disjoint.)  We can also
4388   // treat a copy with a destination index  less that the source index
4389   // as disjoint since a low->high copy will work correctly in this case.
4390   if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4391       dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4392     // both indices are constants
4393     int s_offs = src_offset_inttype->get_con();
4394     int d_offs = dest_offset_inttype->get_con();
4395     int element_size = type2aelembytes(t);
4396     aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4397               ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4398     if (s_offs >= d_offs)  disjoint = true;
4399   } else if (src_offset == dest_offset && src_offset != NULL) {
4400     // This can occur if the offsets are identical non-constants.
4401     disjoint = true;
4402   }
4403 
4404   return select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4405 }
4406 
4407 
4408 //------------------------------inline_arraycopy-----------------------
4409 bool LibraryCallKit::inline_arraycopy() {
4410   // Restore the stack and pop off the arguments.
4411   int nargs = 5;  // 2 oops, 3 ints, no size_t or long
4412   assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4413 
4414   Node *src         = argument(0);
4415   Node *src_offset  = argument(1);
4416   Node *dest        = argument(2);
4417   Node *dest_offset = argument(3);
4418   Node *length      = argument(4);
4419 
4420   // Compile time checks.  If any of these checks cannot be verified at compile time,
4421   // we do not make a fast path for this call.  Instead, we let the call remain as it
4422   // is.  The checks we choose to mandate at compile time are:
4423   //
4424   // (1) src and dest are arrays.
4425   const Type* src_type = src->Value(&_gvn);
4426   const Type* dest_type = dest->Value(&_gvn);
4427   const TypeAryPtr* top_src = src_type->isa_aryptr();
4428   const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4429   if (top_src  == NULL || top_src->klass()  == NULL ||
4430       top_dest == NULL || top_dest->klass() == NULL) {
4431     // Conservatively insert a memory barrier on all memory slices.
4432     // Do not let writes into the source float below the arraycopy.
4433     insert_mem_bar(Op_MemBarCPUOrder);
4434 
4435     // Call StubRoutines::generic_arraycopy stub.
4436     generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4437                        src, src_offset, dest, dest_offset, length);
4438 
4439     // Do not let reads from the destination float above the arraycopy.
4440     // Since we cannot type the arrays, we don't know which slices
4441     // might be affected.  We could restrict this barrier only to those
4442     // memory slices which pertain to array elements--but don't bother.
4443     if (!InsertMemBarAfterArraycopy)
4444       // (If InsertMemBarAfterArraycopy, there is already one in place.)
4445       insert_mem_bar(Op_MemBarCPUOrder);
4446     return true;
4447   }
4448 
4449   // (2) src and dest arrays must have elements of the same BasicType
4450   // Figure out the size and type of the elements we will be copying.
4451   BasicType src_elem  =  top_src->klass()->as_array_klass()->element_type()->basic_type();
4452   BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4453   if (src_elem  == T_ARRAY)  src_elem  = T_OBJECT;
4454   if (dest_elem == T_ARRAY)  dest_elem = T_OBJECT;
4455 
4456   if (src_elem != dest_elem || dest_elem == T_VOID) {
4457     // The component types are not the same or are not recognized.  Punt.
4458     // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4459     generate_slow_arraycopy(TypePtr::BOTTOM,
4460                             src, src_offset, dest, dest_offset, length,
4461                             /*dest_uninitialized*/false);
4462     return true;
4463   }
4464 
4465   //---------------------------------------------------------------------------
4466   // We will make a fast path for this call to arraycopy.
4467 
4468   // We have the following tests left to perform:
4469   //
4470   // (3) src and dest must not be null.
4471   // (4) src_offset must not be negative.
4472   // (5) dest_offset must not be negative.
4473   // (6) length must not be negative.
4474   // (7) src_offset + length must not exceed length of src.
4475   // (8) dest_offset + length must not exceed length of dest.
4476   // (9) each element of an oop array must be assignable
4477 
4478   RegionNode* slow_region = new (C, 1) RegionNode(1);
4479   record_for_igvn(slow_region);
4480 
4481   // (3) operands must not be null
4482   // We currently perform our null checks with the do_null_check routine.
4483   // This means that the null exceptions will be reported in the caller
4484   // rather than (correctly) reported inside of the native arraycopy call.
4485   // This should be corrected, given time.  We do our null check with the
4486   // stack pointer restored.
4487   _sp += nargs;
4488   src  = do_null_check(src,  T_ARRAY);
4489   dest = do_null_check(dest, T_ARRAY);
4490   _sp -= nargs;
4491 
4492   // (4) src_offset must not be negative.
4493   generate_negative_guard(src_offset, slow_region);
4494 
4495   // (5) dest_offset must not be negative.
4496   generate_negative_guard(dest_offset, slow_region);
4497 
4498   // (6) length must not be negative (moved to generate_arraycopy()).
4499   // generate_negative_guard(length, slow_region);
4500 
4501   // (7) src_offset + length must not exceed length of src.
4502   generate_limit_guard(src_offset, length,
4503                        load_array_length(src),
4504                        slow_region);
4505 
4506   // (8) dest_offset + length must not exceed length of dest.
4507   generate_limit_guard(dest_offset, length,
4508                        load_array_length(dest),
4509                        slow_region);
4510 
4511   // (9) each element of an oop array must be assignable
4512   // The generate_arraycopy subroutine checks this.
4513 
4514   // This is where the memory effects are placed:
4515   const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4516   generate_arraycopy(adr_type, dest_elem,
4517                      src, src_offset, dest, dest_offset, length,
4518                      false, false, slow_region);
4519 
4520   return true;
4521 }
4522 
4523 //-----------------------------generate_arraycopy----------------------
4524 // Generate an optimized call to arraycopy.
4525 // Caller must guard against non-arrays.
4526 // Caller must determine a common array basic-type for both arrays.
4527 // Caller must validate offsets against array bounds.
4528 // The slow_region has already collected guard failure paths
4529 // (such as out of bounds length or non-conformable array types).
4530 // The generated code has this shape, in general:
4531 //
4532 //     if (length == 0)  return   // via zero_path
4533 //     slowval = -1
4534 //     if (types unknown) {
4535 //       slowval = call generic copy loop
4536 //       if (slowval == 0)  return  // via checked_path
4537 //     } else if (indexes in bounds) {
4538 //       if ((is object array) && !(array type check)) {
4539 //         slowval = call checked copy loop
4540 //         if (slowval == 0)  return  // via checked_path
4541 //       } else {
4542 //         call bulk copy loop
4543 //         return  // via fast_path
4544 //       }
4545 //     }
4546 //     // adjust params for remaining work:
4547 //     if (slowval != -1) {
4548 //       n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4549 //     }
4550 //   slow_region:
4551 //     call slow arraycopy(src, src_offset, dest, dest_offset, length)
4552 //     return  // via slow_call_path
4553 //
4554 // This routine is used from several intrinsics:  System.arraycopy,
4555 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4556 //
4557 void
4558 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4559                                    BasicType basic_elem_type,
4560                                    Node* src,  Node* src_offset,
4561                                    Node* dest, Node* dest_offset,
4562                                    Node* copy_length,
4563                                    bool disjoint_bases,
4564                                    bool length_never_negative,
4565                                    RegionNode* slow_region) {
4566 
4567   if (slow_region == NULL) {
4568     slow_region = new(C,1) RegionNode(1);
4569     record_for_igvn(slow_region);
4570   }
4571 
4572   Node* original_dest      = dest;
4573   AllocateArrayNode* alloc = NULL;  // used for zeroing, if needed
4574   bool  dest_uninitialized = false;
4575 
4576   // See if this is the initialization of a newly-allocated array.
4577   // If so, we will take responsibility here for initializing it to zero.
4578   // (Note:  Because tightly_coupled_allocation performs checks on the
4579   // out-edges of the dest, we need to avoid making derived pointers
4580   // from it until we have checked its uses.)
4581   if (ReduceBulkZeroing
4582       && !ZeroTLAB              // pointless if already zeroed
4583       && basic_elem_type != T_CONFLICT // avoid corner case
4584       && !_gvn.eqv_uncast(src, dest)
4585       && ((alloc = tightly_coupled_allocation(dest, slow_region))
4586           != NULL)
4587       && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4588       && alloc->maybe_set_complete(&_gvn)) {
4589     // "You break it, you buy it."
4590     InitializeNode* init = alloc->initialization();
4591     assert(init->is_complete(), "we just did this");
4592     assert(dest->is_CheckCastPP(), "sanity");
4593     assert(dest->in(0)->in(0) == init, "dest pinned");
4594     adr_type = TypeRawPtr::BOTTOM;  // all initializations are into raw memory
4595     // From this point on, every exit path is responsible for
4596     // initializing any non-copied parts of the object to zero.
4597     // Also, if this flag is set we make sure that arraycopy interacts properly
4598     // with G1, eliding pre-barriers. See CR 6627983.
4599     dest_uninitialized = true;
4600   } else {
4601     // No zeroing elimination here.
4602     alloc             = NULL;
4603     //original_dest   = dest;
4604     //dest_uninitialized = false;
4605   }
4606 
4607   // Results are placed here:
4608   enum { fast_path        = 1,  // normal void-returning assembly stub
4609          checked_path     = 2,  // special assembly stub with cleanup
4610          slow_call_path   = 3,  // something went wrong; call the VM
4611          zero_path        = 4,  // bypass when length of copy is zero
4612          bcopy_path       = 5,  // copy primitive array by 64-bit blocks
4613          PATH_LIMIT       = 6
4614   };
4615   RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4616   PhiNode*    result_i_o    = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4617   PhiNode*    result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4618   record_for_igvn(result_region);
4619   _gvn.set_type_bottom(result_i_o);
4620   _gvn.set_type_bottom(result_memory);
4621   assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4622 
4623   // The slow_control path:
4624   Node* slow_control;
4625   Node* slow_i_o = i_o();
4626   Node* slow_mem = memory(adr_type);
4627   debug_only(slow_control = (Node*) badAddress);
4628 
4629   // Checked control path:
4630   Node* checked_control = top();
4631   Node* checked_mem     = NULL;
4632   Node* checked_i_o     = NULL;
4633   Node* checked_value   = NULL;
4634 
4635   if (basic_elem_type == T_CONFLICT) {
4636     assert(!dest_uninitialized, "");
4637     Node* cv = generate_generic_arraycopy(adr_type,
4638                                           src, src_offset, dest, dest_offset,
4639                                           copy_length, dest_uninitialized);
4640     if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
4641     checked_control = control();
4642     checked_i_o     = i_o();
4643     checked_mem     = memory(adr_type);
4644     checked_value   = cv;
4645     set_control(top());         // no fast path
4646   }
4647 
4648   Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4649   if (not_pos != NULL) {
4650     PreserveJVMState pjvms(this);
4651     set_control(not_pos);
4652 
4653     // (6) length must not be negative.
4654     if (!length_never_negative) {
4655       generate_negative_guard(copy_length, slow_region);
4656     }
4657 
4658     // copy_length is 0.
4659     if (!stopped() && dest_uninitialized) {
4660       Node* dest_length = alloc->in(AllocateNode::ALength);
4661       if (_gvn.eqv_uncast(copy_length, dest_length)
4662           || _gvn.find_int_con(dest_length, 1) <= 0) {
4663         // There is no zeroing to do. No need for a secondary raw memory barrier.
4664       } else {
4665         // Clear the whole thing since there are no source elements to copy.
4666         generate_clear_array(adr_type, dest, basic_elem_type,
4667                              intcon(0), NULL,
4668                              alloc->in(AllocateNode::AllocSize));
4669         // Use a secondary InitializeNode as raw memory barrier.
4670         // Currently it is needed only on this path since other
4671         // paths have stub or runtime calls as raw memory barriers.
4672         InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4673                                                        Compile::AliasIdxRaw,
4674                                                        top())->as_Initialize();
4675         init->set_complete(&_gvn);  // (there is no corresponding AllocateNode)
4676       }
4677     }
4678 
4679     // Present the results of the fast call.
4680     result_region->init_req(zero_path, control());
4681     result_i_o   ->init_req(zero_path, i_o());
4682     result_memory->init_req(zero_path, memory(adr_type));
4683   }
4684 
4685   if (!stopped() && dest_uninitialized) {
4686     // We have to initialize the *uncopied* part of the array to zero.
4687     // The copy destination is the slice dest[off..off+len].  The other slices
4688     // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4689     Node* dest_size   = alloc->in(AllocateNode::AllocSize);
4690     Node* dest_length = alloc->in(AllocateNode::ALength);
4691     Node* dest_tail   = _gvn.transform( new(C,3) AddINode(dest_offset,
4692                                                           copy_length) );
4693 
4694     // If there is a head section that needs zeroing, do it now.
4695     if (find_int_con(dest_offset, -1) != 0) {
4696       generate_clear_array(adr_type, dest, basic_elem_type,
4697                            intcon(0), dest_offset,
4698                            NULL);
4699     }
4700 
4701     // Next, perform a dynamic check on the tail length.
4702     // It is often zero, and we can win big if we prove this.
4703     // There are two wins:  Avoid generating the ClearArray
4704     // with its attendant messy index arithmetic, and upgrade
4705     // the copy to a more hardware-friendly word size of 64 bits.
4706     Node* tail_ctl = NULL;
4707     if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4708       Node* cmp_lt   = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4709       Node* bol_lt   = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4710       tail_ctl = generate_slow_guard(bol_lt, NULL);
4711       assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4712     }
4713 
4714     // At this point, let's assume there is no tail.
4715     if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4716       // There is no tail.  Try an upgrade to a 64-bit copy.
4717       bool didit = false;
4718       { PreserveJVMState pjvms(this);
4719         didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4720                                          src, src_offset, dest, dest_offset,
4721                                          dest_size, dest_uninitialized);
4722         if (didit) {
4723           // Present the results of the block-copying fast call.
4724           result_region->init_req(bcopy_path, control());
4725           result_i_o   ->init_req(bcopy_path, i_o());
4726           result_memory->init_req(bcopy_path, memory(adr_type));
4727         }
4728       }
4729       if (didit)
4730         set_control(top());     // no regular fast path
4731     }
4732 
4733     // Clear the tail, if any.
4734     if (tail_ctl != NULL) {
4735       Node* notail_ctl = stopped() ? NULL : control();
4736       set_control(tail_ctl);
4737       if (notail_ctl == NULL) {
4738         generate_clear_array(adr_type, dest, basic_elem_type,
4739                              dest_tail, NULL,
4740                              dest_size);
4741       } else {
4742         // Make a local merge.
4743         Node* done_ctl = new(C,3) RegionNode(3);
4744         Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4745         done_ctl->init_req(1, notail_ctl);
4746         done_mem->init_req(1, memory(adr_type));
4747         generate_clear_array(adr_type, dest, basic_elem_type,
4748                              dest_tail, NULL,
4749                              dest_size);
4750         done_ctl->init_req(2, control());
4751         done_mem->init_req(2, memory(adr_type));
4752         set_control( _gvn.transform(done_ctl) );
4753         set_memory(  _gvn.transform(done_mem), adr_type );
4754       }
4755     }
4756   }
4757 
4758   BasicType copy_type = basic_elem_type;
4759   assert(basic_elem_type != T_ARRAY, "caller must fix this");
4760   if (!stopped() && copy_type == T_OBJECT) {
4761     // If src and dest have compatible element types, we can copy bits.
4762     // Types S[] and D[] are compatible if D is a supertype of S.
4763     //
4764     // If they are not, we will use checked_oop_disjoint_arraycopy,
4765     // which performs a fast optimistic per-oop check, and backs off
4766     // further to JVM_ArrayCopy on the first per-oop check that fails.
4767     // (Actually, we don't move raw bits only; the GC requires card marks.)
4768 
4769     // Get the klassOop for both src and dest
4770     Node* src_klass  = load_object_klass(src);
4771     Node* dest_klass = load_object_klass(dest);
4772 
4773     // Generate the subtype check.
4774     // This might fold up statically, or then again it might not.
4775     //
4776     // Non-static example:  Copying List<String>.elements to a new String[].
4777     // The backing store for a List<String> is always an Object[],
4778     // but its elements are always type String, if the generic types
4779     // are correct at the source level.
4780     //
4781     // Test S[] against D[], not S against D, because (probably)
4782     // the secondary supertype cache is less busy for S[] than S.
4783     // This usually only matters when D is an interface.
4784     Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4785     // Plug failing path into checked_oop_disjoint_arraycopy
4786     if (not_subtype_ctrl != top()) {
4787       PreserveJVMState pjvms(this);
4788       set_control(not_subtype_ctrl);
4789       // (At this point we can assume disjoint_bases, since types differ.)
4790       int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4791       Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4792       Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4793       Node* dest_elem_klass = _gvn.transform(n1);
4794       Node* cv = generate_checkcast_arraycopy(adr_type,
4795                                               dest_elem_klass,
4796                                               src, src_offset, dest, dest_offset,
4797                                               ConvI2X(copy_length), dest_uninitialized);
4798       if (cv == NULL)  cv = intcon(-1);  // failure (no stub available)
4799       checked_control = control();
4800       checked_i_o     = i_o();
4801       checked_mem     = memory(adr_type);
4802       checked_value   = cv;
4803     }
4804     // At this point we know we do not need type checks on oop stores.
4805 
4806     // Let's see if we need card marks:
4807     if (alloc != NULL && use_ReduceInitialCardMarks()) {
4808       // If we do not need card marks, copy using the jint or jlong stub.
4809       copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4810       assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4811              "sizes agree");
4812     }
4813   }
4814 
4815   if (!stopped()) {
4816     // Generate the fast path, if possible.
4817     PreserveJVMState pjvms(this);
4818     generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4819                                  src, src_offset, dest, dest_offset,
4820                                  ConvI2X(copy_length), dest_uninitialized);
4821 
4822     // Present the results of the fast call.
4823     result_region->init_req(fast_path, control());
4824     result_i_o   ->init_req(fast_path, i_o());
4825     result_memory->init_req(fast_path, memory(adr_type));
4826   }
4827 
4828   // Here are all the slow paths up to this point, in one bundle:
4829   slow_control = top();
4830   if (slow_region != NULL)
4831     slow_control = _gvn.transform(slow_region);
4832   debug_only(slow_region = (RegionNode*)badAddress);
4833 
4834   set_control(checked_control);
4835   if (!stopped()) {
4836     // Clean up after the checked call.
4837     // The returned value is either 0 or -1^K,
4838     // where K = number of partially transferred array elements.
4839     Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4840     Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4841     IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4842 
4843     // If it is 0, we are done, so transfer to the end.
4844     Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4845     result_region->init_req(checked_path, checks_done);
4846     result_i_o   ->init_req(checked_path, checked_i_o);
4847     result_memory->init_req(checked_path, checked_mem);
4848 
4849     // If it is not zero, merge into the slow call.
4850     set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4851     RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4852     PhiNode*    slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4853     PhiNode*    slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4854     record_for_igvn(slow_reg2);
4855     slow_reg2  ->init_req(1, slow_control);
4856     slow_i_o2  ->init_req(1, slow_i_o);
4857     slow_mem2  ->init_req(1, slow_mem);
4858     slow_reg2  ->init_req(2, control());
4859     slow_i_o2  ->init_req(2, checked_i_o);
4860     slow_mem2  ->init_req(2, checked_mem);
4861 
4862     slow_control = _gvn.transform(slow_reg2);
4863     slow_i_o     = _gvn.transform(slow_i_o2);
4864     slow_mem     = _gvn.transform(slow_mem2);
4865 
4866     if (alloc != NULL) {
4867       // We'll restart from the very beginning, after zeroing the whole thing.
4868       // This can cause double writes, but that's OK since dest is brand new.
4869       // So we ignore the low 31 bits of the value returned from the stub.
4870     } else {
4871       // We must continue the copy exactly where it failed, or else
4872       // another thread might see the wrong number of writes to dest.
4873       Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4874       Node* slow_offset    = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4875       slow_offset->init_req(1, intcon(0));
4876       slow_offset->init_req(2, checked_offset);
4877       slow_offset  = _gvn.transform(slow_offset);
4878 
4879       // Adjust the arguments by the conditionally incoming offset.
4880       Node* src_off_plus  = _gvn.transform( new(C, 3) AddINode(src_offset,  slow_offset) );
4881       Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4882       Node* length_minus  = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4883 
4884       // Tweak the node variables to adjust the code produced below:
4885       src_offset  = src_off_plus;
4886       dest_offset = dest_off_plus;
4887       copy_length = length_minus;
4888     }
4889   }
4890 
4891   set_control(slow_control);
4892   if (!stopped()) {
4893     // Generate the slow path, if needed.
4894     PreserveJVMState pjvms(this);   // replace_in_map may trash the map
4895 
4896     set_memory(slow_mem, adr_type);
4897     set_i_o(slow_i_o);
4898 
4899     if (dest_uninitialized) {
4900       generate_clear_array(adr_type, dest, basic_elem_type,
4901                            intcon(0), NULL,
4902                            alloc->in(AllocateNode::AllocSize));
4903     }
4904 
4905     generate_slow_arraycopy(adr_type,
4906                             src, src_offset, dest, dest_offset,
4907                             copy_length, /*dest_uninitialized*/false);
4908 
4909     result_region->init_req(slow_call_path, control());
4910     result_i_o   ->init_req(slow_call_path, i_o());
4911     result_memory->init_req(slow_call_path, memory(adr_type));
4912   }
4913 
4914   // Remove unused edges.
4915   for (uint i = 1; i < result_region->req(); i++) {
4916     if (result_region->in(i) == NULL)
4917       result_region->init_req(i, top());
4918   }
4919 
4920   // Finished; return the combined state.
4921   set_control( _gvn.transform(result_region) );
4922   set_i_o(     _gvn.transform(result_i_o)    );
4923   set_memory(  _gvn.transform(result_memory), adr_type );
4924 
4925   // The memory edges above are precise in order to model effects around
4926   // array copies accurately to allow value numbering of field loads around
4927   // arraycopy.  Such field loads, both before and after, are common in Java
4928   // collections and similar classes involving header/array data structures.
4929   //
4930   // But with low number of register or when some registers are used or killed
4931   // by arraycopy calls it causes registers spilling on stack. See 6544710.
4932   // The next memory barrier is added to avoid it. If the arraycopy can be
4933   // optimized away (which it can, sometimes) then we can manually remove
4934   // the membar also.
4935   //
4936   // Do not let reads from the cloned object float above the arraycopy.
4937   if (InsertMemBarAfterArraycopy || alloc != NULL)
4938     insert_mem_bar(Op_MemBarCPUOrder);
4939 }
4940 
4941 
4942 // Helper function which determines if an arraycopy immediately follows
4943 // an allocation, with no intervening tests or other escapes for the object.
4944 AllocateArrayNode*
4945 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4946                                            RegionNode* slow_region) {
4947   if (stopped())             return NULL;  // no fast path
4948   if (C->AliasLevel() == 0)  return NULL;  // no MergeMems around
4949 
4950   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4951   if (alloc == NULL)  return NULL;
4952 
4953   Node* rawmem = memory(Compile::AliasIdxRaw);
4954   // Is the allocation's memory state untouched?
4955   if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4956     // Bail out if there have been raw-memory effects since the allocation.
4957     // (Example:  There might have been a call or safepoint.)
4958     return NULL;
4959   }
4960   rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4961   if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4962     return NULL;
4963   }
4964 
4965   // There must be no unexpected observers of this allocation.
4966   for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4967     Node* obs = ptr->fast_out(i);
4968     if (obs != this->map()) {
4969       return NULL;
4970     }
4971   }
4972 
4973   // This arraycopy must unconditionally follow the allocation of the ptr.
4974   Node* alloc_ctl = ptr->in(0);
4975   assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4976 
4977   Node* ctl = control();
4978   while (ctl != alloc_ctl) {
4979     // There may be guards which feed into the slow_region.
4980     // Any other control flow means that we might not get a chance
4981     // to finish initializing the allocated object.
4982     if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4983       IfNode* iff = ctl->in(0)->as_If();
4984       Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4985       assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4986       if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4987         ctl = iff->in(0);       // This test feeds the known slow_region.
4988         continue;
4989       }
4990       // One more try:  Various low-level checks bottom out in
4991       // uncommon traps.  If the debug-info of the trap omits
4992       // any reference to the allocation, as we've already
4993       // observed, then there can be no objection to the trap.
4994       bool found_trap = false;
4995       for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4996         Node* obs = not_ctl->fast_out(j);
4997         if (obs->in(0) == not_ctl && obs->is_Call() &&
4998             (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4999           found_trap = true; break;
5000         }
5001       }
5002       if (found_trap) {
5003         ctl = iff->in(0);       // This test feeds a harmless uncommon trap.
5004         continue;
5005       }
5006     }
5007     return NULL;
5008   }
5009 
5010   // If we get this far, we have an allocation which immediately
5011   // precedes the arraycopy, and we can take over zeroing the new object.
5012   // The arraycopy will finish the initialization, and provide
5013   // a new control state to which we will anchor the destination pointer.
5014 
5015   return alloc;
5016 }
5017 
5018 // Helper for initialization of arrays, creating a ClearArray.
5019 // It writes zero bits in [start..end), within the body of an array object.
5020 // The memory effects are all chained onto the 'adr_type' alias category.
5021 //
5022 // Since the object is otherwise uninitialized, we are free
5023 // to put a little "slop" around the edges of the cleared area,
5024 // as long as it does not go back into the array's header,
5025 // or beyond the array end within the heap.
5026 //
5027 // The lower edge can be rounded down to the nearest jint and the
5028 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5029 //
5030 // Arguments:
5031 //   adr_type           memory slice where writes are generated
5032 //   dest               oop of the destination array
5033 //   basic_elem_type    element type of the destination
5034 //   slice_idx          array index of first element to store
5035 //   slice_len          number of elements to store (or NULL)
5036 //   dest_size          total size in bytes of the array object
5037 //
5038 // Exactly one of slice_len or dest_size must be non-NULL.
5039 // If dest_size is non-NULL, zeroing extends to the end of the object.
5040 // If slice_len is non-NULL, the slice_idx value must be a constant.
5041 void
5042 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5043                                      Node* dest,
5044                                      BasicType basic_elem_type,
5045                                      Node* slice_idx,
5046                                      Node* slice_len,
5047                                      Node* dest_size) {
5048   // one or the other but not both of slice_len and dest_size:
5049   assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5050   if (slice_len == NULL)  slice_len = top();
5051   if (dest_size == NULL)  dest_size = top();
5052 
5053   // operate on this memory slice:
5054   Node* mem = memory(adr_type); // memory slice to operate on
5055 
5056   // scaling and rounding of indexes:
5057   int scale = exact_log2(type2aelembytes(basic_elem_type));
5058   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5059   int clear_low = (-1 << scale) & (BytesPerInt  - 1);
5060   int bump_bit  = (-1 << scale) & BytesPerInt;
5061 
5062   // determine constant starts and ends
5063   const intptr_t BIG_NEG = -128;
5064   assert(BIG_NEG + 2*abase < 0, "neg enough");
5065   intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5066   intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5067   if (slice_len_con == 0) {
5068     return;                     // nothing to do here
5069   }
5070   intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5071   intptr_t end_con   = find_intptr_t_con(dest_size, -1);
5072   if (slice_idx_con >= 0 && slice_len_con >= 0) {
5073     assert(end_con < 0, "not two cons");
5074     end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5075                        BytesPerLong);
5076   }
5077 
5078   if (start_con >= 0 && end_con >= 0) {
5079     // Constant start and end.  Simple.
5080     mem = ClearArrayNode::clear_memory(control(), mem, dest,
5081                                        start_con, end_con, &_gvn);
5082   } else if (start_con >= 0 && dest_size != top()) {
5083     // Constant start, pre-rounded end after the tail of the array.
5084     Node* end = dest_size;
5085     mem = ClearArrayNode::clear_memory(control(), mem, dest,
5086                                        start_con, end, &_gvn);
5087   } else if (start_con >= 0 && slice_len != top()) {
5088     // Constant start, non-constant end.  End needs rounding up.
5089     // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5090     intptr_t end_base  = abase + (slice_idx_con << scale);
5091     int      end_round = (-1 << scale) & (BytesPerLong  - 1);
5092     Node*    end       = ConvI2X(slice_len);
5093     if (scale != 0)
5094       end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5095     end_base += end_round;
5096     end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5097     end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5098     mem = ClearArrayNode::clear_memory(control(), mem, dest,
5099                                        start_con, end, &_gvn);
5100   } else if (start_con < 0 && dest_size != top()) {
5101     // Non-constant start, pre-rounded end after the tail of the array.
5102     // This is almost certainly a "round-to-end" operation.
5103     Node* start = slice_idx;
5104     start = ConvI2X(start);
5105     if (scale != 0)
5106       start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5107     start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5108     if ((bump_bit | clear_low) != 0) {
5109       int to_clear = (bump_bit | clear_low);
5110       // Align up mod 8, then store a jint zero unconditionally
5111       // just before the mod-8 boundary.
5112       if (((abase + bump_bit) & ~to_clear) - bump_bit
5113           < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5114         bump_bit = 0;
5115         assert((abase & to_clear) == 0, "array base must be long-aligned");
5116       } else {
5117         // Bump 'start' up to (or past) the next jint boundary:
5118         start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5119         assert((abase & clear_low) == 0, "array base must be int-aligned");
5120       }
5121       // Round bumped 'start' down to jlong boundary in body of array.
5122       start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5123       if (bump_bit != 0) {
5124         // Store a zero to the immediately preceding jint:
5125         Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5126         Node* p1 = basic_plus_adr(dest, x1);
5127         mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5128         mem = _gvn.transform(mem);
5129       }
5130     }
5131     Node* end = dest_size; // pre-rounded
5132     mem = ClearArrayNode::clear_memory(control(), mem, dest,
5133                                        start, end, &_gvn);
5134   } else {
5135     // Non-constant start, unrounded non-constant end.
5136     // (Nobody zeroes a random midsection of an array using this routine.)
5137     ShouldNotReachHere();       // fix caller
5138   }
5139 
5140   // Done.
5141   set_memory(mem, adr_type);
5142 }
5143 
5144 
5145 bool
5146 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5147                                          BasicType basic_elem_type,
5148                                          AllocateNode* alloc,
5149                                          Node* src,  Node* src_offset,
5150                                          Node* dest, Node* dest_offset,
5151                                          Node* dest_size, bool dest_uninitialized) {
5152   // See if there is an advantage from block transfer.
5153   int scale = exact_log2(type2aelembytes(basic_elem_type));
5154   if (scale >= LogBytesPerLong)
5155     return false;               // it is already a block transfer
5156 
5157   // Look at the alignment of the starting offsets.
5158   int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5159   const intptr_t BIG_NEG = -128;
5160   assert(BIG_NEG + 2*abase < 0, "neg enough");
5161 
5162   intptr_t src_off  = abase + ((intptr_t) find_int_con(src_offset, -1)  << scale);
5163   intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5164   if (src_off < 0 || dest_off < 0)
5165     // At present, we can only understand constants.
5166     return false;
5167 
5168   if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5169     // Non-aligned; too bad.
5170     // One more chance:  Pick off an initial 32-bit word.
5171     // This is a common case, since abase can be odd mod 8.
5172     if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5173         ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5174       Node* sptr = basic_plus_adr(src,  src_off);
5175       Node* dptr = basic_plus_adr(dest, dest_off);
5176       Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5177       store_to_memory(control(), dptr, sval, T_INT, adr_type);
5178       src_off += BytesPerInt;
5179       dest_off += BytesPerInt;
5180     } else {
5181       return false;
5182     }
5183   }
5184   assert(src_off % BytesPerLong == 0, "");
5185   assert(dest_off % BytesPerLong == 0, "");
5186 
5187   // Do this copy by giant steps.
5188   Node* sptr  = basic_plus_adr(src,  src_off);
5189   Node* dptr  = basic_plus_adr(dest, dest_off);
5190   Node* countx = dest_size;
5191   countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5192   countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5193 
5194   bool disjoint_bases = true;   // since alloc != NULL
5195   generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5196                                sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5197 
5198   return true;
5199 }
5200 
5201 
5202 // Helper function; generates code for the slow case.
5203 // We make a call to a runtime method which emulates the native method,
5204 // but without the native wrapper overhead.
5205 void
5206 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5207                                         Node* src,  Node* src_offset,
5208                                         Node* dest, Node* dest_offset,
5209                                         Node* copy_length, bool dest_uninitialized) {
5210   assert(!dest_uninitialized, "Invariant");
5211   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5212                                  OptoRuntime::slow_arraycopy_Type(),
5213                                  OptoRuntime::slow_arraycopy_Java(),
5214                                  "slow_arraycopy", adr_type,
5215                                  src, src_offset, dest, dest_offset,
5216                                  copy_length);
5217 
5218   // Handle exceptions thrown by this fellow:
5219   make_slow_call_ex(call, env()->Throwable_klass(), false);
5220 }
5221 
5222 // Helper function; generates code for cases requiring runtime checks.
5223 Node*
5224 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5225                                              Node* dest_elem_klass,
5226                                              Node* src,  Node* src_offset,
5227                                              Node* dest, Node* dest_offset,
5228                                              Node* copy_length, bool dest_uninitialized) {
5229   if (stopped())  return NULL;
5230 
5231   address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5232   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5233     return NULL;
5234   }
5235 
5236   // Pick out the parameters required to perform a store-check
5237   // for the target array.  This is an optimistic check.  It will
5238   // look in each non-null element's class, at the desired klass's
5239   // super_check_offset, for the desired klass.
5240   int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5241   Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5242   Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5243   Node* check_offset = ConvI2X(_gvn.transform(n3));
5244   Node* check_value  = dest_elem_klass;
5245 
5246   Node* src_start  = array_element_address(src,  src_offset,  T_OBJECT);
5247   Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5248 
5249   // (We know the arrays are never conjoint, because their types differ.)
5250   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5251                                  OptoRuntime::checkcast_arraycopy_Type(),
5252                                  copyfunc_addr, "checkcast_arraycopy", adr_type,
5253                                  // five arguments, of which two are
5254                                  // intptr_t (jlong in LP64)
5255                                  src_start, dest_start,
5256                                  copy_length XTOP,
5257                                  check_offset XTOP,
5258                                  check_value);
5259 
5260   return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5261 }
5262 
5263 
5264 // Helper function; generates code for cases requiring runtime checks.
5265 Node*
5266 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5267                                            Node* src,  Node* src_offset,
5268                                            Node* dest, Node* dest_offset,
5269                                            Node* copy_length, bool dest_uninitialized) {
5270   assert(!dest_uninitialized, "Invariant");
5271   if (stopped())  return NULL;
5272   address copyfunc_addr = StubRoutines::generic_arraycopy();
5273   if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5274     return NULL;
5275   }
5276 
5277   Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5278                     OptoRuntime::generic_arraycopy_Type(),
5279                     copyfunc_addr, "generic_arraycopy", adr_type,
5280                     src, src_offset, dest, dest_offset, copy_length);
5281 
5282   return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5283 }
5284 
5285 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5286 void
5287 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5288                                              BasicType basic_elem_type,
5289                                              bool disjoint_bases,
5290                                              Node* src,  Node* src_offset,
5291                                              Node* dest, Node* dest_offset,
5292                                              Node* copy_length, bool dest_uninitialized) {
5293   if (stopped())  return;               // nothing to do
5294 
5295   Node* src_start  = src;
5296   Node* dest_start = dest;
5297   if (src_offset != NULL || dest_offset != NULL) {
5298     assert(src_offset != NULL && dest_offset != NULL, "");
5299     src_start  = array_element_address(src,  src_offset,  basic_elem_type);
5300     dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5301   }
5302 
5303   // Figure out which arraycopy runtime method to call.
5304   const char* copyfunc_name = "arraycopy";
5305   address     copyfunc_addr =
5306       basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5307                           disjoint_bases, copyfunc_name, dest_uninitialized);
5308 
5309   // Call it.  Note that the count_ix value is not scaled to a byte-size.
5310   make_runtime_call(RC_LEAF|RC_NO_FP,
5311                     OptoRuntime::fast_arraycopy_Type(),
5312                     copyfunc_addr, copyfunc_name, adr_type,
5313                     src_start, dest_start, copy_length XTOP);
5314 }