1 /*
   2  * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.hpp"
  27 #include "asm/macroAssembler.inline.hpp"
  28 #include "ci/ciReplay.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/exceptionHandlerTable.hpp"
  31 #include "code/nmethod.hpp"
  32 #include "compiler/compileBroker.hpp"
  33 #include "compiler/compileLog.hpp"
  34 #include "compiler/disassembler.hpp"
  35 #include "compiler/oopMap.hpp"
  36 #include "memory/resourceArea.hpp"
  37 #include "opto/addnode.hpp"
  38 #include "opto/block.hpp"
  39 #include "opto/c2compiler.hpp"
  40 #include "opto/callGenerator.hpp"
  41 #include "opto/callnode.hpp"
  42 #include "opto/castnode.hpp"
  43 #include "opto/cfgnode.hpp"
  44 #include "opto/chaitin.hpp"
  45 #include "opto/compile.hpp"
  46 #include "opto/connode.hpp"
  47 #include "opto/convertnode.hpp"
  48 #include "opto/divnode.hpp"
  49 #include "opto/escape.hpp"
  50 #include "opto/idealGraphPrinter.hpp"
  51 #include "opto/loopnode.hpp"
  52 #include "opto/machnode.hpp"
  53 #include "opto/macro.hpp"
  54 #include "opto/matcher.hpp"
  55 #include "opto/mathexactnode.hpp"
  56 #include "opto/memnode.hpp"
  57 #include "opto/mulnode.hpp"
  58 #include "opto/narrowptrnode.hpp"
  59 #include "opto/node.hpp"
  60 #include "opto/opcodes.hpp"
  61 #include "opto/output.hpp"
  62 #include "opto/parse.hpp"
  63 #include "opto/phaseX.hpp"
  64 #include "opto/rootnode.hpp"
  65 #include "opto/runtime.hpp"
  66 #include "opto/stringopts.hpp"
  67 #include "opto/type.hpp"
  68 #include "opto/valuetypenode.hpp"
  69 #include "opto/vectornode.hpp"
  70 #include "runtime/arguments.hpp"
  71 #include "runtime/sharedRuntime.hpp"
  72 #include "runtime/signature.hpp"
  73 #include "runtime/stubRoutines.hpp"
  74 #include "runtime/timer.hpp"
  75 #include "utilities/align.hpp"
  76 #include "utilities/copy.hpp"
  77 
  78 
  79 // -------------------- Compile::mach_constant_base_node -----------------------
  80 // Constant table base node singleton.
  81 MachConstantBaseNode* Compile::mach_constant_base_node() {
  82   if (_mach_constant_base_node == NULL) {
  83     _mach_constant_base_node = new MachConstantBaseNode();
  84     _mach_constant_base_node->add_req(C->root());
  85   }
  86   return _mach_constant_base_node;
  87 }
  88 
  89 
  90 /// Support for intrinsics.
  91 
  92 // Return the index at which m must be inserted (or already exists).
  93 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
  94 class IntrinsicDescPair {
  95  private:
  96   ciMethod* _m;
  97   bool _is_virtual;
  98  public:
  99   IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
 100   static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
 101     ciMethod* m= elt->method();
 102     ciMethod* key_m = key->_m;
 103     if (key_m < m)      return -1;
 104     else if (key_m > m) return 1;
 105     else {
 106       bool is_virtual = elt->is_virtual();
 107       bool key_virtual = key->_is_virtual;
 108       if (key_virtual < is_virtual)      return -1;
 109       else if (key_virtual > is_virtual) return 1;
 110       else                               return 0;
 111     }
 112   }
 113 };
 114 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
 115 #ifdef ASSERT
 116   for (int i = 1; i < _intrinsics->length(); i++) {
 117     CallGenerator* cg1 = _intrinsics->at(i-1);
 118     CallGenerator* cg2 = _intrinsics->at(i);
 119     assert(cg1->method() != cg2->method()
 120            ? cg1->method()     < cg2->method()
 121            : cg1->is_virtual() < cg2->is_virtual(),
 122            "compiler intrinsics list must stay sorted");
 123   }
 124 #endif
 125   IntrinsicDescPair pair(m, is_virtual);
 126   return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
 127 }
 128 
 129 void Compile::register_intrinsic(CallGenerator* cg) {
 130   if (_intrinsics == NULL) {
 131     _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 132   }
 133   int len = _intrinsics->length();
 134   bool found = false;
 135   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
 136   assert(!found, "registering twice");
 137   _intrinsics->insert_before(index, cg);
 138   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 139 }
 140 
 141 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 142   assert(m->is_loaded(), "don't try this on unloaded methods");
 143   if (_intrinsics != NULL) {
 144     bool found = false;
 145     int index = intrinsic_insertion_index(m, is_virtual, found);
 146      if (found) {
 147       return _intrinsics->at(index);
 148     }
 149   }
 150   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 151   if (m->intrinsic_id() != vmIntrinsics::_none &&
 152       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 153     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 154     if (cg != NULL) {
 155       // Save it for next time:
 156       register_intrinsic(cg);
 157       return cg;
 158     } else {
 159       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 160     }
 161   }
 162   return NULL;
 163 }
 164 
 165 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 166 // in library_call.cpp.
 167 
 168 
 169 #ifndef PRODUCT
 170 // statistics gathering...
 171 
 172 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 174 
 175 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 176   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 177   int oflags = _intrinsic_hist_flags[id];
 178   assert(flags != 0, "what happened?");
 179   if (is_virtual) {
 180     flags |= _intrinsic_virtual;
 181   }
 182   bool changed = (flags != oflags);
 183   if ((flags & _intrinsic_worked) != 0) {
 184     juint count = (_intrinsic_hist_count[id] += 1);
 185     if (count == 1) {
 186       changed = true;           // first time
 187     }
 188     // increment the overall count also:
 189     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 190   }
 191   if (changed) {
 192     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 193       // Something changed about the intrinsic's virtuality.
 194       if ((flags & _intrinsic_virtual) != 0) {
 195         // This is the first use of this intrinsic as a virtual call.
 196         if (oflags != 0) {
 197           // We already saw it as a non-virtual, so note both cases.
 198           flags |= _intrinsic_both;
 199         }
 200       } else if ((oflags & _intrinsic_both) == 0) {
 201         // This is the first use of this intrinsic as a non-virtual
 202         flags |= _intrinsic_both;
 203       }
 204     }
 205     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 206   }
 207   // update the overall flags also:
 208   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 209   return changed;
 210 }
 211 
 212 static char* format_flags(int flags, char* buf) {
 213   buf[0] = 0;
 214   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 215   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 216   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 217   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 218   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 219   if (buf[0] == 0)  strcat(buf, ",");
 220   assert(buf[0] == ',', "must be");
 221   return &buf[1];
 222 }
 223 
 224 void Compile::print_intrinsic_statistics() {
 225   char flagsbuf[100];
 226   ttyLocker ttyl;
 227   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 228   tty->print_cr("Compiler intrinsic usage:");
 229   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 230   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 231   #define PRINT_STAT_LINE(name, c, f) \
 232     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 233   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 234     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 235     int   flags = _intrinsic_hist_flags[id];
 236     juint count = _intrinsic_hist_count[id];
 237     if ((flags | count) != 0) {
 238       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 239     }
 240   }
 241   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 242   if (xtty != NULL)  xtty->tail("statistics");
 243 }
 244 
 245 void Compile::print_statistics() {
 246   { ttyLocker ttyl;
 247     if (xtty != NULL)  xtty->head("statistics type='opto'");
 248     Parse::print_statistics();
 249     PhaseCCP::print_statistics();
 250     PhaseRegAlloc::print_statistics();
 251     Scheduling::print_statistics();
 252     PhasePeephole::print_statistics();
 253     PhaseIdealLoop::print_statistics();
 254     if (xtty != NULL)  xtty->tail("statistics");
 255   }
 256   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 257     // put this under its own <statistics> element.
 258     print_intrinsic_statistics();
 259   }
 260 }
 261 #endif //PRODUCT
 262 
 263 // Support for bundling info
 264 Bundle* Compile::node_bundling(const Node *n) {
 265   assert(valid_bundle_info(n), "oob");
 266   return &_node_bundling_base[n->_idx];
 267 }
 268 
 269 bool Compile::valid_bundle_info(const Node *n) {
 270   return (_node_bundling_limit > n->_idx);
 271 }
 272 
 273 
 274 void Compile::gvn_replace_by(Node* n, Node* nn) {
 275   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 276     Node* use = n->last_out(i);
 277     bool is_in_table = initial_gvn()->hash_delete(use);
 278     uint uses_found = 0;
 279     for (uint j = 0; j < use->len(); j++) {
 280       if (use->in(j) == n) {
 281         if (j < use->req())
 282           use->set_req(j, nn);
 283         else
 284           use->set_prec(j, nn);
 285         uses_found++;
 286       }
 287     }
 288     if (is_in_table) {
 289       // reinsert into table
 290       initial_gvn()->hash_find_insert(use);
 291     }
 292     record_for_igvn(use);
 293     i -= uses_found;    // we deleted 1 or more copies of this edge
 294   }
 295 }
 296 
 297 
 298 static inline bool not_a_node(const Node* n) {
 299   if (n == NULL)                   return true;
 300   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 301   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 302   return false;
 303 }
 304 
 305 // Identify all nodes that are reachable from below, useful.
 306 // Use breadth-first pass that records state in a Unique_Node_List,
 307 // recursive traversal is slower.
 308 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 309   int estimated_worklist_size = live_nodes();
 310   useful.map( estimated_worklist_size, NULL );  // preallocate space
 311 
 312   // Initialize worklist
 313   if (root() != NULL)     { useful.push(root()); }
 314   // If 'top' is cached, declare it useful to preserve cached node
 315   if( cached_top_node() ) { useful.push(cached_top_node()); }
 316 
 317   // Push all useful nodes onto the list, breadthfirst
 318   for( uint next = 0; next < useful.size(); ++next ) {
 319     assert( next < unique(), "Unique useful nodes < total nodes");
 320     Node *n  = useful.at(next);
 321     uint max = n->len();
 322     for( uint i = 0; i < max; ++i ) {
 323       Node *m = n->in(i);
 324       if (not_a_node(m))  continue;
 325       useful.push(m);
 326     }
 327   }
 328 }
 329 
 330 // Update dead_node_list with any missing dead nodes using useful
 331 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 332 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 333   uint max_idx = unique();
 334   VectorSet& useful_node_set = useful.member_set();
 335 
 336   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 337     // If node with index node_idx is not in useful set,
 338     // mark it as dead in dead node list.
 339     if (! useful_node_set.test(node_idx) ) {
 340       record_dead_node(node_idx);
 341     }
 342   }
 343 }
 344 
 345 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 346   int shift = 0;
 347   for (int i = 0; i < inlines->length(); i++) {
 348     CallGenerator* cg = inlines->at(i);
 349     CallNode* call = cg->call_node();
 350     if (shift > 0) {
 351       inlines->at_put(i-shift, cg);
 352     }
 353     if (!useful.member(call)) {
 354       shift++;
 355     }
 356   }
 357   inlines->trunc_to(inlines->length()-shift);
 358 }
 359 
 360 // Disconnect all useless nodes by disconnecting those at the boundary.
 361 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 362   uint next = 0;
 363   while (next < useful.size()) {
 364     Node *n = useful.at(next++);
 365     if (n->is_SafePoint()) {
 366       // We're done with a parsing phase. Replaced nodes are not valid
 367       // beyond that point.
 368       n->as_SafePoint()->delete_replaced_nodes();
 369     }
 370     // Use raw traversal of out edges since this code removes out edges
 371     int max = n->outcnt();
 372     for (int j = 0; j < max; ++j) {
 373       Node* child = n->raw_out(j);
 374       if (! useful.member(child)) {
 375         assert(!child->is_top() || child != top(),
 376                "If top is cached in Compile object it is in useful list");
 377         // Only need to remove this out-edge to the useless node
 378         n->raw_del_out(j);
 379         --j;
 380         --max;
 381       }
 382     }
 383     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 384       record_for_igvn(n->unique_out());
 385     }
 386   }
 387   // Remove useless macro and predicate opaq nodes
 388   for (int i = C->macro_count()-1; i >= 0; i--) {
 389     Node* n = C->macro_node(i);
 390     if (!useful.member(n)) {
 391       remove_macro_node(n);
 392     }
 393   }
 394   // Remove useless CastII nodes with range check dependency
 395   for (int i = range_check_cast_count() - 1; i >= 0; i--) {
 396     Node* cast = range_check_cast_node(i);
 397     if (!useful.member(cast)) {
 398       remove_range_check_cast(cast);
 399     }
 400   }
 401   // Remove useless expensive node
 402   for (int i = C->expensive_count()-1; i >= 0; i--) {
 403     Node* n = C->expensive_node(i);
 404     if (!useful.member(n)) {
 405       remove_expensive_node(n);
 406     }
 407   }
 408   // Remove useless value type nodes
 409   if (_value_type_nodes != NULL) {
 410     _value_type_nodes->remove_useless_nodes(useful.member_set());
 411   }
 412   // clean up the late inline lists
 413   remove_useless_late_inlines(&_string_late_inlines, useful);
 414   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 415   remove_useless_late_inlines(&_late_inlines, useful);
 416   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 417 }
 418 
 419 //------------------------------frame_size_in_words-----------------------------
 420 // frame_slots in units of words
 421 int Compile::frame_size_in_words() const {
 422   // shift is 0 in LP32 and 1 in LP64
 423   const int shift = (LogBytesPerWord - LogBytesPerInt);
 424   int words = _frame_slots >> shift;
 425   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 426   return words;
 427 }
 428 
 429 // To bang the stack of this compiled method we use the stack size
 430 // that the interpreter would need in case of a deoptimization. This
 431 // removes the need to bang the stack in the deoptimization blob which
 432 // in turn simplifies stack overflow handling.
 433 int Compile::bang_size_in_bytes() const {
 434   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
 435 }
 436 
 437 // ============================================================================
 438 //------------------------------CompileWrapper---------------------------------
 439 class CompileWrapper : public StackObj {
 440   Compile *const _compile;
 441  public:
 442   CompileWrapper(Compile* compile);
 443 
 444   ~CompileWrapper();
 445 };
 446 
 447 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 448   // the Compile* pointer is stored in the current ciEnv:
 449   ciEnv* env = compile->env();
 450   assert(env == ciEnv::current(), "must already be a ciEnv active");
 451   assert(env->compiler_data() == NULL, "compile already active?");
 452   env->set_compiler_data(compile);
 453   assert(compile == Compile::current(), "sanity");
 454 
 455   compile->set_type_dict(NULL);
 456   compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
 457   compile->clone_map().set_clone_idx(0);
 458   compile->set_type_hwm(NULL);
 459   compile->set_type_last_size(0);
 460   compile->set_last_tf(NULL, NULL);
 461   compile->set_indexSet_arena(NULL);
 462   compile->set_indexSet_free_block_list(NULL);
 463   compile->init_type_arena();
 464   Type::Initialize(compile);
 465   _compile->set_scratch_buffer_blob(NULL);
 466   _compile->begin_method();
 467   _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
 468 }
 469 CompileWrapper::~CompileWrapper() {
 470   _compile->end_method();
 471   if (_compile->scratch_buffer_blob() != NULL)
 472     BufferBlob::free(_compile->scratch_buffer_blob());
 473   _compile->env()->set_compiler_data(NULL);
 474 }
 475 
 476 
 477 //----------------------------print_compile_messages---------------------------
 478 void Compile::print_compile_messages() {
 479 #ifndef PRODUCT
 480   // Check if recompiling
 481   if (_subsume_loads == false && PrintOpto) {
 482     // Recompiling without allowing machine instructions to subsume loads
 483     tty->print_cr("*********************************************************");
 484     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 485     tty->print_cr("*********************************************************");
 486   }
 487   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 488     // Recompiling without escape analysis
 489     tty->print_cr("*********************************************************");
 490     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 491     tty->print_cr("*********************************************************");
 492   }
 493   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 494     // Recompiling without boxing elimination
 495     tty->print_cr("*********************************************************");
 496     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 497     tty->print_cr("*********************************************************");
 498   }
 499   if (C->directive()->BreakAtCompileOption) {
 500     // Open the debugger when compiling this method.
 501     tty->print("### Breaking when compiling: ");
 502     method()->print_short_name();
 503     tty->cr();
 504     BREAKPOINT;
 505   }
 506 
 507   if( PrintOpto ) {
 508     if (is_osr_compilation()) {
 509       tty->print("[OSR]%3d", _compile_id);
 510     } else {
 511       tty->print("%3d", _compile_id);
 512     }
 513   }
 514 #endif
 515 }
 516 
 517 
 518 //-----------------------init_scratch_buffer_blob------------------------------
 519 // Construct a temporary BufferBlob and cache it for this compile.
 520 void Compile::init_scratch_buffer_blob(int const_size) {
 521   // If there is already a scratch buffer blob allocated and the
 522   // constant section is big enough, use it.  Otherwise free the
 523   // current and allocate a new one.
 524   BufferBlob* blob = scratch_buffer_blob();
 525   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 526     // Use the current blob.
 527   } else {
 528     if (blob != NULL) {
 529       BufferBlob::free(blob);
 530     }
 531 
 532     ResourceMark rm;
 533     _scratch_const_size = const_size;
 534     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
 535     blob = BufferBlob::create("Compile::scratch_buffer", size);
 536     // Record the buffer blob for next time.
 537     set_scratch_buffer_blob(blob);
 538     // Have we run out of code space?
 539     if (scratch_buffer_blob() == NULL) {
 540       // Let CompilerBroker disable further compilations.
 541       record_failure("Not enough space for scratch buffer in CodeCache");
 542       return;
 543     }
 544   }
 545 
 546   // Initialize the relocation buffers
 547   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 548   set_scratch_locs_memory(locs_buf);
 549 }
 550 
 551 
 552 //-----------------------scratch_emit_size-------------------------------------
 553 // Helper function that computes size by emitting code
 554 uint Compile::scratch_emit_size(const Node* n) {
 555   // Start scratch_emit_size section.
 556   set_in_scratch_emit_size(true);
 557 
 558   // Emit into a trash buffer and count bytes emitted.
 559   // This is a pretty expensive way to compute a size,
 560   // but it works well enough if seldom used.
 561   // All common fixed-size instructions are given a size
 562   // method by the AD file.
 563   // Note that the scratch buffer blob and locs memory are
 564   // allocated at the beginning of the compile task, and
 565   // may be shared by several calls to scratch_emit_size.
 566   // The allocation of the scratch buffer blob is particularly
 567   // expensive, since it has to grab the code cache lock.
 568   BufferBlob* blob = this->scratch_buffer_blob();
 569   assert(blob != NULL, "Initialize BufferBlob at start");
 570   assert(blob->size() > MAX_inst_size, "sanity");
 571   relocInfo* locs_buf = scratch_locs_memory();
 572   address blob_begin = blob->content_begin();
 573   address blob_end   = (address)locs_buf;
 574   assert(blob->contains(blob_end), "sanity");
 575   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 576   buf.initialize_consts_size(_scratch_const_size);
 577   buf.initialize_stubs_size(MAX_stubs_size);
 578   assert(locs_buf != NULL, "sanity");
 579   int lsize = MAX_locs_size / 3;
 580   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 581   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 582   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 583   // Mark as scratch buffer.
 584   buf.consts()->set_scratch_emit();
 585   buf.insts()->set_scratch_emit();
 586   buf.stubs()->set_scratch_emit();
 587 
 588   // Do the emission.
 589 
 590   Label fakeL; // Fake label for branch instructions.
 591   Label*   saveL = NULL;
 592   uint save_bnum = 0;
 593   bool is_branch = n->is_MachBranch();
 594   if (is_branch) {
 595     MacroAssembler masm(&buf);
 596     masm.bind(fakeL);
 597     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 598     n->as_MachBranch()->label_set(&fakeL, 0);
 599   }
 600   n->emit(buf, this->regalloc());
 601 
 602   // Emitting into the scratch buffer should not fail
 603   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
 604 
 605   if (is_branch) // Restore label.
 606     n->as_MachBranch()->label_set(saveL, save_bnum);
 607 
 608   // End scratch_emit_size section.
 609   set_in_scratch_emit_size(false);
 610 
 611   return buf.insts_size();
 612 }
 613 
 614 
 615 // ============================================================================
 616 //------------------------------Compile standard-------------------------------
 617 debug_only( int Compile::_debug_idx = 100000; )
 618 
 619 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 620 // the continuation bci for on stack replacement.
 621 
 622 
 623 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 624                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
 625                 : Phase(Compiler),
 626                   _env(ci_env),
 627                   _directive(directive),
 628                   _log(ci_env->log()),
 629                   _compile_id(ci_env->compile_id()),
 630                   _save_argument_registers(false),
 631                   _stub_name(NULL),
 632                   _stub_function(NULL),
 633                   _stub_entry_point(NULL),
 634                   _method(target),
 635                   _entry_bci(osr_bci),
 636                   _initial_gvn(NULL),
 637                   _for_igvn(NULL),
 638                   _warm_calls(NULL),
 639                   _subsume_loads(subsume_loads),
 640                   _do_escape_analysis(do_escape_analysis),
 641                   _eliminate_boxing(eliminate_boxing),
 642                   _failure_reason(NULL),
 643                   _code_buffer("Compile::Fill_buffer"),
 644                   _orig_pc_slot(0),
 645                   _orig_pc_slot_offset_in_bytes(0),
 646                   _has_method_handle_invokes(false),
 647                   _mach_constant_base_node(NULL),
 648                   _node_bundling_limit(0),
 649                   _node_bundling_base(NULL),
 650                   _java_calls(0),
 651                   _inner_loops(0),
 652                   _scratch_const_size(-1),
 653                   _in_scratch_emit_size(false),
 654                   _dead_node_list(comp_arena()),
 655                   _dead_node_count(0),
 656 #ifndef PRODUCT
 657                   _trace_opto_output(directive->TraceOptoOutputOption),
 658                   _in_dump_cnt(0),
 659                   _printer(IdealGraphPrinter::printer()),
 660 #endif
 661                   _congraph(NULL),
 662                   _comp_arena(mtCompiler),
 663                   _node_arena(mtCompiler),
 664                   _old_arena(mtCompiler),
 665                   _Compile_types(mtCompiler),
 666                   _replay_inline_data(NULL),
 667                   _late_inlines(comp_arena(), 2, 0, NULL),
 668                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 669                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 670                   _late_inlines_pos(0),
 671                   _number_of_mh_late_inlines(0),
 672                   _inlining_progress(false),
 673                   _inlining_incrementally(false),
 674                   _print_inlining_list(NULL),
 675                   _print_inlining_stream(NULL),
 676                   _print_inlining_idx(0),
 677                   _print_inlining_output(NULL),
 678                   _interpreter_frame_size(0),
 679                   _max_node_limit(MaxNodeLimit),
 680                   _has_reserved_stack_access(target->has_reserved_stack_access()) {
 681   C = this;
 682 #ifndef PRODUCT
 683   if (_printer != NULL) {
 684     _printer->set_compile(this);
 685   }
 686 #endif
 687   CompileWrapper cw(this);
 688 
 689   if (CITimeVerbose) {
 690     tty->print(" ");
 691     target->holder()->name()->print();
 692     tty->print(".");
 693     target->print_short_name();
 694     tty->print("  ");
 695   }
 696   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 697   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 698 
 699 #ifndef PRODUCT
 700   bool print_opto_assembly = directive->PrintOptoAssemblyOption;
 701   if (!print_opto_assembly) {
 702     bool print_assembly = directive->PrintAssemblyOption;
 703     if (print_assembly && !Disassembler::can_decode()) {
 704       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 705       print_opto_assembly = true;
 706     }
 707   }
 708   set_print_assembly(print_opto_assembly);
 709   set_parsed_irreducible_loop(false);
 710 
 711   if (directive->ReplayInlineOption) {
 712     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 713   }
 714 #endif
 715   set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
 716   set_print_intrinsics(directive->PrintIntrinsicsOption);
 717   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 718 
 719   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 720     // Make sure the method being compiled gets its own MDO,
 721     // so we can at least track the decompile_count().
 722     // Need MDO to record RTM code generation state.
 723     method()->ensure_method_data();
 724   }
 725 
 726   Init(::AliasLevel);
 727 
 728 
 729   print_compile_messages();
 730 
 731   _ilt = InlineTree::build_inline_tree_root();
 732 
 733   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 734   assert(num_alias_types() >= AliasIdxRaw, "");
 735 
 736 #define MINIMUM_NODE_HASH  1023
 737   // Node list that Iterative GVN will start with
 738   Unique_Node_List for_igvn(comp_arena());
 739   set_for_igvn(&for_igvn);
 740 
 741   // GVN that will be run immediately on new nodes
 742   uint estimated_size = method()->code_size()*4+64;
 743   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 744   PhaseGVN gvn(node_arena(), estimated_size);
 745   set_initial_gvn(&gvn);
 746 
 747   print_inlining_init();
 748   { // Scope for timing the parser
 749     TracePhase tp("parse", &timers[_t_parser]);
 750 
 751     // Put top into the hash table ASAP.
 752     initial_gvn()->transform_no_reclaim(top());
 753 
 754     // Set up tf(), start(), and find a CallGenerator.
 755     CallGenerator* cg = NULL;
 756     if (is_osr_compilation()) {
 757       const TypeTuple *domain = StartOSRNode::osr_domain();
 758       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 759       init_tf(TypeFunc::make(domain, range));
 760       StartNode* s = new StartOSRNode(root(), domain);
 761       initial_gvn()->set_type_bottom(s);
 762       init_start(s);
 763       cg = CallGenerator::for_osr(method(), entry_bci());
 764     } else {
 765       // Normal case.
 766       init_tf(TypeFunc::make(method()));
 767       StartNode* s = new StartNode(root(), tf()->domain_cc());
 768       initial_gvn()->set_type_bottom(s);
 769       init_start(s);
 770       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
 771         // With java.lang.ref.reference.get() we must go through the
 772         // intrinsic when G1 is enabled - even when get() is the root
 773         // method of the compile - so that, if necessary, the value in
 774         // the referent field of the reference object gets recorded by
 775         // the pre-barrier code.
 776         // Specifically, if G1 is enabled, the value in the referent
 777         // field is recorded by the G1 SATB pre barrier. This will
 778         // result in the referent being marked live and the reference
 779         // object removed from the list of discovered references during
 780         // reference processing.
 781         cg = find_intrinsic(method(), false);
 782       }
 783       if (cg == NULL) {
 784         float past_uses = method()->interpreter_invocation_count();
 785         float expected_uses = past_uses;
 786         cg = CallGenerator::for_inline(method(), expected_uses);
 787       }
 788     }
 789     if (failing())  return;
 790     if (cg == NULL) {
 791       record_method_not_compilable("cannot parse method");
 792       return;
 793     }
 794     JVMState* jvms = build_start_state(start(), tf());
 795     if ((jvms = cg->generate(jvms)) == NULL) {
 796       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 797         record_method_not_compilable("method parse failed");
 798       }
 799       return;
 800     }
 801     GraphKit kit(jvms);
 802 
 803     if (!kit.stopped()) {
 804       // Accept return values, and transfer control we know not where.
 805       // This is done by a special, unique ReturnNode bound to root.
 806       return_values(kit.jvms());
 807     }
 808 
 809     if (kit.has_exceptions()) {
 810       // Any exceptions that escape from this call must be rethrown
 811       // to whatever caller is dynamically above us on the stack.
 812       // This is done by a special, unique RethrowNode bound to root.
 813       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 814     }
 815 
 816     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 817 
 818     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 819       inline_string_calls(true);
 820     }
 821 
 822     if (failing())  return;
 823 
 824     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 825 
 826     // Remove clutter produced by parsing.
 827     if (!failing()) {
 828       ResourceMark rm;
 829       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 830     }
 831   }
 832 
 833   // Note:  Large methods are capped off in do_one_bytecode().
 834   if (failing())  return;
 835 
 836   // After parsing, node notes are no longer automagic.
 837   // They must be propagated by register_new_node_with_optimizer(),
 838   // clone(), or the like.
 839   set_default_node_notes(NULL);
 840 
 841   for (;;) {
 842     int successes = Inline_Warm();
 843     if (failing())  return;
 844     if (successes == 0)  break;
 845   }
 846 
 847   // Drain the list.
 848   Finish_Warm();
 849 #ifndef PRODUCT
 850   if (_printer && _printer->should_print(1)) {
 851     _printer->print_inlining();
 852   }
 853 #endif
 854 
 855   if (failing())  return;
 856   NOT_PRODUCT( verify_graph_edges(); )
 857 
 858   // Now optimize
 859   Optimize();
 860   if (failing())  return;
 861   NOT_PRODUCT( verify_graph_edges(); )
 862 
 863 #ifndef PRODUCT
 864   if (PrintIdeal) {
 865     ttyLocker ttyl;  // keep the following output all in one block
 866     // This output goes directly to the tty, not the compiler log.
 867     // To enable tools to match it up with the compilation activity,
 868     // be sure to tag this tty output with the compile ID.
 869     if (xtty != NULL) {
 870       xtty->head("ideal compile_id='%d'%s", compile_id(),
 871                  is_osr_compilation()    ? " compile_kind='osr'" :
 872                  "");
 873     }
 874     root()->dump(9999);
 875     if (xtty != NULL) {
 876       xtty->tail("ideal");
 877     }
 878   }
 879 #endif
 880 
 881   NOT_PRODUCT( verify_barriers(); )
 882 
 883   // Dump compilation data to replay it.
 884   if (directive->DumpReplayOption) {
 885     env()->dump_replay_data(_compile_id);
 886   }
 887   if (directive->DumpInlineOption && (ilt() != NULL)) {
 888     env()->dump_inline_data(_compile_id);
 889   }
 890 
 891   // Now that we know the size of all the monitors we can add a fixed slot
 892   // for the original deopt pc.
 893 
 894   _orig_pc_slot =  fixed_slots();
 895   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 896   set_fixed_slots(next_slot);
 897 
 898   // Compute when to use implicit null checks. Used by matching trap based
 899   // nodes and NullCheck optimization.
 900   set_allowed_deopt_reasons();
 901 
 902   // Now generate code
 903   Code_Gen();
 904   if (failing())  return;
 905 
 906   // Check if we want to skip execution of all compiled code.
 907   {
 908 #ifndef PRODUCT
 909     if (OptoNoExecute) {
 910       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 911       return;
 912     }
 913 #endif
 914     TracePhase tp("install_code", &timers[_t_registerMethod]);
 915 
 916     if (is_osr_compilation()) {
 917       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 918       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 919     } else {
 920       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 921       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 922     }
 923 
 924     env()->register_method(_method, _entry_bci,
 925                            &_code_offsets,
 926                            _orig_pc_slot_offset_in_bytes,
 927                            code_buffer(),
 928                            frame_size_in_words(), _oop_map_set,
 929                            &_handler_table, &_inc_table,
 930                            compiler,
 931                            has_unsafe_access(),
 932                            SharedRuntime::is_wide_vector(max_vector_size()),
 933                            rtm_state()
 934                            );
 935 
 936     if (log() != NULL) // Print code cache state into compiler log
 937       log()->code_cache_state();
 938   }
 939 }
 940 
 941 //------------------------------Compile----------------------------------------
 942 // Compile a runtime stub
 943 Compile::Compile( ciEnv* ci_env,
 944                   TypeFunc_generator generator,
 945                   address stub_function,
 946                   const char *stub_name,
 947                   int is_fancy_jump,
 948                   bool pass_tls,
 949                   bool save_arg_registers,
 950                   bool return_pc,
 951                   DirectiveSet* directive)
 952   : Phase(Compiler),
 953     _env(ci_env),
 954     _directive(directive),
 955     _log(ci_env->log()),
 956     _compile_id(0),
 957     _save_argument_registers(save_arg_registers),
 958     _method(NULL),
 959     _stub_name(stub_name),
 960     _stub_function(stub_function),
 961     _stub_entry_point(NULL),
 962     _entry_bci(InvocationEntryBci),
 963     _initial_gvn(NULL),
 964     _for_igvn(NULL),
 965     _warm_calls(NULL),
 966     _orig_pc_slot(0),
 967     _orig_pc_slot_offset_in_bytes(0),
 968     _subsume_loads(true),
 969     _do_escape_analysis(false),
 970     _eliminate_boxing(false),
 971     _failure_reason(NULL),
 972     _code_buffer("Compile::Fill_buffer"),
 973     _has_method_handle_invokes(false),
 974     _mach_constant_base_node(NULL),
 975     _node_bundling_limit(0),
 976     _node_bundling_base(NULL),
 977     _java_calls(0),
 978     _inner_loops(0),
 979 #ifndef PRODUCT
 980     _trace_opto_output(directive->TraceOptoOutputOption),
 981     _in_dump_cnt(0),
 982     _printer(NULL),
 983 #endif
 984     _comp_arena(mtCompiler),
 985     _node_arena(mtCompiler),
 986     _old_arena(mtCompiler),
 987     _Compile_types(mtCompiler),
 988     _dead_node_list(comp_arena()),
 989     _dead_node_count(0),
 990     _congraph(NULL),
 991     _replay_inline_data(NULL),
 992     _number_of_mh_late_inlines(0),
 993     _inlining_progress(false),
 994     _inlining_incrementally(false),
 995     _print_inlining_list(NULL),
 996     _print_inlining_stream(NULL),
 997     _print_inlining_idx(0),
 998     _print_inlining_output(NULL),
 999     _allowed_reasons(0),
1000     _interpreter_frame_size(0),
1001     _max_node_limit(MaxNodeLimit),
1002     _has_reserved_stack_access(false) {
1003   C = this;
1004 
1005   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1006   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1007 
1008 #ifndef PRODUCT
1009   set_print_assembly(PrintFrameConverterAssembly);
1010   set_parsed_irreducible_loop(false);
1011 #endif
1012   set_has_irreducible_loop(false); // no loops
1013 
1014   CompileWrapper cw(this);
1015   Init(/*AliasLevel=*/ 0);
1016   init_tf((*generator)());
1017 
1018   {
1019     // The following is a dummy for the sake of GraphKit::gen_stub
1020     Unique_Node_List for_igvn(comp_arena());
1021     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1022     PhaseGVN gvn(Thread::current()->resource_area(),255);
1023     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1024     gvn.transform_no_reclaim(top());
1025 
1026     GraphKit kit;
1027     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1028   }
1029 
1030   NOT_PRODUCT( verify_graph_edges(); )
1031   Code_Gen();
1032   if (failing())  return;
1033 
1034 
1035   // Entry point will be accessed using compile->stub_entry_point();
1036   if (code_buffer() == NULL) {
1037     Matcher::soft_match_failure();
1038   } else {
1039     if (PrintAssembly && (WizardMode || Verbose))
1040       tty->print_cr("### Stub::%s", stub_name);
1041 
1042     if (!failing()) {
1043       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1044 
1045       // Make the NMethod
1046       // For now we mark the frame as never safe for profile stackwalking
1047       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1048                                                       code_buffer(),
1049                                                       CodeOffsets::frame_never_safe,
1050                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1051                                                       frame_size_in_words(),
1052                                                       _oop_map_set,
1053                                                       save_arg_registers);
1054       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1055 
1056       _stub_entry_point = rs->entry_point();
1057     }
1058   }
1059 }
1060 
1061 //------------------------------Init-------------------------------------------
1062 // Prepare for a single compilation
1063 void Compile::Init(int aliaslevel) {
1064   _unique  = 0;
1065   _regalloc = NULL;
1066 
1067   _tf      = NULL;  // filled in later
1068   _top     = NULL;  // cached later
1069   _matcher = NULL;  // filled in later
1070   _cfg     = NULL;  // filled in later
1071 
1072   set_24_bit_selection_and_mode(Use24BitFP, false);
1073 
1074   _node_note_array = NULL;
1075   _default_node_notes = NULL;
1076   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1077 
1078   _immutable_memory = NULL; // filled in at first inquiry
1079 
1080   // Globally visible Nodes
1081   // First set TOP to NULL to give safe behavior during creation of RootNode
1082   set_cached_top_node(NULL);
1083   set_root(new RootNode());
1084   // Now that you have a Root to point to, create the real TOP
1085   set_cached_top_node( new ConNode(Type::TOP) );
1086   set_recent_alloc(NULL, NULL);
1087 
1088   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1089   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1090   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1091   env()->set_dependencies(new Dependencies(env()));
1092 
1093   _fixed_slots = 0;
1094   set_has_split_ifs(false);
1095   set_has_loops(has_method() && method()->has_loops()); // first approximation
1096   set_has_stringbuilder(false);
1097   set_has_boxed_value(false);
1098   _trap_can_recompile = false;  // no traps emitted yet
1099   _major_progress = true; // start out assuming good things will happen
1100   set_has_unsafe_access(false);
1101   set_max_vector_size(0);
1102   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1103   set_decompile_count(0);
1104 
1105   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1106   set_num_loop_opts(LoopOptsCount);
1107   set_do_inlining(Inline);
1108   set_max_inline_size(MaxInlineSize);
1109   set_freq_inline_size(FreqInlineSize);
1110   set_do_scheduling(OptoScheduling);
1111   set_do_count_invocations(false);
1112   set_do_method_data_update(false);
1113 
1114   set_do_vector_loop(false);
1115 
1116   if (AllowVectorizeOnDemand) {
1117     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1118       set_do_vector_loop(true);
1119       NOT_PRODUCT(if (do_vector_loop() && Verbose) {tty->print("Compile::Init: do vectorized loops (SIMD like) for method %s\n",  method()->name()->as_quoted_ascii());})
1120     } else if (has_method() && method()->name() != 0 &&
1121                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1122       set_do_vector_loop(true);
1123     }
1124   }
1125   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1126   NOT_PRODUCT(if (use_cmove() && Verbose && has_method()) {tty->print("Compile::Init: use CMove without profitability tests for method %s\n",  method()->name()->as_quoted_ascii());})
1127 
1128   set_age_code(has_method() && method()->profile_aging());
1129   set_rtm_state(NoRTM); // No RTM lock eliding by default
1130   _max_node_limit = _directive->MaxNodeLimitOption;
1131 
1132 #if INCLUDE_RTM_OPT
1133   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1134     int rtm_state = method()->method_data()->rtm_state();
1135     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1136       // Don't generate RTM lock eliding code.
1137       set_rtm_state(NoRTM);
1138     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1139       // Generate RTM lock eliding code without abort ratio calculation code.
1140       set_rtm_state(UseRTM);
1141     } else if (UseRTMDeopt) {
1142       // Generate RTM lock eliding code and include abort ratio calculation
1143       // code if UseRTMDeopt is on.
1144       set_rtm_state(ProfileRTM);
1145     }
1146   }
1147 #endif
1148   if (debug_info()->recording_non_safepoints()) {
1149     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1150                         (comp_arena(), 8, 0, NULL));
1151     set_default_node_notes(Node_Notes::make(this));
1152   }
1153 
1154   // // -- Initialize types before each compile --
1155   // // Update cached type information
1156   // if( _method && _method->constants() )
1157   //   Type::update_loaded_types(_method, _method->constants());
1158 
1159   // Init alias_type map.
1160   if (!_do_escape_analysis && aliaslevel == 3)
1161     aliaslevel = 2;  // No unique types without escape analysis
1162   _AliasLevel = aliaslevel;
1163   const int grow_ats = 16;
1164   _max_alias_types = grow_ats;
1165   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1166   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1167   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1168   {
1169     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1170   }
1171   // Initialize the first few types.
1172   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1173   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1174   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1175   _num_alias_types = AliasIdxRaw+1;
1176   // Zero out the alias type cache.
1177   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1178   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1179   probe_alias_cache(NULL)->_index = AliasIdxTop;
1180 
1181   _intrinsics = NULL;
1182   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1183   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1184   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1185   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1186   _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
1187   register_library_intrinsics();
1188 }
1189 
1190 //---------------------------init_start----------------------------------------
1191 // Install the StartNode on this compile object.
1192 void Compile::init_start(StartNode* s) {
1193   if (failing())
1194     return; // already failing
1195   assert(s == start(), "");
1196 }
1197 
1198 /**
1199  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1200  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1201  * the ideal graph.
1202  */
1203 StartNode* Compile::start() const {
1204   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1205   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1206     Node* start = root()->fast_out(i);
1207     if (start->is_Start()) {
1208       return start->as_Start();
1209     }
1210   }
1211   fatal("Did not find Start node!");
1212   return NULL;
1213 }
1214 
1215 //-------------------------------immutable_memory-------------------------------------
1216 // Access immutable memory
1217 Node* Compile::immutable_memory() {
1218   if (_immutable_memory != NULL) {
1219     return _immutable_memory;
1220   }
1221   StartNode* s = start();
1222   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1223     Node *p = s->fast_out(i);
1224     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1225       _immutable_memory = p;
1226       return _immutable_memory;
1227     }
1228   }
1229   ShouldNotReachHere();
1230   return NULL;
1231 }
1232 
1233 //----------------------set_cached_top_node------------------------------------
1234 // Install the cached top node, and make sure Node::is_top works correctly.
1235 void Compile::set_cached_top_node(Node* tn) {
1236   if (tn != NULL)  verify_top(tn);
1237   Node* old_top = _top;
1238   _top = tn;
1239   // Calling Node::setup_is_top allows the nodes the chance to adjust
1240   // their _out arrays.
1241   if (_top != NULL)     _top->setup_is_top();
1242   if (old_top != NULL)  old_top->setup_is_top();
1243   assert(_top == NULL || top()->is_top(), "");
1244 }
1245 
1246 #ifdef ASSERT
1247 uint Compile::count_live_nodes_by_graph_walk() {
1248   Unique_Node_List useful(comp_arena());
1249   // Get useful node list by walking the graph.
1250   identify_useful_nodes(useful);
1251   return useful.size();
1252 }
1253 
1254 void Compile::print_missing_nodes() {
1255 
1256   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1257   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1258     return;
1259   }
1260 
1261   // This is an expensive function. It is executed only when the user
1262   // specifies VerifyIdealNodeCount option or otherwise knows the
1263   // additional work that needs to be done to identify reachable nodes
1264   // by walking the flow graph and find the missing ones using
1265   // _dead_node_list.
1266 
1267   Unique_Node_List useful(comp_arena());
1268   // Get useful node list by walking the graph.
1269   identify_useful_nodes(useful);
1270 
1271   uint l_nodes = C->live_nodes();
1272   uint l_nodes_by_walk = useful.size();
1273 
1274   if (l_nodes != l_nodes_by_walk) {
1275     if (_log != NULL) {
1276       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1277       _log->stamp();
1278       _log->end_head();
1279     }
1280     VectorSet& useful_member_set = useful.member_set();
1281     int last_idx = l_nodes_by_walk;
1282     for (int i = 0; i < last_idx; i++) {
1283       if (useful_member_set.test(i)) {
1284         if (_dead_node_list.test(i)) {
1285           if (_log != NULL) {
1286             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1287           }
1288           if (PrintIdealNodeCount) {
1289             // Print the log message to tty
1290               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1291               useful.at(i)->dump();
1292           }
1293         }
1294       }
1295       else if (! _dead_node_list.test(i)) {
1296         if (_log != NULL) {
1297           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1298         }
1299         if (PrintIdealNodeCount) {
1300           // Print the log message to tty
1301           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1302         }
1303       }
1304     }
1305     if (_log != NULL) {
1306       _log->tail("mismatched_nodes");
1307     }
1308   }
1309 }
1310 void Compile::record_modified_node(Node* n) {
1311   if (_modified_nodes != NULL && !_inlining_incrementally &&
1312       n->outcnt() != 0 && !n->is_Con()) {
1313     _modified_nodes->push(n);
1314   }
1315 }
1316 
1317 void Compile::remove_modified_node(Node* n) {
1318   if (_modified_nodes != NULL) {
1319     _modified_nodes->remove(n);
1320   }
1321 }
1322 #endif
1323 
1324 #ifndef PRODUCT
1325 void Compile::verify_top(Node* tn) const {
1326   if (tn != NULL) {
1327     assert(tn->is_Con(), "top node must be a constant");
1328     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1329     assert(tn->in(0) != NULL, "must have live top node");
1330   }
1331 }
1332 #endif
1333 
1334 
1335 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1336 
1337 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1338   guarantee(arr != NULL, "");
1339   int num_blocks = arr->length();
1340   if (grow_by < num_blocks)  grow_by = num_blocks;
1341   int num_notes = grow_by * _node_notes_block_size;
1342   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1343   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1344   while (num_notes > 0) {
1345     arr->append(notes);
1346     notes     += _node_notes_block_size;
1347     num_notes -= _node_notes_block_size;
1348   }
1349   assert(num_notes == 0, "exact multiple, please");
1350 }
1351 
1352 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1353   if (source == NULL || dest == NULL)  return false;
1354 
1355   if (dest->is_Con())
1356     return false;               // Do not push debug info onto constants.
1357 
1358 #ifdef ASSERT
1359   // Leave a bread crumb trail pointing to the original node:
1360   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1361     dest->set_debug_orig(source);
1362   }
1363 #endif
1364 
1365   if (node_note_array() == NULL)
1366     return false;               // Not collecting any notes now.
1367 
1368   // This is a copy onto a pre-existing node, which may already have notes.
1369   // If both nodes have notes, do not overwrite any pre-existing notes.
1370   Node_Notes* source_notes = node_notes_at(source->_idx);
1371   if (source_notes == NULL || source_notes->is_clear())  return false;
1372   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1373   if (dest_notes == NULL || dest_notes->is_clear()) {
1374     return set_node_notes_at(dest->_idx, source_notes);
1375   }
1376 
1377   Node_Notes merged_notes = (*source_notes);
1378   // The order of operations here ensures that dest notes will win...
1379   merged_notes.update_from(dest_notes);
1380   return set_node_notes_at(dest->_idx, &merged_notes);
1381 }
1382 
1383 
1384 //--------------------------allow_range_check_smearing-------------------------
1385 // Gating condition for coalescing similar range checks.
1386 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1387 // single covering check that is at least as strong as any of them.
1388 // If the optimization succeeds, the simplified (strengthened) range check
1389 // will always succeed.  If it fails, we will deopt, and then give up
1390 // on the optimization.
1391 bool Compile::allow_range_check_smearing() const {
1392   // If this method has already thrown a range-check,
1393   // assume it was because we already tried range smearing
1394   // and it failed.
1395   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1396   return !already_trapped;
1397 }
1398 
1399 
1400 //------------------------------flatten_alias_type-----------------------------
1401 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1402   int offset = tj->offset();
1403   TypePtr::PTR ptr = tj->ptr();
1404 
1405   // Known instance (scalarizable allocation) alias only with itself.
1406   bool is_known_inst = tj->isa_oopptr() != NULL &&
1407                        tj->is_oopptr()->is_known_instance();
1408 
1409   // Process weird unsafe references.
1410   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1411     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1412     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1413     tj = TypeOopPtr::BOTTOM;
1414     ptr = tj->ptr();
1415     offset = tj->offset();
1416   }
1417 
1418   // Array pointers need some flattening
1419   const TypeAryPtr *ta = tj->isa_aryptr();
1420   if (ta && ta->is_stable()) {
1421     // Erase stability property for alias analysis.
1422     tj = ta = ta->cast_to_stable(false);
1423   }
1424   if( ta && is_known_inst ) {
1425     if ( offset != Type::OffsetBot &&
1426          offset > arrayOopDesc::length_offset_in_bytes() ) {
1427       offset = Type::OffsetBot; // Flatten constant access into array body only
1428       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id());
1429     }
1430   } else if( ta && _AliasLevel >= 2 ) {
1431     // For arrays indexed by constant indices, we flatten the alias
1432     // space to include all of the array body.  Only the header, klass
1433     // and array length can be accessed un-aliased.
1434     // For flattened value type array, each field has its own slice so
1435     // we must include the field offset.
1436     if( offset != Type::OffsetBot ) {
1437       if( ta->const_oop() ) { // MethodData* or Method*
1438         offset = Type::OffsetBot;   // Flatten constant access into array body
1439         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1440       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1441         // range is OK as-is.
1442         tj = ta = TypeAryPtr::RANGE;
1443       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1444         tj = TypeInstPtr::KLASS; // all klass loads look alike
1445         ta = TypeAryPtr::RANGE; // generic ignored junk
1446         ptr = TypePtr::BotPTR;
1447       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1448         tj = TypeInstPtr::MARK;
1449         ta = TypeAryPtr::RANGE; // generic ignored junk
1450         ptr = TypePtr::BotPTR;
1451       } else {                  // Random constant offset into array body
1452         offset = Type::OffsetBot;   // Flatten constant access into array body
1453         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1454       }
1455     }
1456     // Arrays of fixed size alias with arrays of unknown size.
1457     if (ta->size() != TypeInt::POS) {
1458       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1459       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset());
1460     }
1461     // Arrays of known objects become arrays of unknown objects.
1462     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1463       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1464       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1465     }
1466     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1467       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1468       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1469     }
1470     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1471     // cannot be distinguished by bytecode alone.
1472     if (ta->elem() == TypeInt::BOOL) {
1473       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1474       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1475       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1476     }
1477     // During the 2nd round of IterGVN, NotNull castings are removed.
1478     // Make sure the Bottom and NotNull variants alias the same.
1479     // Also, make sure exact and non-exact variants alias the same.
1480     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1481       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1482     }
1483   }
1484 
1485   // Oop pointers need some flattening
1486   const TypeInstPtr *to = tj->isa_instptr();
1487   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1488     ciInstanceKlass *k = to->klass()->as_instance_klass();
1489     if( ptr == TypePtr::Constant ) {
1490       if (to->klass() != ciEnv::current()->Class_klass() ||
1491           offset < k->size_helper() * wordSize) {
1492         // No constant oop pointers (such as Strings); they alias with
1493         // unknown strings.
1494         assert(!is_known_inst, "not scalarizable allocation");
1495         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1496       }
1497     } else if( is_known_inst ) {
1498       tj = to; // Keep NotNull and klass_is_exact for instance type
1499     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1500       // During the 2nd round of IterGVN, NotNull castings are removed.
1501       // Make sure the Bottom and NotNull variants alias the same.
1502       // Also, make sure exact and non-exact variants alias the same.
1503       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1504     }
1505     if (to->speculative() != NULL) {
1506       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id());
1507     }
1508     // Canonicalize the holder of this field
1509     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1510       // First handle header references such as a LoadKlassNode, even if the
1511       // object's klass is unloaded at compile time (4965979).
1512       if (!is_known_inst) { // Do it only for non-instance types
1513         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset));
1514       }
1515     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1516       // Static fields are in the space above the normal instance
1517       // fields in the java.lang.Class instance.
1518       if (to->klass() != ciEnv::current()->Class_klass()) {
1519         to = NULL;
1520         tj = TypeOopPtr::BOTTOM;
1521         offset = tj->offset();
1522       }
1523     } else {
1524       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1525       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1526         if( is_known_inst ) {
1527           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id());
1528         } else {
1529           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
1530         }
1531       }
1532     }
1533   }
1534 
1535   // Value type pointers need flattening
1536   const TypeValueTypePtr* tv = tj->isa_valuetypeptr();
1537   if (tv != NULL && _AliasLevel >= 2) {
1538     assert(tv->speculative() == NULL, "should not have speculative type information");
1539     ciValueKlass* vk = tv->klass()->as_value_klass();
1540     if (ptr == TypePtr::Constant) {
1541       assert(!is_known_inst, "not scalarizable allocation");
1542       tj = tv = TypeValueTypePtr::make(TypePtr::BotPTR, tv->value_klass(), NULL, Type::Offset(offset));
1543     } else if (is_known_inst) {
1544       tj = tv; // Keep NotNull and klass_is_exact for instance type
1545     } else if (ptr == TypePtr::NotNull || tv->klass_is_exact()) {
1546       // During the 2nd round of IterGVN, NotNull castings are removed.
1547       // Make sure the Bottom and NotNull variants alias the same.
1548       // Also, make sure exact and non-exact variants alias the same.
1549       tj = tv = TypeValueTypePtr::make(TypePtr::BotPTR, tv->value_klass(), tv->const_oop(), Type::Offset(offset));
1550     }
1551     // Canonicalize the holder of this field
1552     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1553       // First handle header references such as a LoadKlassNode, even if the
1554       // object's klass is unloaded at compile time (4965979).
1555       if (!is_known_inst) { // Do it only for non-instance types
1556         tj = tv = TypeValueTypePtr::make(TypePtr::BotPTR, env()->___Value_klass()->as_value_klass(), NULL, Type::Offset(offset));
1557       }
1558     } else if (offset < 0 || offset >= vk->size_helper() * wordSize) {
1559       // Static fields are in the space above the normal instance
1560       // fields in the java.lang.Class instance.
1561       tv = NULL;
1562       tj = TypeOopPtr::BOTTOM;
1563       offset = tj->offset();
1564     } else {
1565       ciInstanceKlass* canonical_holder = vk->get_canonical_holder(offset);
1566       assert(vk->equals(canonical_holder), "value types should not inherit fields");
1567     }
1568   }
1569 
1570   // Klass pointers to object array klasses need some flattening
1571   const TypeKlassPtr *tk = tj->isa_klassptr();
1572   if( tk ) {
1573     // If we are referencing a field within a Klass, we need
1574     // to assume the worst case of an Object.  Both exact and
1575     // inexact types must flatten to the same alias class so
1576     // use NotNull as the PTR.
1577     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1578 
1579       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1580                                    TypeKlassPtr::OBJECT->klass(),
1581                                    Type::Offset(offset));
1582     }
1583 
1584     ciKlass* klass = tk->klass();
1585     if (klass != NULL && klass->is_obj_array_klass()) {
1586       ciKlass* k = TypeAryPtr::OOPS->klass();
1587       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1588         k = TypeInstPtr::BOTTOM->klass();
1589       tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset));
1590     }
1591 
1592     // Check for precise loads from the primary supertype array and force them
1593     // to the supertype cache alias index.  Check for generic array loads from
1594     // the primary supertype array and also force them to the supertype cache
1595     // alias index.  Since the same load can reach both, we need to merge
1596     // these 2 disparate memories into the same alias class.  Since the
1597     // primary supertype array is read-only, there's no chance of confusion
1598     // where we bypass an array load and an array store.
1599     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1600     if (offset == Type::OffsetBot ||
1601         (offset >= primary_supers_offset &&
1602          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1603         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1604       offset = in_bytes(Klass::secondary_super_cache_offset());
1605       tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset));
1606     }
1607   }
1608 
1609   // Flatten all Raw pointers together.
1610   if (tj->base() == Type::RawPtr)
1611     tj = TypeRawPtr::BOTTOM;
1612 
1613   if (tj->base() == Type::AnyPtr)
1614     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1615 
1616   // Flatten all to bottom for now
1617   switch( _AliasLevel ) {
1618   case 0:
1619     tj = TypePtr::BOTTOM;
1620     break;
1621   case 1:                       // Flatten to: oop, static, field or array
1622     switch (tj->base()) {
1623     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1624     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1625     case Type::AryPtr:   // do not distinguish arrays at all
1626     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1627     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1628     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1629     default: ShouldNotReachHere();
1630     }
1631     break;
1632   case 2:                       // No collapsing at level 2; keep all splits
1633   case 3:                       // No collapsing at level 3; keep all splits
1634     break;
1635   default:
1636     Unimplemented();
1637   }
1638 
1639   offset = tj->offset();
1640   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1641 
1642   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1643           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1644           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1645           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1646           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1647           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1648           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr)  ,
1649           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1650   assert( tj->ptr() != TypePtr::TopPTR &&
1651           tj->ptr() != TypePtr::AnyNull &&
1652           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1653 //    assert( tj->ptr() != TypePtr::Constant ||
1654 //            tj->base() == Type::RawPtr ||
1655 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1656 
1657   return tj;
1658 }
1659 
1660 void Compile::AliasType::Init(int i, const TypePtr* at) {
1661   _index = i;
1662   _adr_type = at;
1663   _field = NULL;
1664   _element = NULL;
1665   _is_rewritable = true; // default
1666   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1667   if (atoop != NULL && atoop->is_known_instance()) {
1668     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1669     _general_index = Compile::current()->get_alias_index(gt);
1670   } else {
1671     _general_index = 0;
1672   }
1673 }
1674 
1675 BasicType Compile::AliasType::basic_type() const {
1676   if (element() != NULL) {
1677     const Type* element = adr_type()->is_aryptr()->elem();
1678     return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1679   } if (field() != NULL) {
1680     return field()->layout_type();
1681   } else {
1682     return T_ILLEGAL; // unknown
1683   }
1684 }
1685 
1686 //---------------------------------print_on------------------------------------
1687 #ifndef PRODUCT
1688 void Compile::AliasType::print_on(outputStream* st) {
1689   if (index() < 10)
1690         st->print("@ <%d> ", index());
1691   else  st->print("@ <%d>",  index());
1692   st->print(is_rewritable() ? "   " : " RO");
1693   int offset = adr_type()->offset();
1694   if (offset == Type::OffsetBot)
1695         st->print(" +any");
1696   else  st->print(" +%-3d", offset);
1697   st->print(" in ");
1698   adr_type()->dump_on(st);
1699   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1700   if (field() != NULL && tjp) {
1701     if (tjp->klass()  != field()->holder() ||
1702         tjp->offset() != field()->offset_in_bytes()) {
1703       st->print(" != ");
1704       field()->print();
1705       st->print(" ***");
1706     }
1707   }
1708 }
1709 
1710 void print_alias_types() {
1711   Compile* C = Compile::current();
1712   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1713   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1714     C->alias_type(idx)->print_on(tty);
1715     tty->cr();
1716   }
1717 }
1718 #endif
1719 
1720 
1721 //----------------------------probe_alias_cache--------------------------------
1722 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1723   intptr_t key = (intptr_t) adr_type;
1724   key ^= key >> logAliasCacheSize;
1725   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1726 }
1727 
1728 
1729 //-----------------------------grow_alias_types--------------------------------
1730 void Compile::grow_alias_types() {
1731   const int old_ats  = _max_alias_types; // how many before?
1732   const int new_ats  = old_ats;          // how many more?
1733   const int grow_ats = old_ats+new_ats;  // how many now?
1734   _max_alias_types = grow_ats;
1735   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1736   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1737   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1738   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1739 }
1740 
1741 
1742 //--------------------------------find_alias_type------------------------------
1743 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1744   if (_AliasLevel == 0)
1745     return alias_type(AliasIdxBot);
1746 
1747   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1748   if (ace->_adr_type == adr_type) {
1749     return alias_type(ace->_index);
1750   }
1751 
1752   // Handle special cases.
1753   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1754   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1755 
1756   // Do it the slow way.
1757   const TypePtr* flat = flatten_alias_type(adr_type);
1758 
1759 #ifdef ASSERT
1760   {
1761     ResourceMark rm;
1762     assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1763            Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1764     assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1765            Type::str(adr_type));
1766     if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1767       const TypeOopPtr* foop = flat->is_oopptr();
1768       // Scalarizable allocations have exact klass always.
1769       bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1770       const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1771       assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1772              Type::str(foop), Type::str(xoop));
1773     }
1774   }
1775 #endif
1776 
1777   int idx = AliasIdxTop;
1778   for (int i = 0; i < num_alias_types(); i++) {
1779     if (alias_type(i)->adr_type() == flat) {
1780       idx = i;
1781       break;
1782     }
1783   }
1784 
1785   if (idx == AliasIdxTop) {
1786     if (no_create)  return NULL;
1787     // Grow the array if necessary.
1788     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1789     // Add a new alias type.
1790     idx = _num_alias_types++;
1791     _alias_types[idx]->Init(idx, flat);
1792     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1793     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1794     if (flat->isa_instptr()) {
1795       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1796           && flat->is_instptr()->klass() == env()->Class_klass())
1797         alias_type(idx)->set_rewritable(false);
1798     }
1799     if (flat->isa_aryptr()) {
1800 #ifdef ASSERT
1801       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1802       // (T_BYTE has the weakest alignment and size restrictions...)
1803       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1804 #endif
1805       if (flat->offset() == TypePtr::OffsetBot) {
1806         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1807       }
1808     }
1809     if (flat->isa_klassptr()) {
1810       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1811         alias_type(idx)->set_rewritable(false);
1812       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1813         alias_type(idx)->set_rewritable(false);
1814       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1815         alias_type(idx)->set_rewritable(false);
1816       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1817         alias_type(idx)->set_rewritable(false);
1818     }
1819     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1820     // but the base pointer type is not distinctive enough to identify
1821     // references into JavaThread.)
1822 
1823     // Check for final fields.
1824     const TypeInstPtr* tinst = flat->isa_instptr();
1825     const TypeValueTypePtr* vtptr = flat->isa_valuetypeptr();
1826     ciField* field = NULL;
1827     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1828       if (tinst->const_oop() != NULL &&
1829           tinst->klass() == ciEnv::current()->Class_klass() &&
1830           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1831         // static field
1832         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1833         field = k->get_field_by_offset(tinst->offset(), true);
1834       } else {
1835         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1836         field = k->get_field_by_offset(tinst->offset(), false);
1837       }
1838     } else if (vtptr) {
1839       // Value type field
1840       ciValueKlass* vk = vtptr->klass()->as_value_klass();
1841       field = vk->get_field_by_offset(vtptr->offset(), false);
1842     }
1843     assert(field == NULL ||
1844            original_field == NULL ||
1845            (field->holder() == original_field->holder() &&
1846             field->offset() == original_field->offset() &&
1847             field->is_static() == original_field->is_static()), "wrong field?");
1848     // Set field() and is_rewritable() attributes.
1849     if (field != NULL)  alias_type(idx)->set_field(field);
1850   }
1851 
1852   // Fill the cache for next time.
1853   ace->_adr_type = adr_type;
1854   ace->_index    = idx;
1855   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1856 
1857   // Might as well try to fill the cache for the flattened version, too.
1858   AliasCacheEntry* face = probe_alias_cache(flat);
1859   if (face->_adr_type == NULL) {
1860     face->_adr_type = flat;
1861     face->_index    = idx;
1862     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1863   }
1864 
1865   return alias_type(idx);
1866 }
1867 
1868 
1869 Compile::AliasType* Compile::alias_type(ciField* field) {
1870   const TypeOopPtr* t;
1871   if (field->is_static())
1872     t = TypeInstPtr::make(field->holder()->java_mirror());
1873   else
1874     t = TypeOopPtr::make_from_klass_raw(field->holder());
1875   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1876   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1877   return atp;
1878 }
1879 
1880 
1881 //------------------------------have_alias_type--------------------------------
1882 bool Compile::have_alias_type(const TypePtr* adr_type) {
1883   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1884   if (ace->_adr_type == adr_type) {
1885     return true;
1886   }
1887 
1888   // Handle special cases.
1889   if (adr_type == NULL)             return true;
1890   if (adr_type == TypePtr::BOTTOM)  return true;
1891 
1892   return find_alias_type(adr_type, true, NULL) != NULL;
1893 }
1894 
1895 //-----------------------------must_alias--------------------------------------
1896 // True if all values of the given address type are in the given alias category.
1897 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1898   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1899   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1900   if (alias_idx == AliasIdxTop)         return false; // the empty category
1901   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1902 
1903   // the only remaining possible overlap is identity
1904   int adr_idx = get_alias_index(adr_type);
1905   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1906   assert(adr_idx == alias_idx ||
1907          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1908           && adr_type                       != TypeOopPtr::BOTTOM),
1909          "should not be testing for overlap with an unsafe pointer");
1910   return adr_idx == alias_idx;
1911 }
1912 
1913 //------------------------------can_alias--------------------------------------
1914 // True if any values of the given address type are in the given alias category.
1915 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1916   if (alias_idx == AliasIdxTop)         return false; // the empty category
1917   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1918   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1919   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1920 
1921   // the only remaining possible overlap is identity
1922   int adr_idx = get_alias_index(adr_type);
1923   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1924   return adr_idx == alias_idx;
1925 }
1926 
1927 
1928 
1929 //---------------------------pop_warm_call-------------------------------------
1930 WarmCallInfo* Compile::pop_warm_call() {
1931   WarmCallInfo* wci = _warm_calls;
1932   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1933   return wci;
1934 }
1935 
1936 //----------------------------Inline_Warm--------------------------------------
1937 int Compile::Inline_Warm() {
1938   // If there is room, try to inline some more warm call sites.
1939   // %%% Do a graph index compaction pass when we think we're out of space?
1940   if (!InlineWarmCalls)  return 0;
1941 
1942   int calls_made_hot = 0;
1943   int room_to_grow   = NodeCountInliningCutoff - unique();
1944   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1945   int amount_grown   = 0;
1946   WarmCallInfo* call;
1947   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1948     int est_size = (int)call->size();
1949     if (est_size > (room_to_grow - amount_grown)) {
1950       // This one won't fit anyway.  Get rid of it.
1951       call->make_cold();
1952       continue;
1953     }
1954     call->make_hot();
1955     calls_made_hot++;
1956     amount_grown   += est_size;
1957     amount_to_grow -= est_size;
1958   }
1959 
1960   if (calls_made_hot > 0)  set_major_progress();
1961   return calls_made_hot;
1962 }
1963 
1964 
1965 //----------------------------Finish_Warm--------------------------------------
1966 void Compile::Finish_Warm() {
1967   if (!InlineWarmCalls)  return;
1968   if (failing())  return;
1969   if (warm_calls() == NULL)  return;
1970 
1971   // Clean up loose ends, if we are out of space for inlining.
1972   WarmCallInfo* call;
1973   while ((call = pop_warm_call()) != NULL) {
1974     call->make_cold();
1975   }
1976 }
1977 
1978 //---------------------cleanup_loop_predicates-----------------------
1979 // Remove the opaque nodes that protect the predicates so that all unused
1980 // checks and uncommon_traps will be eliminated from the ideal graph
1981 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1982   if (predicate_count()==0) return;
1983   for (int i = predicate_count(); i > 0; i--) {
1984     Node * n = predicate_opaque1_node(i-1);
1985     assert(n->Opcode() == Op_Opaque1, "must be");
1986     igvn.replace_node(n, n->in(1));
1987   }
1988   assert(predicate_count()==0, "should be clean!");
1989 }
1990 
1991 void Compile::add_range_check_cast(Node* n) {
1992   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1993   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1994   _range_check_casts->append(n);
1995 }
1996 
1997 // Remove all range check dependent CastIINodes.
1998 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1999   for (int i = range_check_cast_count(); i > 0; i--) {
2000     Node* cast = range_check_cast_node(i-1);
2001     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2002     igvn.replace_node(cast, cast->in(1));
2003   }
2004   assert(range_check_cast_count() == 0, "should be empty");
2005 }
2006 
2007 void Compile::add_value_type(Node* n) {
2008   assert(n->is_ValueTypeBase(), "unexpected node");
2009   if (_value_type_nodes != NULL) {
2010     _value_type_nodes->push(n);
2011   }
2012 }
2013 
2014 void Compile::remove_value_type(Node* n) {
2015   assert(n->is_ValueTypeBase(), "unexpected node");
2016   if (_value_type_nodes != NULL) {
2017     _value_type_nodes->remove(n);
2018   }
2019 }
2020 
2021 void Compile::process_value_types(PhaseIterGVN &igvn) {
2022   // Make value types scalar in safepoints
2023   while (_value_type_nodes->size() != 0) {
2024     ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase();
2025     vt->make_scalar_in_safepoints(igvn.C->root(), &igvn);
2026     if (vt->is_ValueTypePtr()) {
2027       igvn.replace_node(vt, vt->get_oop());
2028     }
2029   }
2030   _value_type_nodes = NULL;
2031   igvn.optimize();
2032 }
2033 
2034 // StringOpts and late inlining of string methods
2035 void Compile::inline_string_calls(bool parse_time) {
2036   {
2037     // remove useless nodes to make the usage analysis simpler
2038     ResourceMark rm;
2039     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2040   }
2041 
2042   {
2043     ResourceMark rm;
2044     print_method(PHASE_BEFORE_STRINGOPTS, 3);
2045     PhaseStringOpts pso(initial_gvn(), for_igvn());
2046     print_method(PHASE_AFTER_STRINGOPTS, 3);
2047   }
2048 
2049   // now inline anything that we skipped the first time around
2050   if (!parse_time) {
2051     _late_inlines_pos = _late_inlines.length();
2052   }
2053 
2054   while (_string_late_inlines.length() > 0) {
2055     CallGenerator* cg = _string_late_inlines.pop();
2056     cg->do_late_inline();
2057     if (failing())  return;
2058   }
2059   _string_late_inlines.trunc_to(0);
2060 }
2061 
2062 // Late inlining of boxing methods
2063 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2064   if (_boxing_late_inlines.length() > 0) {
2065     assert(has_boxed_value(), "inconsistent");
2066 
2067     PhaseGVN* gvn = initial_gvn();
2068     set_inlining_incrementally(true);
2069 
2070     assert( igvn._worklist.size() == 0, "should be done with igvn" );
2071     for_igvn()->clear();
2072     gvn->replace_with(&igvn);
2073 
2074     _late_inlines_pos = _late_inlines.length();
2075 
2076     while (_boxing_late_inlines.length() > 0) {
2077       CallGenerator* cg = _boxing_late_inlines.pop();
2078       cg->do_late_inline();
2079       if (failing())  return;
2080     }
2081     _boxing_late_inlines.trunc_to(0);
2082 
2083     {
2084       ResourceMark rm;
2085       PhaseRemoveUseless pru(gvn, for_igvn());
2086     }
2087 
2088     igvn = PhaseIterGVN(gvn);
2089     igvn.optimize();
2090 
2091     set_inlining_progress(false);
2092     set_inlining_incrementally(false);
2093   }
2094 }
2095 
2096 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2097   assert(IncrementalInline, "incremental inlining should be on");
2098   PhaseGVN* gvn = initial_gvn();
2099 
2100   set_inlining_progress(false);
2101   for_igvn()->clear();
2102   gvn->replace_with(&igvn);
2103 
2104   {
2105     TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2106     int i = 0;
2107     for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2108       CallGenerator* cg = _late_inlines.at(i);
2109       _late_inlines_pos = i+1;
2110       cg->do_late_inline();
2111       if (failing())  return;
2112     }
2113     int j = 0;
2114     for (; i < _late_inlines.length(); i++, j++) {
2115       _late_inlines.at_put(j, _late_inlines.at(i));
2116     }
2117     _late_inlines.trunc_to(j);
2118   }
2119 
2120   {
2121     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2122     ResourceMark rm;
2123     PhaseRemoveUseless pru(gvn, for_igvn());
2124   }
2125 
2126   {
2127     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2128     igvn = PhaseIterGVN(gvn);
2129   }
2130 }
2131 
2132 // Perform incremental inlining until bound on number of live nodes is reached
2133 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2134   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2135 
2136   PhaseGVN* gvn = initial_gvn();
2137 
2138   set_inlining_incrementally(true);
2139   set_inlining_progress(true);
2140   uint low_live_nodes = 0;
2141 
2142   while(inlining_progress() && _late_inlines.length() > 0) {
2143 
2144     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2145       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2146         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2147         // PhaseIdealLoop is expensive so we only try it once we are
2148         // out of live nodes and we only try it again if the previous
2149         // helped got the number of nodes down significantly
2150         PhaseIdealLoop ideal_loop( igvn, false, true );
2151         if (failing())  return;
2152         low_live_nodes = live_nodes();
2153         _major_progress = true;
2154       }
2155 
2156       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2157         break;
2158       }
2159     }
2160 
2161     inline_incrementally_one(igvn);
2162 
2163     if (failing())  return;
2164 
2165     {
2166       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2167       igvn.optimize();
2168     }
2169 
2170     if (failing())  return;
2171   }
2172 
2173   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2174 
2175   if (_string_late_inlines.length() > 0) {
2176     assert(has_stringbuilder(), "inconsistent");
2177     for_igvn()->clear();
2178     initial_gvn()->replace_with(&igvn);
2179 
2180     inline_string_calls(false);
2181 
2182     if (failing())  return;
2183 
2184     {
2185       TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2186       ResourceMark rm;
2187       PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2188     }
2189 
2190     {
2191       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2192       igvn = PhaseIterGVN(gvn);
2193       igvn.optimize();
2194     }
2195   }
2196 
2197   set_inlining_incrementally(false);
2198 }
2199 
2200 
2201 //------------------------------Optimize---------------------------------------
2202 // Given a graph, optimize it.
2203 void Compile::Optimize() {
2204   TracePhase tp("optimizer", &timers[_t_optimizer]);
2205 
2206 #ifndef PRODUCT
2207   if (_directive->BreakAtCompileOption) {
2208     BREAKPOINT;
2209   }
2210 
2211 #endif
2212 
2213   ResourceMark rm;
2214   int          loop_opts_cnt;
2215 
2216   print_inlining_reinit();
2217 
2218   NOT_PRODUCT( verify_graph_edges(); )
2219 
2220   print_method(PHASE_AFTER_PARSING);
2221 
2222  {
2223   // Iterative Global Value Numbering, including ideal transforms
2224   // Initialize IterGVN with types and values from parse-time GVN
2225   PhaseIterGVN igvn(initial_gvn());
2226 #ifdef ASSERT
2227   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2228 #endif
2229   {
2230     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2231     igvn.optimize();
2232   }
2233 
2234   print_method(PHASE_ITER_GVN1, 2);
2235 
2236   if (failing())  return;
2237 
2238   inline_incrementally(igvn);
2239 
2240   print_method(PHASE_INCREMENTAL_INLINE, 2);
2241 
2242   if (failing())  return;
2243 
2244   if (eliminate_boxing()) {
2245     // Inline valueOf() methods now.
2246     inline_boxing_calls(igvn);
2247 
2248     if (AlwaysIncrementalInline) {
2249       inline_incrementally(igvn);
2250     }
2251 
2252     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2253 
2254     if (failing())  return;
2255   }
2256 
2257   // Remove the speculative part of types and clean up the graph from
2258   // the extra CastPP nodes whose only purpose is to carry them. Do
2259   // that early so that optimizations are not disrupted by the extra
2260   // CastPP nodes.
2261   remove_speculative_types(igvn);
2262 
2263   // No more new expensive nodes will be added to the list from here
2264   // so keep only the actual candidates for optimizations.
2265   cleanup_expensive_nodes(igvn);
2266 
2267   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2268     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2269     initial_gvn()->replace_with(&igvn);
2270     for_igvn()->clear();
2271     Unique_Node_List new_worklist(C->comp_arena());
2272     {
2273       ResourceMark rm;
2274       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2275     }
2276     set_for_igvn(&new_worklist);
2277     igvn = PhaseIterGVN(initial_gvn());
2278     igvn.optimize();
2279   }
2280 
2281   if (_value_type_nodes->size() > 0) {
2282     // Do this once all inlining is over to avoid getting inconsistent debug info
2283     process_value_types(igvn);
2284   }
2285 
2286   // Perform escape analysis
2287   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2288     if (has_loops()) {
2289       // Cleanup graph (remove dead nodes).
2290       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2291       PhaseIdealLoop ideal_loop( igvn, false, true );
2292       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2293       if (failing())  return;
2294     }
2295     ConnectionGraph::do_analysis(this, &igvn);
2296 
2297     if (failing())  return;
2298 
2299     // Optimize out fields loads from scalar replaceable allocations.
2300     igvn.optimize();
2301     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2302 
2303     if (failing())  return;
2304 
2305     if (congraph() != NULL && macro_count() > 0) {
2306       TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2307       PhaseMacroExpand mexp(igvn);
2308       mexp.eliminate_macro_nodes();
2309       igvn.set_delay_transform(false);
2310 
2311       igvn.optimize();
2312       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2313 
2314       if (failing())  return;
2315     }
2316   }
2317 
2318   // Loop transforms on the ideal graph.  Range Check Elimination,
2319   // peeling, unrolling, etc.
2320 
2321   // Set loop opts counter
2322   loop_opts_cnt = num_loop_opts();
2323   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2324     {
2325       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2326       PhaseIdealLoop ideal_loop( igvn, true );
2327       loop_opts_cnt--;
2328       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2329       if (failing())  return;
2330     }
2331     // Loop opts pass if partial peeling occurred in previous pass
2332     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2333       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2334       PhaseIdealLoop ideal_loop( igvn, false );
2335       loop_opts_cnt--;
2336       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2337       if (failing())  return;
2338     }
2339     // Loop opts pass for loop-unrolling before CCP
2340     if(major_progress() && (loop_opts_cnt > 0)) {
2341       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2342       PhaseIdealLoop ideal_loop( igvn, false );
2343       loop_opts_cnt--;
2344       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2345     }
2346     if (!failing()) {
2347       // Verify that last round of loop opts produced a valid graph
2348       TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2349       PhaseIdealLoop::verify(igvn);
2350     }
2351   }
2352   if (failing())  return;
2353 
2354   // Conditional Constant Propagation;
2355   PhaseCCP ccp( &igvn );
2356   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2357   {
2358     TracePhase tp("ccp", &timers[_t_ccp]);
2359     ccp.do_transform();
2360   }
2361   print_method(PHASE_CPP1, 2);
2362 
2363   assert( true, "Break here to ccp.dump_old2new_map()");
2364 
2365   // Iterative Global Value Numbering, including ideal transforms
2366   {
2367     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2368     igvn = ccp;
2369     igvn.optimize();
2370   }
2371 
2372   print_method(PHASE_ITER_GVN2, 2);
2373 
2374   if (failing())  return;
2375 
2376   // Loop transforms on the ideal graph.  Range Check Elimination,
2377   // peeling, unrolling, etc.
2378   if(loop_opts_cnt > 0) {
2379     debug_only( int cnt = 0; );
2380     while(major_progress() && (loop_opts_cnt > 0)) {
2381       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2382       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2383       PhaseIdealLoop ideal_loop( igvn, true);
2384       loop_opts_cnt--;
2385       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2386       if (failing())  return;
2387     }
2388   }
2389   // Ensure that major progress is now clear
2390   C->clear_major_progress();
2391 
2392   {
2393     // Verify that all previous optimizations produced a valid graph
2394     // at least to this point, even if no loop optimizations were done.
2395     TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2396     PhaseIdealLoop::verify(igvn);
2397   }
2398 
2399   if (range_check_cast_count() > 0) {
2400     // No more loop optimizations. Remove all range check dependent CastIINodes.
2401     C->remove_range_check_casts(igvn);
2402     igvn.optimize();
2403   }
2404 
2405   {
2406     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2407     PhaseMacroExpand  mex(igvn);
2408     if (mex.expand_macro_nodes()) {
2409       assert(failing(), "must bail out w/ explicit message");
2410       return;
2411     }
2412   }
2413 
2414   DEBUG_ONLY( _modified_nodes = NULL; )
2415  } // (End scope of igvn; run destructor if necessary for asserts.)
2416 
2417  process_print_inlining();
2418  // A method with only infinite loops has no edges entering loops from root
2419  {
2420    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2421    if (final_graph_reshaping()) {
2422      assert(failing(), "must bail out w/ explicit message");
2423      return;
2424    }
2425  }
2426 
2427  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2428 }
2429 
2430 //------------------------------Code_Gen---------------------------------------
2431 // Given a graph, generate code for it
2432 void Compile::Code_Gen() {
2433   if (failing()) {
2434     return;
2435   }
2436 
2437   // Perform instruction selection.  You might think we could reclaim Matcher
2438   // memory PDQ, but actually the Matcher is used in generating spill code.
2439   // Internals of the Matcher (including some VectorSets) must remain live
2440   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2441   // set a bit in reclaimed memory.
2442 
2443   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2444   // nodes.  Mapping is only valid at the root of each matched subtree.
2445   NOT_PRODUCT( verify_graph_edges(); )
2446 
2447   Matcher matcher;
2448   _matcher = &matcher;
2449   {
2450     TracePhase tp("matcher", &timers[_t_matcher]);
2451     matcher.match();
2452   }
2453   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2454   // nodes.  Mapping is only valid at the root of each matched subtree.
2455   NOT_PRODUCT( verify_graph_edges(); )
2456 
2457   // If you have too many nodes, or if matching has failed, bail out
2458   check_node_count(0, "out of nodes matching instructions");
2459   if (failing()) {
2460     return;
2461   }
2462 
2463   // Build a proper-looking CFG
2464   PhaseCFG cfg(node_arena(), root(), matcher);
2465   _cfg = &cfg;
2466   {
2467     TracePhase tp("scheduler", &timers[_t_scheduler]);
2468     bool success = cfg.do_global_code_motion();
2469     if (!success) {
2470       return;
2471     }
2472 
2473     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2474     NOT_PRODUCT( verify_graph_edges(); )
2475     debug_only( cfg.verify(); )
2476   }
2477 
2478   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2479   _regalloc = &regalloc;
2480   {
2481     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2482     // Perform register allocation.  After Chaitin, use-def chains are
2483     // no longer accurate (at spill code) and so must be ignored.
2484     // Node->LRG->reg mappings are still accurate.
2485     _regalloc->Register_Allocate();
2486 
2487     // Bail out if the allocator builds too many nodes
2488     if (failing()) {
2489       return;
2490     }
2491   }
2492 
2493   // Prior to register allocation we kept empty basic blocks in case the
2494   // the allocator needed a place to spill.  After register allocation we
2495   // are not adding any new instructions.  If any basic block is empty, we
2496   // can now safely remove it.
2497   {
2498     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2499     cfg.remove_empty_blocks();
2500     if (do_freq_based_layout()) {
2501       PhaseBlockLayout layout(cfg);
2502     } else {
2503       cfg.set_loop_alignment();
2504     }
2505     cfg.fixup_flow();
2506   }
2507 
2508   // Apply peephole optimizations
2509   if( OptoPeephole ) {
2510     TracePhase tp("peephole", &timers[_t_peephole]);
2511     PhasePeephole peep( _regalloc, cfg);
2512     peep.do_transform();
2513   }
2514 
2515   // Do late expand if CPU requires this.
2516   if (Matcher::require_postalloc_expand) {
2517     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2518     cfg.postalloc_expand(_regalloc);
2519   }
2520 
2521   // Convert Nodes to instruction bits in a buffer
2522   {
2523     TraceTime tp("output", &timers[_t_output], CITime);
2524     Output();
2525   }
2526 
2527   print_method(PHASE_FINAL_CODE);
2528 
2529   // He's dead, Jim.
2530   _cfg     = (PhaseCFG*)0xdeadbeef;
2531   _regalloc = (PhaseChaitin*)0xdeadbeef;
2532 }
2533 
2534 
2535 //------------------------------dump_asm---------------------------------------
2536 // Dump formatted assembly
2537 #ifndef PRODUCT
2538 void Compile::dump_asm(int *pcs, uint pc_limit) {
2539   bool cut_short = false;
2540   tty->print_cr("#");
2541   tty->print("#  ");  _tf->dump();  tty->cr();
2542   tty->print_cr("#");
2543 
2544   // For all blocks
2545   int pc = 0x0;                 // Program counter
2546   char starts_bundle = ' ';
2547   _regalloc->dump_frame();
2548 
2549   Node *n = NULL;
2550   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2551     if (VMThread::should_terminate()) {
2552       cut_short = true;
2553       break;
2554     }
2555     Block* block = _cfg->get_block(i);
2556     if (block->is_connector() && !Verbose) {
2557       continue;
2558     }
2559     n = block->head();
2560     if (pcs && n->_idx < pc_limit) {
2561       tty->print("%3.3x   ", pcs[n->_idx]);
2562     } else {
2563       tty->print("      ");
2564     }
2565     block->dump_head(_cfg);
2566     if (block->is_connector()) {
2567       tty->print_cr("        # Empty connector block");
2568     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2569       tty->print_cr("        # Block is sole successor of call");
2570     }
2571 
2572     // For all instructions
2573     Node *delay = NULL;
2574     for (uint j = 0; j < block->number_of_nodes(); j++) {
2575       if (VMThread::should_terminate()) {
2576         cut_short = true;
2577         break;
2578       }
2579       n = block->get_node(j);
2580       if (valid_bundle_info(n)) {
2581         Bundle* bundle = node_bundling(n);
2582         if (bundle->used_in_unconditional_delay()) {
2583           delay = n;
2584           continue;
2585         }
2586         if (bundle->starts_bundle()) {
2587           starts_bundle = '+';
2588         }
2589       }
2590 
2591       if (WizardMode) {
2592         n->dump();
2593       }
2594 
2595       if( !n->is_Region() &&    // Dont print in the Assembly
2596           !n->is_Phi() &&       // a few noisely useless nodes
2597           !n->is_Proj() &&
2598           !n->is_MachTemp() &&
2599           !n->is_SafePointScalarObject() &&
2600           !n->is_Catch() &&     // Would be nice to print exception table targets
2601           !n->is_MergeMem() &&  // Not very interesting
2602           !n->is_top() &&       // Debug info table constants
2603           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2604           ) {
2605         if (pcs && n->_idx < pc_limit)
2606           tty->print("%3.3x", pcs[n->_idx]);
2607         else
2608           tty->print("   ");
2609         tty->print(" %c ", starts_bundle);
2610         starts_bundle = ' ';
2611         tty->print("\t");
2612         n->format(_regalloc, tty);
2613         tty->cr();
2614       }
2615 
2616       // If we have an instruction with a delay slot, and have seen a delay,
2617       // then back up and print it
2618       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2619         assert(delay != NULL, "no unconditional delay instruction");
2620         if (WizardMode) delay->dump();
2621 
2622         if (node_bundling(delay)->starts_bundle())
2623           starts_bundle = '+';
2624         if (pcs && n->_idx < pc_limit)
2625           tty->print("%3.3x", pcs[n->_idx]);
2626         else
2627           tty->print("   ");
2628         tty->print(" %c ", starts_bundle);
2629         starts_bundle = ' ';
2630         tty->print("\t");
2631         delay->format(_regalloc, tty);
2632         tty->cr();
2633         delay = NULL;
2634       }
2635 
2636       // Dump the exception table as well
2637       if( n->is_Catch() && (Verbose || WizardMode) ) {
2638         // Print the exception table for this offset
2639         _handler_table.print_subtable_for(pc);
2640       }
2641     }
2642 
2643     if (pcs && n->_idx < pc_limit)
2644       tty->print_cr("%3.3x", pcs[n->_idx]);
2645     else
2646       tty->cr();
2647 
2648     assert(cut_short || delay == NULL, "no unconditional delay branch");
2649 
2650   } // End of per-block dump
2651   tty->cr();
2652 
2653   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2654 }
2655 #endif
2656 
2657 //------------------------------Final_Reshape_Counts---------------------------
2658 // This class defines counters to help identify when a method
2659 // may/must be executed using hardware with only 24-bit precision.
2660 struct Final_Reshape_Counts : public StackObj {
2661   int  _call_count;             // count non-inlined 'common' calls
2662   int  _float_count;            // count float ops requiring 24-bit precision
2663   int  _double_count;           // count double ops requiring more precision
2664   int  _java_call_count;        // count non-inlined 'java' calls
2665   int  _inner_loop_count;       // count loops which need alignment
2666   VectorSet _visited;           // Visitation flags
2667   Node_List _tests;             // Set of IfNodes & PCTableNodes
2668 
2669   Final_Reshape_Counts() :
2670     _call_count(0), _float_count(0), _double_count(0),
2671     _java_call_count(0), _inner_loop_count(0),
2672     _visited( Thread::current()->resource_area() ) { }
2673 
2674   void inc_call_count  () { _call_count  ++; }
2675   void inc_float_count () { _float_count ++; }
2676   void inc_double_count() { _double_count++; }
2677   void inc_java_call_count() { _java_call_count++; }
2678   void inc_inner_loop_count() { _inner_loop_count++; }
2679 
2680   int  get_call_count  () const { return _call_count  ; }
2681   int  get_float_count () const { return _float_count ; }
2682   int  get_double_count() const { return _double_count; }
2683   int  get_java_call_count() const { return _java_call_count; }
2684   int  get_inner_loop_count() const { return _inner_loop_count; }
2685 };
2686 
2687 #ifdef ASSERT
2688 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2689   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2690   // Make sure the offset goes inside the instance layout.
2691   return k->contains_field_offset(tp->offset());
2692   // Note that OffsetBot and OffsetTop are very negative.
2693 }
2694 #endif
2695 
2696 // Eliminate trivially redundant StoreCMs and accumulate their
2697 // precedence edges.
2698 void Compile::eliminate_redundant_card_marks(Node* n) {
2699   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2700   if (n->in(MemNode::Address)->outcnt() > 1) {
2701     // There are multiple users of the same address so it might be
2702     // possible to eliminate some of the StoreCMs
2703     Node* mem = n->in(MemNode::Memory);
2704     Node* adr = n->in(MemNode::Address);
2705     Node* val = n->in(MemNode::ValueIn);
2706     Node* prev = n;
2707     bool done = false;
2708     // Walk the chain of StoreCMs eliminating ones that match.  As
2709     // long as it's a chain of single users then the optimization is
2710     // safe.  Eliminating partially redundant StoreCMs would require
2711     // cloning copies down the other paths.
2712     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2713       if (adr == mem->in(MemNode::Address) &&
2714           val == mem->in(MemNode::ValueIn)) {
2715         // redundant StoreCM
2716         if (mem->req() > MemNode::OopStore) {
2717           // Hasn't been processed by this code yet.
2718           n->add_prec(mem->in(MemNode::OopStore));
2719         } else {
2720           // Already converted to precedence edge
2721           for (uint i = mem->req(); i < mem->len(); i++) {
2722             // Accumulate any precedence edges
2723             if (mem->in(i) != NULL) {
2724               n->add_prec(mem->in(i));
2725             }
2726           }
2727           // Everything above this point has been processed.
2728           done = true;
2729         }
2730         // Eliminate the previous StoreCM
2731         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2732         assert(mem->outcnt() == 0, "should be dead");
2733         mem->disconnect_inputs(NULL, this);
2734       } else {
2735         prev = mem;
2736       }
2737       mem = prev->in(MemNode::Memory);
2738     }
2739   }
2740 }
2741 
2742 
2743 //------------------------------final_graph_reshaping_impl----------------------
2744 // Implement items 1-5 from final_graph_reshaping below.
2745 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2746 
2747   if ( n->outcnt() == 0 ) return; // dead node
2748   uint nop = n->Opcode();
2749 
2750   // Check for 2-input instruction with "last use" on right input.
2751   // Swap to left input.  Implements item (2).
2752   if( n->req() == 3 &&          // two-input instruction
2753       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2754       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2755       n->in(2)->outcnt() == 1 &&// right use IS a last use
2756       !n->in(2)->is_Con() ) {   // right use is not a constant
2757     // Check for commutative opcode
2758     switch( nop ) {
2759     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2760     case Op_MaxI:  case Op_MinI:
2761     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2762     case Op_AndL:  case Op_XorL:  case Op_OrL:
2763     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2764       // Move "last use" input to left by swapping inputs
2765       n->swap_edges(1, 2);
2766       break;
2767     }
2768     default:
2769       break;
2770     }
2771   }
2772 
2773 #ifdef ASSERT
2774   if( n->is_Mem() ) {
2775     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2776     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2777             // oop will be recorded in oop map if load crosses safepoint
2778             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2779                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2780             "raw memory operations should have control edge");
2781   }
2782 #endif
2783   // Count FPU ops and common calls, implements item (3)
2784   switch( nop ) {
2785   // Count all float operations that may use FPU
2786   case Op_AddF:
2787   case Op_SubF:
2788   case Op_MulF:
2789   case Op_DivF:
2790   case Op_NegF:
2791   case Op_ModF:
2792   case Op_ConvI2F:
2793   case Op_ConF:
2794   case Op_CmpF:
2795   case Op_CmpF3:
2796   // case Op_ConvL2F: // longs are split into 32-bit halves
2797     frc.inc_float_count();
2798     break;
2799 
2800   case Op_ConvF2D:
2801   case Op_ConvD2F:
2802     frc.inc_float_count();
2803     frc.inc_double_count();
2804     break;
2805 
2806   // Count all double operations that may use FPU
2807   case Op_AddD:
2808   case Op_SubD:
2809   case Op_MulD:
2810   case Op_DivD:
2811   case Op_NegD:
2812   case Op_ModD:
2813   case Op_ConvI2D:
2814   case Op_ConvD2I:
2815   // case Op_ConvL2D: // handled by leaf call
2816   // case Op_ConvD2L: // handled by leaf call
2817   case Op_ConD:
2818   case Op_CmpD:
2819   case Op_CmpD3:
2820     frc.inc_double_count();
2821     break;
2822   case Op_Opaque1:              // Remove Opaque Nodes before matching
2823   case Op_Opaque2:              // Remove Opaque Nodes before matching
2824   case Op_Opaque3:
2825     n->subsume_by(n->in(1), this);
2826     break;
2827   case Op_CallStaticJava:
2828   case Op_CallJava:
2829   case Op_CallDynamicJava:
2830     frc.inc_java_call_count(); // Count java call site;
2831   case Op_CallRuntime:
2832   case Op_CallLeaf:
2833   case Op_CallLeafNoFP: {
2834     assert( n->is_Call(), "" );
2835     CallNode *call = n->as_Call();
2836     // Count call sites where the FP mode bit would have to be flipped.
2837     // Do not count uncommon runtime calls:
2838     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2839     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2840     if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2841       frc.inc_call_count();   // Count the call site
2842     } else {                  // See if uncommon argument is shared
2843       Node *n = call->in(TypeFunc::Parms);
2844       int nop = n->Opcode();
2845       // Clone shared simple arguments to uncommon calls, item (1).
2846       if( n->outcnt() > 1 &&
2847           !n->is_Proj() &&
2848           nop != Op_CreateEx &&
2849           nop != Op_CheckCastPP &&
2850           nop != Op_DecodeN &&
2851           nop != Op_DecodeNKlass &&
2852           !n->is_Mem() ) {
2853         Node *x = n->clone();
2854         call->set_req( TypeFunc::Parms, x );
2855       }
2856     }
2857     break;
2858   }
2859 
2860   case Op_StoreD:
2861   case Op_LoadD:
2862   case Op_LoadD_unaligned:
2863     frc.inc_double_count();
2864     goto handle_mem;
2865   case Op_StoreF:
2866   case Op_LoadF:
2867     frc.inc_float_count();
2868     goto handle_mem;
2869 
2870   case Op_StoreCM:
2871     {
2872       // Convert OopStore dependence into precedence edge
2873       Node* prec = n->in(MemNode::OopStore);
2874       n->del_req(MemNode::OopStore);
2875       n->add_prec(prec);
2876       eliminate_redundant_card_marks(n);
2877     }
2878 
2879     // fall through
2880 
2881   case Op_StoreB:
2882   case Op_StoreC:
2883   case Op_StorePConditional:
2884   case Op_StoreI:
2885   case Op_StoreL:
2886   case Op_StoreIConditional:
2887   case Op_StoreLConditional:
2888   case Op_CompareAndSwapB:
2889   case Op_CompareAndSwapS:
2890   case Op_CompareAndSwapI:
2891   case Op_CompareAndSwapL:
2892   case Op_CompareAndSwapP:
2893   case Op_CompareAndSwapN:
2894   case Op_WeakCompareAndSwapB:
2895   case Op_WeakCompareAndSwapS:
2896   case Op_WeakCompareAndSwapI:
2897   case Op_WeakCompareAndSwapL:
2898   case Op_WeakCompareAndSwapP:
2899   case Op_WeakCompareAndSwapN:
2900   case Op_CompareAndExchangeB:
2901   case Op_CompareAndExchangeS:
2902   case Op_CompareAndExchangeI:
2903   case Op_CompareAndExchangeL:
2904   case Op_CompareAndExchangeP:
2905   case Op_CompareAndExchangeN:
2906   case Op_GetAndAddS:
2907   case Op_GetAndAddB:
2908   case Op_GetAndAddI:
2909   case Op_GetAndAddL:
2910   case Op_GetAndSetS:
2911   case Op_GetAndSetB:
2912   case Op_GetAndSetI:
2913   case Op_GetAndSetL:
2914   case Op_GetAndSetP:
2915   case Op_GetAndSetN:
2916   case Op_StoreP:
2917   case Op_StoreN:
2918   case Op_StoreNKlass:
2919   case Op_LoadB:
2920   case Op_LoadUB:
2921   case Op_LoadUS:
2922   case Op_LoadI:
2923   case Op_LoadKlass:
2924   case Op_LoadNKlass:
2925   case Op_LoadL:
2926   case Op_LoadL_unaligned:
2927   case Op_LoadPLocked:
2928   case Op_LoadP:
2929   case Op_LoadN:
2930   case Op_LoadRange:
2931   case Op_LoadS: {
2932   handle_mem:
2933 #ifdef ASSERT
2934     if( VerifyOptoOopOffsets ) {
2935       assert( n->is_Mem(), "" );
2936       MemNode *mem  = (MemNode*)n;
2937       // Check to see if address types have grounded out somehow.
2938       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2939       assert( !tp || oop_offset_is_sane(tp), "" );
2940     }
2941 #endif
2942     break;
2943   }
2944 
2945   case Op_AddP: {               // Assert sane base pointers
2946     Node *addp = n->in(AddPNode::Address);
2947     assert( !addp->is_AddP() ||
2948             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2949             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2950             "Base pointers must match (addp %u)", addp->_idx );
2951 #ifdef _LP64
2952     if ((UseCompressedOops || UseCompressedClassPointers) &&
2953         addp->Opcode() == Op_ConP &&
2954         addp == n->in(AddPNode::Base) &&
2955         n->in(AddPNode::Offset)->is_Con()) {
2956       // If the transformation of ConP to ConN+DecodeN is beneficial depends
2957       // on the platform and on the compressed oops mode.
2958       // Use addressing with narrow klass to load with offset on x86.
2959       // Some platforms can use the constant pool to load ConP.
2960       // Do this transformation here since IGVN will convert ConN back to ConP.
2961       const Type* t = addp->bottom_type();
2962       bool is_oop   = t->isa_oopptr() != NULL;
2963       bool is_klass = t->isa_klassptr() != NULL;
2964 
2965       if ((is_oop   && Matcher::const_oop_prefer_decode()  ) ||
2966           (is_klass && Matcher::const_klass_prefer_decode())) {
2967         Node* nn = NULL;
2968 
2969         int op = is_oop ? Op_ConN : Op_ConNKlass;
2970 
2971         // Look for existing ConN node of the same exact type.
2972         Node* r  = root();
2973         uint cnt = r->outcnt();
2974         for (uint i = 0; i < cnt; i++) {
2975           Node* m = r->raw_out(i);
2976           if (m!= NULL && m->Opcode() == op &&
2977               m->bottom_type()->make_ptr() == t) {
2978             nn = m;
2979             break;
2980           }
2981         }
2982         if (nn != NULL) {
2983           // Decode a narrow oop to match address
2984           // [R12 + narrow_oop_reg<<3 + offset]
2985           if (is_oop) {
2986             nn = new DecodeNNode(nn, t);
2987           } else {
2988             nn = new DecodeNKlassNode(nn, t);
2989           }
2990           // Check for succeeding AddP which uses the same Base.
2991           // Otherwise we will run into the assertion above when visiting that guy.
2992           for (uint i = 0; i < n->outcnt(); ++i) {
2993             Node *out_i = n->raw_out(i);
2994             if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
2995               out_i->set_req(AddPNode::Base, nn);
2996 #ifdef ASSERT
2997               for (uint j = 0; j < out_i->outcnt(); ++j) {
2998                 Node *out_j = out_i->raw_out(j);
2999                 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3000                        "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3001               }
3002 #endif
3003             }
3004           }
3005           n->set_req(AddPNode::Base, nn);
3006           n->set_req(AddPNode::Address, nn);
3007           if (addp->outcnt() == 0) {
3008             addp->disconnect_inputs(NULL, this);
3009           }
3010         }
3011       }
3012     }
3013 #endif
3014     // platform dependent reshaping of the address expression
3015     reshape_address(n->as_AddP());
3016     break;
3017   }
3018 
3019   case Op_CastPP: {
3020     // Remove CastPP nodes to gain more freedom during scheduling but
3021     // keep the dependency they encode as control or precedence edges
3022     // (if control is set already) on memory operations. Some CastPP
3023     // nodes don't have a control (don't carry a dependency): skip
3024     // those.
3025     if (n->in(0) != NULL) {
3026       ResourceMark rm;
3027       Unique_Node_List wq;
3028       wq.push(n);
3029       for (uint next = 0; next < wq.size(); ++next) {
3030         Node *m = wq.at(next);
3031         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3032           Node* use = m->fast_out(i);
3033           if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3034             use->ensure_control_or_add_prec(n->in(0));
3035           } else {
3036             switch(use->Opcode()) {
3037             case Op_AddP:
3038             case Op_DecodeN:
3039             case Op_DecodeNKlass:
3040             case Op_CheckCastPP:
3041             case Op_CastPP:
3042               wq.push(use);
3043               break;
3044             }
3045           }
3046         }
3047       }
3048     }
3049     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3050     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3051       Node* in1 = n->in(1);
3052       const Type* t = n->bottom_type();
3053       Node* new_in1 = in1->clone();
3054       new_in1->as_DecodeN()->set_type(t);
3055 
3056       if (!Matcher::narrow_oop_use_complex_address()) {
3057         //
3058         // x86, ARM and friends can handle 2 adds in addressing mode
3059         // and Matcher can fold a DecodeN node into address by using
3060         // a narrow oop directly and do implicit NULL check in address:
3061         //
3062         // [R12 + narrow_oop_reg<<3 + offset]
3063         // NullCheck narrow_oop_reg
3064         //
3065         // On other platforms (Sparc) we have to keep new DecodeN node and
3066         // use it to do implicit NULL check in address:
3067         //
3068         // decode_not_null narrow_oop_reg, base_reg
3069         // [base_reg + offset]
3070         // NullCheck base_reg
3071         //
3072         // Pin the new DecodeN node to non-null path on these platform (Sparc)
3073         // to keep the information to which NULL check the new DecodeN node
3074         // corresponds to use it as value in implicit_null_check().
3075         //
3076         new_in1->set_req(0, n->in(0));
3077       }
3078 
3079       n->subsume_by(new_in1, this);
3080       if (in1->outcnt() == 0) {
3081         in1->disconnect_inputs(NULL, this);
3082       }
3083     } else {
3084       n->subsume_by(n->in(1), this);
3085       if (n->outcnt() == 0) {
3086         n->disconnect_inputs(NULL, this);
3087       }
3088     }
3089     break;
3090   }
3091 #ifdef _LP64
3092   case Op_CmpP:
3093     // Do this transformation here to preserve CmpPNode::sub() and
3094     // other TypePtr related Ideal optimizations (for example, ptr nullness).
3095     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3096       Node* in1 = n->in(1);
3097       Node* in2 = n->in(2);
3098       if (!in1->is_DecodeNarrowPtr()) {
3099         in2 = in1;
3100         in1 = n->in(2);
3101       }
3102       assert(in1->is_DecodeNarrowPtr(), "sanity");
3103 
3104       Node* new_in2 = NULL;
3105       if (in2->is_DecodeNarrowPtr()) {
3106         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3107         new_in2 = in2->in(1);
3108       } else if (in2->Opcode() == Op_ConP) {
3109         const Type* t = in2->bottom_type();
3110         if (t == TypePtr::NULL_PTR) {
3111           assert(in1->is_DecodeN(), "compare klass to null?");
3112           // Don't convert CmpP null check into CmpN if compressed
3113           // oops implicit null check is not generated.
3114           // This will allow to generate normal oop implicit null check.
3115           if (Matcher::gen_narrow_oop_implicit_null_checks())
3116             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3117           //
3118           // This transformation together with CastPP transformation above
3119           // will generated code for implicit NULL checks for compressed oops.
3120           //
3121           // The original code after Optimize()
3122           //
3123           //    LoadN memory, narrow_oop_reg
3124           //    decode narrow_oop_reg, base_reg
3125           //    CmpP base_reg, NULL
3126           //    CastPP base_reg // NotNull
3127           //    Load [base_reg + offset], val_reg
3128           //
3129           // after these transformations will be
3130           //
3131           //    LoadN memory, narrow_oop_reg
3132           //    CmpN narrow_oop_reg, NULL
3133           //    decode_not_null narrow_oop_reg, base_reg
3134           //    Load [base_reg + offset], val_reg
3135           //
3136           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3137           // since narrow oops can be used in debug info now (see the code in
3138           // final_graph_reshaping_walk()).
3139           //
3140           // At the end the code will be matched to
3141           // on x86:
3142           //
3143           //    Load_narrow_oop memory, narrow_oop_reg
3144           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3145           //    NullCheck narrow_oop_reg
3146           //
3147           // and on sparc:
3148           //
3149           //    Load_narrow_oop memory, narrow_oop_reg
3150           //    decode_not_null narrow_oop_reg, base_reg
3151           //    Load [base_reg + offset], val_reg
3152           //    NullCheck base_reg
3153           //
3154         } else if (t->isa_oopptr()) {
3155           new_in2 = ConNode::make(t->make_narrowoop());
3156         } else if (t->isa_klassptr()) {
3157           new_in2 = ConNode::make(t->make_narrowklass());
3158         }
3159       }
3160       if (new_in2 != NULL) {
3161         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3162         n->subsume_by(cmpN, this);
3163         if (in1->outcnt() == 0) {
3164           in1->disconnect_inputs(NULL, this);
3165         }
3166         if (in2->outcnt() == 0) {
3167           in2->disconnect_inputs(NULL, this);
3168         }
3169       }
3170     }
3171     break;
3172 
3173   case Op_DecodeN:
3174   case Op_DecodeNKlass:
3175     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3176     // DecodeN could be pinned when it can't be fold into
3177     // an address expression, see the code for Op_CastPP above.
3178     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3179     break;
3180 
3181   case Op_EncodeP:
3182   case Op_EncodePKlass: {
3183     Node* in1 = n->in(1);
3184     if (in1->is_DecodeNarrowPtr()) {
3185       n->subsume_by(in1->in(1), this);
3186     } else if (in1->Opcode() == Op_ConP) {
3187       const Type* t = in1->bottom_type();
3188       if (t == TypePtr::NULL_PTR) {
3189         assert(t->isa_oopptr(), "null klass?");
3190         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3191       } else if (t->isa_oopptr()) {
3192         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3193       } else if (t->isa_klassptr()) {
3194         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3195       }
3196     }
3197     if (in1->outcnt() == 0) {
3198       in1->disconnect_inputs(NULL, this);
3199     }
3200     break;
3201   }
3202 
3203   case Op_Proj: {
3204     if (OptimizeStringConcat) {
3205       ProjNode* p = n->as_Proj();
3206       if (p->_is_io_use) {
3207         // Separate projections were used for the exception path which
3208         // are normally removed by a late inline.  If it wasn't inlined
3209         // then they will hang around and should just be replaced with
3210         // the original one.
3211         Node* proj = NULL;
3212         // Replace with just one
3213         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3214           Node *use = i.get();
3215           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3216             proj = use;
3217             break;
3218           }
3219         }
3220         assert(proj != NULL, "must be found");
3221         p->subsume_by(proj, this);
3222       }
3223     }
3224     break;
3225   }
3226 
3227   case Op_Phi:
3228     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3229       // The EncodeP optimization may create Phi with the same edges
3230       // for all paths. It is not handled well by Register Allocator.
3231       Node* unique_in = n->in(1);
3232       assert(unique_in != NULL, "");
3233       uint cnt = n->req();
3234       for (uint i = 2; i < cnt; i++) {
3235         Node* m = n->in(i);
3236         assert(m != NULL, "");
3237         if (unique_in != m)
3238           unique_in = NULL;
3239       }
3240       if (unique_in != NULL) {
3241         n->subsume_by(unique_in, this);
3242       }
3243     }
3244     break;
3245 
3246 #endif
3247 
3248 #ifdef ASSERT
3249   case Op_CastII:
3250     // Verify that all range check dependent CastII nodes were removed.
3251     if (n->isa_CastII()->has_range_check()) {
3252       n->dump(3);
3253       assert(false, "Range check dependent CastII node was not removed");
3254     }
3255     break;
3256 #endif
3257 
3258   case Op_ModI:
3259     if (UseDivMod) {
3260       // Check if a%b and a/b both exist
3261       Node* d = n->find_similar(Op_DivI);
3262       if (d) {
3263         // Replace them with a fused divmod if supported
3264         if (Matcher::has_match_rule(Op_DivModI)) {
3265           DivModINode* divmod = DivModINode::make(n);
3266           d->subsume_by(divmod->div_proj(), this);
3267           n->subsume_by(divmod->mod_proj(), this);
3268         } else {
3269           // replace a%b with a-((a/b)*b)
3270           Node* mult = new MulINode(d, d->in(2));
3271           Node* sub  = new SubINode(d->in(1), mult);
3272           n->subsume_by(sub, this);
3273         }
3274       }
3275     }
3276     break;
3277 
3278   case Op_ModL:
3279     if (UseDivMod) {
3280       // Check if a%b and a/b both exist
3281       Node* d = n->find_similar(Op_DivL);
3282       if (d) {
3283         // Replace them with a fused divmod if supported
3284         if (Matcher::has_match_rule(Op_DivModL)) {
3285           DivModLNode* divmod = DivModLNode::make(n);
3286           d->subsume_by(divmod->div_proj(), this);
3287           n->subsume_by(divmod->mod_proj(), this);
3288         } else {
3289           // replace a%b with a-((a/b)*b)
3290           Node* mult = new MulLNode(d, d->in(2));
3291           Node* sub  = new SubLNode(d->in(1), mult);
3292           n->subsume_by(sub, this);
3293         }
3294       }
3295     }
3296     break;
3297 
3298   case Op_LoadVector:
3299   case Op_StoreVector:
3300     break;
3301 
3302   case Op_AddReductionVI:
3303   case Op_AddReductionVL:
3304   case Op_AddReductionVF:
3305   case Op_AddReductionVD:
3306   case Op_MulReductionVI:
3307   case Op_MulReductionVL:
3308   case Op_MulReductionVF:
3309   case Op_MulReductionVD:
3310     break;
3311 
3312   case Op_PackB:
3313   case Op_PackS:
3314   case Op_PackI:
3315   case Op_PackF:
3316   case Op_PackL:
3317   case Op_PackD:
3318     if (n->req()-1 > 2) {
3319       // Replace many operand PackNodes with a binary tree for matching
3320       PackNode* p = (PackNode*) n;
3321       Node* btp = p->binary_tree_pack(1, n->req());
3322       n->subsume_by(btp, this);
3323     }
3324     break;
3325   case Op_Loop:
3326   case Op_CountedLoop:
3327     if (n->as_Loop()->is_inner_loop()) {
3328       frc.inc_inner_loop_count();
3329     }
3330     break;
3331   case Op_LShiftI:
3332   case Op_RShiftI:
3333   case Op_URShiftI:
3334   case Op_LShiftL:
3335   case Op_RShiftL:
3336   case Op_URShiftL:
3337     if (Matcher::need_masked_shift_count) {
3338       // The cpu's shift instructions don't restrict the count to the
3339       // lower 5/6 bits. We need to do the masking ourselves.
3340       Node* in2 = n->in(2);
3341       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3342       const TypeInt* t = in2->find_int_type();
3343       if (t != NULL && t->is_con()) {
3344         juint shift = t->get_con();
3345         if (shift > mask) { // Unsigned cmp
3346           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3347         }
3348       } else {
3349         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3350           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3351           n->set_req(2, shift);
3352         }
3353       }
3354       if (in2->outcnt() == 0) { // Remove dead node
3355         in2->disconnect_inputs(NULL, this);
3356       }
3357     }
3358     break;
3359   case Op_MemBarStoreStore:
3360   case Op_MemBarRelease:
3361     // Break the link with AllocateNode: it is no longer useful and
3362     // confuses register allocation.
3363     if (n->req() > MemBarNode::Precedent) {
3364       n->set_req(MemBarNode::Precedent, top());
3365     }
3366     break;
3367   case Op_RangeCheck: {
3368     RangeCheckNode* rc = n->as_RangeCheck();
3369     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3370     n->subsume_by(iff, this);
3371     frc._tests.push(iff);
3372     break;
3373   }
3374   case Op_ConvI2L: {
3375     if (!Matcher::convi2l_type_required) {
3376       // Code generation on some platforms doesn't need accurate
3377       // ConvI2L types. Widening the type can help remove redundant
3378       // address computations.
3379       n->as_Type()->set_type(TypeLong::INT);
3380       ResourceMark rm;
3381       Node_List wq;
3382       wq.push(n);
3383       for (uint next = 0; next < wq.size(); next++) {
3384         Node *m = wq.at(next);
3385 
3386         for(;;) {
3387           // Loop over all nodes with identical inputs edges as m
3388           Node* k = m->find_similar(m->Opcode());
3389           if (k == NULL) {
3390             break;
3391           }
3392           // Push their uses so we get a chance to remove node made
3393           // redundant
3394           for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3395             Node* u = k->fast_out(i);
3396             assert(!wq.contains(u), "shouldn't process one node several times");
3397             if (u->Opcode() == Op_LShiftL ||
3398                 u->Opcode() == Op_AddL ||
3399                 u->Opcode() == Op_SubL ||
3400                 u->Opcode() == Op_AddP) {
3401               wq.push(u);
3402             }
3403           }
3404           // Replace all nodes with identical edges as m with m
3405           k->subsume_by(m, this);
3406         }
3407       }
3408     }
3409     break;
3410   }
3411 #ifdef ASSERT
3412   case Op_ValueTypePtr:
3413   case Op_ValueType: {
3414     n->dump(-1);
3415     assert(false, "value type node was not removed");
3416     break;
3417   }
3418 #endif
3419   default:
3420     assert( !n->is_Call(), "" );
3421     assert( !n->is_Mem(), "" );
3422     assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3423     break;
3424   }
3425 
3426   // Collect CFG split points
3427   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3428     frc._tests.push(n);
3429   }
3430 }
3431 
3432 //------------------------------final_graph_reshaping_walk---------------------
3433 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3434 // requires that the walk visits a node's inputs before visiting the node.
3435 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3436   ResourceArea *area = Thread::current()->resource_area();
3437   Unique_Node_List sfpt(area);
3438 
3439   frc._visited.set(root->_idx); // first, mark node as visited
3440   uint cnt = root->req();
3441   Node *n = root;
3442   uint  i = 0;
3443   while (true) {
3444     if (i < cnt) {
3445       // Place all non-visited non-null inputs onto stack
3446       Node* m = n->in(i);
3447       ++i;
3448       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3449         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3450           // compute worst case interpreter size in case of a deoptimization
3451           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3452 
3453           sfpt.push(m);
3454         }
3455         cnt = m->req();
3456         nstack.push(n, i); // put on stack parent and next input's index
3457         n = m;
3458         i = 0;
3459       }
3460     } else {
3461       // Now do post-visit work
3462       final_graph_reshaping_impl( n, frc );
3463       if (nstack.is_empty())
3464         break;             // finished
3465       n = nstack.node();   // Get node from stack
3466       cnt = n->req();
3467       i = nstack.index();
3468       nstack.pop();        // Shift to the next node on stack
3469     }
3470   }
3471 
3472   // Skip next transformation if compressed oops are not used.
3473   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3474       (!UseCompressedOops && !UseCompressedClassPointers))
3475     return;
3476 
3477   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3478   // It could be done for an uncommon traps or any safepoints/calls
3479   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3480   while (sfpt.size() > 0) {
3481     n = sfpt.pop();
3482     JVMState *jvms = n->as_SafePoint()->jvms();
3483     assert(jvms != NULL, "sanity");
3484     int start = jvms->debug_start();
3485     int end   = n->req();
3486     bool is_uncommon = (n->is_CallStaticJava() &&
3487                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3488     for (int j = start; j < end; j++) {
3489       Node* in = n->in(j);
3490       if (in->is_DecodeNarrowPtr()) {
3491         bool safe_to_skip = true;
3492         if (!is_uncommon ) {
3493           // Is it safe to skip?
3494           for (uint i = 0; i < in->outcnt(); i++) {
3495             Node* u = in->raw_out(i);
3496             if (!u->is_SafePoint() ||
3497                 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3498               safe_to_skip = false;
3499             }
3500           }
3501         }
3502         if (safe_to_skip) {
3503           n->set_req(j, in->in(1));
3504         }
3505         if (in->outcnt() == 0) {
3506           in->disconnect_inputs(NULL, this);
3507         }
3508       }
3509     }
3510   }
3511 }
3512 
3513 //------------------------------final_graph_reshaping--------------------------
3514 // Final Graph Reshaping.
3515 //
3516 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3517 //     and not commoned up and forced early.  Must come after regular
3518 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3519 //     inputs to Loop Phis; these will be split by the allocator anyways.
3520 //     Remove Opaque nodes.
3521 // (2) Move last-uses by commutative operations to the left input to encourage
3522 //     Intel update-in-place two-address operations and better register usage
3523 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3524 //     calls canonicalizing them back.
3525 // (3) Count the number of double-precision FP ops, single-precision FP ops
3526 //     and call sites.  On Intel, we can get correct rounding either by
3527 //     forcing singles to memory (requires extra stores and loads after each
3528 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3529 //     clearing the mode bit around call sites).  The mode bit is only used
3530 //     if the relative frequency of single FP ops to calls is low enough.
3531 //     This is a key transform for SPEC mpeg_audio.
3532 // (4) Detect infinite loops; blobs of code reachable from above but not
3533 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3534 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3535 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3536 //     Detection is by looking for IfNodes where only 1 projection is
3537 //     reachable from below or CatchNodes missing some targets.
3538 // (5) Assert for insane oop offsets in debug mode.
3539 
3540 bool Compile::final_graph_reshaping() {
3541   // an infinite loop may have been eliminated by the optimizer,
3542   // in which case the graph will be empty.
3543   if (root()->req() == 1) {
3544     record_method_not_compilable("trivial infinite loop");
3545     return true;
3546   }
3547 
3548   // Expensive nodes have their control input set to prevent the GVN
3549   // from freely commoning them. There's no GVN beyond this point so
3550   // no need to keep the control input. We want the expensive nodes to
3551   // be freely moved to the least frequent code path by gcm.
3552   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3553   for (int i = 0; i < expensive_count(); i++) {
3554     _expensive_nodes->at(i)->set_req(0, NULL);
3555   }
3556 
3557   Final_Reshape_Counts frc;
3558 
3559   // Visit everybody reachable!
3560   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3561   Node_Stack nstack(live_nodes() >> 1);
3562   final_graph_reshaping_walk(nstack, root(), frc);
3563 
3564   // Check for unreachable (from below) code (i.e., infinite loops).
3565   for( uint i = 0; i < frc._tests.size(); i++ ) {
3566     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3567     // Get number of CFG targets.
3568     // Note that PCTables include exception targets after calls.
3569     uint required_outcnt = n->required_outcnt();
3570     if (n->outcnt() != required_outcnt) {
3571       // Check for a few special cases.  Rethrow Nodes never take the
3572       // 'fall-thru' path, so expected kids is 1 less.
3573       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3574         if (n->in(0)->in(0)->is_Call()) {
3575           CallNode *call = n->in(0)->in(0)->as_Call();
3576           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3577             required_outcnt--;      // Rethrow always has 1 less kid
3578           } else if (call->req() > TypeFunc::Parms &&
3579                      call->is_CallDynamicJava()) {
3580             // Check for null receiver. In such case, the optimizer has
3581             // detected that the virtual call will always result in a null
3582             // pointer exception. The fall-through projection of this CatchNode
3583             // will not be populated.
3584             Node *arg0 = call->in(TypeFunc::Parms);
3585             if (arg0->is_Type() &&
3586                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3587               required_outcnt--;
3588             }
3589           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3590                      call->req() > TypeFunc::Parms+1 &&
3591                      call->is_CallStaticJava()) {
3592             // Check for negative array length. In such case, the optimizer has
3593             // detected that the allocation attempt will always result in an
3594             // exception. There is no fall-through projection of this CatchNode .
3595             Node *arg1 = call->in(TypeFunc::Parms+1);
3596             if (arg1->is_Type() &&
3597                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3598               required_outcnt--;
3599             }
3600           }
3601         }
3602       }
3603       // Recheck with a better notion of 'required_outcnt'
3604       if (n->outcnt() != required_outcnt) {
3605         assert(false, "malformed control flow");
3606         record_method_not_compilable("malformed control flow");
3607         return true;            // Not all targets reachable!
3608       }
3609     }
3610     // Check that I actually visited all kids.  Unreached kids
3611     // must be infinite loops.
3612     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3613       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3614         record_method_not_compilable("infinite loop");
3615         assert(false, "infinite loop");
3616         return true;            // Found unvisited kid; must be unreach
3617       }
3618   }
3619 
3620   // If original bytecodes contained a mixture of floats and doubles
3621   // check if the optimizer has made it homogenous, item (3).
3622   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3623       frc.get_float_count() > 32 &&
3624       frc.get_double_count() == 0 &&
3625       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3626     set_24_bit_selection_and_mode( false,  true );
3627   }
3628 
3629   set_java_calls(frc.get_java_call_count());
3630   set_inner_loops(frc.get_inner_loop_count());
3631 
3632   // No infinite loops, no reason to bail out.
3633   return false;
3634 }
3635 
3636 //-----------------------------too_many_traps----------------------------------
3637 // Report if there are too many traps at the current method and bci.
3638 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3639 bool Compile::too_many_traps(ciMethod* method,
3640                              int bci,
3641                              Deoptimization::DeoptReason reason) {
3642   ciMethodData* md = method->method_data();
3643   if (md->is_empty()) {
3644     // Assume the trap has not occurred, or that it occurred only
3645     // because of a transient condition during start-up in the interpreter.
3646     return false;
3647   }
3648   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3649   if (md->has_trap_at(bci, m, reason) != 0) {
3650     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3651     // Also, if there are multiple reasons, or if there is no per-BCI record,
3652     // assume the worst.
3653     if (log())
3654       log()->elem("observe trap='%s' count='%d'",
3655                   Deoptimization::trap_reason_name(reason),
3656                   md->trap_count(reason));
3657     return true;
3658   } else {
3659     // Ignore method/bci and see if there have been too many globally.
3660     return too_many_traps(reason, md);
3661   }
3662 }
3663 
3664 // Less-accurate variant which does not require a method and bci.
3665 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3666                              ciMethodData* logmd) {
3667   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3668     // Too many traps globally.
3669     // Note that we use cumulative trap_count, not just md->trap_count.
3670     if (log()) {
3671       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3672       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3673                   Deoptimization::trap_reason_name(reason),
3674                   mcount, trap_count(reason));
3675     }
3676     return true;
3677   } else {
3678     // The coast is clear.
3679     return false;
3680   }
3681 }
3682 
3683 //--------------------------too_many_recompiles--------------------------------
3684 // Report if there are too many recompiles at the current method and bci.
3685 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3686 // Is not eager to return true, since this will cause the compiler to use
3687 // Action_none for a trap point, to avoid too many recompilations.
3688 bool Compile::too_many_recompiles(ciMethod* method,
3689                                   int bci,
3690                                   Deoptimization::DeoptReason reason) {
3691   ciMethodData* md = method->method_data();
3692   if (md->is_empty()) {
3693     // Assume the trap has not occurred, or that it occurred only
3694     // because of a transient condition during start-up in the interpreter.
3695     return false;
3696   }
3697   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3698   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3699   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3700   Deoptimization::DeoptReason per_bc_reason
3701     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3702   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3703   if ((per_bc_reason == Deoptimization::Reason_none
3704        || md->has_trap_at(bci, m, reason) != 0)
3705       // The trap frequency measure we care about is the recompile count:
3706       && md->trap_recompiled_at(bci, m)
3707       && md->overflow_recompile_count() >= bc_cutoff) {
3708     // Do not emit a trap here if it has already caused recompilations.
3709     // Also, if there are multiple reasons, or if there is no per-BCI record,
3710     // assume the worst.
3711     if (log())
3712       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3713                   Deoptimization::trap_reason_name(reason),
3714                   md->trap_count(reason),
3715                   md->overflow_recompile_count());
3716     return true;
3717   } else if (trap_count(reason) != 0
3718              && decompile_count() >= m_cutoff) {
3719     // Too many recompiles globally, and we have seen this sort of trap.
3720     // Use cumulative decompile_count, not just md->decompile_count.
3721     if (log())
3722       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3723                   Deoptimization::trap_reason_name(reason),
3724                   md->trap_count(reason), trap_count(reason),
3725                   md->decompile_count(), decompile_count());
3726     return true;
3727   } else {
3728     // The coast is clear.
3729     return false;
3730   }
3731 }
3732 
3733 // Compute when not to trap. Used by matching trap based nodes and
3734 // NullCheck optimization.
3735 void Compile::set_allowed_deopt_reasons() {
3736   _allowed_reasons = 0;
3737   if (is_method_compilation()) {
3738     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3739       assert(rs < BitsPerInt, "recode bit map");
3740       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3741         _allowed_reasons |= nth_bit(rs);
3742       }
3743     }
3744   }
3745 }
3746 
3747 #ifndef PRODUCT
3748 //------------------------------verify_graph_edges---------------------------
3749 // Walk the Graph and verify that there is a one-to-one correspondence
3750 // between Use-Def edges and Def-Use edges in the graph.
3751 void Compile::verify_graph_edges(bool no_dead_code) {
3752   if (VerifyGraphEdges) {
3753     ResourceArea *area = Thread::current()->resource_area();
3754     Unique_Node_List visited(area);
3755     // Call recursive graph walk to check edges
3756     _root->verify_edges(visited);
3757     if (no_dead_code) {
3758       // Now make sure that no visited node is used by an unvisited node.
3759       bool dead_nodes = false;
3760       Unique_Node_List checked(area);
3761       while (visited.size() > 0) {
3762         Node* n = visited.pop();
3763         checked.push(n);
3764         for (uint i = 0; i < n->outcnt(); i++) {
3765           Node* use = n->raw_out(i);
3766           if (checked.member(use))  continue;  // already checked
3767           if (visited.member(use))  continue;  // already in the graph
3768           if (use->is_Con())        continue;  // a dead ConNode is OK
3769           // At this point, we have found a dead node which is DU-reachable.
3770           if (!dead_nodes) {
3771             tty->print_cr("*** Dead nodes reachable via DU edges:");
3772             dead_nodes = true;
3773           }
3774           use->dump(2);
3775           tty->print_cr("---");
3776           checked.push(use);  // No repeats; pretend it is now checked.
3777         }
3778       }
3779       assert(!dead_nodes, "using nodes must be reachable from root");
3780     }
3781   }
3782 }
3783 
3784 // Verify GC barriers consistency
3785 // Currently supported:
3786 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3787 void Compile::verify_barriers() {
3788   if (UseG1GC) {
3789     // Verify G1 pre-barriers
3790     const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active());
3791 
3792     ResourceArea *area = Thread::current()->resource_area();
3793     Unique_Node_List visited(area);
3794     Node_List worklist(area);
3795     // We're going to walk control flow backwards starting from the Root
3796     worklist.push(_root);
3797     while (worklist.size() > 0) {
3798       Node* x = worklist.pop();
3799       if (x == NULL || x == top()) continue;
3800       if (visited.member(x)) {
3801         continue;
3802       } else {
3803         visited.push(x);
3804       }
3805 
3806       if (x->is_Region()) {
3807         for (uint i = 1; i < x->req(); i++) {
3808           worklist.push(x->in(i));
3809         }
3810       } else {
3811         worklist.push(x->in(0));
3812         // We are looking for the pattern:
3813         //                            /->ThreadLocal
3814         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3815         //              \->ConI(0)
3816         // We want to verify that the If and the LoadB have the same control
3817         // See GraphKit::g1_write_barrier_pre()
3818         if (x->is_If()) {
3819           IfNode *iff = x->as_If();
3820           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3821             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3822             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3823                 && cmp->in(1)->is_Load()) {
3824               LoadNode* load = cmp->in(1)->as_Load();
3825               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3826                   && load->in(2)->in(3)->is_Con()
3827                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3828 
3829                 Node* if_ctrl = iff->in(0);
3830                 Node* load_ctrl = load->in(0);
3831 
3832                 if (if_ctrl != load_ctrl) {
3833                   // Skip possible CProj->NeverBranch in infinite loops
3834                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3835                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3836                     if_ctrl = if_ctrl->in(0)->in(0);
3837                   }
3838                 }
3839                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3840               }
3841             }
3842           }
3843         }
3844       }
3845     }
3846   }
3847 }
3848 
3849 #endif
3850 
3851 // The Compile object keeps track of failure reasons separately from the ciEnv.
3852 // This is required because there is not quite a 1-1 relation between the
3853 // ciEnv and its compilation task and the Compile object.  Note that one
3854 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3855 // to backtrack and retry without subsuming loads.  Other than this backtracking
3856 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3857 // by the logic in C2Compiler.
3858 void Compile::record_failure(const char* reason) {
3859   if (log() != NULL) {
3860     log()->elem("failure reason='%s' phase='compile'", reason);
3861   }
3862   if (_failure_reason == NULL) {
3863     // Record the first failure reason.
3864     _failure_reason = reason;
3865   }
3866 
3867   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3868     C->print_method(PHASE_FAILURE);
3869   }
3870   _root = NULL;  // flush the graph, too
3871 }
3872 
3873 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3874   : TraceTime(name, accumulator, CITime, CITimeVerbose),
3875     _phase_name(name), _dolog(CITimeVerbose)
3876 {
3877   if (_dolog) {
3878     C = Compile::current();
3879     _log = C->log();
3880   } else {
3881     C = NULL;
3882     _log = NULL;
3883   }
3884   if (_log != NULL) {
3885     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3886     _log->stamp();
3887     _log->end_head();
3888   }
3889 }
3890 
3891 Compile::TracePhase::~TracePhase() {
3892 
3893   C = Compile::current();
3894   if (_dolog) {
3895     _log = C->log();
3896   } else {
3897     _log = NULL;
3898   }
3899 
3900 #ifdef ASSERT
3901   if (PrintIdealNodeCount) {
3902     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3903                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3904   }
3905 
3906   if (VerifyIdealNodeCount) {
3907     Compile::current()->print_missing_nodes();
3908   }
3909 #endif
3910 
3911   if (_log != NULL) {
3912     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3913   }
3914 }
3915 
3916 //=============================================================================
3917 // Two Constant's are equal when the type and the value are equal.
3918 bool Compile::Constant::operator==(const Constant& other) {
3919   if (type()          != other.type()         )  return false;
3920   if (can_be_reused() != other.can_be_reused())  return false;
3921   // For floating point values we compare the bit pattern.
3922   switch (type()) {
3923   case T_INT:
3924   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3925   case T_LONG:
3926   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
3927   case T_OBJECT:
3928   case T_ADDRESS: return (_v._value.l == other._v._value.l);
3929   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
3930   case T_METADATA: return (_v._metadata == other._v._metadata);
3931   default: ShouldNotReachHere(); return false;
3932   }
3933 }
3934 
3935 static int type_to_size_in_bytes(BasicType t) {
3936   switch (t) {
3937   case T_INT:     return sizeof(jint   );
3938   case T_LONG:    return sizeof(jlong  );
3939   case T_FLOAT:   return sizeof(jfloat );
3940   case T_DOUBLE:  return sizeof(jdouble);
3941   case T_METADATA: return sizeof(Metadata*);
3942     // We use T_VOID as marker for jump-table entries (labels) which
3943     // need an internal word relocation.
3944   case T_VOID:
3945   case T_ADDRESS:
3946   case T_OBJECT:  return sizeof(jobject);
3947   default:
3948     ShouldNotReachHere();
3949     return -1;
3950   }
3951 }
3952 
3953 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3954   // sort descending
3955   if (a->freq() > b->freq())  return -1;
3956   if (a->freq() < b->freq())  return  1;
3957   return 0;
3958 }
3959 
3960 void Compile::ConstantTable::calculate_offsets_and_size() {
3961   // First, sort the array by frequencies.
3962   _constants.sort(qsort_comparator);
3963 
3964 #ifdef ASSERT
3965   // Make sure all jump-table entries were sorted to the end of the
3966   // array (they have a negative frequency).
3967   bool found_void = false;
3968   for (int i = 0; i < _constants.length(); i++) {
3969     Constant con = _constants.at(i);
3970     if (con.type() == T_VOID)
3971       found_void = true;  // jump-tables
3972     else
3973       assert(!found_void, "wrong sorting");
3974   }
3975 #endif
3976 
3977   int offset = 0;
3978   for (int i = 0; i < _constants.length(); i++) {
3979     Constant* con = _constants.adr_at(i);
3980 
3981     // Align offset for type.
3982     int typesize = type_to_size_in_bytes(con->type());
3983     offset = align_up(offset, typesize);
3984     con->set_offset(offset);   // set constant's offset
3985 
3986     if (con->type() == T_VOID) {
3987       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3988       offset = offset + typesize * n->outcnt();  // expand jump-table
3989     } else {
3990       offset = offset + typesize;
3991     }
3992   }
3993 
3994   // Align size up to the next section start (which is insts; see
3995   // CodeBuffer::align_at_start).
3996   assert(_size == -1, "already set?");
3997   _size = align_up(offset, (int)CodeEntryAlignment);
3998 }
3999 
4000 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4001   MacroAssembler _masm(&cb);
4002   for (int i = 0; i < _constants.length(); i++) {
4003     Constant con = _constants.at(i);
4004     address constant_addr = NULL;
4005     switch (con.type()) {
4006     case T_INT:    constant_addr = _masm.int_constant(   con.get_jint()   ); break;
4007     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
4008     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4009     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4010     case T_OBJECT: {
4011       jobject obj = con.get_jobject();
4012       int oop_index = _masm.oop_recorder()->find_index(obj);
4013       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4014       break;
4015     }
4016     case T_ADDRESS: {
4017       address addr = (address) con.get_jobject();
4018       constant_addr = _masm.address_constant(addr);
4019       break;
4020     }
4021     // We use T_VOID as marker for jump-table entries (labels) which
4022     // need an internal word relocation.
4023     case T_VOID: {
4024       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4025       // Fill the jump-table with a dummy word.  The real value is
4026       // filled in later in fill_jump_table.
4027       address dummy = (address) n;
4028       constant_addr = _masm.address_constant(dummy);
4029       // Expand jump-table
4030       for (uint i = 1; i < n->outcnt(); i++) {
4031         address temp_addr = _masm.address_constant(dummy + i);
4032         assert(temp_addr, "consts section too small");
4033       }
4034       break;
4035     }
4036     case T_METADATA: {
4037       Metadata* obj = con.get_metadata();
4038       int metadata_index = _masm.oop_recorder()->find_index(obj);
4039       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4040       break;
4041     }
4042     default: ShouldNotReachHere();
4043     }
4044     assert(constant_addr, "consts section too small");
4045     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4046             "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4047   }
4048 }
4049 
4050 int Compile::ConstantTable::find_offset(Constant& con) const {
4051   int idx = _constants.find(con);
4052   assert(idx != -1, "constant must be in constant table");
4053   int offset = _constants.at(idx).offset();
4054   assert(offset != -1, "constant table not emitted yet?");
4055   return offset;
4056 }
4057 
4058 void Compile::ConstantTable::add(Constant& con) {
4059   if (con.can_be_reused()) {
4060     int idx = _constants.find(con);
4061     if (idx != -1 && _constants.at(idx).can_be_reused()) {
4062       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
4063       return;
4064     }
4065   }
4066   (void) _constants.append(con);
4067 }
4068 
4069 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4070   Block* b = Compile::current()->cfg()->get_block_for_node(n);
4071   Constant con(type, value, b->_freq);
4072   add(con);
4073   return con;
4074 }
4075 
4076 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4077   Constant con(metadata);
4078   add(con);
4079   return con;
4080 }
4081 
4082 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4083   jvalue value;
4084   BasicType type = oper->type()->basic_type();
4085   switch (type) {
4086   case T_LONG:    value.j = oper->constantL(); break;
4087   case T_FLOAT:   value.f = oper->constantF(); break;
4088   case T_DOUBLE:  value.d = oper->constantD(); break;
4089   case T_OBJECT:
4090   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4091   case T_METADATA: return add((Metadata*)oper->constant()); break;
4092   default: guarantee(false, "unhandled type: %s", type2name(type));
4093   }
4094   return add(n, type, value);
4095 }
4096 
4097 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4098   jvalue value;
4099   // We can use the node pointer here to identify the right jump-table
4100   // as this method is called from Compile::Fill_buffer right before
4101   // the MachNodes are emitted and the jump-table is filled (means the
4102   // MachNode pointers do not change anymore).
4103   value.l = (jobject) n;
4104   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
4105   add(con);
4106   return con;
4107 }
4108 
4109 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4110   // If called from Compile::scratch_emit_size do nothing.
4111   if (Compile::current()->in_scratch_emit_size())  return;
4112 
4113   assert(labels.is_nonempty(), "must be");
4114   assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4115 
4116   // Since MachConstantNode::constant_offset() also contains
4117   // table_base_offset() we need to subtract the table_base_offset()
4118   // to get the plain offset into the constant table.
4119   int offset = n->constant_offset() - table_base_offset();
4120 
4121   MacroAssembler _masm(&cb);
4122   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4123 
4124   for (uint i = 0; i < n->outcnt(); i++) {
4125     address* constant_addr = &jump_table_base[i];
4126     assert(*constant_addr == (((address) n) + i), "all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i));
4127     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4128     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4129   }
4130 }
4131 
4132 //----------------------------static_subtype_check-----------------------------
4133 // Shortcut important common cases when superklass is exact:
4134 // (0) superklass is java.lang.Object (can occur in reflective code)
4135 // (1) subklass is already limited to a subtype of superklass => always ok
4136 // (2) subklass does not overlap with superklass => always fail
4137 // (3) superklass has NO subtypes and we can check with a simple compare.
4138 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4139   if (StressReflectiveCode || superk == NULL || subk == NULL) {
4140     return SSC_full_test;       // Let caller generate the general case.
4141   }
4142 
4143   if (!EnableMVT && !EnableValhalla && superk == env()->Object_klass()) {
4144     return SSC_always_true;     // (0) this test cannot fail
4145   }
4146 
4147   ciType* superelem = superk;
4148   if (superelem->is_array_klass())
4149     superelem = superelem->as_array_klass()->base_element_type();
4150 
4151   if (!subk->is_interface()) {  // cannot trust static interface types yet
4152     if (subk->is_subtype_of(superk)) {
4153       return SSC_always_true;   // (1) false path dead; no dynamic test needed
4154     }
4155     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4156         !superk->is_subtype_of(subk)) {
4157       return SSC_always_false;
4158     }
4159   }
4160 
4161   // If casting to an instance klass, it must have no subtypes
4162   if (superk->is_interface()) {
4163     // Cannot trust interfaces yet.
4164     // %%% S.B. superk->nof_implementors() == 1
4165   } else if (superelem->is_instance_klass()) {
4166     ciInstanceKlass* ik = superelem->as_instance_klass();
4167     if (!ik->has_subklass() && !ik->is_interface()) {
4168       if (!ik->is_final()) {
4169         // Add a dependency if there is a chance of a later subclass.
4170         dependencies()->assert_leaf_type(ik);
4171       }
4172       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
4173     }
4174   } else {
4175     // A primitive array type has no subtypes.
4176     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
4177   }
4178 
4179   return SSC_full_test;
4180 }
4181 
4182 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4183 #ifdef _LP64
4184   // The scaled index operand to AddP must be a clean 64-bit value.
4185   // Java allows a 32-bit int to be incremented to a negative
4186   // value, which appears in a 64-bit register as a large
4187   // positive number.  Using that large positive number as an
4188   // operand in pointer arithmetic has bad consequences.
4189   // On the other hand, 32-bit overflow is rare, and the possibility
4190   // can often be excluded, if we annotate the ConvI2L node with
4191   // a type assertion that its value is known to be a small positive
4192   // number.  (The prior range check has ensured this.)
4193   // This assertion is used by ConvI2LNode::Ideal.
4194   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4195   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4196   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4197   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4198 #endif
4199   return idx;
4200 }
4201 
4202 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4203 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4204   if (ctrl != NULL) {
4205     // Express control dependency by a CastII node with a narrow type.
4206     value = new CastIINode(value, itype, false, true /* range check dependency */);
4207     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4208     // node from floating above the range check during loop optimizations. Otherwise, the
4209     // ConvI2L node may be eliminated independently of the range check, causing the data path
4210     // to become TOP while the control path is still there (although it's unreachable).
4211     value->set_req(0, ctrl);
4212     // Save CastII node to remove it after loop optimizations.
4213     phase->C->add_range_check_cast(value);
4214     value = phase->transform(value);
4215   }
4216   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4217   return phase->transform(new ConvI2LNode(value, ltype));
4218 }
4219 
4220 // The message about the current inlining is accumulated in
4221 // _print_inlining_stream and transfered into the _print_inlining_list
4222 // once we know whether inlining succeeds or not. For regular
4223 // inlining, messages are appended to the buffer pointed by
4224 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4225 // a new buffer is added after _print_inlining_idx in the list. This
4226 // way we can update the inlining message for late inlining call site
4227 // when the inlining is attempted again.
4228 void Compile::print_inlining_init() {
4229   if (print_inlining() || print_intrinsics()) {
4230     _print_inlining_stream = new stringStream();
4231     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4232   }
4233 }
4234 
4235 void Compile::print_inlining_reinit() {
4236   if (print_inlining() || print_intrinsics()) {
4237     // Re allocate buffer when we change ResourceMark
4238     _print_inlining_stream = new stringStream();
4239   }
4240 }
4241 
4242 void Compile::print_inlining_reset() {
4243   _print_inlining_stream->reset();
4244 }
4245 
4246 void Compile::print_inlining_commit() {
4247   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4248   // Transfer the message from _print_inlining_stream to the current
4249   // _print_inlining_list buffer and clear _print_inlining_stream.
4250   _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4251   print_inlining_reset();
4252 }
4253 
4254 void Compile::print_inlining_push() {
4255   // Add new buffer to the _print_inlining_list at current position
4256   _print_inlining_idx++;
4257   _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4258 }
4259 
4260 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4261   return _print_inlining_list->at(_print_inlining_idx);
4262 }
4263 
4264 void Compile::print_inlining_update(CallGenerator* cg) {
4265   if (print_inlining() || print_intrinsics()) {
4266     if (!cg->is_late_inline()) {
4267       if (print_inlining_current().cg() != NULL) {
4268         print_inlining_push();
4269       }
4270       print_inlining_commit();
4271     } else {
4272       if (print_inlining_current().cg() != cg &&
4273           (print_inlining_current().cg() != NULL ||
4274            print_inlining_current().ss()->size() != 0)) {
4275         print_inlining_push();
4276       }
4277       print_inlining_commit();
4278       print_inlining_current().set_cg(cg);
4279     }
4280   }
4281 }
4282 
4283 void Compile::print_inlining_move_to(CallGenerator* cg) {
4284   // We resume inlining at a late inlining call site. Locate the
4285   // corresponding inlining buffer so that we can update it.
4286   if (print_inlining()) {
4287     for (int i = 0; i < _print_inlining_list->length(); i++) {
4288       if (_print_inlining_list->adr_at(i)->cg() == cg) {
4289         _print_inlining_idx = i;
4290         return;
4291       }
4292     }
4293     ShouldNotReachHere();
4294   }
4295 }
4296 
4297 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4298   if (print_inlining()) {
4299     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4300     assert(print_inlining_current().cg() == cg, "wrong entry");
4301     // replace message with new message
4302     _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4303     print_inlining_commit();
4304     print_inlining_current().set_cg(cg);
4305   }
4306 }
4307 
4308 void Compile::print_inlining_assert_ready() {
4309   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4310 }
4311 
4312 void Compile::process_print_inlining() {
4313   bool do_print_inlining = print_inlining() || print_intrinsics();
4314   if (do_print_inlining || log() != NULL) {
4315     // Print inlining message for candidates that we couldn't inline
4316     // for lack of space
4317     for (int i = 0; i < _late_inlines.length(); i++) {
4318       CallGenerator* cg = _late_inlines.at(i);
4319       if (!cg->is_mh_late_inline()) {
4320         const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4321         if (do_print_inlining) {
4322           cg->print_inlining_late(msg);
4323         }
4324         log_late_inline_failure(cg, msg);
4325       }
4326     }
4327   }
4328   if (do_print_inlining) {
4329     ResourceMark rm;
4330     stringStream ss;
4331     for (int i = 0; i < _print_inlining_list->length(); i++) {
4332       ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4333     }
4334     size_t end = ss.size();
4335     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4336     strncpy(_print_inlining_output, ss.base(), end+1);
4337     _print_inlining_output[end] = 0;
4338   }
4339 }
4340 
4341 void Compile::dump_print_inlining() {
4342   if (_print_inlining_output != NULL) {
4343     tty->print_raw(_print_inlining_output);
4344   }
4345 }
4346 
4347 void Compile::log_late_inline(CallGenerator* cg) {
4348   if (log() != NULL) {
4349     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4350                 cg->unique_id());
4351     JVMState* p = cg->call_node()->jvms();
4352     while (p != NULL) {
4353       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4354       p = p->caller();
4355     }
4356     log()->tail("late_inline");
4357   }
4358 }
4359 
4360 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4361   log_late_inline(cg);
4362   if (log() != NULL) {
4363     log()->inline_fail(msg);
4364   }
4365 }
4366 
4367 void Compile::log_inline_id(CallGenerator* cg) {
4368   if (log() != NULL) {
4369     // The LogCompilation tool needs a unique way to identify late
4370     // inline call sites. This id must be unique for this call site in
4371     // this compilation. Try to have it unique across compilations as
4372     // well because it can be convenient when grepping through the log
4373     // file.
4374     // Distinguish OSR compilations from others in case CICountOSR is
4375     // on.
4376     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4377     cg->set_unique_id(id);
4378     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4379   }
4380 }
4381 
4382 void Compile::log_inline_failure(const char* msg) {
4383   if (C->log() != NULL) {
4384     C->log()->inline_fail(msg);
4385   }
4386 }
4387 
4388 
4389 // Dump inlining replay data to the stream.
4390 // Don't change thread state and acquire any locks.
4391 void Compile::dump_inline_data(outputStream* out) {
4392   InlineTree* inl_tree = ilt();
4393   if (inl_tree != NULL) {
4394     out->print(" inline %d", inl_tree->count());
4395     inl_tree->dump_replay_data(out);
4396   }
4397 }
4398 
4399 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4400   if (n1->Opcode() < n2->Opcode())      return -1;
4401   else if (n1->Opcode() > n2->Opcode()) return 1;
4402 
4403   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4404   for (uint i = 1; i < n1->req(); i++) {
4405     if (n1->in(i) < n2->in(i))      return -1;
4406     else if (n1->in(i) > n2->in(i)) return 1;
4407   }
4408 
4409   return 0;
4410 }
4411 
4412 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4413   Node* n1 = *n1p;
4414   Node* n2 = *n2p;
4415 
4416   return cmp_expensive_nodes(n1, n2);
4417 }
4418 
4419 void Compile::sort_expensive_nodes() {
4420   if (!expensive_nodes_sorted()) {
4421     _expensive_nodes->sort(cmp_expensive_nodes);
4422   }
4423 }
4424 
4425 bool Compile::expensive_nodes_sorted() const {
4426   for (int i = 1; i < _expensive_nodes->length(); i++) {
4427     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4428       return false;
4429     }
4430   }
4431   return true;
4432 }
4433 
4434 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4435   if (_expensive_nodes->length() == 0) {
4436     return false;
4437   }
4438 
4439   assert(OptimizeExpensiveOps, "optimization off?");
4440 
4441   // Take this opportunity to remove dead nodes from the list
4442   int j = 0;
4443   for (int i = 0; i < _expensive_nodes->length(); i++) {
4444     Node* n = _expensive_nodes->at(i);
4445     if (!n->is_unreachable(igvn)) {
4446       assert(n->is_expensive(), "should be expensive");
4447       _expensive_nodes->at_put(j, n);
4448       j++;
4449     }
4450   }
4451   _expensive_nodes->trunc_to(j);
4452 
4453   // Then sort the list so that similar nodes are next to each other
4454   // and check for at least two nodes of identical kind with same data
4455   // inputs.
4456   sort_expensive_nodes();
4457 
4458   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4459     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4460       return true;
4461     }
4462   }
4463 
4464   return false;
4465 }
4466 
4467 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4468   if (_expensive_nodes->length() == 0) {
4469     return;
4470   }
4471 
4472   assert(OptimizeExpensiveOps, "optimization off?");
4473 
4474   // Sort to bring similar nodes next to each other and clear the
4475   // control input of nodes for which there's only a single copy.
4476   sort_expensive_nodes();
4477 
4478   int j = 0;
4479   int identical = 0;
4480   int i = 0;
4481   bool modified = false;
4482   for (; i < _expensive_nodes->length()-1; i++) {
4483     assert(j <= i, "can't write beyond current index");
4484     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4485       identical++;
4486       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4487       continue;
4488     }
4489     if (identical > 0) {
4490       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4491       identical = 0;
4492     } else {
4493       Node* n = _expensive_nodes->at(i);
4494       igvn.replace_input_of(n, 0, NULL);
4495       igvn.hash_insert(n);
4496       modified = true;
4497     }
4498   }
4499   if (identical > 0) {
4500     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4501   } else if (_expensive_nodes->length() >= 1) {
4502     Node* n = _expensive_nodes->at(i);
4503     igvn.replace_input_of(n, 0, NULL);
4504     igvn.hash_insert(n);
4505     modified = true;
4506   }
4507   _expensive_nodes->trunc_to(j);
4508   if (modified) {
4509     igvn.optimize();
4510   }
4511 }
4512 
4513 void Compile::add_expensive_node(Node * n) {
4514   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4515   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4516   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4517   if (OptimizeExpensiveOps) {
4518     _expensive_nodes->append(n);
4519   } else {
4520     // Clear control input and let IGVN optimize expensive nodes if
4521     // OptimizeExpensiveOps is off.
4522     n->set_req(0, NULL);
4523   }
4524 }
4525 
4526 /**
4527  * Remove the speculative part of types and clean up the graph
4528  */
4529 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4530   if (UseTypeSpeculation) {
4531     Unique_Node_List worklist;
4532     worklist.push(root());
4533     int modified = 0;
4534     // Go over all type nodes that carry a speculative type, drop the
4535     // speculative part of the type and enqueue the node for an igvn
4536     // which may optimize it out.
4537     for (uint next = 0; next < worklist.size(); ++next) {
4538       Node *n  = worklist.at(next);
4539       if (n->is_Type()) {
4540         TypeNode* tn = n->as_Type();
4541         const Type* t = tn->type();
4542         const Type* t_no_spec = t->remove_speculative();
4543         if (t_no_spec != t) {
4544           bool in_hash = igvn.hash_delete(n);
4545           assert(in_hash, "node should be in igvn hash table");
4546           tn->set_type(t_no_spec);
4547           igvn.hash_insert(n);
4548           igvn._worklist.push(n); // give it a chance to go away
4549           modified++;
4550         }
4551       }
4552       uint max = n->len();
4553       for( uint i = 0; i < max; ++i ) {
4554         Node *m = n->in(i);
4555         if (not_a_node(m))  continue;
4556         worklist.push(m);
4557       }
4558     }
4559     // Drop the speculative part of all types in the igvn's type table
4560     igvn.remove_speculative_types();
4561     if (modified > 0) {
4562       igvn.optimize();
4563     }
4564 #ifdef ASSERT
4565     // Verify that after the IGVN is over no speculative type has resurfaced
4566     worklist.clear();
4567     worklist.push(root());
4568     for (uint next = 0; next < worklist.size(); ++next) {
4569       Node *n  = worklist.at(next);
4570       const Type* t = igvn.type_or_null(n);
4571       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4572       if (n->is_Type()) {
4573         t = n->as_Type()->type();
4574         assert(t == t->remove_speculative(), "no more speculative types");
4575       }
4576       uint max = n->len();
4577       for( uint i = 0; i < max; ++i ) {
4578         Node *m = n->in(i);
4579         if (not_a_node(m))  continue;
4580         worklist.push(m);
4581       }
4582     }
4583     igvn.check_no_speculative_types();
4584 #endif
4585   }
4586 }
4587 
4588 // Auxiliary method to support randomized stressing/fuzzing.
4589 //
4590 // This method can be called the arbitrary number of times, with current count
4591 // as the argument. The logic allows selecting a single candidate from the
4592 // running list of candidates as follows:
4593 //    int count = 0;
4594 //    Cand* selected = null;
4595 //    while(cand = cand->next()) {
4596 //      if (randomized_select(++count)) {
4597 //        selected = cand;
4598 //      }
4599 //    }
4600 //
4601 // Including count equalizes the chances any candidate is "selected".
4602 // This is useful when we don't have the complete list of candidates to choose
4603 // from uniformly. In this case, we need to adjust the randomicity of the
4604 // selection, or else we will end up biasing the selection towards the latter
4605 // candidates.
4606 //
4607 // Quick back-envelope calculation shows that for the list of n candidates
4608 // the equal probability for the candidate to persist as "best" can be
4609 // achieved by replacing it with "next" k-th candidate with the probability
4610 // of 1/k. It can be easily shown that by the end of the run, the
4611 // probability for any candidate is converged to 1/n, thus giving the
4612 // uniform distribution among all the candidates.
4613 //
4614 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4615 #define RANDOMIZED_DOMAIN_POW 29
4616 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4617 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4618 bool Compile::randomized_select(int count) {
4619   assert(count > 0, "only positive");
4620   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4621 }
4622 
4623 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4624 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4625 
4626 void NodeCloneInfo::dump() const {
4627   tty->print(" {%d:%d} ", idx(), gen());
4628 }
4629 
4630 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4631   uint64_t val = value(old->_idx);
4632   NodeCloneInfo cio(val);
4633   assert(val != 0, "old node should be in the map");
4634   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4635   insert(nnn->_idx, cin.get());
4636 #ifndef PRODUCT
4637   if (is_debug()) {
4638     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4639   }
4640 #endif
4641 }
4642 
4643 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4644   NodeCloneInfo cio(value(old->_idx));
4645   if (cio.get() == 0) {
4646     cio.set(old->_idx, 0);
4647     insert(old->_idx, cio.get());
4648 #ifndef PRODUCT
4649     if (is_debug()) {
4650       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4651     }
4652 #endif
4653   }
4654   clone(old, nnn, gen);
4655 }
4656 
4657 int CloneMap::max_gen() const {
4658   int g = 0;
4659   DictI di(_dict);
4660   for(; di.test(); ++di) {
4661     int t = gen(di._key);
4662     if (g < t) {
4663       g = t;
4664 #ifndef PRODUCT
4665       if (is_debug()) {
4666         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4667       }
4668 #endif
4669     }
4670   }
4671   return g;
4672 }
4673 
4674 void CloneMap::dump(node_idx_t key) const {
4675   uint64_t val = value(key);
4676   if (val != 0) {
4677     NodeCloneInfo ni(val);
4678     ni.dump();
4679   }
4680 }
4681 
4682 // Helper function for enforcing certain bytecodes to reexecute if
4683 // deoptimization happens
4684 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
4685   ciMethod* cur_method = jvms->method();
4686   int       cur_bci   = jvms->bci();
4687   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
4688     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
4689     return Interpreter::bytecode_should_reexecute(code) ||
4690            (is_anewarray && (code == Bytecodes::_multianewarray));
4691     // Reexecute _multianewarray bytecode which was replaced with
4692     // sequence of [a]newarray. See Parse::do_multianewarray().
4693     //
4694     // Note: interpreter should not have it set since this optimization
4695     // is limited by dimensions and guarded by flag so in some cases
4696     // multianewarray() runtime calls will be generated and
4697     // the bytecode should not be reexecutes (stack will not be reset).
4698   } else
4699     return false;
4700 }
4701 
4702 void Compile::add_safepoint_edges(SafePointNode* call, JVMState* youngest_jvms, bool can_prune_locals, uint stack_slots_not_pruned) {
4703   // do not scribble on the input jvms
4704   JVMState* out_jvms = youngest_jvms->clone_deep(C);
4705   call->set_jvms(out_jvms); // Start jvms list for call node
4706 
4707   // For a known set of bytecodes, the interpreter should reexecute them if
4708   // deoptimization happens. We set the reexecute state for them here
4709   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
4710       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
4711     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
4712   }
4713 
4714   // Presize the call:
4715   DEBUG_ONLY(uint non_debug_edges = call->req());
4716   call->add_req_batch(top(), youngest_jvms->debug_depth());
4717   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
4718 
4719   // Set up edges so that the call looks like this:
4720   //  Call [state:] ctl io mem fptr retadr
4721   //       [parms:] parm0 ... parmN
4722   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
4723   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
4724   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
4725   // Note that caller debug info precedes callee debug info.
4726 
4727   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
4728   uint debug_ptr = call->req();
4729 
4730   // Loop over the map input edges associated with jvms, add them
4731   // to the call node, & reset all offsets to match call node array.
4732   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
4733     uint debug_end   = debug_ptr;
4734     uint debug_start = debug_ptr - in_jvms->debug_size();
4735     debug_ptr = debug_start;  // back up the ptr
4736 
4737     uint p = debug_start;  // walks forward in [debug_start, debug_end)
4738     uint j, k, l;
4739     SafePointNode* in_map = in_jvms->map();
4740     out_jvms->set_map(call);
4741 
4742     if (can_prune_locals) {
4743       assert(in_jvms->method() == out_jvms->method(), "sanity");
4744       // If the current throw can reach an exception handler in this JVMS,
4745       // then we must keep everything live that can reach that handler.
4746       // As a quick and dirty approximation, we look for any handlers at all.
4747       if (in_jvms->method()->has_exception_handlers()) {
4748         can_prune_locals = false;
4749       }
4750     }
4751 
4752     // Add the Locals
4753     k = in_jvms->locoff();
4754     l = in_jvms->loc_size();
4755     out_jvms->set_locoff(p);
4756     if (!can_prune_locals) {
4757       for (j = 0; j < l; j++)
4758         call->set_req(p++, in_map->in(k+j));
4759     } else {
4760       p += l;  // already set to top above by add_req_batch
4761     }
4762 
4763     // Add the Expression Stack
4764     k = in_jvms->stkoff();
4765     l = in_jvms->sp();
4766     out_jvms->set_stkoff(p);
4767     if (!can_prune_locals) {
4768       for (j = 0; j < l; j++)
4769         call->set_req(p++, in_map->in(k+j));
4770     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
4771       // Divide stack into {S0,...,S1}, where S0 is set to top.
4772       uint s1 = stack_slots_not_pruned;
4773       stack_slots_not_pruned = 0;  // for next iteration
4774       if (s1 > l)  s1 = l;
4775       uint s0 = l - s1;
4776       p += s0;  // skip the tops preinstalled by add_req_batch
4777       for (j = s0; j < l; j++)
4778         call->set_req(p++, in_map->in(k+j));
4779     } else {
4780       p += l;  // already set to top above by add_req_batch
4781     }
4782 
4783     // Add the Monitors
4784     k = in_jvms->monoff();
4785     l = in_jvms->mon_size();
4786     out_jvms->set_monoff(p);
4787     for (j = 0; j < l; j++)
4788       call->set_req(p++, in_map->in(k+j));
4789 
4790     // Copy any scalar object fields.
4791     k = in_jvms->scloff();
4792     l = in_jvms->scl_size();
4793     out_jvms->set_scloff(p);
4794     for (j = 0; j < l; j++)
4795       call->set_req(p++, in_map->in(k+j));
4796 
4797     // Finish the new jvms.
4798     out_jvms->set_endoff(p);
4799 
4800     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
4801     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
4802     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
4803     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
4804     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
4805     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
4806 
4807     // Update the two tail pointers in parallel.
4808     out_jvms = out_jvms->caller();
4809     in_jvms  = in_jvms->caller();
4810   }
4811 
4812   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
4813 
4814   // Test the correctness of JVMState::debug_xxx accessors:
4815   assert(call->jvms()->debug_start() == non_debug_edges, "");
4816   assert(call->jvms()->debug_end()   == call->req(), "");
4817   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
4818 }