1 /*
   2  * Copyright (c) 1997, 2018, 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 "gc/shared/barrierSet.hpp"
  37 #include "gc/shared/c2/barrierSetC2.hpp"
  38 #include "memory/resourceArea.hpp"
  39 #include "opto/addnode.hpp"
  40 #include "opto/block.hpp"
  41 #include "opto/c2compiler.hpp"
  42 #include "opto/callGenerator.hpp"
  43 #include "opto/callnode.hpp"
  44 #include "opto/castnode.hpp"
  45 #include "opto/cfgnode.hpp"
  46 #include "opto/chaitin.hpp"
  47 #include "opto/compile.hpp"
  48 #include "opto/connode.hpp"
  49 #include "opto/convertnode.hpp"
  50 #include "opto/divnode.hpp"
  51 #include "opto/escape.hpp"
  52 #include "opto/idealGraphPrinter.hpp"
  53 #include "opto/loopnode.hpp"
  54 #include "opto/machnode.hpp"
  55 #include "opto/macro.hpp"
  56 #include "opto/matcher.hpp"
  57 #include "opto/mathexactnode.hpp"
  58 #include "opto/memnode.hpp"
  59 #include "opto/mulnode.hpp"
  60 #include "opto/narrowptrnode.hpp"
  61 #include "opto/node.hpp"
  62 #include "opto/opcodes.hpp"
  63 #include "opto/output.hpp"
  64 #include "opto/parse.hpp"
  65 #include "opto/phaseX.hpp"
  66 #include "opto/rootnode.hpp"
  67 #include "opto/runtime.hpp"
  68 #include "opto/stringopts.hpp"
  69 #include "opto/type.hpp"
  70 #include "opto/vectornode.hpp"
  71 #include "runtime/arguments.hpp"
  72 #include "runtime/sharedRuntime.hpp"
  73 #include "runtime/signature.hpp"
  74 #include "runtime/stubRoutines.hpp"
  75 #include "runtime/timer.hpp"
  76 #include "utilities/align.hpp"
  77 #include "utilities/copy.hpp"
  78 #include "utilities/macros.hpp"
  79 #if INCLUDE_G1GC
  80 #include "gc/g1/g1ThreadLocalData.hpp"
  81 #endif // INCLUDE_G1GC
  82 #if INCLUDE_ZGC
  83 #include "gc/z/c2/zBarrierSetC2.hpp"
  84 #endif
  85 #if INCLUDE_SHENANDOAHGC
  86 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
  87 #endif
  88 
  89 
  90 // -------------------- Compile::mach_constant_base_node -----------------------
  91 // Constant table base node singleton.
  92 MachConstantBaseNode* Compile::mach_constant_base_node() {
  93   if (_mach_constant_base_node == NULL) {
  94     _mach_constant_base_node = new MachConstantBaseNode();
  95     _mach_constant_base_node->add_req(C->root());
  96   }
  97   return _mach_constant_base_node;
  98 }
  99 
 100 
 101 /// Support for intrinsics.
 102 
 103 // Return the index at which m must be inserted (or already exists).
 104 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
 105 class IntrinsicDescPair {
 106  private:
 107   ciMethod* _m;
 108   bool _is_virtual;
 109  public:
 110   IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
 111   static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
 112     ciMethod* m= elt->method();
 113     ciMethod* key_m = key->_m;
 114     if (key_m < m)      return -1;
 115     else if (key_m > m) return 1;
 116     else {
 117       bool is_virtual = elt->is_virtual();
 118       bool key_virtual = key->_is_virtual;
 119       if (key_virtual < is_virtual)      return -1;
 120       else if (key_virtual > is_virtual) return 1;
 121       else                               return 0;
 122     }
 123   }
 124 };
 125 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
 126 #ifdef ASSERT
 127   for (int i = 1; i < _intrinsics->length(); i++) {
 128     CallGenerator* cg1 = _intrinsics->at(i-1);
 129     CallGenerator* cg2 = _intrinsics->at(i);
 130     assert(cg1->method() != cg2->method()
 131            ? cg1->method()     < cg2->method()
 132            : cg1->is_virtual() < cg2->is_virtual(),
 133            "compiler intrinsics list must stay sorted");
 134   }
 135 #endif
 136   IntrinsicDescPair pair(m, is_virtual);
 137   return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
 138 }
 139 
 140 void Compile::register_intrinsic(CallGenerator* cg) {
 141   if (_intrinsics == NULL) {
 142     _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 143   }
 144   int len = _intrinsics->length();
 145   bool found = false;
 146   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
 147   assert(!found, "registering twice");
 148   _intrinsics->insert_before(index, cg);
 149   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 150 }
 151 
 152 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 153   assert(m->is_loaded(), "don't try this on unloaded methods");
 154   if (_intrinsics != NULL) {
 155     bool found = false;
 156     int index = intrinsic_insertion_index(m, is_virtual, found);
 157      if (found) {
 158       return _intrinsics->at(index);
 159     }
 160   }
 161   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 162   if (m->intrinsic_id() != vmIntrinsics::_none &&
 163       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 164     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 165     if (cg != NULL) {
 166       // Save it for next time:
 167       register_intrinsic(cg);
 168       return cg;
 169     } else {
 170       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 171     }
 172   }
 173   return NULL;
 174 }
 175 
 176 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 177 // in library_call.cpp.
 178 
 179 
 180 #ifndef PRODUCT
 181 // statistics gathering...
 182 
 183 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 184 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 185 
 186 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 187   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 188   int oflags = _intrinsic_hist_flags[id];
 189   assert(flags != 0, "what happened?");
 190   if (is_virtual) {
 191     flags |= _intrinsic_virtual;
 192   }
 193   bool changed = (flags != oflags);
 194   if ((flags & _intrinsic_worked) != 0) {
 195     juint count = (_intrinsic_hist_count[id] += 1);
 196     if (count == 1) {
 197       changed = true;           // first time
 198     }
 199     // increment the overall count also:
 200     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 201   }
 202   if (changed) {
 203     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 204       // Something changed about the intrinsic's virtuality.
 205       if ((flags & _intrinsic_virtual) != 0) {
 206         // This is the first use of this intrinsic as a virtual call.
 207         if (oflags != 0) {
 208           // We already saw it as a non-virtual, so note both cases.
 209           flags |= _intrinsic_both;
 210         }
 211       } else if ((oflags & _intrinsic_both) == 0) {
 212         // This is the first use of this intrinsic as a non-virtual
 213         flags |= _intrinsic_both;
 214       }
 215     }
 216     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 217   }
 218   // update the overall flags also:
 219   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 220   return changed;
 221 }
 222 
 223 static char* format_flags(int flags, char* buf) {
 224   buf[0] = 0;
 225   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 226   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 227   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 228   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 229   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 230   if (buf[0] == 0)  strcat(buf, ",");
 231   assert(buf[0] == ',', "must be");
 232   return &buf[1];
 233 }
 234 
 235 void Compile::print_intrinsic_statistics() {
 236   char flagsbuf[100];
 237   ttyLocker ttyl;
 238   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 239   tty->print_cr("Compiler intrinsic usage:");
 240   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 241   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 242   #define PRINT_STAT_LINE(name, c, f) \
 243     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 244   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 245     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 246     int   flags = _intrinsic_hist_flags[id];
 247     juint count = _intrinsic_hist_count[id];
 248     if ((flags | count) != 0) {
 249       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 250     }
 251   }
 252   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 253   if (xtty != NULL)  xtty->tail("statistics");
 254 }
 255 
 256 void Compile::print_statistics() {
 257   { ttyLocker ttyl;
 258     if (xtty != NULL)  xtty->head("statistics type='opto'");
 259     Parse::print_statistics();
 260     PhaseCCP::print_statistics();
 261     PhaseRegAlloc::print_statistics();
 262     Scheduling::print_statistics();
 263     PhasePeephole::print_statistics();
 264     PhaseIdealLoop::print_statistics();
 265     if (xtty != NULL)  xtty->tail("statistics");
 266   }
 267   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 268     // put this under its own <statistics> element.
 269     print_intrinsic_statistics();
 270   }
 271 }
 272 #endif //PRODUCT
 273 
 274 // Support for bundling info
 275 Bundle* Compile::node_bundling(const Node *n) {
 276   assert(valid_bundle_info(n), "oob");
 277   return &_node_bundling_base[n->_idx];
 278 }
 279 
 280 bool Compile::valid_bundle_info(const Node *n) {
 281   return (_node_bundling_limit > n->_idx);
 282 }
 283 
 284 
 285 void Compile::gvn_replace_by(Node* n, Node* nn) {
 286   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 287     Node* use = n->last_out(i);
 288     bool is_in_table = initial_gvn()->hash_delete(use);
 289     uint uses_found = 0;
 290     for (uint j = 0; j < use->len(); j++) {
 291       if (use->in(j) == n) {
 292         if (j < use->req())
 293           use->set_req(j, nn);
 294         else
 295           use->set_prec(j, nn);
 296         uses_found++;
 297       }
 298     }
 299     if (is_in_table) {
 300       // reinsert into table
 301       initial_gvn()->hash_find_insert(use);
 302     }
 303     record_for_igvn(use);
 304     i -= uses_found;    // we deleted 1 or more copies of this edge
 305   }
 306 }
 307 
 308 
 309 static inline bool not_a_node(const Node* n) {
 310   if (n == NULL)                   return true;
 311   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 312   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 313   return false;
 314 }
 315 
 316 // Identify all nodes that are reachable from below, useful.
 317 // Use breadth-first pass that records state in a Unique_Node_List,
 318 // recursive traversal is slower.
 319 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 320   int estimated_worklist_size = live_nodes();
 321   useful.map( estimated_worklist_size, NULL );  // preallocate space
 322 
 323   // Initialize worklist
 324   if (root() != NULL)     { useful.push(root()); }
 325   // If 'top' is cached, declare it useful to preserve cached node
 326   if( cached_top_node() ) { useful.push(cached_top_node()); }
 327 
 328   // Push all useful nodes onto the list, breadthfirst
 329   for( uint next = 0; next < useful.size(); ++next ) {
 330     assert( next < unique(), "Unique useful nodes < total nodes");
 331     Node *n  = useful.at(next);
 332     uint max = n->len();
 333     for( uint i = 0; i < max; ++i ) {
 334       Node *m = n->in(i);
 335       if (not_a_node(m))  continue;
 336       useful.push(m);
 337     }
 338   }
 339 }
 340 
 341 // Update dead_node_list with any missing dead nodes using useful
 342 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 343 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 344   uint max_idx = unique();
 345   VectorSet& useful_node_set = useful.member_set();
 346 
 347   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 348     // If node with index node_idx is not in useful set,
 349     // mark it as dead in dead node list.
 350     if (! useful_node_set.test(node_idx) ) {
 351       record_dead_node(node_idx);
 352     }
 353   }
 354 }
 355 
 356 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 357   int shift = 0;
 358   for (int i = 0; i < inlines->length(); i++) {
 359     CallGenerator* cg = inlines->at(i);
 360     CallNode* call = cg->call_node();
 361     if (shift > 0) {
 362       inlines->at_put(i-shift, cg);
 363     }
 364     if (!useful.member(call)) {
 365       shift++;
 366     }
 367   }
 368   inlines->trunc_to(inlines->length()-shift);
 369 }
 370 
 371 // Disconnect all useless nodes by disconnecting those at the boundary.
 372 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 373   uint next = 0;
 374   while (next < useful.size()) {
 375     Node *n = useful.at(next++);
 376     if (n->is_SafePoint()) {
 377       // We're done with a parsing phase. Replaced nodes are not valid
 378       // beyond that point.
 379       n->as_SafePoint()->delete_replaced_nodes();
 380     }
 381     // Use raw traversal of out edges since this code removes out edges
 382     int max = n->outcnt();
 383     for (int j = 0; j < max; ++j) {
 384       Node* child = n->raw_out(j);
 385       if (! useful.member(child)) {
 386         assert(!child->is_top() || child != top(),
 387                "If top is cached in Compile object it is in useful list");
 388         // Only need to remove this out-edge to the useless node
 389         n->raw_del_out(j);
 390         --j;
 391         --max;
 392       }
 393     }
 394     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 395       record_for_igvn(n->unique_out());
 396     }
 397 #if INCLUDE_SHENANDOAHGC
 398     // TODO: Move into below eliminate_useless_gc_barriers(..) below
 399     if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
 400       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 401         record_for_igvn(n->fast_out(i));
 402       }
 403     }
 404 #endif
 405   }
 406   // Remove useless macro and predicate opaq nodes
 407   for (int i = C->macro_count()-1; i >= 0; i--) {
 408     Node* n = C->macro_node(i);
 409     if (!useful.member(n)) {
 410       remove_macro_node(n);
 411     }
 412   }
 413   // Remove useless CastII nodes with range check dependency
 414   for (int i = range_check_cast_count() - 1; i >= 0; i--) {
 415     Node* cast = range_check_cast_node(i);
 416     if (!useful.member(cast)) {
 417       remove_range_check_cast(cast);
 418     }
 419   }
 420   // Remove useless expensive nodes
 421   for (int i = C->expensive_count()-1; i >= 0; i--) {
 422     Node* n = C->expensive_node(i);
 423     if (!useful.member(n)) {
 424       remove_expensive_node(n);
 425     }
 426   }
 427   // Remove useless Opaque4 nodes
 428   for (int i = opaque4_count() - 1; i >= 0; i--) {
 429     Node* opaq = opaque4_node(i);
 430     if (!useful.member(opaq)) {
 431       remove_opaque4_node(opaq);
 432     }
 433   }
 434   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 435   bs->eliminate_useless_gc_barriers(useful);
 436   // clean up the late inline lists
 437   remove_useless_late_inlines(&_string_late_inlines, useful);
 438   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 439   remove_useless_late_inlines(&_late_inlines, useful);
 440   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 441 }
 442 
 443 //------------------------------frame_size_in_words-----------------------------
 444 // frame_slots in units of words
 445 int Compile::frame_size_in_words() const {
 446   // shift is 0 in LP32 and 1 in LP64
 447   const int shift = (LogBytesPerWord - LogBytesPerInt);
 448   int words = _frame_slots >> shift;
 449   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 450   return words;
 451 }
 452 
 453 // To bang the stack of this compiled method we use the stack size
 454 // that the interpreter would need in case of a deoptimization. This
 455 // removes the need to bang the stack in the deoptimization blob which
 456 // in turn simplifies stack overflow handling.
 457 int Compile::bang_size_in_bytes() const {
 458   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
 459 }
 460 
 461 // ============================================================================
 462 //------------------------------CompileWrapper---------------------------------
 463 class CompileWrapper : public StackObj {
 464   Compile *const _compile;
 465  public:
 466   CompileWrapper(Compile* compile);
 467 
 468   ~CompileWrapper();
 469 };
 470 
 471 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 472   // the Compile* pointer is stored in the current ciEnv:
 473   ciEnv* env = compile->env();
 474   assert(env == ciEnv::current(), "must already be a ciEnv active");
 475   assert(env->compiler_data() == NULL, "compile already active?");
 476   env->set_compiler_data(compile);
 477   assert(compile == Compile::current(), "sanity");
 478 
 479   compile->set_type_dict(NULL);
 480   compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
 481   compile->clone_map().set_clone_idx(0);
 482   compile->set_type_hwm(NULL);
 483   compile->set_type_last_size(0);
 484   compile->set_last_tf(NULL, NULL);
 485   compile->set_indexSet_arena(NULL);
 486   compile->set_indexSet_free_block_list(NULL);
 487   compile->init_type_arena();
 488   Type::Initialize(compile);
 489   _compile->set_scratch_buffer_blob(NULL);
 490   _compile->begin_method();
 491   _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
 492 }
 493 CompileWrapper::~CompileWrapper() {
 494   _compile->end_method();
 495   if (_compile->scratch_buffer_blob() != NULL)
 496     BufferBlob::free(_compile->scratch_buffer_blob());
 497   _compile->env()->set_compiler_data(NULL);
 498 }
 499 
 500 
 501 //----------------------------print_compile_messages---------------------------
 502 void Compile::print_compile_messages() {
 503 #ifndef PRODUCT
 504   // Check if recompiling
 505   if (_subsume_loads == false && PrintOpto) {
 506     // Recompiling without allowing machine instructions to subsume loads
 507     tty->print_cr("*********************************************************");
 508     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 509     tty->print_cr("*********************************************************");
 510   }
 511   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 512     // Recompiling without escape analysis
 513     tty->print_cr("*********************************************************");
 514     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 515     tty->print_cr("*********************************************************");
 516   }
 517   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 518     // Recompiling without boxing elimination
 519     tty->print_cr("*********************************************************");
 520     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 521     tty->print_cr("*********************************************************");
 522   }
 523   if (C->directive()->BreakAtCompileOption) {
 524     // Open the debugger when compiling this method.
 525     tty->print("### Breaking when compiling: ");
 526     method()->print_short_name();
 527     tty->cr();
 528     BREAKPOINT;
 529   }
 530 
 531   if( PrintOpto ) {
 532     if (is_osr_compilation()) {
 533       tty->print("[OSR]%3d", _compile_id);
 534     } else {
 535       tty->print("%3d", _compile_id);
 536     }
 537   }
 538 #endif
 539 }
 540 
 541 
 542 //-----------------------init_scratch_buffer_blob------------------------------
 543 // Construct a temporary BufferBlob and cache it for this compile.
 544 void Compile::init_scratch_buffer_blob(int const_size) {
 545   // If there is already a scratch buffer blob allocated and the
 546   // constant section is big enough, use it.  Otherwise free the
 547   // current and allocate a new one.
 548   BufferBlob* blob = scratch_buffer_blob();
 549   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 550     // Use the current blob.
 551   } else {
 552     if (blob != NULL) {
 553       BufferBlob::free(blob);
 554     }
 555 
 556     ResourceMark rm;
 557     _scratch_const_size = const_size;
 558     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
 559     blob = BufferBlob::create("Compile::scratch_buffer", size);
 560     // Record the buffer blob for next time.
 561     set_scratch_buffer_blob(blob);
 562     // Have we run out of code space?
 563     if (scratch_buffer_blob() == NULL) {
 564       // Let CompilerBroker disable further compilations.
 565       record_failure("Not enough space for scratch buffer in CodeCache");
 566       return;
 567     }
 568   }
 569 
 570   // Initialize the relocation buffers
 571   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 572   set_scratch_locs_memory(locs_buf);
 573 }
 574 
 575 
 576 //-----------------------scratch_emit_size-------------------------------------
 577 // Helper function that computes size by emitting code
 578 uint Compile::scratch_emit_size(const Node* n) {
 579   // Start scratch_emit_size section.
 580   set_in_scratch_emit_size(true);
 581 
 582   // Emit into a trash buffer and count bytes emitted.
 583   // This is a pretty expensive way to compute a size,
 584   // but it works well enough if seldom used.
 585   // All common fixed-size instructions are given a size
 586   // method by the AD file.
 587   // Note that the scratch buffer blob and locs memory are
 588   // allocated at the beginning of the compile task, and
 589   // may be shared by several calls to scratch_emit_size.
 590   // The allocation of the scratch buffer blob is particularly
 591   // expensive, since it has to grab the code cache lock.
 592   BufferBlob* blob = this->scratch_buffer_blob();
 593   assert(blob != NULL, "Initialize BufferBlob at start");
 594   assert(blob->size() > MAX_inst_size, "sanity");
 595   relocInfo* locs_buf = scratch_locs_memory();
 596   address blob_begin = blob->content_begin();
 597   address blob_end   = (address)locs_buf;
 598   assert(blob->contains(blob_end), "sanity");
 599   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 600   buf.initialize_consts_size(_scratch_const_size);
 601   buf.initialize_stubs_size(MAX_stubs_size);
 602   assert(locs_buf != NULL, "sanity");
 603   int lsize = MAX_locs_size / 3;
 604   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 605   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 606   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 607   // Mark as scratch buffer.
 608   buf.consts()->set_scratch_emit();
 609   buf.insts()->set_scratch_emit();
 610   buf.stubs()->set_scratch_emit();
 611 
 612   // Do the emission.
 613 
 614   Label fakeL; // Fake label for branch instructions.
 615   Label*   saveL = NULL;
 616   uint save_bnum = 0;
 617   bool is_branch = n->is_MachBranch();
 618   if (is_branch) {
 619     MacroAssembler masm(&buf);
 620     masm.bind(fakeL);
 621     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 622     n->as_MachBranch()->label_set(&fakeL, 0);
 623   }
 624   n->emit(buf, this->regalloc());
 625 
 626   // Emitting into the scratch buffer should not fail
 627   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
 628 
 629   if (is_branch) // Restore label.
 630     n->as_MachBranch()->label_set(saveL, save_bnum);
 631 
 632   // End scratch_emit_size section.
 633   set_in_scratch_emit_size(false);
 634 
 635   return buf.insts_size();
 636 }
 637 
 638 
 639 // ============================================================================
 640 //------------------------------Compile standard-------------------------------
 641 debug_only( int Compile::_debug_idx = 100000; )
 642 
 643 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 644 // the continuation bci for on stack replacement.
 645 
 646 
 647 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 648                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
 649                 : Phase(Compiler),
 650                   _env(ci_env),
 651                   _directive(directive),
 652                   _log(ci_env->log()),
 653                   _compile_id(ci_env->compile_id()),
 654                   _save_argument_registers(false),
 655                   _stub_name(NULL),
 656                   _stub_function(NULL),
 657                   _stub_entry_point(NULL),
 658                   _method(target),
 659                   _entry_bci(osr_bci),
 660                   _initial_gvn(NULL),
 661                   _for_igvn(NULL),
 662                   _warm_calls(NULL),
 663                   _subsume_loads(subsume_loads),
 664                   _do_escape_analysis(do_escape_analysis),
 665                   _eliminate_boxing(eliminate_boxing),
 666                   _failure_reason(NULL),
 667                   _code_buffer("Compile::Fill_buffer"),
 668                   _orig_pc_slot(0),
 669                   _orig_pc_slot_offset_in_bytes(0),
 670                   _has_method_handle_invokes(false),
 671                   _mach_constant_base_node(NULL),
 672                   _node_bundling_limit(0),
 673                   _node_bundling_base(NULL),
 674                   _java_calls(0),
 675                   _inner_loops(0),
 676                   _scratch_const_size(-1),
 677                   _in_scratch_emit_size(false),
 678                   _dead_node_list(comp_arena()),
 679                   _dead_node_count(0),
 680 #ifndef PRODUCT
 681                   _trace_opto_output(directive->TraceOptoOutputOption),
 682                   _in_dump_cnt(0),
 683                   _printer(IdealGraphPrinter::printer()),
 684 #endif
 685                   _congraph(NULL),
 686                   _comp_arena(mtCompiler),
 687                   _node_arena(mtCompiler),
 688                   _old_arena(mtCompiler),
 689                   _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
 690                   _Compile_types(mtCompiler),
 691                   _replay_inline_data(NULL),
 692                   _late_inlines(comp_arena(), 2, 0, NULL),
 693                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 694                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 695                   _late_inlines_pos(0),
 696                   _number_of_mh_late_inlines(0),
 697                   _inlining_progress(false),
 698                   _inlining_incrementally(false),
 699                   _print_inlining_list(NULL),
 700                   _print_inlining_stream(NULL),
 701                   _print_inlining_idx(0),
 702                   _print_inlining_output(NULL),
 703                   _interpreter_frame_size(0),
 704                   _max_node_limit(MaxNodeLimit),
 705                   _has_reserved_stack_access(target->has_reserved_stack_access()) {
 706   C = this;
 707 #ifndef PRODUCT
 708   if (_printer != NULL) {
 709     _printer->set_compile(this);
 710   }
 711 #endif
 712   CompileWrapper cw(this);
 713 
 714   if (CITimeVerbose) {
 715     tty->print(" ");
 716     target->holder()->name()->print();
 717     tty->print(".");
 718     target->print_short_name();
 719     tty->print("  ");
 720   }
 721   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 722   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 723 
 724 #ifndef PRODUCT
 725   bool print_opto_assembly = directive->PrintOptoAssemblyOption;
 726   if (!print_opto_assembly) {
 727     bool print_assembly = directive->PrintAssemblyOption;
 728     if (print_assembly && !Disassembler::can_decode()) {
 729       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 730       print_opto_assembly = true;
 731     }
 732   }
 733   set_print_assembly(print_opto_assembly);
 734   set_parsed_irreducible_loop(false);
 735 
 736   if (directive->ReplayInlineOption) {
 737     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 738   }
 739 #endif
 740   set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
 741   set_print_intrinsics(directive->PrintIntrinsicsOption);
 742   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 743 
 744   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 745     // Make sure the method being compiled gets its own MDO,
 746     // so we can at least track the decompile_count().
 747     // Need MDO to record RTM code generation state.
 748     method()->ensure_method_data();
 749   }
 750 
 751   Init(::AliasLevel);
 752 
 753 
 754   print_compile_messages();
 755 
 756   _ilt = InlineTree::build_inline_tree_root();
 757 
 758   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 759   assert(num_alias_types() >= AliasIdxRaw, "");
 760 
 761 #define MINIMUM_NODE_HASH  1023
 762   // Node list that Iterative GVN will start with
 763   Unique_Node_List for_igvn(comp_arena());
 764   set_for_igvn(&for_igvn);
 765 
 766   // GVN that will be run immediately on new nodes
 767   uint estimated_size = method()->code_size()*4+64;
 768   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 769   PhaseGVN gvn(node_arena(), estimated_size);
 770   set_initial_gvn(&gvn);
 771 
 772   print_inlining_init();
 773   { // Scope for timing the parser
 774     TracePhase tp("parse", &timers[_t_parser]);
 775 
 776     // Put top into the hash table ASAP.
 777     initial_gvn()->transform_no_reclaim(top());
 778 
 779     // Set up tf(), start(), and find a CallGenerator.
 780     CallGenerator* cg = NULL;
 781     if (is_osr_compilation()) {
 782       const TypeTuple *domain = StartOSRNode::osr_domain();
 783       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 784       init_tf(TypeFunc::make(domain, range));
 785       StartNode* s = new StartOSRNode(root(), domain);
 786       initial_gvn()->set_type_bottom(s);
 787       init_start(s);
 788       cg = CallGenerator::for_osr(method(), entry_bci());
 789     } else {
 790       // Normal case.
 791       init_tf(TypeFunc::make(method()));
 792       StartNode* s = new StartNode(root(), tf()->domain());
 793       initial_gvn()->set_type_bottom(s);
 794       init_start(s);
 795       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
 796         // With java.lang.ref.reference.get() we must go through the
 797         // intrinsic - even when get() is the root
 798         // method of the compile - so that, if necessary, the value in
 799         // the referent field of the reference object gets recorded by
 800         // the pre-barrier code.
 801         cg = find_intrinsic(method(), false);
 802       }
 803       if (cg == NULL) {
 804         float past_uses = method()->interpreter_invocation_count();
 805         float expected_uses = past_uses;
 806         cg = CallGenerator::for_inline(method(), expected_uses);
 807       }
 808     }
 809     if (failing())  return;
 810     if (cg == NULL) {
 811       record_method_not_compilable("cannot parse method");
 812       return;
 813     }
 814     JVMState* jvms = build_start_state(start(), tf());
 815     if ((jvms = cg->generate(jvms)) == NULL) {
 816       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 817         record_method_not_compilable("method parse failed");
 818       }
 819       return;
 820     }
 821     GraphKit kit(jvms);
 822 
 823     if (!kit.stopped()) {
 824       // Accept return values, and transfer control we know not where.
 825       // This is done by a special, unique ReturnNode bound to root.
 826       return_values(kit.jvms());
 827     }
 828 
 829     if (kit.has_exceptions()) {
 830       // Any exceptions that escape from this call must be rethrown
 831       // to whatever caller is dynamically above us on the stack.
 832       // This is done by a special, unique RethrowNode bound to root.
 833       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 834     }
 835 
 836     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 837 
 838     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 839       inline_string_calls(true);
 840     }
 841 
 842     if (failing())  return;
 843 
 844     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 845 
 846     // Remove clutter produced by parsing.
 847     if (!failing()) {
 848       ResourceMark rm;
 849       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 850     }
 851   }
 852 
 853   // Note:  Large methods are capped off in do_one_bytecode().
 854   if (failing())  return;
 855 
 856   // After parsing, node notes are no longer automagic.
 857   // They must be propagated by register_new_node_with_optimizer(),
 858   // clone(), or the like.
 859   set_default_node_notes(NULL);
 860 
 861   for (;;) {
 862     int successes = Inline_Warm();
 863     if (failing())  return;
 864     if (successes == 0)  break;
 865   }
 866 
 867   // Drain the list.
 868   Finish_Warm();
 869 #ifndef PRODUCT
 870   if (_printer && _printer->should_print(1)) {
 871     _printer->print_inlining();
 872   }
 873 #endif
 874 
 875   if (failing())  return;
 876   NOT_PRODUCT( verify_graph_edges(); )
 877 
 878   // Now optimize
 879   Optimize();
 880   if (failing())  return;
 881   NOT_PRODUCT( verify_graph_edges(); )
 882 
 883 #ifndef PRODUCT
 884   if (PrintIdeal) {
 885     ttyLocker ttyl;  // keep the following output all in one block
 886     // This output goes directly to the tty, not the compiler log.
 887     // To enable tools to match it up with the compilation activity,
 888     // be sure to tag this tty output with the compile ID.
 889     if (xtty != NULL) {
 890       xtty->head("ideal compile_id='%d'%s", compile_id(),
 891                  is_osr_compilation()    ? " compile_kind='osr'" :
 892                  "");
 893     }
 894     root()->dump(9999);
 895     if (xtty != NULL) {
 896       xtty->tail("ideal");
 897     }
 898   }
 899 #endif
 900 
 901   NOT_PRODUCT( verify_barriers(); )
 902 
 903   // Dump compilation data to replay it.
 904   if (directive->DumpReplayOption) {
 905     env()->dump_replay_data(_compile_id);
 906   }
 907   if (directive->DumpInlineOption && (ilt() != NULL)) {
 908     env()->dump_inline_data(_compile_id);
 909   }
 910 
 911   // Now that we know the size of all the monitors we can add a fixed slot
 912   // for the original deopt pc.
 913 
 914   _orig_pc_slot =  fixed_slots();
 915   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 916   set_fixed_slots(next_slot);
 917 
 918   // Compute when to use implicit null checks. Used by matching trap based
 919   // nodes and NullCheck optimization.
 920   set_allowed_deopt_reasons();
 921 
 922   // Now generate code
 923   Code_Gen();
 924   if (failing())  return;
 925 
 926   // Check if we want to skip execution of all compiled code.
 927   {
 928 #ifndef PRODUCT
 929     if (OptoNoExecute) {
 930       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 931       return;
 932     }
 933 #endif
 934     TracePhase tp("install_code", &timers[_t_registerMethod]);
 935 
 936     if (is_osr_compilation()) {
 937       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 938       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 939     } else {
 940       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 941       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 942     }
 943 
 944     env()->register_method(_method, _entry_bci,
 945                            &_code_offsets,
 946                            _orig_pc_slot_offset_in_bytes,
 947                            code_buffer(),
 948                            frame_size_in_words(), _oop_map_set,
 949                            &_handler_table, &_inc_table,
 950                            compiler,
 951                            has_unsafe_access(),
 952                            SharedRuntime::is_wide_vector(max_vector_size()),
 953                            rtm_state()
 954                            );
 955 
 956     if (log() != NULL) // Print code cache state into compiler log
 957       log()->code_cache_state();
 958   }
 959 }
 960 
 961 //------------------------------Compile----------------------------------------
 962 // Compile a runtime stub
 963 Compile::Compile( ciEnv* ci_env,
 964                   TypeFunc_generator generator,
 965                   address stub_function,
 966                   const char *stub_name,
 967                   int is_fancy_jump,
 968                   bool pass_tls,
 969                   bool save_arg_registers,
 970                   bool return_pc,
 971                   DirectiveSet* directive)
 972   : Phase(Compiler),
 973     _env(ci_env),
 974     _directive(directive),
 975     _log(ci_env->log()),
 976     _compile_id(0),
 977     _save_argument_registers(save_arg_registers),
 978     _method(NULL),
 979     _stub_name(stub_name),
 980     _stub_function(stub_function),
 981     _stub_entry_point(NULL),
 982     _entry_bci(InvocationEntryBci),
 983     _initial_gvn(NULL),
 984     _for_igvn(NULL),
 985     _warm_calls(NULL),
 986     _orig_pc_slot(0),
 987     _orig_pc_slot_offset_in_bytes(0),
 988     _subsume_loads(true),
 989     _do_escape_analysis(false),
 990     _eliminate_boxing(false),
 991     _failure_reason(NULL),
 992     _code_buffer("Compile::Fill_buffer"),
 993     _has_method_handle_invokes(false),
 994     _mach_constant_base_node(NULL),
 995     _node_bundling_limit(0),
 996     _node_bundling_base(NULL),
 997     _java_calls(0),
 998     _inner_loops(0),
 999 #ifndef PRODUCT
1000     _trace_opto_output(directive->TraceOptoOutputOption),
1001     _in_dump_cnt(0),
1002     _printer(NULL),
1003 #endif
1004     _comp_arena(mtCompiler),
1005     _node_arena(mtCompiler),
1006     _old_arena(mtCompiler),
1007     _Compile_types(mtCompiler),
1008     _dead_node_list(comp_arena()),
1009     _dead_node_count(0),
1010     _congraph(NULL),
1011     _replay_inline_data(NULL),
1012     _number_of_mh_late_inlines(0),
1013     _inlining_progress(false),
1014     _inlining_incrementally(false),
1015     _print_inlining_list(NULL),
1016     _print_inlining_stream(NULL),
1017     _print_inlining_idx(0),
1018     _print_inlining_output(NULL),
1019     _allowed_reasons(0),
1020     _interpreter_frame_size(0),
1021     _max_node_limit(MaxNodeLimit),
1022     _has_reserved_stack_access(false) {
1023   C = this;
1024 
1025   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1026   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1027 
1028 #ifndef PRODUCT
1029   set_print_assembly(PrintFrameConverterAssembly);
1030   set_parsed_irreducible_loop(false);
1031 #endif
1032   set_has_irreducible_loop(false); // no loops
1033 
1034   CompileWrapper cw(this);
1035   Init(/*AliasLevel=*/ 0);
1036   init_tf((*generator)());
1037 
1038   {
1039     // The following is a dummy for the sake of GraphKit::gen_stub
1040     Unique_Node_List for_igvn(comp_arena());
1041     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1042     PhaseGVN gvn(Thread::current()->resource_area(),255);
1043     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1044     gvn.transform_no_reclaim(top());
1045 
1046     GraphKit kit;
1047     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1048   }
1049 
1050   NOT_PRODUCT( verify_graph_edges(); )
1051   Code_Gen();
1052   if (failing())  return;
1053 
1054 
1055   // Entry point will be accessed using compile->stub_entry_point();
1056   if (code_buffer() == NULL) {
1057     Matcher::soft_match_failure();
1058   } else {
1059     if (PrintAssembly && (WizardMode || Verbose))
1060       tty->print_cr("### Stub::%s", stub_name);
1061 
1062     if (!failing()) {
1063       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1064 
1065       // Make the NMethod
1066       // For now we mark the frame as never safe for profile stackwalking
1067       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1068                                                       code_buffer(),
1069                                                       CodeOffsets::frame_never_safe,
1070                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1071                                                       frame_size_in_words(),
1072                                                       _oop_map_set,
1073                                                       save_arg_registers);
1074       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1075 
1076       _stub_entry_point = rs->entry_point();
1077     }
1078   }
1079 }
1080 
1081 //------------------------------Init-------------------------------------------
1082 // Prepare for a single compilation
1083 void Compile::Init(int aliaslevel) {
1084   _unique  = 0;
1085   _regalloc = NULL;
1086 
1087   _tf      = NULL;  // filled in later
1088   _top     = NULL;  // cached later
1089   _matcher = NULL;  // filled in later
1090   _cfg     = NULL;  // filled in later
1091 
1092   set_24_bit_selection_and_mode(Use24BitFP, false);
1093 
1094   _node_note_array = NULL;
1095   _default_node_notes = NULL;
1096   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1097 
1098   _immutable_memory = NULL; // filled in at first inquiry
1099 
1100   // Globally visible Nodes
1101   // First set TOP to NULL to give safe behavior during creation of RootNode
1102   set_cached_top_node(NULL);
1103   set_root(new RootNode());
1104   // Now that you have a Root to point to, create the real TOP
1105   set_cached_top_node( new ConNode(Type::TOP) );
1106   set_recent_alloc(NULL, NULL);
1107 
1108   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1109   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1110   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1111   env()->set_dependencies(new Dependencies(env()));
1112 
1113   _fixed_slots = 0;
1114   set_has_split_ifs(false);
1115   set_has_loops(has_method() && method()->has_loops()); // first approximation
1116   set_has_stringbuilder(false);
1117   set_has_boxed_value(false);
1118   _trap_can_recompile = false;  // no traps emitted yet
1119   _major_progress = true; // start out assuming good things will happen
1120   set_has_unsafe_access(false);
1121   set_max_vector_size(0);
1122   set_clear_upper_avx(false);  //false as default for clear upper bits of ymm registers
1123   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1124   set_decompile_count(0);
1125 
1126   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1127   set_num_loop_opts(LoopOptsCount);
1128   set_do_inlining(Inline);
1129   set_max_inline_size(MaxInlineSize);
1130   set_freq_inline_size(FreqInlineSize);
1131   set_do_scheduling(OptoScheduling);
1132   set_do_count_invocations(false);
1133   set_do_method_data_update(false);
1134 
1135   set_do_vector_loop(false);
1136 
1137   if (AllowVectorizeOnDemand) {
1138     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1139       set_do_vector_loop(true);
1140       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());})
1141     } else if (has_method() && method()->name() != 0 &&
1142                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1143       set_do_vector_loop(true);
1144     }
1145   }
1146   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1147   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());})
1148 
1149   set_age_code(has_method() && method()->profile_aging());
1150   set_rtm_state(NoRTM); // No RTM lock eliding by default
1151   _max_node_limit = _directive->MaxNodeLimitOption;
1152 
1153 #if INCLUDE_RTM_OPT
1154   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1155     int rtm_state = method()->method_data()->rtm_state();
1156     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1157       // Don't generate RTM lock eliding code.
1158       set_rtm_state(NoRTM);
1159     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1160       // Generate RTM lock eliding code without abort ratio calculation code.
1161       set_rtm_state(UseRTM);
1162     } else if (UseRTMDeopt) {
1163       // Generate RTM lock eliding code and include abort ratio calculation
1164       // code if UseRTMDeopt is on.
1165       set_rtm_state(ProfileRTM);
1166     }
1167   }
1168 #endif
1169   if (debug_info()->recording_non_safepoints()) {
1170     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1171                         (comp_arena(), 8, 0, NULL));
1172     set_default_node_notes(Node_Notes::make(this));
1173   }
1174 
1175   // // -- Initialize types before each compile --
1176   // // Update cached type information
1177   // if( _method && _method->constants() )
1178   //   Type::update_loaded_types(_method, _method->constants());
1179 
1180   // Init alias_type map.
1181   if (!_do_escape_analysis && aliaslevel == 3)
1182     aliaslevel = 2;  // No unique types without escape analysis
1183   _AliasLevel = aliaslevel;
1184   const int grow_ats = 16;
1185   _max_alias_types = grow_ats;
1186   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1187   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1188   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1189   {
1190     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1191   }
1192   // Initialize the first few types.
1193   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1194   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1195   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1196   _num_alias_types = AliasIdxRaw+1;
1197   // Zero out the alias type cache.
1198   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1199   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1200   probe_alias_cache(NULL)->_index = AliasIdxTop;
1201 
1202   _intrinsics = NULL;
1203   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1204   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1205   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1206   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1207   _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1208   register_library_intrinsics();
1209 }
1210 
1211 //---------------------------init_start----------------------------------------
1212 // Install the StartNode on this compile object.
1213 void Compile::init_start(StartNode* s) {
1214   if (failing())
1215     return; // already failing
1216   assert(s == start(), "");
1217 }
1218 
1219 /**
1220  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1221  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1222  * the ideal graph.
1223  */
1224 StartNode* Compile::start() const {
1225   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1226   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1227     Node* start = root()->fast_out(i);
1228     if (start->is_Start()) {
1229       return start->as_Start();
1230     }
1231   }
1232   fatal("Did not find Start node!");
1233   return NULL;
1234 }
1235 
1236 //-------------------------------immutable_memory-------------------------------------
1237 // Access immutable memory
1238 Node* Compile::immutable_memory() {
1239   if (_immutable_memory != NULL) {
1240     return _immutable_memory;
1241   }
1242   StartNode* s = start();
1243   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1244     Node *p = s->fast_out(i);
1245     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1246       _immutable_memory = p;
1247       return _immutable_memory;
1248     }
1249   }
1250   ShouldNotReachHere();
1251   return NULL;
1252 }
1253 
1254 //----------------------set_cached_top_node------------------------------------
1255 // Install the cached top node, and make sure Node::is_top works correctly.
1256 void Compile::set_cached_top_node(Node* tn) {
1257   if (tn != NULL)  verify_top(tn);
1258   Node* old_top = _top;
1259   _top = tn;
1260   // Calling Node::setup_is_top allows the nodes the chance to adjust
1261   // their _out arrays.
1262   if (_top != NULL)     _top->setup_is_top();
1263   if (old_top != NULL)  old_top->setup_is_top();
1264   assert(_top == NULL || top()->is_top(), "");
1265 }
1266 
1267 #ifdef ASSERT
1268 uint Compile::count_live_nodes_by_graph_walk() {
1269   Unique_Node_List useful(comp_arena());
1270   // Get useful node list by walking the graph.
1271   identify_useful_nodes(useful);
1272   return useful.size();
1273 }
1274 
1275 void Compile::print_missing_nodes() {
1276 
1277   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1278   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1279     return;
1280   }
1281 
1282   // This is an expensive function. It is executed only when the user
1283   // specifies VerifyIdealNodeCount option or otherwise knows the
1284   // additional work that needs to be done to identify reachable nodes
1285   // by walking the flow graph and find the missing ones using
1286   // _dead_node_list.
1287 
1288   Unique_Node_List useful(comp_arena());
1289   // Get useful node list by walking the graph.
1290   identify_useful_nodes(useful);
1291 
1292   uint l_nodes = C->live_nodes();
1293   uint l_nodes_by_walk = useful.size();
1294 
1295   if (l_nodes != l_nodes_by_walk) {
1296     if (_log != NULL) {
1297       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1298       _log->stamp();
1299       _log->end_head();
1300     }
1301     VectorSet& useful_member_set = useful.member_set();
1302     int last_idx = l_nodes_by_walk;
1303     for (int i = 0; i < last_idx; i++) {
1304       if (useful_member_set.test(i)) {
1305         if (_dead_node_list.test(i)) {
1306           if (_log != NULL) {
1307             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1308           }
1309           if (PrintIdealNodeCount) {
1310             // Print the log message to tty
1311               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1312               useful.at(i)->dump();
1313           }
1314         }
1315       }
1316       else if (! _dead_node_list.test(i)) {
1317         if (_log != NULL) {
1318           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1319         }
1320         if (PrintIdealNodeCount) {
1321           // Print the log message to tty
1322           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1323         }
1324       }
1325     }
1326     if (_log != NULL) {
1327       _log->tail("mismatched_nodes");
1328     }
1329   }
1330 }
1331 void Compile::record_modified_node(Node* n) {
1332   if (_modified_nodes != NULL && !_inlining_incrementally &&
1333       n->outcnt() != 0 && !n->is_Con()) {
1334     _modified_nodes->push(n);
1335   }
1336 }
1337 
1338 void Compile::remove_modified_node(Node* n) {
1339   if (_modified_nodes != NULL) {
1340     _modified_nodes->remove(n);
1341   }
1342 }
1343 #endif
1344 
1345 #ifndef PRODUCT
1346 void Compile::verify_top(Node* tn) const {
1347   if (tn != NULL) {
1348     assert(tn->is_Con(), "top node must be a constant");
1349     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1350     assert(tn->in(0) != NULL, "must have live top node");
1351   }
1352 }
1353 #endif
1354 
1355 
1356 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1357 
1358 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1359   guarantee(arr != NULL, "");
1360   int num_blocks = arr->length();
1361   if (grow_by < num_blocks)  grow_by = num_blocks;
1362   int num_notes = grow_by * _node_notes_block_size;
1363   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1364   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1365   while (num_notes > 0) {
1366     arr->append(notes);
1367     notes     += _node_notes_block_size;
1368     num_notes -= _node_notes_block_size;
1369   }
1370   assert(num_notes == 0, "exact multiple, please");
1371 }
1372 
1373 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1374   if (source == NULL || dest == NULL)  return false;
1375 
1376   if (dest->is_Con())
1377     return false;               // Do not push debug info onto constants.
1378 
1379 #ifdef ASSERT
1380   // Leave a bread crumb trail pointing to the original node:
1381   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1382     dest->set_debug_orig(source);
1383   }
1384 #endif
1385 
1386   if (node_note_array() == NULL)
1387     return false;               // Not collecting any notes now.
1388 
1389   // This is a copy onto a pre-existing node, which may already have notes.
1390   // If both nodes have notes, do not overwrite any pre-existing notes.
1391   Node_Notes* source_notes = node_notes_at(source->_idx);
1392   if (source_notes == NULL || source_notes->is_clear())  return false;
1393   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1394   if (dest_notes == NULL || dest_notes->is_clear()) {
1395     return set_node_notes_at(dest->_idx, source_notes);
1396   }
1397 
1398   Node_Notes merged_notes = (*source_notes);
1399   // The order of operations here ensures that dest notes will win...
1400   merged_notes.update_from(dest_notes);
1401   return set_node_notes_at(dest->_idx, &merged_notes);
1402 }
1403 
1404 
1405 //--------------------------allow_range_check_smearing-------------------------
1406 // Gating condition for coalescing similar range checks.
1407 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1408 // single covering check that is at least as strong as any of them.
1409 // If the optimization succeeds, the simplified (strengthened) range check
1410 // will always succeed.  If it fails, we will deopt, and then give up
1411 // on the optimization.
1412 bool Compile::allow_range_check_smearing() const {
1413   // If this method has already thrown a range-check,
1414   // assume it was because we already tried range smearing
1415   // and it failed.
1416   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1417   return !already_trapped;
1418 }
1419 
1420 
1421 //------------------------------flatten_alias_type-----------------------------
1422 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1423   int offset = tj->offset();
1424   TypePtr::PTR ptr = tj->ptr();
1425 
1426   // Known instance (scalarizable allocation) alias only with itself.
1427   bool is_known_inst = tj->isa_oopptr() != NULL &&
1428                        tj->is_oopptr()->is_known_instance();
1429 
1430   // Process weird unsafe references.
1431   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1432     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1433     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1434     tj = TypeOopPtr::BOTTOM;
1435     ptr = tj->ptr();
1436     offset = tj->offset();
1437   }
1438 
1439   // Array pointers need some flattening
1440   const TypeAryPtr *ta = tj->isa_aryptr();
1441   if (ta && ta->is_stable()) {
1442     // Erase stability property for alias analysis.
1443     tj = ta = ta->cast_to_stable(false);
1444   }
1445   if( ta && is_known_inst ) {
1446     if ( offset != Type::OffsetBot &&
1447          offset > arrayOopDesc::length_offset_in_bytes() ) {
1448       offset = Type::OffsetBot; // Flatten constant access into array body only
1449       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1450     }
1451   } else if( ta && _AliasLevel >= 2 ) {
1452     // For arrays indexed by constant indices, we flatten the alias
1453     // space to include all of the array body.  Only the header, klass
1454     // and array length can be accessed un-aliased.
1455     if( offset != Type::OffsetBot ) {
1456       if( ta->const_oop() ) { // MethodData* or Method*
1457         offset = Type::OffsetBot;   // Flatten constant access into array body
1458         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1459       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1460         // range is OK as-is.
1461         tj = ta = TypeAryPtr::RANGE;
1462       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1463         tj = TypeInstPtr::KLASS; // all klass loads look alike
1464         ta = TypeAryPtr::RANGE; // generic ignored junk
1465         ptr = TypePtr::BotPTR;
1466       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1467         tj = TypeInstPtr::MARK;
1468         ta = TypeAryPtr::RANGE; // generic ignored junk
1469         ptr = TypePtr::BotPTR;
1470 #if INCLUDE_SHENANDOAHGC
1471       } else if (offset == ShenandoahBrooksPointer::byte_offset() && UseShenandoahGC) {
1472         // Need to distinguish brooks ptr as is.
1473         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1474 #endif
1475       } else {                  // Random constant offset into array body
1476         offset = Type::OffsetBot;   // Flatten constant access into array body
1477         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1478       }
1479     }
1480     // Arrays of fixed size alias with arrays of unknown size.
1481     if (ta->size() != TypeInt::POS) {
1482       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1483       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1484     }
1485     // Arrays of known objects become arrays of unknown objects.
1486     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1487       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1488       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1489     }
1490     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1491       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1492       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1493     }
1494     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1495     // cannot be distinguished by bytecode alone.
1496     if (ta->elem() == TypeInt::BOOL) {
1497       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1498       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1499       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1500     }
1501     // During the 2nd round of IterGVN, NotNull castings are removed.
1502     // Make sure the Bottom and NotNull variants alias the same.
1503     // Also, make sure exact and non-exact variants alias the same.
1504     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1505       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1506     }
1507   }
1508 
1509   // Oop pointers need some flattening
1510   const TypeInstPtr *to = tj->isa_instptr();
1511   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1512     ciInstanceKlass *k = to->klass()->as_instance_klass();
1513     if( ptr == TypePtr::Constant ) {
1514       if (to->klass() != ciEnv::current()->Class_klass() ||
1515           offset < k->size_helper() * wordSize) {
1516         // No constant oop pointers (such as Strings); they alias with
1517         // unknown strings.
1518         assert(!is_known_inst, "not scalarizable allocation");
1519         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1520       }
1521     } else if( is_known_inst ) {
1522       tj = to; // Keep NotNull and klass_is_exact for instance type
1523     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1524       // During the 2nd round of IterGVN, NotNull castings are removed.
1525       // Make sure the Bottom and NotNull variants alias the same.
1526       // Also, make sure exact and non-exact variants alias the same.
1527       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1528     }
1529     if (to->speculative() != NULL) {
1530       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1531     }
1532     // Canonicalize the holder of this field
1533     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1534       // First handle header references such as a LoadKlassNode, even if the
1535       // object's klass is unloaded at compile time (4965979).
1536       if (!is_known_inst) { // Do it only for non-instance types
1537         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1538       }
1539     } else if (SHENANDOAHGC_ONLY((offset != ShenandoahBrooksPointer::byte_offset() || !UseShenandoahGC) &&) (offset < 0 || offset >= k->size_helper() * wordSize)) {
1540       // Static fields are in the space above the normal instance
1541       // fields in the java.lang.Class instance.
1542       if (to->klass() != ciEnv::current()->Class_klass()) {
1543         to = NULL;
1544         tj = TypeOopPtr::BOTTOM;
1545         offset = tj->offset();
1546       }
1547     } else {
1548       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1549       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1550         if( is_known_inst ) {
1551           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1552         } else {
1553           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1554         }
1555       }
1556     }
1557   }
1558 
1559   // Klass pointers to object array klasses need some flattening
1560   const TypeKlassPtr *tk = tj->isa_klassptr();
1561   if( tk ) {
1562     // If we are referencing a field within a Klass, we need
1563     // to assume the worst case of an Object.  Both exact and
1564     // inexact types must flatten to the same alias class so
1565     // use NotNull as the PTR.
1566     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1567 
1568       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1569                                    TypeKlassPtr::OBJECT->klass(),
1570                                    offset);
1571     }
1572 
1573     ciKlass* klass = tk->klass();
1574     if( klass->is_obj_array_klass() ) {
1575       ciKlass* k = TypeAryPtr::OOPS->klass();
1576       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1577         k = TypeInstPtr::BOTTOM->klass();
1578       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1579     }
1580 
1581     // Check for precise loads from the primary supertype array and force them
1582     // to the supertype cache alias index.  Check for generic array loads from
1583     // the primary supertype array and also force them to the supertype cache
1584     // alias index.  Since the same load can reach both, we need to merge
1585     // these 2 disparate memories into the same alias class.  Since the
1586     // primary supertype array is read-only, there's no chance of confusion
1587     // where we bypass an array load and an array store.
1588     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1589     if (offset == Type::OffsetBot ||
1590         (offset >= primary_supers_offset &&
1591          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1592         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1593       offset = in_bytes(Klass::secondary_super_cache_offset());
1594       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1595     }
1596   }
1597 
1598   // Flatten all Raw pointers together.
1599   if (tj->base() == Type::RawPtr)
1600     tj = TypeRawPtr::BOTTOM;
1601 
1602   if (tj->base() == Type::AnyPtr)
1603     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1604 
1605   // Flatten all to bottom for now
1606   switch( _AliasLevel ) {
1607   case 0:
1608     tj = TypePtr::BOTTOM;
1609     break;
1610   case 1:                       // Flatten to: oop, static, field or array
1611     switch (tj->base()) {
1612     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1613     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1614     case Type::AryPtr:   // do not distinguish arrays at all
1615     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1616     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1617     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1618     default: ShouldNotReachHere();
1619     }
1620     break;
1621   case 2:                       // No collapsing at level 2; keep all splits
1622   case 3:                       // No collapsing at level 3; keep all splits
1623     break;
1624   default:
1625     Unimplemented();
1626   }
1627 
1628   offset = tj->offset();
1629   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1630 
1631   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1632           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1633           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1634           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1635           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1636           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1637           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1638           (UseShenandoahGC SHENANDOAHGC_ONLY(&& offset == ShenandoahBrooksPointer::byte_offset() && tj->base() == Type::AryPtr)),
1639           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1640   assert( tj->ptr() != TypePtr::TopPTR &&
1641           tj->ptr() != TypePtr::AnyNull &&
1642           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1643 //    assert( tj->ptr() != TypePtr::Constant ||
1644 //            tj->base() == Type::RawPtr ||
1645 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1646 
1647   return tj;
1648 }
1649 
1650 void Compile::AliasType::Init(int i, const TypePtr* at) {
1651   _index = i;
1652   _adr_type = at;
1653   _field = NULL;
1654   _element = NULL;
1655   _is_rewritable = true; // default
1656   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1657   if (atoop != NULL && atoop->is_known_instance()) {
1658     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1659     _general_index = Compile::current()->get_alias_index(gt);
1660   } else {
1661     _general_index = 0;
1662   }
1663 }
1664 
1665 BasicType Compile::AliasType::basic_type() const {
1666   if (element() != NULL) {
1667     const Type* element = adr_type()->is_aryptr()->elem();
1668     return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1669   } if (field() != NULL) {
1670     return field()->layout_type();
1671   } else {
1672     return T_ILLEGAL; // unknown
1673   }
1674 }
1675 
1676 //---------------------------------print_on------------------------------------
1677 #ifndef PRODUCT
1678 void Compile::AliasType::print_on(outputStream* st) {
1679   if (index() < 10)
1680         st->print("@ <%d> ", index());
1681   else  st->print("@ <%d>",  index());
1682   st->print(is_rewritable() ? "   " : " RO");
1683   int offset = adr_type()->offset();
1684   if (offset == Type::OffsetBot)
1685         st->print(" +any");
1686   else  st->print(" +%-3d", offset);
1687   st->print(" in ");
1688   adr_type()->dump_on(st);
1689   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1690   if (field() != NULL && tjp) {
1691     if (tjp->klass()  != field()->holder() ||
1692         tjp->offset() != field()->offset_in_bytes()) {
1693       st->print(" != ");
1694       field()->print();
1695       st->print(" ***");
1696     }
1697   }
1698 }
1699 
1700 void print_alias_types() {
1701   Compile* C = Compile::current();
1702   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1703   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1704     C->alias_type(idx)->print_on(tty);
1705     tty->cr();
1706   }
1707 }
1708 #endif
1709 
1710 
1711 //----------------------------probe_alias_cache--------------------------------
1712 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1713   intptr_t key = (intptr_t) adr_type;
1714   key ^= key >> logAliasCacheSize;
1715   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1716 }
1717 
1718 
1719 //-----------------------------grow_alias_types--------------------------------
1720 void Compile::grow_alias_types() {
1721   const int old_ats  = _max_alias_types; // how many before?
1722   const int new_ats  = old_ats;          // how many more?
1723   const int grow_ats = old_ats+new_ats;  // how many now?
1724   _max_alias_types = grow_ats;
1725   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1726   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1727   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1728   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1729 }
1730 
1731 
1732 //--------------------------------find_alias_type------------------------------
1733 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1734   if (_AliasLevel == 0)
1735     return alias_type(AliasIdxBot);
1736 
1737   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1738   if (ace->_adr_type == adr_type) {
1739     return alias_type(ace->_index);
1740   }
1741 
1742   // Handle special cases.
1743   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1744   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1745 
1746   // Do it the slow way.
1747   const TypePtr* flat = flatten_alias_type(adr_type);
1748 
1749 #ifdef ASSERT
1750   {
1751     ResourceMark rm;
1752     assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1753            Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1754     assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1755            Type::str(adr_type));
1756     if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1757       const TypeOopPtr* foop = flat->is_oopptr();
1758       // Scalarizable allocations have exact klass always.
1759       bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1760       const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1761       assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1762              Type::str(foop), Type::str(xoop));
1763     }
1764   }
1765 #endif
1766 
1767   int idx = AliasIdxTop;
1768   for (int i = 0; i < num_alias_types(); i++) {
1769     if (alias_type(i)->adr_type() == flat) {
1770       idx = i;
1771       break;
1772     }
1773   }
1774 
1775   if (idx == AliasIdxTop) {
1776     if (no_create)  return NULL;
1777     // Grow the array if necessary.
1778     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1779     // Add a new alias type.
1780     idx = _num_alias_types++;
1781     _alias_types[idx]->Init(idx, flat);
1782     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1783     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1784     if (flat->isa_instptr()) {
1785       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1786           && flat->is_instptr()->klass() == env()->Class_klass())
1787         alias_type(idx)->set_rewritable(false);
1788     }
1789     if (flat->isa_aryptr()) {
1790 #ifdef ASSERT
1791       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1792       // (T_BYTE has the weakest alignment and size restrictions...)
1793       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1794 #endif
1795       if (flat->offset() == TypePtr::OffsetBot) {
1796         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1797       }
1798     }
1799     if (flat->isa_klassptr()) {
1800       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1801         alias_type(idx)->set_rewritable(false);
1802       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1803         alias_type(idx)->set_rewritable(false);
1804       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1805         alias_type(idx)->set_rewritable(false);
1806       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1807         alias_type(idx)->set_rewritable(false);
1808     }
1809     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1810     // but the base pointer type is not distinctive enough to identify
1811     // references into JavaThread.)
1812 
1813     // Check for final fields.
1814     const TypeInstPtr* tinst = flat->isa_instptr();
1815     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1816       ciField* field;
1817       if (tinst->const_oop() != NULL &&
1818           tinst->klass() == ciEnv::current()->Class_klass() &&
1819           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1820         // static field
1821         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1822         field = k->get_field_by_offset(tinst->offset(), true);
1823       } else {
1824         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1825         field = k->get_field_by_offset(tinst->offset(), false);
1826       }
1827       assert(field == NULL ||
1828              original_field == NULL ||
1829              (field->holder() == original_field->holder() &&
1830               field->offset() == original_field->offset() &&
1831               field->is_static() == original_field->is_static()), "wrong field?");
1832       // Set field() and is_rewritable() attributes.
1833       if (field != NULL)  alias_type(idx)->set_field(field);
1834     }
1835   }
1836 
1837   // Fill the cache for next time.
1838   ace->_adr_type = adr_type;
1839   ace->_index    = idx;
1840   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1841 
1842   // Might as well try to fill the cache for the flattened version, too.
1843   AliasCacheEntry* face = probe_alias_cache(flat);
1844   if (face->_adr_type == NULL) {
1845     face->_adr_type = flat;
1846     face->_index    = idx;
1847     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1848   }
1849 
1850   return alias_type(idx);
1851 }
1852 
1853 
1854 Compile::AliasType* Compile::alias_type(ciField* field) {
1855   const TypeOopPtr* t;
1856   if (field->is_static())
1857     t = TypeInstPtr::make(field->holder()->java_mirror());
1858   else
1859     t = TypeOopPtr::make_from_klass_raw(field->holder());
1860   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1861   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1862   return atp;
1863 }
1864 
1865 
1866 //------------------------------have_alias_type--------------------------------
1867 bool Compile::have_alias_type(const TypePtr* adr_type) {
1868   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1869   if (ace->_adr_type == adr_type) {
1870     return true;
1871   }
1872 
1873   // Handle special cases.
1874   if (adr_type == NULL)             return true;
1875   if (adr_type == TypePtr::BOTTOM)  return true;
1876 
1877   return find_alias_type(adr_type, true, NULL) != NULL;
1878 }
1879 
1880 //-----------------------------must_alias--------------------------------------
1881 // True if all values of the given address type are in the given alias category.
1882 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1883   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1884   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1885   if (alias_idx == AliasIdxTop)         return false; // the empty category
1886   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1887 
1888   // the only remaining possible overlap is identity
1889   int adr_idx = get_alias_index(adr_type);
1890   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1891   assert(adr_idx == alias_idx ||
1892          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1893           && adr_type                       != TypeOopPtr::BOTTOM),
1894          "should not be testing for overlap with an unsafe pointer");
1895   return adr_idx == alias_idx;
1896 }
1897 
1898 //------------------------------can_alias--------------------------------------
1899 // True if any values of the given address type are in the given alias category.
1900 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1901   if (alias_idx == AliasIdxTop)         return false; // the empty category
1902   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1903   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1904   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1905 
1906   // the only remaining possible overlap is identity
1907   int adr_idx = get_alias_index(adr_type);
1908   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1909   return adr_idx == alias_idx;
1910 }
1911 
1912 
1913 
1914 //---------------------------pop_warm_call-------------------------------------
1915 WarmCallInfo* Compile::pop_warm_call() {
1916   WarmCallInfo* wci = _warm_calls;
1917   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1918   return wci;
1919 }
1920 
1921 //----------------------------Inline_Warm--------------------------------------
1922 int Compile::Inline_Warm() {
1923   // If there is room, try to inline some more warm call sites.
1924   // %%% Do a graph index compaction pass when we think we're out of space?
1925   if (!InlineWarmCalls)  return 0;
1926 
1927   int calls_made_hot = 0;
1928   int room_to_grow   = NodeCountInliningCutoff - unique();
1929   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1930   int amount_grown   = 0;
1931   WarmCallInfo* call;
1932   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1933     int est_size = (int)call->size();
1934     if (est_size > (room_to_grow - amount_grown)) {
1935       // This one won't fit anyway.  Get rid of it.
1936       call->make_cold();
1937       continue;
1938     }
1939     call->make_hot();
1940     calls_made_hot++;
1941     amount_grown   += est_size;
1942     amount_to_grow -= est_size;
1943   }
1944 
1945   if (calls_made_hot > 0)  set_major_progress();
1946   return calls_made_hot;
1947 }
1948 
1949 
1950 //----------------------------Finish_Warm--------------------------------------
1951 void Compile::Finish_Warm() {
1952   if (!InlineWarmCalls)  return;
1953   if (failing())  return;
1954   if (warm_calls() == NULL)  return;
1955 
1956   // Clean up loose ends, if we are out of space for inlining.
1957   WarmCallInfo* call;
1958   while ((call = pop_warm_call()) != NULL) {
1959     call->make_cold();
1960   }
1961 }
1962 
1963 //---------------------cleanup_loop_predicates-----------------------
1964 // Remove the opaque nodes that protect the predicates so that all unused
1965 // checks and uncommon_traps will be eliminated from the ideal graph
1966 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1967   if (predicate_count()==0) return;
1968   for (int i = predicate_count(); i > 0; i--) {
1969     Node * n = predicate_opaque1_node(i-1);
1970     assert(n->Opcode() == Op_Opaque1, "must be");
1971     igvn.replace_node(n, n->in(1));
1972   }
1973   assert(predicate_count()==0, "should be clean!");
1974 }
1975 
1976 void Compile::add_range_check_cast(Node* n) {
1977   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1978   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1979   _range_check_casts->append(n);
1980 }
1981 
1982 // Remove all range check dependent CastIINodes.
1983 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1984   for (int i = range_check_cast_count(); i > 0; i--) {
1985     Node* cast = range_check_cast_node(i-1);
1986     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1987     igvn.replace_node(cast, cast->in(1));
1988   }
1989   assert(range_check_cast_count() == 0, "should be empty");
1990 }
1991 
1992 void Compile::add_opaque4_node(Node* n) {
1993   assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
1994   assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
1995   _opaque4_nodes->append(n);
1996 }
1997 
1998 // Remove all Opaque4 nodes.
1999 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2000   for (int i = opaque4_count(); i > 0; i--) {
2001     Node* opaq = opaque4_node(i-1);
2002     assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2003     igvn.replace_node(opaq, opaq->in(2));
2004   }
2005   assert(opaque4_count() == 0, "should be empty");
2006 }
2007 
2008 // StringOpts and late inlining of string methods
2009 void Compile::inline_string_calls(bool parse_time) {
2010   {
2011     // remove useless nodes to make the usage analysis simpler
2012     ResourceMark rm;
2013     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2014   }
2015 
2016   {
2017     ResourceMark rm;
2018     print_method(PHASE_BEFORE_STRINGOPTS, 3);
2019     PhaseStringOpts pso(initial_gvn(), for_igvn());
2020     print_method(PHASE_AFTER_STRINGOPTS, 3);
2021   }
2022 
2023   // now inline anything that we skipped the first time around
2024   if (!parse_time) {
2025     _late_inlines_pos = _late_inlines.length();
2026   }
2027 
2028   while (_string_late_inlines.length() > 0) {
2029     CallGenerator* cg = _string_late_inlines.pop();
2030     cg->do_late_inline();
2031     if (failing())  return;
2032   }
2033   _string_late_inlines.trunc_to(0);
2034 }
2035 
2036 // Late inlining of boxing methods
2037 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2038   if (_boxing_late_inlines.length() > 0) {
2039     assert(has_boxed_value(), "inconsistent");
2040 
2041     PhaseGVN* gvn = initial_gvn();
2042     set_inlining_incrementally(true);
2043 
2044     assert( igvn._worklist.size() == 0, "should be done with igvn" );
2045     for_igvn()->clear();
2046     gvn->replace_with(&igvn);
2047 
2048     _late_inlines_pos = _late_inlines.length();
2049 
2050     while (_boxing_late_inlines.length() > 0) {
2051       CallGenerator* cg = _boxing_late_inlines.pop();
2052       cg->do_late_inline();
2053       if (failing())  return;
2054     }
2055     _boxing_late_inlines.trunc_to(0);
2056 
2057     {
2058       ResourceMark rm;
2059       PhaseRemoveUseless pru(gvn, for_igvn());
2060     }
2061 
2062     igvn = PhaseIterGVN(gvn);
2063     igvn.optimize();
2064 
2065     set_inlining_progress(false);
2066     set_inlining_incrementally(false);
2067   }
2068 }
2069 
2070 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2071   assert(IncrementalInline, "incremental inlining should be on");
2072   PhaseGVN* gvn = initial_gvn();
2073 
2074   set_inlining_progress(false);
2075   for_igvn()->clear();
2076   gvn->replace_with(&igvn);
2077 
2078   {
2079     TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2080     int i = 0;
2081     for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2082       CallGenerator* cg = _late_inlines.at(i);
2083       _late_inlines_pos = i+1;
2084       cg->do_late_inline();
2085       if (failing())  return;
2086     }
2087     int j = 0;
2088     for (; i < _late_inlines.length(); i++, j++) {
2089       _late_inlines.at_put(j, _late_inlines.at(i));
2090     }
2091     _late_inlines.trunc_to(j);
2092   }
2093 
2094   {
2095     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2096     ResourceMark rm;
2097     PhaseRemoveUseless pru(gvn, for_igvn());
2098   }
2099 
2100   {
2101     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2102     igvn = PhaseIterGVN(gvn);
2103   }
2104 }
2105 
2106 // Perform incremental inlining until bound on number of live nodes is reached
2107 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2108   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2109 
2110   PhaseGVN* gvn = initial_gvn();
2111 
2112   set_inlining_incrementally(true);
2113   set_inlining_progress(true);
2114   uint low_live_nodes = 0;
2115 
2116   while(inlining_progress() && _late_inlines.length() > 0) {
2117 
2118     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2119       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2120         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2121         // PhaseIdealLoop is expensive so we only try it once we are
2122         // out of live nodes and we only try it again if the previous
2123         // helped got the number of nodes down significantly
2124         PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2125         if (failing())  return;
2126         low_live_nodes = live_nodes();
2127         _major_progress = true;
2128       }
2129 
2130       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2131         break;
2132       }
2133     }
2134 
2135     inline_incrementally_one(igvn);
2136 
2137     if (failing())  return;
2138 
2139     {
2140       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2141       igvn.optimize();
2142     }
2143 
2144     if (failing())  return;
2145   }
2146 
2147   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2148 
2149   if (_string_late_inlines.length() > 0) {
2150     assert(has_stringbuilder(), "inconsistent");
2151     for_igvn()->clear();
2152     initial_gvn()->replace_with(&igvn);
2153 
2154     inline_string_calls(false);
2155 
2156     if (failing())  return;
2157 
2158     {
2159       TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2160       ResourceMark rm;
2161       PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2162     }
2163 
2164     {
2165       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2166       igvn = PhaseIterGVN(gvn);
2167       igvn.optimize();
2168     }
2169   }
2170 
2171   set_inlining_incrementally(false);
2172 }
2173 
2174 
2175 bool Compile::optimize_loops(int& loop_opts_cnt, PhaseIterGVN& igvn, LoopOptsMode mode) {
2176   if(loop_opts_cnt > 0) {
2177     debug_only( int cnt = 0; );
2178     while(major_progress() && (loop_opts_cnt > 0)) {
2179       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2180       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2181       PhaseIdealLoop ideal_loop(igvn, mode);
2182       loop_opts_cnt--;
2183       if (failing())  return false;
2184       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2185     }
2186   }
2187   return true;
2188 }
2189 
2190 //------------------------------Optimize---------------------------------------
2191 // Given a graph, optimize it.
2192 void Compile::Optimize() {
2193   TracePhase tp("optimizer", &timers[_t_optimizer]);
2194 
2195 #ifndef PRODUCT
2196   if (_directive->BreakAtCompileOption) {
2197     BREAKPOINT;
2198   }
2199 
2200 #endif
2201 
2202 #ifdef ASSERT
2203   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2204   bs->verify_gc_barriers(true);
2205 #endif
2206 
2207   ResourceMark rm;
2208   int          loop_opts_cnt;
2209 
2210   print_inlining_reinit();
2211 
2212   NOT_PRODUCT( verify_graph_edges(); )
2213 
2214   print_method(PHASE_AFTER_PARSING);
2215 
2216  {
2217   // Iterative Global Value Numbering, including ideal transforms
2218   // Initialize IterGVN with types and values from parse-time GVN
2219   PhaseIterGVN igvn(initial_gvn());
2220 #ifdef ASSERT
2221   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2222 #endif
2223   {
2224     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2225     igvn.optimize();
2226   }
2227 
2228   if (failing())  return;
2229 
2230   print_method(PHASE_ITER_GVN1, 2);
2231 
2232   inline_incrementally(igvn);
2233 
2234   print_method(PHASE_INCREMENTAL_INLINE, 2);
2235 
2236   if (failing())  return;
2237 
2238   if (eliminate_boxing()) {
2239     // Inline valueOf() methods now.
2240     inline_boxing_calls(igvn);
2241 
2242     if (AlwaysIncrementalInline) {
2243       inline_incrementally(igvn);
2244     }
2245 
2246     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2247 
2248     if (failing())  return;
2249   }
2250 
2251   // Remove the speculative part of types and clean up the graph from
2252   // the extra CastPP nodes whose only purpose is to carry them. Do
2253   // that early so that optimizations are not disrupted by the extra
2254   // CastPP nodes.
2255   remove_speculative_types(igvn);
2256 
2257   // No more new expensive nodes will be added to the list from here
2258   // so keep only the actual candidates for optimizations.
2259   cleanup_expensive_nodes(igvn);
2260 
2261   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2262     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2263     initial_gvn()->replace_with(&igvn);
2264     for_igvn()->clear();
2265     Unique_Node_List new_worklist(C->comp_arena());
2266     {
2267       ResourceMark rm;
2268       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2269     }
2270     set_for_igvn(&new_worklist);
2271     igvn = PhaseIterGVN(initial_gvn());
2272     igvn.optimize();
2273   }
2274 
2275   // Perform escape analysis
2276   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2277     if (has_loops()) {
2278       // Cleanup graph (remove dead nodes).
2279       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2280       PhaseIdealLoop ideal_loop(igvn, LoopOptsNone);
2281       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2282       if (failing())  return;
2283     }
2284     ConnectionGraph::do_analysis(this, &igvn);
2285 
2286     if (failing())  return;
2287 
2288     // Optimize out fields loads from scalar replaceable allocations.
2289     igvn.optimize();
2290     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2291 
2292     if (failing())  return;
2293 
2294     if (congraph() != NULL && macro_count() > 0) {
2295       TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2296       PhaseMacroExpand mexp(igvn);
2297       mexp.eliminate_macro_nodes();
2298       igvn.set_delay_transform(false);
2299 
2300       igvn.optimize();
2301       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2302 
2303       if (failing())  return;
2304     }
2305   }
2306 
2307   // Loop transforms on the ideal graph.  Range Check Elimination,
2308   // peeling, unrolling, etc.
2309 
2310   // Set loop opts counter
2311   loop_opts_cnt = num_loop_opts();
2312   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2313     {
2314       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2315       PhaseIdealLoop ideal_loop(igvn, LoopOptsDefault);
2316       loop_opts_cnt--;
2317       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2318       if (failing())  return;
2319     }
2320     // Loop opts pass if partial peeling occurred in previous pass
2321     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2322       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2323       PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2324       loop_opts_cnt--;
2325       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2326       if (failing())  return;
2327     }
2328     // Loop opts pass for loop-unrolling before CCP
2329     if(major_progress() && (loop_opts_cnt > 0)) {
2330       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2331       PhaseIdealLoop ideal_loop(igvn, LoopOptsSkipSplitIf);
2332       loop_opts_cnt--;
2333       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2334     }
2335     if (!failing()) {
2336       // Verify that last round of loop opts produced a valid graph
2337       TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2338       PhaseIdealLoop::verify(igvn);
2339     }
2340   }
2341   if (failing())  return;
2342 
2343   // Conditional Constant Propagation;
2344   PhaseCCP ccp( &igvn );
2345   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2346   {
2347     TracePhase tp("ccp", &timers[_t_ccp]);
2348     ccp.do_transform();
2349   }
2350   print_method(PHASE_CPP1, 2);
2351 
2352   assert( true, "Break here to ccp.dump_old2new_map()");
2353 
2354   // Iterative Global Value Numbering, including ideal transforms
2355   {
2356     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2357     igvn = ccp;
2358     igvn.optimize();
2359   }
2360 
2361   print_method(PHASE_ITER_GVN2, 2);
2362 
2363   if (failing())  return;
2364 
2365   // Loop transforms on the ideal graph.  Range Check Elimination,
2366   // peeling, unrolling, etc.
2367   if (!optimize_loops(loop_opts_cnt, igvn, LoopOptsDefault)) {
2368     return;
2369   }
2370 
2371 #if INCLUDE_ZGC
2372   if (UseZGC) {
2373     ZBarrierSetC2::find_dominating_barriers(igvn);
2374   }
2375 #endif
2376 
2377   if (failing())  return;
2378 
2379   // Ensure that major progress is now clear
2380   C->clear_major_progress();
2381 
2382   {
2383     // Verify that all previous optimizations produced a valid graph
2384     // at least to this point, even if no loop optimizations were done.
2385     TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2386     PhaseIdealLoop::verify(igvn);
2387   }
2388 
2389   if (range_check_cast_count() > 0) {
2390     // No more loop optimizations. Remove all range check dependent CastIINodes.
2391     C->remove_range_check_casts(igvn);
2392     igvn.optimize();
2393   }
2394 
2395 #ifdef ASSERT
2396   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2397   bs->verify_gc_barriers(false);
2398 #endif
2399 
2400   {
2401     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2402     PhaseMacroExpand  mex(igvn);
2403     print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2404     if (mex.expand_macro_nodes()) {
2405       assert(failing(), "must bail out w/ explicit message");
2406       return;
2407     }
2408   }
2409 
2410   print_method(PHASE_BEFORE_BARRIER_EXPAND, 2);
2411 
2412 #if INCLUDE_SHENANDOAHGC
2413   if (UseShenandoahGC && !ShenandoahWriteBarrierNode::expand(this, igvn, loop_opts_cnt)) {
2414     assert(failing(), "must bail out w/ explicit message");
2415     return;
2416   }
2417 #endif
2418 
2419   if (opaque4_count() > 0) {
2420     C->remove_opaque4_nodes(igvn);
2421     igvn.optimize();
2422   }
2423 
2424   DEBUG_ONLY( _modified_nodes = NULL; )
2425  } // (End scope of igvn; run destructor if necessary for asserts.)
2426 
2427  process_print_inlining();
2428  // A method with only infinite loops has no edges entering loops from root
2429  {
2430    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2431    if (final_graph_reshaping()) {
2432      assert(failing(), "must bail out w/ explicit message");
2433      return;
2434    }
2435  }
2436 
2437  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2438 }
2439 
2440 
2441 //------------------------------Code_Gen---------------------------------------
2442 // Given a graph, generate code for it
2443 void Compile::Code_Gen() {
2444   if (failing()) {
2445     return;
2446   }
2447 
2448   // Perform instruction selection.  You might think we could reclaim Matcher
2449   // memory PDQ, but actually the Matcher is used in generating spill code.
2450   // Internals of the Matcher (including some VectorSets) must remain live
2451   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2452   // set a bit in reclaimed memory.
2453 
2454   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2455   // nodes.  Mapping is only valid at the root of each matched subtree.
2456   NOT_PRODUCT( verify_graph_edges(); )
2457 
2458   Matcher matcher;
2459   _matcher = &matcher;
2460   {
2461     TracePhase tp("matcher", &timers[_t_matcher]);
2462     matcher.match();
2463   }
2464   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2465   // nodes.  Mapping is only valid at the root of each matched subtree.
2466   NOT_PRODUCT( verify_graph_edges(); )
2467 
2468   // If you have too many nodes, or if matching has failed, bail out
2469   check_node_count(0, "out of nodes matching instructions");
2470   if (failing()) {
2471     return;
2472   }
2473 
2474   print_method(PHASE_MATCHING, 2);
2475 
2476   // Build a proper-looking CFG
2477   PhaseCFG cfg(node_arena(), root(), matcher);
2478   _cfg = &cfg;
2479   {
2480     TracePhase tp("scheduler", &timers[_t_scheduler]);
2481     bool success = cfg.do_global_code_motion();
2482     if (!success) {
2483       return;
2484     }
2485 
2486     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2487     NOT_PRODUCT( verify_graph_edges(); )
2488     debug_only( cfg.verify(); )
2489   }
2490 
2491   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2492   _regalloc = &regalloc;
2493   {
2494     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2495     // Perform register allocation.  After Chaitin, use-def chains are
2496     // no longer accurate (at spill code) and so must be ignored.
2497     // Node->LRG->reg mappings are still accurate.
2498     _regalloc->Register_Allocate();
2499 
2500     // Bail out if the allocator builds too many nodes
2501     if (failing()) {
2502       return;
2503     }
2504   }
2505 
2506   // Prior to register allocation we kept empty basic blocks in case the
2507   // the allocator needed a place to spill.  After register allocation we
2508   // are not adding any new instructions.  If any basic block is empty, we
2509   // can now safely remove it.
2510   {
2511     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2512     cfg.remove_empty_blocks();
2513     if (do_freq_based_layout()) {
2514       PhaseBlockLayout layout(cfg);
2515     } else {
2516       cfg.set_loop_alignment();
2517     }
2518     cfg.fixup_flow();
2519   }
2520 
2521   // Apply peephole optimizations
2522   if( OptoPeephole ) {
2523     TracePhase tp("peephole", &timers[_t_peephole]);
2524     PhasePeephole peep( _regalloc, cfg);
2525     peep.do_transform();
2526   }
2527 
2528   // Do late expand if CPU requires this.
2529   if (Matcher::require_postalloc_expand) {
2530     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2531     cfg.postalloc_expand(_regalloc);
2532   }
2533 
2534   // Convert Nodes to instruction bits in a buffer
2535   {
2536     TraceTime tp("output", &timers[_t_output], CITime);
2537     Output();
2538   }
2539 
2540   print_method(PHASE_FINAL_CODE);
2541 
2542   // He's dead, Jim.
2543   _cfg     = (PhaseCFG*)((intptr_t)0xdeadbeef);
2544   _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2545 }
2546 
2547 
2548 //------------------------------dump_asm---------------------------------------
2549 // Dump formatted assembly
2550 #ifndef PRODUCT
2551 void Compile::dump_asm(int *pcs, uint pc_limit) {
2552   bool cut_short = false;
2553   tty->print_cr("#");
2554   tty->print("#  ");  _tf->dump();  tty->cr();
2555   tty->print_cr("#");
2556 
2557   // For all blocks
2558   int pc = 0x0;                 // Program counter
2559   char starts_bundle = ' ';
2560   _regalloc->dump_frame();
2561 
2562   Node *n = NULL;
2563   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2564     if (VMThread::should_terminate()) {
2565       cut_short = true;
2566       break;
2567     }
2568     Block* block = _cfg->get_block(i);
2569     if (block->is_connector() && !Verbose) {
2570       continue;
2571     }
2572     n = block->head();
2573     if (pcs && n->_idx < pc_limit) {
2574       tty->print("%3.3x   ", pcs[n->_idx]);
2575     } else {
2576       tty->print("      ");
2577     }
2578     block->dump_head(_cfg);
2579     if (block->is_connector()) {
2580       tty->print_cr("        # Empty connector block");
2581     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2582       tty->print_cr("        # Block is sole successor of call");
2583     }
2584 
2585     // For all instructions
2586     Node *delay = NULL;
2587     for (uint j = 0; j < block->number_of_nodes(); j++) {
2588       if (VMThread::should_terminate()) {
2589         cut_short = true;
2590         break;
2591       }
2592       n = block->get_node(j);
2593       if (valid_bundle_info(n)) {
2594         Bundle* bundle = node_bundling(n);
2595         if (bundle->used_in_unconditional_delay()) {
2596           delay = n;
2597           continue;
2598         }
2599         if (bundle->starts_bundle()) {
2600           starts_bundle = '+';
2601         }
2602       }
2603 
2604       if (WizardMode) {
2605         n->dump();
2606       }
2607 
2608       if( !n->is_Region() &&    // Dont print in the Assembly
2609           !n->is_Phi() &&       // a few noisely useless nodes
2610           !n->is_Proj() &&
2611           !n->is_MachTemp() &&
2612           !n->is_SafePointScalarObject() &&
2613           !n->is_Catch() &&     // Would be nice to print exception table targets
2614           !n->is_MergeMem() &&  // Not very interesting
2615           !n->is_top() &&       // Debug info table constants
2616           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2617           ) {
2618         if (pcs && n->_idx < pc_limit)
2619           tty->print("%3.3x", pcs[n->_idx]);
2620         else
2621           tty->print("   ");
2622         tty->print(" %c ", starts_bundle);
2623         starts_bundle = ' ';
2624         tty->print("\t");
2625         n->format(_regalloc, tty);
2626         tty->cr();
2627       }
2628 
2629       // If we have an instruction with a delay slot, and have seen a delay,
2630       // then back up and print it
2631       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2632         assert(delay != NULL, "no unconditional delay instruction");
2633         if (WizardMode) delay->dump();
2634 
2635         if (node_bundling(delay)->starts_bundle())
2636           starts_bundle = '+';
2637         if (pcs && n->_idx < pc_limit)
2638           tty->print("%3.3x", pcs[n->_idx]);
2639         else
2640           tty->print("   ");
2641         tty->print(" %c ", starts_bundle);
2642         starts_bundle = ' ';
2643         tty->print("\t");
2644         delay->format(_regalloc, tty);
2645         tty->cr();
2646         delay = NULL;
2647       }
2648 
2649       // Dump the exception table as well
2650       if( n->is_Catch() && (Verbose || WizardMode) ) {
2651         // Print the exception table for this offset
2652         _handler_table.print_subtable_for(pc);
2653       }
2654     }
2655 
2656     if (pcs && n->_idx < pc_limit)
2657       tty->print_cr("%3.3x", pcs[n->_idx]);
2658     else
2659       tty->cr();
2660 
2661     assert(cut_short || delay == NULL, "no unconditional delay branch");
2662 
2663   } // End of per-block dump
2664   tty->cr();
2665 
2666   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2667 }
2668 #endif
2669 
2670 //------------------------------Final_Reshape_Counts---------------------------
2671 // This class defines counters to help identify when a method
2672 // may/must be executed using hardware with only 24-bit precision.
2673 struct Final_Reshape_Counts : public StackObj {
2674   int  _call_count;             // count non-inlined 'common' calls
2675   int  _float_count;            // count float ops requiring 24-bit precision
2676   int  _double_count;           // count double ops requiring more precision
2677   int  _java_call_count;        // count non-inlined 'java' calls
2678   int  _inner_loop_count;       // count loops which need alignment
2679   VectorSet _visited;           // Visitation flags
2680   Node_List _tests;             // Set of IfNodes & PCTableNodes
2681 
2682   Final_Reshape_Counts() :
2683     _call_count(0), _float_count(0), _double_count(0),
2684     _java_call_count(0), _inner_loop_count(0),
2685     _visited( Thread::current()->resource_area() ) { }
2686 
2687   void inc_call_count  () { _call_count  ++; }
2688   void inc_float_count () { _float_count ++; }
2689   void inc_double_count() { _double_count++; }
2690   void inc_java_call_count() { _java_call_count++; }
2691   void inc_inner_loop_count() { _inner_loop_count++; }
2692 
2693   int  get_call_count  () const { return _call_count  ; }
2694   int  get_float_count () const { return _float_count ; }
2695   int  get_double_count() const { return _double_count; }
2696   int  get_java_call_count() const { return _java_call_count; }
2697   int  get_inner_loop_count() const { return _inner_loop_count; }
2698 };
2699 
2700 #ifdef ASSERT
2701 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2702   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2703   // Make sure the offset goes inside the instance layout.
2704   return k->contains_field_offset(tp->offset());
2705   // Note that OffsetBot and OffsetTop are very negative.
2706 }
2707 #endif
2708 
2709 // Eliminate trivially redundant StoreCMs and accumulate their
2710 // precedence edges.
2711 void Compile::eliminate_redundant_card_marks(Node* n) {
2712   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2713   if (n->in(MemNode::Address)->outcnt() > 1) {
2714     // There are multiple users of the same address so it might be
2715     // possible to eliminate some of the StoreCMs
2716     Node* mem = n->in(MemNode::Memory);
2717     Node* adr = n->in(MemNode::Address);
2718     Node* val = n->in(MemNode::ValueIn);
2719     Node* prev = n;
2720     bool done = false;
2721     // Walk the chain of StoreCMs eliminating ones that match.  As
2722     // long as it's a chain of single users then the optimization is
2723     // safe.  Eliminating partially redundant StoreCMs would require
2724     // cloning copies down the other paths.
2725     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2726       if (adr == mem->in(MemNode::Address) &&
2727           val == mem->in(MemNode::ValueIn)) {
2728         // redundant StoreCM
2729         if (mem->req() > MemNode::OopStore) {
2730           // Hasn't been processed by this code yet.
2731           n->add_prec(mem->in(MemNode::OopStore));
2732         } else {
2733           // Already converted to precedence edge
2734           for (uint i = mem->req(); i < mem->len(); i++) {
2735             // Accumulate any precedence edges
2736             if (mem->in(i) != NULL) {
2737               n->add_prec(mem->in(i));
2738             }
2739           }
2740           // Everything above this point has been processed.
2741           done = true;
2742         }
2743         // Eliminate the previous StoreCM
2744         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2745         assert(mem->outcnt() == 0, "should be dead");
2746         mem->disconnect_inputs(NULL, this);
2747       } else {
2748         prev = mem;
2749       }
2750       mem = prev->in(MemNode::Memory);
2751     }
2752   }
2753 }
2754 
2755 //------------------------------final_graph_reshaping_impl----------------------
2756 // Implement items 1-5 from final_graph_reshaping below.
2757 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2758 
2759   if ( n->outcnt() == 0 ) return; // dead node
2760   uint nop = n->Opcode();
2761 
2762   // Check for 2-input instruction with "last use" on right input.
2763   // Swap to left input.  Implements item (2).
2764   if( n->req() == 3 &&          // two-input instruction
2765       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2766       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2767       n->in(2)->outcnt() == 1 &&// right use IS a last use
2768       !n->in(2)->is_Con() ) {   // right use is not a constant
2769     // Check for commutative opcode
2770     switch( nop ) {
2771     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2772     case Op_MaxI:  case Op_MinI:
2773     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2774     case Op_AndL:  case Op_XorL:  case Op_OrL:
2775     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2776       // Move "last use" input to left by swapping inputs
2777       n->swap_edges(1, 2);
2778       break;
2779     }
2780     default:
2781       break;
2782     }
2783   }
2784 
2785 #ifdef ASSERT
2786   if( n->is_Mem() ) {
2787     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2788     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2789             // oop will be recorded in oop map if load crosses safepoint
2790             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2791                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2792             "raw memory operations should have control edge");
2793   }
2794 #endif
2795   // Count FPU ops and common calls, implements item (3)
2796   switch( nop ) {
2797   // Count all float operations that may use FPU
2798   case Op_AddF:
2799   case Op_SubF:
2800   case Op_MulF:
2801   case Op_DivF:
2802   case Op_NegF:
2803   case Op_ModF:
2804   case Op_ConvI2F:
2805   case Op_ConF:
2806   case Op_CmpF:
2807   case Op_CmpF3:
2808   // case Op_ConvL2F: // longs are split into 32-bit halves
2809     frc.inc_float_count();
2810     break;
2811 
2812   case Op_ConvF2D:
2813   case Op_ConvD2F:
2814     frc.inc_float_count();
2815     frc.inc_double_count();
2816     break;
2817 
2818   // Count all double operations that may use FPU
2819   case Op_AddD:
2820   case Op_SubD:
2821   case Op_MulD:
2822   case Op_DivD:
2823   case Op_NegD:
2824   case Op_ModD:
2825   case Op_ConvI2D:
2826   case Op_ConvD2I:
2827   // case Op_ConvL2D: // handled by leaf call
2828   // case Op_ConvD2L: // handled by leaf call
2829   case Op_ConD:
2830   case Op_CmpD:
2831   case Op_CmpD3:
2832     frc.inc_double_count();
2833     break;
2834   case Op_Opaque1:              // Remove Opaque Nodes before matching
2835   case Op_Opaque2:              // Remove Opaque Nodes before matching
2836   case Op_Opaque3:
2837     n->subsume_by(n->in(1), this);
2838     break;
2839   case Op_CallStaticJava:
2840   case Op_CallJava:
2841   case Op_CallDynamicJava:
2842     frc.inc_java_call_count(); // Count java call site;
2843   case Op_CallRuntime:
2844   case Op_CallLeaf:
2845   case Op_CallLeafNoFP: {
2846     assert (n->is_Call(), "");
2847     CallNode *call = n->as_Call();
2848 #if INCLUDE_SHENANDOAHGC
2849     if (UseShenandoahGC && ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
2850       uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
2851       if (call->req() > cnt) {
2852         assert(call->req() == cnt+1, "only one extra input");
2853         Node* addp = call->in(cnt);
2854         assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
2855         call->del_req(cnt);
2856       }
2857     }
2858 #endif
2859     // Count call sites where the FP mode bit would have to be flipped.
2860     // Do not count uncommon runtime calls:
2861     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2862     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2863     if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2864       frc.inc_call_count();   // Count the call site
2865     } else {                  // See if uncommon argument is shared
2866       Node *n = call->in(TypeFunc::Parms);
2867       int nop = n->Opcode();
2868       // Clone shared simple arguments to uncommon calls, item (1).
2869       if (n->outcnt() > 1 &&
2870           !n->is_Proj() &&
2871           nop != Op_CreateEx &&
2872           nop != Op_CheckCastPP &&
2873           nop != Op_DecodeN &&
2874           nop != Op_DecodeNKlass &&
2875           !n->is_Mem() &&
2876           !n->is_Phi()) {
2877         Node *x = n->clone();
2878         call->set_req(TypeFunc::Parms, x);
2879       }
2880     }
2881     break;
2882   }
2883 
2884   case Op_StoreD:
2885   case Op_LoadD:
2886   case Op_LoadD_unaligned:
2887     frc.inc_double_count();
2888     goto handle_mem;
2889   case Op_StoreF:
2890   case Op_LoadF:
2891     frc.inc_float_count();
2892     goto handle_mem;
2893 
2894   case Op_StoreCM:
2895     {
2896       // Convert OopStore dependence into precedence edge
2897       Node* prec = n->in(MemNode::OopStore);
2898       n->del_req(MemNode::OopStore);
2899       n->add_prec(prec);
2900       eliminate_redundant_card_marks(n);
2901     }
2902 
2903     // fall through
2904 
2905   case Op_StoreB:
2906   case Op_StoreC:
2907   case Op_StorePConditional:
2908   case Op_StoreI:
2909   case Op_StoreL:
2910   case Op_StoreIConditional:
2911   case Op_StoreLConditional:
2912   case Op_CompareAndSwapB:
2913   case Op_CompareAndSwapS:
2914   case Op_CompareAndSwapI:
2915   case Op_CompareAndSwapL:
2916   case Op_CompareAndSwapP:
2917   case Op_CompareAndSwapN:
2918   case Op_WeakCompareAndSwapB:
2919   case Op_WeakCompareAndSwapS:
2920   case Op_WeakCompareAndSwapI:
2921   case Op_WeakCompareAndSwapL:
2922   case Op_WeakCompareAndSwapP:
2923   case Op_WeakCompareAndSwapN:
2924   case Op_CompareAndExchangeB:
2925   case Op_CompareAndExchangeS:
2926   case Op_CompareAndExchangeI:
2927   case Op_CompareAndExchangeL:
2928   case Op_CompareAndExchangeP:
2929   case Op_CompareAndExchangeN:
2930   case Op_GetAndAddS:
2931   case Op_GetAndAddB:
2932   case Op_GetAndAddI:
2933   case Op_GetAndAddL:
2934   case Op_GetAndSetS:
2935   case Op_GetAndSetB:
2936   case Op_GetAndSetI:
2937   case Op_GetAndSetL:
2938   case Op_GetAndSetP:
2939   case Op_GetAndSetN:
2940   case Op_StoreP:
2941   case Op_StoreN:
2942   case Op_StoreNKlass:
2943   case Op_LoadB:
2944   case Op_LoadUB:
2945   case Op_LoadUS:
2946   case Op_LoadI:
2947   case Op_LoadKlass:
2948   case Op_LoadNKlass:
2949   case Op_LoadL:
2950   case Op_LoadL_unaligned:
2951   case Op_LoadPLocked:
2952   case Op_LoadP:
2953 #if INCLUDE_ZGC
2954   case Op_LoadBarrierSlowReg:
2955   case Op_LoadBarrierWeakSlowReg:
2956 #endif
2957   case Op_LoadN:
2958   case Op_LoadRange:
2959   case Op_LoadS: {
2960   handle_mem:
2961 #ifdef ASSERT
2962     if( VerifyOptoOopOffsets ) {
2963       assert( n->is_Mem(), "" );
2964       MemNode *mem  = (MemNode*)n;
2965       // Check to see if address types have grounded out somehow.
2966       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2967       assert( !tp || oop_offset_is_sane(tp), "" );
2968     }
2969 #endif
2970     break;
2971   }
2972 
2973   case Op_AddP: {               // Assert sane base pointers
2974     Node *addp = n->in(AddPNode::Address);
2975     assert( !addp->is_AddP() ||
2976             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2977             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2978             "Base pointers must match (addp %u)", addp->_idx );
2979 #ifdef _LP64
2980     if ((UseCompressedOops || UseCompressedClassPointers) &&
2981         addp->Opcode() == Op_ConP &&
2982         addp == n->in(AddPNode::Base) &&
2983         n->in(AddPNode::Offset)->is_Con()) {
2984       // If the transformation of ConP to ConN+DecodeN is beneficial depends
2985       // on the platform and on the compressed oops mode.
2986       // Use addressing with narrow klass to load with offset on x86.
2987       // Some platforms can use the constant pool to load ConP.
2988       // Do this transformation here since IGVN will convert ConN back to ConP.
2989       const Type* t = addp->bottom_type();
2990       bool is_oop   = t->isa_oopptr() != NULL;
2991       bool is_klass = t->isa_klassptr() != NULL;
2992 
2993       if ((is_oop   && Matcher::const_oop_prefer_decode()  ) ||
2994           (is_klass && Matcher::const_klass_prefer_decode())) {
2995         Node* nn = NULL;
2996 
2997         int op = is_oop ? Op_ConN : Op_ConNKlass;
2998 
2999         // Look for existing ConN node of the same exact type.
3000         Node* r  = root();
3001         uint cnt = r->outcnt();
3002         for (uint i = 0; i < cnt; i++) {
3003           Node* m = r->raw_out(i);
3004           if (m!= NULL && m->Opcode() == op &&
3005               m->bottom_type()->make_ptr() == t) {
3006             nn = m;
3007             break;
3008           }
3009         }
3010         if (nn != NULL) {
3011           // Decode a narrow oop to match address
3012           // [R12 + narrow_oop_reg<<3 + offset]
3013           if (is_oop) {
3014             nn = new DecodeNNode(nn, t);
3015           } else {
3016             nn = new DecodeNKlassNode(nn, t);
3017           }
3018           // Check for succeeding AddP which uses the same Base.
3019           // Otherwise we will run into the assertion above when visiting that guy.
3020           for (uint i = 0; i < n->outcnt(); ++i) {
3021             Node *out_i = n->raw_out(i);
3022             if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3023               out_i->set_req(AddPNode::Base, nn);
3024 #ifdef ASSERT
3025               for (uint j = 0; j < out_i->outcnt(); ++j) {
3026                 Node *out_j = out_i->raw_out(j);
3027                 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3028                        "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3029               }
3030 #endif
3031             }
3032           }
3033           n->set_req(AddPNode::Base, nn);
3034           n->set_req(AddPNode::Address, nn);
3035           if (addp->outcnt() == 0) {
3036             addp->disconnect_inputs(NULL, this);
3037           }
3038         }
3039       }
3040     }
3041 #endif
3042     // platform dependent reshaping of the address expression
3043     reshape_address(n->as_AddP());
3044     break;
3045   }
3046 
3047   case Op_CastPP: {
3048     // Remove CastPP nodes to gain more freedom during scheduling but
3049     // keep the dependency they encode as control or precedence edges
3050     // (if control is set already) on memory operations. Some CastPP
3051     // nodes don't have a control (don't carry a dependency): skip
3052     // those.
3053     if (n->in(0) != NULL) {
3054       ResourceMark rm;
3055       Unique_Node_List wq;
3056       wq.push(n);
3057       for (uint next = 0; next < wq.size(); ++next) {
3058         Node *m = wq.at(next);
3059         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3060           Node* use = m->fast_out(i);
3061           if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->is_ShenandoahBarrier()) {
3062             use->ensure_control_or_add_prec(n->in(0));
3063           } else {
3064             switch(use->Opcode()) {
3065             case Op_AddP:
3066             case Op_DecodeN:
3067             case Op_DecodeNKlass:
3068             case Op_CheckCastPP:
3069             case Op_CastPP:
3070               wq.push(use);
3071               break;
3072             }
3073           }
3074         }
3075       }
3076     }
3077     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3078     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3079       Node* in1 = n->in(1);
3080       const Type* t = n->bottom_type();
3081       Node* new_in1 = in1->clone();
3082       new_in1->as_DecodeN()->set_type(t);
3083 
3084       if (!Matcher::narrow_oop_use_complex_address()) {
3085         //
3086         // x86, ARM and friends can handle 2 adds in addressing mode
3087         // and Matcher can fold a DecodeN node into address by using
3088         // a narrow oop directly and do implicit NULL check in address:
3089         //
3090         // [R12 + narrow_oop_reg<<3 + offset]
3091         // NullCheck narrow_oop_reg
3092         //
3093         // On other platforms (Sparc) we have to keep new DecodeN node and
3094         // use it to do implicit NULL check in address:
3095         //
3096         // decode_not_null narrow_oop_reg, base_reg
3097         // [base_reg + offset]
3098         // NullCheck base_reg
3099         //
3100         // Pin the new DecodeN node to non-null path on these platform (Sparc)
3101         // to keep the information to which NULL check the new DecodeN node
3102         // corresponds to use it as value in implicit_null_check().
3103         //
3104         new_in1->set_req(0, n->in(0));
3105       }
3106 
3107       n->subsume_by(new_in1, this);
3108       if (in1->outcnt() == 0) {
3109         in1->disconnect_inputs(NULL, this);
3110       }
3111     } else {
3112       n->subsume_by(n->in(1), this);
3113       if (n->outcnt() == 0) {
3114         n->disconnect_inputs(NULL, this);
3115       }
3116     }
3117     break;
3118   }
3119 #ifdef _LP64
3120   case Op_CmpP:
3121     // Do this transformation here to preserve CmpPNode::sub() and
3122     // other TypePtr related Ideal optimizations (for example, ptr nullness).
3123     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3124       Node* in1 = n->in(1);
3125       Node* in2 = n->in(2);
3126       if (!in1->is_DecodeNarrowPtr()) {
3127         in2 = in1;
3128         in1 = n->in(2);
3129       }
3130       assert(in1->is_DecodeNarrowPtr(), "sanity");
3131 
3132       Node* new_in2 = NULL;
3133       if (in2->is_DecodeNarrowPtr()) {
3134         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3135         new_in2 = in2->in(1);
3136       } else if (in2->Opcode() == Op_ConP) {
3137         const Type* t = in2->bottom_type();
3138         if (t == TypePtr::NULL_PTR) {
3139           assert(in1->is_DecodeN(), "compare klass to null?");
3140           // Don't convert CmpP null check into CmpN if compressed
3141           // oops implicit null check is not generated.
3142           // This will allow to generate normal oop implicit null check.
3143           if (Matcher::gen_narrow_oop_implicit_null_checks())
3144             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3145           //
3146           // This transformation together with CastPP transformation above
3147           // will generated code for implicit NULL checks for compressed oops.
3148           //
3149           // The original code after Optimize()
3150           //
3151           //    LoadN memory, narrow_oop_reg
3152           //    decode narrow_oop_reg, base_reg
3153           //    CmpP base_reg, NULL
3154           //    CastPP base_reg // NotNull
3155           //    Load [base_reg + offset], val_reg
3156           //
3157           // after these transformations will be
3158           //
3159           //    LoadN memory, narrow_oop_reg
3160           //    CmpN narrow_oop_reg, NULL
3161           //    decode_not_null narrow_oop_reg, base_reg
3162           //    Load [base_reg + offset], val_reg
3163           //
3164           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3165           // since narrow oops can be used in debug info now (see the code in
3166           // final_graph_reshaping_walk()).
3167           //
3168           // At the end the code will be matched to
3169           // on x86:
3170           //
3171           //    Load_narrow_oop memory, narrow_oop_reg
3172           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3173           //    NullCheck narrow_oop_reg
3174           //
3175           // and on sparc:
3176           //
3177           //    Load_narrow_oop memory, narrow_oop_reg
3178           //    decode_not_null narrow_oop_reg, base_reg
3179           //    Load [base_reg + offset], val_reg
3180           //    NullCheck base_reg
3181           //
3182         } else if (t->isa_oopptr()) {
3183           new_in2 = ConNode::make(t->make_narrowoop());
3184         } else if (t->isa_klassptr()) {
3185           new_in2 = ConNode::make(t->make_narrowklass());
3186         }
3187       }
3188       if (new_in2 != NULL) {
3189         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3190         n->subsume_by(cmpN, this);
3191         if (in1->outcnt() == 0) {
3192           in1->disconnect_inputs(NULL, this);
3193         }
3194         if (in2->outcnt() == 0) {
3195           in2->disconnect_inputs(NULL, this);
3196         }
3197       }
3198     }
3199     break;
3200 
3201   case Op_DecodeN:
3202   case Op_DecodeNKlass:
3203     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3204     // DecodeN could be pinned when it can't be fold into
3205     // an address expression, see the code for Op_CastPP above.
3206     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3207     break;
3208 
3209   case Op_EncodeP:
3210   case Op_EncodePKlass: {
3211     Node* in1 = n->in(1);
3212     if (in1->is_DecodeNarrowPtr()) {
3213       n->subsume_by(in1->in(1), this);
3214     } else if (in1->Opcode() == Op_ConP) {
3215       const Type* t = in1->bottom_type();
3216       if (t == TypePtr::NULL_PTR) {
3217         assert(t->isa_oopptr(), "null klass?");
3218         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3219       } else if (t->isa_oopptr()) {
3220         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3221       } else if (t->isa_klassptr()) {
3222         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3223       }
3224     }
3225     if (in1->outcnt() == 0) {
3226       in1->disconnect_inputs(NULL, this);
3227     }
3228     break;
3229   }
3230 
3231   case Op_Proj: {
3232     if (OptimizeStringConcat) {
3233       ProjNode* p = n->as_Proj();
3234       if (p->_is_io_use) {
3235         // Separate projections were used for the exception path which
3236         // are normally removed by a late inline.  If it wasn't inlined
3237         // then they will hang around and should just be replaced with
3238         // the original one.
3239         Node* proj = NULL;
3240         // Replace with just one
3241         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3242           Node *use = i.get();
3243           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3244             proj = use;
3245             break;
3246           }
3247         }
3248         assert(proj != NULL, "must be found");
3249         p->subsume_by(proj, this);
3250       }
3251     }
3252     break;
3253   }
3254 
3255   case Op_Phi:
3256     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3257       // The EncodeP optimization may create Phi with the same edges
3258       // for all paths. It is not handled well by Register Allocator.
3259       Node* unique_in = n->in(1);
3260       assert(unique_in != NULL, "");
3261       uint cnt = n->req();
3262       for (uint i = 2; i < cnt; i++) {
3263         Node* m = n->in(i);
3264         assert(m != NULL, "");
3265         if (unique_in != m)
3266           unique_in = NULL;
3267       }
3268       if (unique_in != NULL) {
3269         n->subsume_by(unique_in, this);
3270       }
3271     }
3272     break;
3273 
3274 #endif
3275 
3276 #ifdef ASSERT
3277   case Op_CastII:
3278     // Verify that all range check dependent CastII nodes were removed.
3279     if (n->isa_CastII()->has_range_check()) {
3280       n->dump(3);
3281       assert(false, "Range check dependent CastII node was not removed");
3282     }
3283     break;
3284 #endif
3285 
3286   case Op_ModI:
3287     if (UseDivMod) {
3288       // Check if a%b and a/b both exist
3289       Node* d = n->find_similar(Op_DivI);
3290       if (d) {
3291         // Replace them with a fused divmod if supported
3292         if (Matcher::has_match_rule(Op_DivModI)) {
3293           DivModINode* divmod = DivModINode::make(n);
3294           d->subsume_by(divmod->div_proj(), this);
3295           n->subsume_by(divmod->mod_proj(), this);
3296         } else {
3297           // replace a%b with a-((a/b)*b)
3298           Node* mult = new MulINode(d, d->in(2));
3299           Node* sub  = new SubINode(d->in(1), mult);
3300           n->subsume_by(sub, this);
3301         }
3302       }
3303     }
3304     break;
3305 
3306   case Op_ModL:
3307     if (UseDivMod) {
3308       // Check if a%b and a/b both exist
3309       Node* d = n->find_similar(Op_DivL);
3310       if (d) {
3311         // Replace them with a fused divmod if supported
3312         if (Matcher::has_match_rule(Op_DivModL)) {
3313           DivModLNode* divmod = DivModLNode::make(n);
3314           d->subsume_by(divmod->div_proj(), this);
3315           n->subsume_by(divmod->mod_proj(), this);
3316         } else {
3317           // replace a%b with a-((a/b)*b)
3318           Node* mult = new MulLNode(d, d->in(2));
3319           Node* sub  = new SubLNode(d->in(1), mult);
3320           n->subsume_by(sub, this);
3321         }
3322       }
3323     }
3324     break;
3325 
3326   case Op_LoadVector:
3327   case Op_StoreVector:
3328     break;
3329 
3330   case Op_AddReductionVI:
3331   case Op_AddReductionVL:
3332   case Op_AddReductionVF:
3333   case Op_AddReductionVD:
3334   case Op_MulReductionVI:
3335   case Op_MulReductionVL:
3336   case Op_MulReductionVF:
3337   case Op_MulReductionVD:
3338     break;
3339 
3340   case Op_PackB:
3341   case Op_PackS:
3342   case Op_PackI:
3343   case Op_PackF:
3344   case Op_PackL:
3345   case Op_PackD:
3346     if (n->req()-1 > 2) {
3347       // Replace many operand PackNodes with a binary tree for matching
3348       PackNode* p = (PackNode*) n;
3349       Node* btp = p->binary_tree_pack(1, n->req());
3350       n->subsume_by(btp, this);
3351     }
3352     break;
3353   case Op_Loop:
3354   case Op_CountedLoop:
3355   case Op_OuterStripMinedLoop:
3356     if (n->as_Loop()->is_inner_loop()) {
3357       frc.inc_inner_loop_count();
3358     }
3359     n->as_Loop()->verify_strip_mined(0);
3360     break;
3361   case Op_LShiftI:
3362   case Op_RShiftI:
3363   case Op_URShiftI:
3364   case Op_LShiftL:
3365   case Op_RShiftL:
3366   case Op_URShiftL:
3367     if (Matcher::need_masked_shift_count) {
3368       // The cpu's shift instructions don't restrict the count to the
3369       // lower 5/6 bits. We need to do the masking ourselves.
3370       Node* in2 = n->in(2);
3371       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3372       const TypeInt* t = in2->find_int_type();
3373       if (t != NULL && t->is_con()) {
3374         juint shift = t->get_con();
3375         if (shift > mask) { // Unsigned cmp
3376           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3377         }
3378       } else {
3379         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3380           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3381           n->set_req(2, shift);
3382         }
3383       }
3384       if (in2->outcnt() == 0) { // Remove dead node
3385         in2->disconnect_inputs(NULL, this);
3386       }
3387     }
3388     break;
3389   case Op_MemBarStoreStore:
3390   case Op_MemBarRelease:
3391     // Break the link with AllocateNode: it is no longer useful and
3392     // confuses register allocation.
3393     if (n->req() > MemBarNode::Precedent) {
3394       n->set_req(MemBarNode::Precedent, top());
3395     }
3396     break;
3397 #if INCLUDE_SHENANDOAHGC
3398   case Op_ShenandoahReadBarrier:
3399     break;
3400   case Op_ShenandoahWriteBarrier:
3401     assert(false, "should have been expanded already");
3402     break;
3403 #endif
3404   case Op_RangeCheck: {
3405     RangeCheckNode* rc = n->as_RangeCheck();
3406     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3407     n->subsume_by(iff, this);
3408     frc._tests.push(iff);
3409     break;
3410   }
3411   case Op_ConvI2L: {
3412     if (!Matcher::convi2l_type_required) {
3413       // Code generation on some platforms doesn't need accurate
3414       // ConvI2L types. Widening the type can help remove redundant
3415       // address computations.
3416       n->as_Type()->set_type(TypeLong::INT);
3417       ResourceMark rm;
3418       Node_List wq;
3419       wq.push(n);
3420       for (uint next = 0; next < wq.size(); next++) {
3421         Node *m = wq.at(next);
3422 
3423         for(;;) {
3424           // Loop over all nodes with identical inputs edges as m
3425           Node* k = m->find_similar(m->Opcode());
3426           if (k == NULL) {
3427             break;
3428           }
3429           // Push their uses so we get a chance to remove node made
3430           // redundant
3431           for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3432             Node* u = k->fast_out(i);
3433             assert(!wq.contains(u), "shouldn't process one node several times");
3434             if (u->Opcode() == Op_LShiftL ||
3435                 u->Opcode() == Op_AddL ||
3436                 u->Opcode() == Op_SubL ||
3437                 u->Opcode() == Op_AddP) {
3438               wq.push(u);
3439             }
3440           }
3441           // Replace all nodes with identical edges as m with m
3442           k->subsume_by(m, this);
3443         }
3444       }
3445     }
3446     break;
3447   }
3448   case Op_CmpUL: {
3449     if (!Matcher::has_match_rule(Op_CmpUL)) {
3450       // No support for unsigned long comparisons
3451       ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3452       Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3453       Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3454       ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3455       Node* andl = new AndLNode(orl, remove_sign_mask);
3456       Node* cmp = new CmpLNode(andl, n->in(2));
3457       n->subsume_by(cmp, this);
3458     }
3459     break;
3460   }
3461   default:
3462     assert( !n->is_Call(), "" );
3463     assert( !n->is_Mem(), "" );
3464     assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3465     break;
3466   }
3467 
3468   // Collect CFG split points
3469   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3470     frc._tests.push(n);
3471   }
3472 }
3473 
3474 //------------------------------final_graph_reshaping_walk---------------------
3475 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3476 // requires that the walk visits a node's inputs before visiting the node.
3477 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3478   ResourceArea *area = Thread::current()->resource_area();
3479   Unique_Node_List sfpt(area);
3480 
3481   frc._visited.set(root->_idx); // first, mark node as visited
3482   uint cnt = root->req();
3483   Node *n = root;
3484   uint  i = 0;
3485   while (true) {
3486     if (i < cnt) {
3487       // Place all non-visited non-null inputs onto stack
3488       Node* m = n->in(i);
3489       ++i;
3490       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3491         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3492           // compute worst case interpreter size in case of a deoptimization
3493           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3494 
3495           sfpt.push(m);
3496         }
3497         cnt = m->req();
3498         nstack.push(n, i); // put on stack parent and next input's index
3499         n = m;
3500         i = 0;
3501       }
3502     } else {
3503       // Now do post-visit work
3504       final_graph_reshaping_impl( n, frc );
3505       if (nstack.is_empty())
3506         break;             // finished
3507       n = nstack.node();   // Get node from stack
3508       cnt = n->req();
3509       i = nstack.index();
3510       nstack.pop();        // Shift to the next node on stack
3511     }
3512   }
3513 
3514   // Skip next transformation if compressed oops are not used.
3515   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3516       (!UseCompressedOops && !UseCompressedClassPointers))
3517     return;
3518 
3519   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3520   // It could be done for an uncommon traps or any safepoints/calls
3521   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3522   while (sfpt.size() > 0) {
3523     n = sfpt.pop();
3524     JVMState *jvms = n->as_SafePoint()->jvms();
3525     assert(jvms != NULL, "sanity");
3526     int start = jvms->debug_start();
3527     int end   = n->req();
3528     bool is_uncommon = (n->is_CallStaticJava() &&
3529                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3530     for (int j = start; j < end; j++) {
3531       Node* in = n->in(j);
3532       if (in->is_DecodeNarrowPtr()) {
3533         bool safe_to_skip = true;
3534         if (!is_uncommon ) {
3535           // Is it safe to skip?
3536           for (uint i = 0; i < in->outcnt(); i++) {
3537             Node* u = in->raw_out(i);
3538             if (!u->is_SafePoint() ||
3539                 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3540               safe_to_skip = false;
3541             }
3542           }
3543         }
3544         if (safe_to_skip) {
3545           n->set_req(j, in->in(1));
3546         }
3547         if (in->outcnt() == 0) {
3548           in->disconnect_inputs(NULL, this);
3549         }
3550       }
3551     }
3552   }
3553 }
3554 
3555 //------------------------------final_graph_reshaping--------------------------
3556 // Final Graph Reshaping.
3557 //
3558 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3559 //     and not commoned up and forced early.  Must come after regular
3560 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3561 //     inputs to Loop Phis; these will be split by the allocator anyways.
3562 //     Remove Opaque nodes.
3563 // (2) Move last-uses by commutative operations to the left input to encourage
3564 //     Intel update-in-place two-address operations and better register usage
3565 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3566 //     calls canonicalizing them back.
3567 // (3) Count the number of double-precision FP ops, single-precision FP ops
3568 //     and call sites.  On Intel, we can get correct rounding either by
3569 //     forcing singles to memory (requires extra stores and loads after each
3570 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3571 //     clearing the mode bit around call sites).  The mode bit is only used
3572 //     if the relative frequency of single FP ops to calls is low enough.
3573 //     This is a key transform for SPEC mpeg_audio.
3574 // (4) Detect infinite loops; blobs of code reachable from above but not
3575 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3576 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3577 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3578 //     Detection is by looking for IfNodes where only 1 projection is
3579 //     reachable from below or CatchNodes missing some targets.
3580 // (5) Assert for insane oop offsets in debug mode.
3581 
3582 bool Compile::final_graph_reshaping() {
3583   // an infinite loop may have been eliminated by the optimizer,
3584   // in which case the graph will be empty.
3585   if (root()->req() == 1) {
3586     record_method_not_compilable("trivial infinite loop");
3587     return true;
3588   }
3589 
3590   // Expensive nodes have their control input set to prevent the GVN
3591   // from freely commoning them. There's no GVN beyond this point so
3592   // no need to keep the control input. We want the expensive nodes to
3593   // be freely moved to the least frequent code path by gcm.
3594   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3595   for (int i = 0; i < expensive_count(); i++) {
3596     _expensive_nodes->at(i)->set_req(0, NULL);
3597   }
3598 
3599   Final_Reshape_Counts frc;
3600 
3601   // Visit everybody reachable!
3602   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3603   Node_Stack nstack(live_nodes() >> 1);
3604   final_graph_reshaping_walk(nstack, root(), frc);
3605 
3606   // Check for unreachable (from below) code (i.e., infinite loops).
3607   for( uint i = 0; i < frc._tests.size(); i++ ) {
3608     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3609     // Get number of CFG targets.
3610     // Note that PCTables include exception targets after calls.
3611     uint required_outcnt = n->required_outcnt();
3612     if (n->outcnt() != required_outcnt) {
3613       // Check for a few special cases.  Rethrow Nodes never take the
3614       // 'fall-thru' path, so expected kids is 1 less.
3615       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3616         if (n->in(0)->in(0)->is_Call()) {
3617           CallNode *call = n->in(0)->in(0)->as_Call();
3618           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3619             required_outcnt--;      // Rethrow always has 1 less kid
3620           } else if (call->req() > TypeFunc::Parms &&
3621                      call->is_CallDynamicJava()) {
3622             // Check for null receiver. In such case, the optimizer has
3623             // detected that the virtual call will always result in a null
3624             // pointer exception. The fall-through projection of this CatchNode
3625             // will not be populated.
3626             Node *arg0 = call->in(TypeFunc::Parms);
3627             if (arg0->is_Type() &&
3628                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3629               required_outcnt--;
3630             }
3631           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3632                      call->req() > TypeFunc::Parms+1 &&
3633                      call->is_CallStaticJava()) {
3634             // Check for negative array length. In such case, the optimizer has
3635             // detected that the allocation attempt will always result in an
3636             // exception. There is no fall-through projection of this CatchNode .
3637             Node *arg1 = call->in(TypeFunc::Parms+1);
3638             if (arg1->is_Type() &&
3639                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3640               required_outcnt--;
3641             }
3642           }
3643         }
3644       }
3645       // Recheck with a better notion of 'required_outcnt'
3646       if (n->outcnt() != required_outcnt) {
3647         record_method_not_compilable("malformed control flow");
3648         return true;            // Not all targets reachable!
3649       }
3650     }
3651     // Check that I actually visited all kids.  Unreached kids
3652     // must be infinite loops.
3653     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3654       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3655         record_method_not_compilable("infinite loop");
3656         return true;            // Found unvisited kid; must be unreach
3657       }
3658 
3659     // Here so verification code in final_graph_reshaping_walk()
3660     // always see an OuterStripMinedLoopEnd
3661     if (n->is_OuterStripMinedLoopEnd()) {
3662       IfNode* init_iff = n->as_If();
3663       Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3664       n->subsume_by(iff, this);
3665     }
3666   }
3667 
3668   // If original bytecodes contained a mixture of floats and doubles
3669   // check if the optimizer has made it homogenous, item (3).
3670   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3671       frc.get_float_count() > 32 &&
3672       frc.get_double_count() == 0 &&
3673       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3674     set_24_bit_selection_and_mode( false,  true );
3675   }
3676 
3677   set_java_calls(frc.get_java_call_count());
3678   set_inner_loops(frc.get_inner_loop_count());
3679 
3680   // No infinite loops, no reason to bail out.
3681   return false;
3682 }
3683 
3684 //-----------------------------too_many_traps----------------------------------
3685 // Report if there are too many traps at the current method and bci.
3686 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3687 bool Compile::too_many_traps(ciMethod* method,
3688                              int bci,
3689                              Deoptimization::DeoptReason reason) {
3690   ciMethodData* md = method->method_data();
3691   if (md->is_empty()) {
3692     // Assume the trap has not occurred, or that it occurred only
3693     // because of a transient condition during start-up in the interpreter.
3694     return false;
3695   }
3696   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3697   if (md->has_trap_at(bci, m, reason) != 0) {
3698     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3699     // Also, if there are multiple reasons, or if there is no per-BCI record,
3700     // assume the worst.
3701     if (log())
3702       log()->elem("observe trap='%s' count='%d'",
3703                   Deoptimization::trap_reason_name(reason),
3704                   md->trap_count(reason));
3705     return true;
3706   } else {
3707     // Ignore method/bci and see if there have been too many globally.
3708     return too_many_traps(reason, md);
3709   }
3710 }
3711 
3712 // Less-accurate variant which does not require a method and bci.
3713 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3714                              ciMethodData* logmd) {
3715   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3716     // Too many traps globally.
3717     // Note that we use cumulative trap_count, not just md->trap_count.
3718     if (log()) {
3719       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3720       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3721                   Deoptimization::trap_reason_name(reason),
3722                   mcount, trap_count(reason));
3723     }
3724     return true;
3725   } else {
3726     // The coast is clear.
3727     return false;
3728   }
3729 }
3730 
3731 //--------------------------too_many_recompiles--------------------------------
3732 // Report if there are too many recompiles at the current method and bci.
3733 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3734 // Is not eager to return true, since this will cause the compiler to use
3735 // Action_none for a trap point, to avoid too many recompilations.
3736 bool Compile::too_many_recompiles(ciMethod* method,
3737                                   int bci,
3738                                   Deoptimization::DeoptReason reason) {
3739   ciMethodData* md = method->method_data();
3740   if (md->is_empty()) {
3741     // Assume the trap has not occurred, or that it occurred only
3742     // because of a transient condition during start-up in the interpreter.
3743     return false;
3744   }
3745   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3746   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3747   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3748   Deoptimization::DeoptReason per_bc_reason
3749     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3750   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3751   if ((per_bc_reason == Deoptimization::Reason_none
3752        || md->has_trap_at(bci, m, reason) != 0)
3753       // The trap frequency measure we care about is the recompile count:
3754       && md->trap_recompiled_at(bci, m)
3755       && md->overflow_recompile_count() >= bc_cutoff) {
3756     // Do not emit a trap here if it has already caused recompilations.
3757     // Also, if there are multiple reasons, or if there is no per-BCI record,
3758     // assume the worst.
3759     if (log())
3760       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3761                   Deoptimization::trap_reason_name(reason),
3762                   md->trap_count(reason),
3763                   md->overflow_recompile_count());
3764     return true;
3765   } else if (trap_count(reason) != 0
3766              && decompile_count() >= m_cutoff) {
3767     // Too many recompiles globally, and we have seen this sort of trap.
3768     // Use cumulative decompile_count, not just md->decompile_count.
3769     if (log())
3770       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3771                   Deoptimization::trap_reason_name(reason),
3772                   md->trap_count(reason), trap_count(reason),
3773                   md->decompile_count(), decompile_count());
3774     return true;
3775   } else {
3776     // The coast is clear.
3777     return false;
3778   }
3779 }
3780 
3781 // Compute when not to trap. Used by matching trap based nodes and
3782 // NullCheck optimization.
3783 void Compile::set_allowed_deopt_reasons() {
3784   _allowed_reasons = 0;
3785   if (is_method_compilation()) {
3786     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3787       assert(rs < BitsPerInt, "recode bit map");
3788       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3789         _allowed_reasons |= nth_bit(rs);
3790       }
3791     }
3792   }
3793 }
3794 
3795 bool Compile::is_compiling_clinit_for(ciKlass* k) {
3796   ciMethod* root = method(); // the root method of compilation
3797   return root->is_static_initializer() && root->holder() == k; // access in the context of clinit
3798 }
3799 
3800 #ifndef PRODUCT
3801 //------------------------------verify_graph_edges---------------------------
3802 // Walk the Graph and verify that there is a one-to-one correspondence
3803 // between Use-Def edges and Def-Use edges in the graph.
3804 void Compile::verify_graph_edges(bool no_dead_code) {
3805   if (VerifyGraphEdges) {
3806     ResourceArea *area = Thread::current()->resource_area();
3807     Unique_Node_List visited(area);
3808     // Call recursive graph walk to check edges
3809     _root->verify_edges(visited);
3810     if (no_dead_code) {
3811       // Now make sure that no visited node is used by an unvisited node.
3812       bool dead_nodes = false;
3813       Unique_Node_List checked(area);
3814       while (visited.size() > 0) {
3815         Node* n = visited.pop();
3816         checked.push(n);
3817         for (uint i = 0; i < n->outcnt(); i++) {
3818           Node* use = n->raw_out(i);
3819           if (checked.member(use))  continue;  // already checked
3820           if (visited.member(use))  continue;  // already in the graph
3821           if (use->is_Con())        continue;  // a dead ConNode is OK
3822           // At this point, we have found a dead node which is DU-reachable.
3823           if (!dead_nodes) {
3824             tty->print_cr("*** Dead nodes reachable via DU edges:");
3825             dead_nodes = true;
3826           }
3827           use->dump(2);
3828           tty->print_cr("---");
3829           checked.push(use);  // No repeats; pretend it is now checked.
3830         }
3831       }
3832       assert(!dead_nodes, "using nodes must be reachable from root");
3833     }
3834   }
3835 }
3836 
3837 // Verify GC barriers consistency
3838 // Currently supported:
3839 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3840 void Compile::verify_barriers() {
3841 #if INCLUDE_G1GC || INCLUDE_SHENANDOAHGC
3842   if (UseG1GC || UseShenandoahGC) {
3843     // Verify G1 pre-barriers
3844 
3845 #if INCLUDE_G1GC && INCLUDE_SHENANDOAHGC
3846     const int marking_offset = in_bytes(UseG1GC ? G1ThreadLocalData::satb_mark_queue_active_offset()
3847                                                 : ShenandoahThreadLocalData::satb_mark_queue_active_offset());
3848 #elif INCLUDE_G1GC
3849     const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
3850 #else
3851     const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
3852 #endif
3853 
3854     ResourceArea *area = Thread::current()->resource_area();
3855     Unique_Node_List visited(area);
3856     Node_List worklist(area);
3857     // We're going to walk control flow backwards starting from the Root
3858     worklist.push(_root);
3859     while (worklist.size() > 0) {
3860       Node* x = worklist.pop();
3861       if (x == NULL || x == top()) continue;
3862       if (visited.member(x)) {
3863         continue;
3864       } else {
3865         visited.push(x);
3866       }
3867 
3868       if (x->is_Region()) {
3869         for (uint i = 1; i < x->req(); i++) {
3870           worklist.push(x->in(i));
3871         }
3872       } else {
3873         worklist.push(x->in(0));
3874         // We are looking for the pattern:
3875         //                            /->ThreadLocal
3876         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3877         //              \->ConI(0)
3878         // We want to verify that the If and the LoadB have the same control
3879         // See GraphKit::g1_write_barrier_pre()
3880         if (x->is_If()) {
3881           IfNode *iff = x->as_If();
3882           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3883             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3884             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3885                 && cmp->in(1)->is_Load()) {
3886               LoadNode* load = cmp->in(1)->as_Load();
3887               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3888                   && load->in(2)->in(3)->is_Con()
3889                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3890 
3891                 Node* if_ctrl = iff->in(0);
3892                 Node* load_ctrl = load->in(0);
3893 
3894                 if (if_ctrl != load_ctrl) {
3895                   // Skip possible CProj->NeverBranch in infinite loops
3896                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3897                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3898                     if_ctrl = if_ctrl->in(0)->in(0);
3899                   }
3900                 }
3901                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3902               }
3903             }
3904           }
3905         }
3906       }
3907     }
3908   }
3909 #endif
3910 }
3911 
3912 #endif
3913 
3914 // The Compile object keeps track of failure reasons separately from the ciEnv.
3915 // This is required because there is not quite a 1-1 relation between the
3916 // ciEnv and its compilation task and the Compile object.  Note that one
3917 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3918 // to backtrack and retry without subsuming loads.  Other than this backtracking
3919 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3920 // by the logic in C2Compiler.
3921 void Compile::record_failure(const char* reason) {
3922   if (log() != NULL) {
3923     log()->elem("failure reason='%s' phase='compile'", reason);
3924   }
3925   if (_failure_reason == NULL) {
3926     // Record the first failure reason.
3927     _failure_reason = reason;
3928   }
3929 
3930   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3931     C->print_method(PHASE_FAILURE);
3932   }
3933   _root = NULL;  // flush the graph, too
3934 }
3935 
3936 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3937   : TraceTime(name, accumulator, CITime, CITimeVerbose),
3938     _phase_name(name), _dolog(CITimeVerbose)
3939 {
3940   if (_dolog) {
3941     C = Compile::current();
3942     _log = C->log();
3943   } else {
3944     C = NULL;
3945     _log = NULL;
3946   }
3947   if (_log != NULL) {
3948     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3949     _log->stamp();
3950     _log->end_head();
3951   }
3952 }
3953 
3954 Compile::TracePhase::~TracePhase() {
3955 
3956   C = Compile::current();
3957   if (_dolog) {
3958     _log = C->log();
3959   } else {
3960     _log = NULL;
3961   }
3962 
3963 #ifdef ASSERT
3964   if (PrintIdealNodeCount) {
3965     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3966                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3967   }
3968 
3969   if (VerifyIdealNodeCount) {
3970     Compile::current()->print_missing_nodes();
3971   }
3972 #endif
3973 
3974   if (_log != NULL) {
3975     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3976   }
3977 }
3978 
3979 //=============================================================================
3980 // Two Constant's are equal when the type and the value are equal.
3981 bool Compile::Constant::operator==(const Constant& other) {
3982   if (type()          != other.type()         )  return false;
3983   if (can_be_reused() != other.can_be_reused())  return false;
3984   // For floating point values we compare the bit pattern.
3985   switch (type()) {
3986   case T_INT:
3987   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3988   case T_LONG:
3989   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
3990   case T_OBJECT:
3991   case T_ADDRESS: return (_v._value.l == other._v._value.l);
3992   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
3993   case T_METADATA: return (_v._metadata == other._v._metadata);
3994   default: ShouldNotReachHere(); return false;
3995   }
3996 }
3997 
3998 static int type_to_size_in_bytes(BasicType t) {
3999   switch (t) {
4000   case T_INT:     return sizeof(jint   );
4001   case T_LONG:    return sizeof(jlong  );
4002   case T_FLOAT:   return sizeof(jfloat );
4003   case T_DOUBLE:  return sizeof(jdouble);
4004   case T_METADATA: return sizeof(Metadata*);
4005     // We use T_VOID as marker for jump-table entries (labels) which
4006     // need an internal word relocation.
4007   case T_VOID:
4008   case T_ADDRESS:
4009   case T_OBJECT:  return sizeof(jobject);
4010   default:
4011     ShouldNotReachHere();
4012     return -1;
4013   }
4014 }
4015 
4016 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4017   // sort descending
4018   if (a->freq() > b->freq())  return -1;
4019   if (a->freq() < b->freq())  return  1;
4020   return 0;
4021 }
4022 
4023 void Compile::ConstantTable::calculate_offsets_and_size() {
4024   // First, sort the array by frequencies.
4025   _constants.sort(qsort_comparator);
4026 
4027 #ifdef ASSERT
4028   // Make sure all jump-table entries were sorted to the end of the
4029   // array (they have a negative frequency).
4030   bool found_void = false;
4031   for (int i = 0; i < _constants.length(); i++) {
4032     Constant con = _constants.at(i);
4033     if (con.type() == T_VOID)
4034       found_void = true;  // jump-tables
4035     else
4036       assert(!found_void, "wrong sorting");
4037   }
4038 #endif
4039 
4040   int offset = 0;
4041   for (int i = 0; i < _constants.length(); i++) {
4042     Constant* con = _constants.adr_at(i);
4043 
4044     // Align offset for type.
4045     int typesize = type_to_size_in_bytes(con->type());
4046     offset = align_up(offset, typesize);
4047     con->set_offset(offset);   // set constant's offset
4048 
4049     if (con->type() == T_VOID) {
4050       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4051       offset = offset + typesize * n->outcnt();  // expand jump-table
4052     } else {
4053       offset = offset + typesize;
4054     }
4055   }
4056 
4057   // Align size up to the next section start (which is insts; see
4058   // CodeBuffer::align_at_start).
4059   assert(_size == -1, "already set?");
4060   _size = align_up(offset, (int)CodeEntryAlignment);
4061 }
4062 
4063 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4064   MacroAssembler _masm(&cb);
4065   for (int i = 0; i < _constants.length(); i++) {
4066     Constant con = _constants.at(i);
4067     address constant_addr = NULL;
4068     switch (con.type()) {
4069     case T_INT:    constant_addr = _masm.int_constant(   con.get_jint()   ); break;
4070     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
4071     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4072     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4073     case T_OBJECT: {
4074       jobject obj = con.get_jobject();
4075       int oop_index = _masm.oop_recorder()->find_index(obj);
4076       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4077       break;
4078     }
4079     case T_ADDRESS: {
4080       address addr = (address) con.get_jobject();
4081       constant_addr = _masm.address_constant(addr);
4082       break;
4083     }
4084     // We use T_VOID as marker for jump-table entries (labels) which
4085     // need an internal word relocation.
4086     case T_VOID: {
4087       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4088       // Fill the jump-table with a dummy word.  The real value is
4089       // filled in later in fill_jump_table.
4090       address dummy = (address) n;
4091       constant_addr = _masm.address_constant(dummy);
4092       // Expand jump-table
4093       for (uint i = 1; i < n->outcnt(); i++) {
4094         address temp_addr = _masm.address_constant(dummy + i);
4095         assert(temp_addr, "consts section too small");
4096       }
4097       break;
4098     }
4099     case T_METADATA: {
4100       Metadata* obj = con.get_metadata();
4101       int metadata_index = _masm.oop_recorder()->find_index(obj);
4102       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4103       break;
4104     }
4105     default: ShouldNotReachHere();
4106     }
4107     assert(constant_addr, "consts section too small");
4108     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4109             "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4110   }
4111 }
4112 
4113 int Compile::ConstantTable::find_offset(Constant& con) const {
4114   int idx = _constants.find(con);
4115   guarantee(idx != -1, "constant must be in constant table");
4116   int offset = _constants.at(idx).offset();
4117   guarantee(offset != -1, "constant table not emitted yet?");
4118   return offset;
4119 }
4120 
4121 void Compile::ConstantTable::add(Constant& con) {
4122   if (con.can_be_reused()) {
4123     int idx = _constants.find(con);
4124     if (idx != -1 && _constants.at(idx).can_be_reused()) {
4125       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
4126       return;
4127     }
4128   }
4129   (void) _constants.append(con);
4130 }
4131 
4132 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4133   Block* b = Compile::current()->cfg()->get_block_for_node(n);
4134   Constant con(type, value, b->_freq);
4135   add(con);
4136   return con;
4137 }
4138 
4139 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4140   Constant con(metadata);
4141   add(con);
4142   return con;
4143 }
4144 
4145 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4146   jvalue value;
4147   BasicType type = oper->type()->basic_type();
4148   switch (type) {
4149   case T_LONG:    value.j = oper->constantL(); break;
4150   case T_FLOAT:   value.f = oper->constantF(); break;
4151   case T_DOUBLE:  value.d = oper->constantD(); break;
4152   case T_OBJECT:
4153   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4154   case T_METADATA: return add((Metadata*)oper->constant()); break;
4155   default: guarantee(false, "unhandled type: %s", type2name(type));
4156   }
4157   return add(n, type, value);
4158 }
4159 
4160 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4161   jvalue value;
4162   // We can use the node pointer here to identify the right jump-table
4163   // as this method is called from Compile::Fill_buffer right before
4164   // the MachNodes are emitted and the jump-table is filled (means the
4165   // MachNode pointers do not change anymore).
4166   value.l = (jobject) n;
4167   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
4168   add(con);
4169   return con;
4170 }
4171 
4172 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4173   // If called from Compile::scratch_emit_size do nothing.
4174   if (Compile::current()->in_scratch_emit_size())  return;
4175 
4176   assert(labels.is_nonempty(), "must be");
4177   assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4178 
4179   // Since MachConstantNode::constant_offset() also contains
4180   // table_base_offset() we need to subtract the table_base_offset()
4181   // to get the plain offset into the constant table.
4182   int offset = n->constant_offset() - table_base_offset();
4183 
4184   MacroAssembler _masm(&cb);
4185   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4186 
4187   for (uint i = 0; i < n->outcnt(); i++) {
4188     address* constant_addr = &jump_table_base[i];
4189     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));
4190     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4191     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4192   }
4193 }
4194 
4195 //----------------------------static_subtype_check-----------------------------
4196 // Shortcut important common cases when superklass is exact:
4197 // (0) superklass is java.lang.Object (can occur in reflective code)
4198 // (1) subklass is already limited to a subtype of superklass => always ok
4199 // (2) subklass does not overlap with superklass => always fail
4200 // (3) superklass has NO subtypes and we can check with a simple compare.
4201 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4202   if (StressReflectiveCode) {
4203     return SSC_full_test;       // Let caller generate the general case.
4204   }
4205 
4206   if (superk == env()->Object_klass()) {
4207     return SSC_always_true;     // (0) this test cannot fail
4208   }
4209 
4210   ciType* superelem = superk;
4211   if (superelem->is_array_klass())
4212     superelem = superelem->as_array_klass()->base_element_type();
4213 
4214   if (!subk->is_interface()) {  // cannot trust static interface types yet
4215     if (subk->is_subtype_of(superk)) {
4216       return SSC_always_true;   // (1) false path dead; no dynamic test needed
4217     }
4218     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4219         !superk->is_subtype_of(subk)) {
4220       return SSC_always_false;
4221     }
4222   }
4223 
4224   // If casting to an instance klass, it must have no subtypes
4225   if (superk->is_interface()) {
4226     // Cannot trust interfaces yet.
4227     // %%% S.B. superk->nof_implementors() == 1
4228   } else if (superelem->is_instance_klass()) {
4229     ciInstanceKlass* ik = superelem->as_instance_klass();
4230     if (!ik->has_subklass() && !ik->is_interface()) {
4231       if (!ik->is_final()) {
4232         // Add a dependency if there is a chance of a later subclass.
4233         dependencies()->assert_leaf_type(ik);
4234       }
4235       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
4236     }
4237   } else {
4238     // A primitive array type has no subtypes.
4239     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
4240   }
4241 
4242   return SSC_full_test;
4243 }
4244 
4245 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4246 #ifdef _LP64
4247   // The scaled index operand to AddP must be a clean 64-bit value.
4248   // Java allows a 32-bit int to be incremented to a negative
4249   // value, which appears in a 64-bit register as a large
4250   // positive number.  Using that large positive number as an
4251   // operand in pointer arithmetic has bad consequences.
4252   // On the other hand, 32-bit overflow is rare, and the possibility
4253   // can often be excluded, if we annotate the ConvI2L node with
4254   // a type assertion that its value is known to be a small positive
4255   // number.  (The prior range check has ensured this.)
4256   // This assertion is used by ConvI2LNode::Ideal.
4257   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4258   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4259   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4260   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4261 #endif
4262   return idx;
4263 }
4264 
4265 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4266 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4267   if (ctrl != NULL) {
4268     // Express control dependency by a CastII node with a narrow type.
4269     value = new CastIINode(value, itype, false, true /* range check dependency */);
4270     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4271     // node from floating above the range check during loop optimizations. Otherwise, the
4272     // ConvI2L node may be eliminated independently of the range check, causing the data path
4273     // to become TOP while the control path is still there (although it's unreachable).
4274     value->set_req(0, ctrl);
4275     // Save CastII node to remove it after loop optimizations.
4276     phase->C->add_range_check_cast(value);
4277     value = phase->transform(value);
4278   }
4279   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4280   return phase->transform(new ConvI2LNode(value, ltype));
4281 }
4282 
4283 // The message about the current inlining is accumulated in
4284 // _print_inlining_stream and transfered into the _print_inlining_list
4285 // once we know whether inlining succeeds or not. For regular
4286 // inlining, messages are appended to the buffer pointed by
4287 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4288 // a new buffer is added after _print_inlining_idx in the list. This
4289 // way we can update the inlining message for late inlining call site
4290 // when the inlining is attempted again.
4291 void Compile::print_inlining_init() {
4292   if (print_inlining() || print_intrinsics()) {
4293     _print_inlining_stream = new stringStream();
4294     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4295   }
4296 }
4297 
4298 void Compile::print_inlining_reinit() {
4299   if (print_inlining() || print_intrinsics()) {
4300     // Re allocate buffer when we change ResourceMark
4301     _print_inlining_stream = new stringStream();
4302   }
4303 }
4304 
4305 void Compile::print_inlining_reset() {
4306   _print_inlining_stream->reset();
4307 }
4308 
4309 void Compile::print_inlining_commit() {
4310   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4311   // Transfer the message from _print_inlining_stream to the current
4312   // _print_inlining_list buffer and clear _print_inlining_stream.
4313   _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4314   print_inlining_reset();
4315 }
4316 
4317 void Compile::print_inlining_push() {
4318   // Add new buffer to the _print_inlining_list at current position
4319   _print_inlining_idx++;
4320   _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4321 }
4322 
4323 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4324   return _print_inlining_list->at(_print_inlining_idx);
4325 }
4326 
4327 void Compile::print_inlining_update(CallGenerator* cg) {
4328   if (print_inlining() || print_intrinsics()) {
4329     if (!cg->is_late_inline()) {
4330       if (print_inlining_current().cg() != NULL) {
4331         print_inlining_push();
4332       }
4333       print_inlining_commit();
4334     } else {
4335       if (print_inlining_current().cg() != cg &&
4336           (print_inlining_current().cg() != NULL ||
4337            print_inlining_current().ss()->size() != 0)) {
4338         print_inlining_push();
4339       }
4340       print_inlining_commit();
4341       print_inlining_current().set_cg(cg);
4342     }
4343   }
4344 }
4345 
4346 void Compile::print_inlining_move_to(CallGenerator* cg) {
4347   // We resume inlining at a late inlining call site. Locate the
4348   // corresponding inlining buffer so that we can update it.
4349   if (print_inlining()) {
4350     for (int i = 0; i < _print_inlining_list->length(); i++) {
4351       if (_print_inlining_list->adr_at(i)->cg() == cg) {
4352         _print_inlining_idx = i;
4353         return;
4354       }
4355     }
4356     ShouldNotReachHere();
4357   }
4358 }
4359 
4360 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4361   if (print_inlining()) {
4362     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4363     assert(print_inlining_current().cg() == cg, "wrong entry");
4364     // replace message with new message
4365     _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4366     print_inlining_commit();
4367     print_inlining_current().set_cg(cg);
4368   }
4369 }
4370 
4371 void Compile::print_inlining_assert_ready() {
4372   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4373 }
4374 
4375 void Compile::process_print_inlining() {
4376   bool do_print_inlining = print_inlining() || print_intrinsics();
4377   if (do_print_inlining || log() != NULL) {
4378     // Print inlining message for candidates that we couldn't inline
4379     // for lack of space
4380     for (int i = 0; i < _late_inlines.length(); i++) {
4381       CallGenerator* cg = _late_inlines.at(i);
4382       if (!cg->is_mh_late_inline()) {
4383         const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4384         if (do_print_inlining) {
4385           cg->print_inlining_late(msg);
4386         }
4387         log_late_inline_failure(cg, msg);
4388       }
4389     }
4390   }
4391   if (do_print_inlining) {
4392     ResourceMark rm;
4393     stringStream ss;
4394     for (int i = 0; i < _print_inlining_list->length(); i++) {
4395       ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4396     }
4397     size_t end = ss.size();
4398     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4399     strncpy(_print_inlining_output, ss.base(), end+1);
4400     _print_inlining_output[end] = 0;
4401   }
4402 }
4403 
4404 void Compile::dump_print_inlining() {
4405   if (_print_inlining_output != NULL) {
4406     tty->print_raw(_print_inlining_output);
4407   }
4408 }
4409 
4410 void Compile::log_late_inline(CallGenerator* cg) {
4411   if (log() != NULL) {
4412     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4413                 cg->unique_id());
4414     JVMState* p = cg->call_node()->jvms();
4415     while (p != NULL) {
4416       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4417       p = p->caller();
4418     }
4419     log()->tail("late_inline");
4420   }
4421 }
4422 
4423 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4424   log_late_inline(cg);
4425   if (log() != NULL) {
4426     log()->inline_fail(msg);
4427   }
4428 }
4429 
4430 void Compile::log_inline_id(CallGenerator* cg) {
4431   if (log() != NULL) {
4432     // The LogCompilation tool needs a unique way to identify late
4433     // inline call sites. This id must be unique for this call site in
4434     // this compilation. Try to have it unique across compilations as
4435     // well because it can be convenient when grepping through the log
4436     // file.
4437     // Distinguish OSR compilations from others in case CICountOSR is
4438     // on.
4439     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4440     cg->set_unique_id(id);
4441     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4442   }
4443 }
4444 
4445 void Compile::log_inline_failure(const char* msg) {
4446   if (C->log() != NULL) {
4447     C->log()->inline_fail(msg);
4448   }
4449 }
4450 
4451 
4452 // Dump inlining replay data to the stream.
4453 // Don't change thread state and acquire any locks.
4454 void Compile::dump_inline_data(outputStream* out) {
4455   InlineTree* inl_tree = ilt();
4456   if (inl_tree != NULL) {
4457     out->print(" inline %d", inl_tree->count());
4458     inl_tree->dump_replay_data(out);
4459   }
4460 }
4461 
4462 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4463   if (n1->Opcode() < n2->Opcode())      return -1;
4464   else if (n1->Opcode() > n2->Opcode()) return 1;
4465 
4466   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4467   for (uint i = 1; i < n1->req(); i++) {
4468     if (n1->in(i) < n2->in(i))      return -1;
4469     else if (n1->in(i) > n2->in(i)) return 1;
4470   }
4471 
4472   return 0;
4473 }
4474 
4475 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4476   Node* n1 = *n1p;
4477   Node* n2 = *n2p;
4478 
4479   return cmp_expensive_nodes(n1, n2);
4480 }
4481 
4482 void Compile::sort_expensive_nodes() {
4483   if (!expensive_nodes_sorted()) {
4484     _expensive_nodes->sort(cmp_expensive_nodes);
4485   }
4486 }
4487 
4488 bool Compile::expensive_nodes_sorted() const {
4489   for (int i = 1; i < _expensive_nodes->length(); i++) {
4490     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4491       return false;
4492     }
4493   }
4494   return true;
4495 }
4496 
4497 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4498   if (_expensive_nodes->length() == 0) {
4499     return false;
4500   }
4501 
4502   assert(OptimizeExpensiveOps, "optimization off?");
4503 
4504   // Take this opportunity to remove dead nodes from the list
4505   int j = 0;
4506   for (int i = 0; i < _expensive_nodes->length(); i++) {
4507     Node* n = _expensive_nodes->at(i);
4508     if (!n->is_unreachable(igvn)) {
4509       assert(n->is_expensive(), "should be expensive");
4510       _expensive_nodes->at_put(j, n);
4511       j++;
4512     }
4513   }
4514   _expensive_nodes->trunc_to(j);
4515 
4516   // Then sort the list so that similar nodes are next to each other
4517   // and check for at least two nodes of identical kind with same data
4518   // inputs.
4519   sort_expensive_nodes();
4520 
4521   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4522     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4523       return true;
4524     }
4525   }
4526 
4527   return false;
4528 }
4529 
4530 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4531   if (_expensive_nodes->length() == 0) {
4532     return;
4533   }
4534 
4535   assert(OptimizeExpensiveOps, "optimization off?");
4536 
4537   // Sort to bring similar nodes next to each other and clear the
4538   // control input of nodes for which there's only a single copy.
4539   sort_expensive_nodes();
4540 
4541   int j = 0;
4542   int identical = 0;
4543   int i = 0;
4544   bool modified = false;
4545   for (; i < _expensive_nodes->length()-1; i++) {
4546     assert(j <= i, "can't write beyond current index");
4547     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4548       identical++;
4549       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4550       continue;
4551     }
4552     if (identical > 0) {
4553       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4554       identical = 0;
4555     } else {
4556       Node* n = _expensive_nodes->at(i);
4557       igvn.replace_input_of(n, 0, NULL);
4558       igvn.hash_insert(n);
4559       modified = true;
4560     }
4561   }
4562   if (identical > 0) {
4563     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4564   } else if (_expensive_nodes->length() >= 1) {
4565     Node* n = _expensive_nodes->at(i);
4566     igvn.replace_input_of(n, 0, NULL);
4567     igvn.hash_insert(n);
4568     modified = true;
4569   }
4570   _expensive_nodes->trunc_to(j);
4571   if (modified) {
4572     igvn.optimize();
4573   }
4574 }
4575 
4576 void Compile::add_expensive_node(Node * n) {
4577   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4578   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4579   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4580   if (OptimizeExpensiveOps) {
4581     _expensive_nodes->append(n);
4582   } else {
4583     // Clear control input and let IGVN optimize expensive nodes if
4584     // OptimizeExpensiveOps is off.
4585     n->set_req(0, NULL);
4586   }
4587 }
4588 
4589 /**
4590  * Remove the speculative part of types and clean up the graph
4591  */
4592 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4593   if (UseTypeSpeculation) {
4594     Unique_Node_List worklist;
4595     worklist.push(root());
4596     int modified = 0;
4597     // Go over all type nodes that carry a speculative type, drop the
4598     // speculative part of the type and enqueue the node for an igvn
4599     // which may optimize it out.
4600     for (uint next = 0; next < worklist.size(); ++next) {
4601       Node *n  = worklist.at(next);
4602       if (n->is_Type()) {
4603         TypeNode* tn = n->as_Type();
4604         const Type* t = tn->type();
4605         const Type* t_no_spec = t->remove_speculative();
4606         if (t_no_spec != t) {
4607           bool in_hash = igvn.hash_delete(n);
4608           assert(in_hash || (UseShenandoahGC && n->hash() == Node::NO_HASH), "node should be in igvn hash table");
4609           tn->set_type(t_no_spec);
4610           igvn.hash_insert(n);
4611           igvn._worklist.push(n); // give it a chance to go away
4612           modified++;
4613         }
4614       }
4615       uint max = n->len();
4616       for( uint i = 0; i < max; ++i ) {
4617         Node *m = n->in(i);
4618         if (not_a_node(m))  continue;
4619         worklist.push(m);
4620       }
4621     }
4622     // Drop the speculative part of all types in the igvn's type table
4623     igvn.remove_speculative_types();
4624     if (modified > 0) {
4625       igvn.optimize();
4626     }
4627 #ifdef ASSERT
4628     // Verify that after the IGVN is over no speculative type has resurfaced
4629     worklist.clear();
4630     worklist.push(root());
4631     for (uint next = 0; next < worklist.size(); ++next) {
4632       Node *n  = worklist.at(next);
4633       const Type* t = igvn.type_or_null(n);
4634       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4635       if (n->is_Type()) {
4636         t = n->as_Type()->type();
4637         assert(t == t->remove_speculative(), "no more speculative types");
4638       }
4639       uint max = n->len();
4640       for( uint i = 0; i < max; ++i ) {
4641         Node *m = n->in(i);
4642         if (not_a_node(m))  continue;
4643         worklist.push(m);
4644       }
4645     }
4646     igvn.check_no_speculative_types();
4647 #endif
4648   }
4649 }
4650 
4651 // Auxiliary method to support randomized stressing/fuzzing.
4652 //
4653 // This method can be called the arbitrary number of times, with current count
4654 // as the argument. The logic allows selecting a single candidate from the
4655 // running list of candidates as follows:
4656 //    int count = 0;
4657 //    Cand* selected = null;
4658 //    while(cand = cand->next()) {
4659 //      if (randomized_select(++count)) {
4660 //        selected = cand;
4661 //      }
4662 //    }
4663 //
4664 // Including count equalizes the chances any candidate is "selected".
4665 // This is useful when we don't have the complete list of candidates to choose
4666 // from uniformly. In this case, we need to adjust the randomicity of the
4667 // selection, or else we will end up biasing the selection towards the latter
4668 // candidates.
4669 //
4670 // Quick back-envelope calculation shows that for the list of n candidates
4671 // the equal probability for the candidate to persist as "best" can be
4672 // achieved by replacing it with "next" k-th candidate with the probability
4673 // of 1/k. It can be easily shown that by the end of the run, the
4674 // probability for any candidate is converged to 1/n, thus giving the
4675 // uniform distribution among all the candidates.
4676 //
4677 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4678 #define RANDOMIZED_DOMAIN_POW 29
4679 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4680 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4681 bool Compile::randomized_select(int count) {
4682   assert(count > 0, "only positive");
4683   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4684 }
4685 
4686 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4687 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4688 
4689 void NodeCloneInfo::dump() const {
4690   tty->print(" {%d:%d} ", idx(), gen());
4691 }
4692 
4693 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4694   uint64_t val = value(old->_idx);
4695   NodeCloneInfo cio(val);
4696   assert(val != 0, "old node should be in the map");
4697   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4698   insert(nnn->_idx, cin.get());
4699 #ifndef PRODUCT
4700   if (is_debug()) {
4701     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4702   }
4703 #endif
4704 }
4705 
4706 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4707   NodeCloneInfo cio(value(old->_idx));
4708   if (cio.get() == 0) {
4709     cio.set(old->_idx, 0);
4710     insert(old->_idx, cio.get());
4711 #ifndef PRODUCT
4712     if (is_debug()) {
4713       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4714     }
4715 #endif
4716   }
4717   clone(old, nnn, gen);
4718 }
4719 
4720 int CloneMap::max_gen() const {
4721   int g = 0;
4722   DictI di(_dict);
4723   for(; di.test(); ++di) {
4724     int t = gen(di._key);
4725     if (g < t) {
4726       g = t;
4727 #ifndef PRODUCT
4728       if (is_debug()) {
4729         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4730       }
4731 #endif
4732     }
4733   }
4734   return g;
4735 }
4736 
4737 void CloneMap::dump(node_idx_t key) const {
4738   uint64_t val = value(key);
4739   if (val != 0) {
4740     NodeCloneInfo ni(val);
4741     ni.dump();
4742   }
4743 }