1 /*
   2  * Copyright (c) 1997, 2023, 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/javaClasses.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 "jfr/jfrEvents.hpp"
  39 #include "jvm_io.h"
  40 #include "memory/allocation.hpp"
  41 #include "memory/resourceArea.hpp"
  42 #include "opto/addnode.hpp"
  43 #include "opto/block.hpp"
  44 #include "opto/c2compiler.hpp"
  45 #include "opto/callGenerator.hpp"
  46 #include "opto/callnode.hpp"
  47 #include "opto/castnode.hpp"
  48 #include "opto/cfgnode.hpp"
  49 #include "opto/chaitin.hpp"
  50 #include "opto/compile.hpp"
  51 #include "opto/connode.hpp"
  52 #include "opto/convertnode.hpp"
  53 #include "opto/divnode.hpp"
  54 #include "opto/escape.hpp"
  55 #include "opto/idealGraphPrinter.hpp"
  56 #include "opto/loopnode.hpp"
  57 #include "opto/machnode.hpp"
  58 #include "opto/macro.hpp"
  59 #include "opto/matcher.hpp"
  60 #include "opto/mathexactnode.hpp"
  61 #include "opto/memnode.hpp"
  62 #include "opto/mulnode.hpp"
  63 #include "opto/narrowptrnode.hpp"
  64 #include "opto/node.hpp"
  65 #include "opto/opcodes.hpp"
  66 #include "opto/output.hpp"
  67 #include "opto/parse.hpp"
  68 #include "opto/phaseX.hpp"
  69 #include "opto/rootnode.hpp"
  70 #include "opto/runtime.hpp"
  71 #include "opto/stringopts.hpp"
  72 #include "opto/type.hpp"
  73 #include "opto/vector.hpp"
  74 #include "opto/vectornode.hpp"
  75 #include "runtime/globals_extension.hpp"
  76 #include "runtime/sharedRuntime.hpp"
  77 #include "runtime/signature.hpp"
  78 #include "runtime/stubRoutines.hpp"
  79 #include "runtime/timer.hpp"
  80 #include "utilities/align.hpp"
  81 #include "utilities/copy.hpp"
  82 #include "utilities/macros.hpp"
  83 #include "utilities/resourceHash.hpp"
  84 
  85 // -------------------- Compile::mach_constant_base_node -----------------------
  86 // Constant table base node singleton.
  87 MachConstantBaseNode* Compile::mach_constant_base_node() {
  88   if (_mach_constant_base_node == NULL) {
  89     _mach_constant_base_node = new MachConstantBaseNode();
  90     _mach_constant_base_node->add_req(C->root());
  91   }
  92   return _mach_constant_base_node;
  93 }
  94 
  95 
  96 /// Support for intrinsics.
  97 
  98 // Return the index at which m must be inserted (or already exists).
  99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
 100 class IntrinsicDescPair {
 101  private:
 102   ciMethod* _m;
 103   bool _is_virtual;
 104  public:
 105   IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
 106   static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
 107     ciMethod* m= elt->method();
 108     ciMethod* key_m = key->_m;
 109     if (key_m < m)      return -1;
 110     else if (key_m > m) return 1;
 111     else {
 112       bool is_virtual = elt->is_virtual();
 113       bool key_virtual = key->_is_virtual;
 114       if (key_virtual < is_virtual)      return -1;
 115       else if (key_virtual > is_virtual) return 1;
 116       else                               return 0;
 117     }
 118   }
 119 };
 120 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
 121 #ifdef ASSERT
 122   for (int i = 1; i < _intrinsics.length(); i++) {
 123     CallGenerator* cg1 = _intrinsics.at(i-1);
 124     CallGenerator* cg2 = _intrinsics.at(i);
 125     assert(cg1->method() != cg2->method()
 126            ? cg1->method()     < cg2->method()
 127            : cg1->is_virtual() < cg2->is_virtual(),
 128            "compiler intrinsics list must stay sorted");
 129   }
 130 #endif
 131   IntrinsicDescPair pair(m, is_virtual);
 132   return _intrinsics.find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
 133 }
 134 
 135 void Compile::register_intrinsic(CallGenerator* cg) {
 136   bool found = false;
 137   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
 138   assert(!found, "registering twice");
 139   _intrinsics.insert_before(index, cg);
 140   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 141 }
 142 
 143 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 144   assert(m->is_loaded(), "don't try this on unloaded methods");
 145   if (_intrinsics.length() > 0) {
 146     bool found = false;
 147     int index = intrinsic_insertion_index(m, is_virtual, found);
 148      if (found) {
 149       return _intrinsics.at(index);
 150     }
 151   }
 152   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 153   if (m->intrinsic_id() != vmIntrinsics::_none &&
 154       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 155     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 156     if (cg != NULL) {
 157       // Save it for next time:
 158       register_intrinsic(cg);
 159       return cg;
 160     } else {
 161       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 162     }
 163   }
 164   return NULL;
 165 }
 166 
 167 // Compile::make_vm_intrinsic is defined in library_call.cpp.
 168 
 169 #ifndef PRODUCT
 170 // statistics gathering...
 171 
 172 juint  Compile::_intrinsic_hist_count[vmIntrinsics::number_of_intrinsics()] = {0};
 173 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::number_of_intrinsics()] = {0};
 174 
 175 inline int as_int(vmIntrinsics::ID id) {
 176   return vmIntrinsics::as_int(id);
 177 }
 178 
 179 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 180   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 181   int oflags = _intrinsic_hist_flags[as_int(id)];
 182   assert(flags != 0, "what happened?");
 183   if (is_virtual) {
 184     flags |= _intrinsic_virtual;
 185   }
 186   bool changed = (flags != oflags);
 187   if ((flags & _intrinsic_worked) != 0) {
 188     juint count = (_intrinsic_hist_count[as_int(id)] += 1);
 189     if (count == 1) {
 190       changed = true;           // first time
 191     }
 192     // increment the overall count also:
 193     _intrinsic_hist_count[as_int(vmIntrinsics::_none)] += 1;
 194   }
 195   if (changed) {
 196     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 197       // Something changed about the intrinsic's virtuality.
 198       if ((flags & _intrinsic_virtual) != 0) {
 199         // This is the first use of this intrinsic as a virtual call.
 200         if (oflags != 0) {
 201           // We already saw it as a non-virtual, so note both cases.
 202           flags |= _intrinsic_both;
 203         }
 204       } else if ((oflags & _intrinsic_both) == 0) {
 205         // This is the first use of this intrinsic as a non-virtual
 206         flags |= _intrinsic_both;
 207       }
 208     }
 209     _intrinsic_hist_flags[as_int(id)] = (jubyte) (oflags | flags);
 210   }
 211   // update the overall flags also:
 212   _intrinsic_hist_flags[as_int(vmIntrinsics::_none)] |= (jubyte) flags;
 213   return changed;
 214 }
 215 
 216 static char* format_flags(int flags, char* buf) {
 217   buf[0] = 0;
 218   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 219   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 220   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 221   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 222   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 223   if (buf[0] == 0)  strcat(buf, ",");
 224   assert(buf[0] == ',', "must be");
 225   return &buf[1];
 226 }
 227 
 228 void Compile::print_intrinsic_statistics() {
 229   char flagsbuf[100];
 230   ttyLocker ttyl;
 231   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 232   tty->print_cr("Compiler intrinsic usage:");
 233   juint total = _intrinsic_hist_count[as_int(vmIntrinsics::_none)];
 234   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 235   #define PRINT_STAT_LINE(name, c, f) \
 236     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 237   for (auto id : EnumRange<vmIntrinsicID>{}) {
 238     int   flags = _intrinsic_hist_flags[as_int(id)];
 239     juint count = _intrinsic_hist_count[as_int(id)];
 240     if ((flags | count) != 0) {
 241       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 242     }
 243   }
 244   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[as_int(vmIntrinsics::_none)], flagsbuf));
 245   if (xtty != NULL)  xtty->tail("statistics");
 246 }
 247 
 248 void Compile::print_statistics() {
 249   { ttyLocker ttyl;
 250     if (xtty != NULL)  xtty->head("statistics type='opto'");
 251     Parse::print_statistics();
 252     PhaseStringOpts::print_statistics();
 253     PhaseCCP::print_statistics();
 254     PhaseRegAlloc::print_statistics();
 255     PhaseOutput::print_statistics();
 256     PhasePeephole::print_statistics();
 257     PhaseIdealLoop::print_statistics();
 258     ConnectionGraph::print_statistics();
 259     PhaseMacroExpand::print_statistics();
 260     if (xtty != NULL)  xtty->tail("statistics");
 261   }
 262   if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
 263     // put this under its own <statistics> element.
 264     print_intrinsic_statistics();
 265   }
 266 }
 267 #endif //PRODUCT
 268 
 269 void Compile::gvn_replace_by(Node* n, Node* nn) {
 270   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 271     Node* use = n->last_out(i);
 272     bool is_in_table = initial_gvn()->hash_delete(use);
 273     uint uses_found = 0;
 274     for (uint j = 0; j < use->len(); j++) {
 275       if (use->in(j) == n) {
 276         if (j < use->req())
 277           use->set_req(j, nn);
 278         else
 279           use->set_prec(j, nn);
 280         uses_found++;
 281       }
 282     }
 283     if (is_in_table) {
 284       // reinsert into table
 285       initial_gvn()->hash_find_insert(use);
 286     }
 287     record_for_igvn(use);
 288     i -= uses_found;    // we deleted 1 or more copies of this edge
 289   }
 290 }
 291 
 292 
 293 // Identify all nodes that are reachable from below, useful.
 294 // Use breadth-first pass that records state in a Unique_Node_List,
 295 // recursive traversal is slower.
 296 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 297   int estimated_worklist_size = live_nodes();
 298   useful.map( estimated_worklist_size, NULL );  // preallocate space
 299 
 300   // Initialize worklist
 301   if (root() != NULL)     { useful.push(root()); }
 302   // If 'top' is cached, declare it useful to preserve cached node
 303   if( cached_top_node() ) { useful.push(cached_top_node()); }
 304 
 305   // Push all useful nodes onto the list, breadthfirst
 306   for( uint next = 0; next < useful.size(); ++next ) {
 307     assert( next < unique(), "Unique useful nodes < total nodes");
 308     Node *n  = useful.at(next);
 309     uint max = n->len();
 310     for( uint i = 0; i < max; ++i ) {
 311       Node *m = n->in(i);
 312       if (not_a_node(m))  continue;
 313       useful.push(m);
 314     }
 315   }
 316 }
 317 
 318 // Update dead_node_list with any missing dead nodes using useful
 319 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 320 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 321   uint max_idx = unique();
 322   VectorSet& useful_node_set = useful.member_set();
 323 
 324   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 325     // If node with index node_idx is not in useful set,
 326     // mark it as dead in dead node list.
 327     if (!useful_node_set.test(node_idx)) {
 328       record_dead_node(node_idx);
 329     }
 330   }
 331 }
 332 
 333 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 334   int shift = 0;
 335   for (int i = 0; i < inlines->length(); i++) {
 336     CallGenerator* cg = inlines->at(i);
 337     if (useful.member(cg->call_node())) {
 338       if (shift > 0) {
 339         inlines->at_put(i - shift, cg);
 340       }
 341     } else {
 342       shift++; // skip over the dead element
 343     }
 344   }
 345   if (shift > 0) {
 346     inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
 347   }
 348 }
 349 
 350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
 351   assert(dead != NULL && dead->is_Call(), "sanity");
 352   int found = 0;
 353   for (int i = 0; i < inlines->length(); i++) {
 354     if (inlines->at(i)->call_node() == dead) {
 355       inlines->remove_at(i);
 356       found++;
 357       NOT_DEBUG( break; ) // elements are unique, so exit early
 358     }
 359   }
 360   assert(found <= 1, "not unique");
 361 }
 362 
 363 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
 364   for (int i = node_list.length() - 1; i >= 0; i--) {
 365     Node* n = node_list.at(i);
 366     if (!useful.member(n)) {
 367       node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
 368     }
 369   }
 370 }
 371 
 372 void Compile::remove_useless_node(Node* dead) {
 373   remove_modified_node(dead);
 374 
 375   // Constant node that has no out-edges and has only one in-edge from
 376   // root is usually dead. However, sometimes reshaping walk makes
 377   // it reachable by adding use edges. So, we will NOT count Con nodes
 378   // as dead to be conservative about the dead node count at any
 379   // given time.
 380   if (!dead->is_Con()) {
 381     record_dead_node(dead->_idx);
 382   }
 383   if (dead->is_macro()) {
 384     remove_macro_node(dead);
 385   }
 386   if (dead->is_expensive()) {
 387     remove_expensive_node(dead);
 388   }
 389   if (dead->Opcode() == Op_Opaque4) {
 390     remove_skeleton_predicate_opaq(dead);
 391   }
 392   if (dead->for_post_loop_opts_igvn()) {
 393     remove_from_post_loop_opts_igvn(dead);
 394   }
 395   if (dead->is_Call()) {
 396     remove_useless_late_inlines(                &_late_inlines, dead);
 397     remove_useless_late_inlines(         &_string_late_inlines, dead);
 398     remove_useless_late_inlines(         &_boxing_late_inlines, dead);
 399     remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
 400 
 401     if (dead->is_CallStaticJava()) {
 402       remove_unstable_if_trap(dead->as_CallStaticJava(), false);
 403     }
 404   }
 405   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 406   bs->unregister_potential_barrier_node(dead);
 407 }
 408 
 409 // Disconnect all useless nodes by disconnecting those at the boundary.
 410 void Compile::disconnect_useless_nodes(Unique_Node_List &useful, Unique_Node_List* worklist) {
 411   uint next = 0;
 412   while (next < useful.size()) {
 413     Node *n = useful.at(next++);
 414     if (n->is_SafePoint()) {
 415       // We're done with a parsing phase. Replaced nodes are not valid
 416       // beyond that point.
 417       n->as_SafePoint()->delete_replaced_nodes();
 418     }
 419     // Use raw traversal of out edges since this code removes out edges
 420     int max = n->outcnt();
 421     for (int j = 0; j < max; ++j) {
 422       Node* child = n->raw_out(j);
 423       if (!useful.member(child)) {
 424         assert(!child->is_top() || child != top(),
 425                "If top is cached in Compile object it is in useful list");
 426         // Only need to remove this out-edge to the useless node
 427         n->raw_del_out(j);
 428         --j;
 429         --max;
 430       }
 431     }
 432     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 433       worklist->push(n->unique_out());
 434     }
 435   }
 436 
 437   remove_useless_nodes(_macro_nodes,        useful); // remove useless macro nodes
 438   remove_useless_nodes(_predicate_opaqs,    useful); // remove useless predicate opaque nodes
 439   remove_useless_nodes(_skeleton_predicate_opaqs, useful);
 440   remove_useless_nodes(_expensive_nodes,    useful); // remove useless expensive nodes
 441   remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
 442   remove_useless_unstable_if_traps(useful);          // remove useless unstable_if traps
 443   remove_useless_coarsened_locks(useful);            // remove useless coarsened locks nodes
 444 #ifdef ASSERT
 445   if (_modified_nodes != NULL) {
 446     _modified_nodes->remove_useless_nodes(useful.member_set());
 447   }
 448 #endif
 449 
 450   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 451   bs->eliminate_useless_gc_barriers(useful, this);
 452   // clean up the late inline lists
 453   remove_useless_late_inlines(                &_late_inlines, useful);
 454   remove_useless_late_inlines(         &_string_late_inlines, useful);
 455   remove_useless_late_inlines(         &_boxing_late_inlines, useful);
 456   remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
 457   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 458 }
 459 
 460 // ============================================================================
 461 //------------------------------CompileWrapper---------------------------------
 462 class CompileWrapper : public StackObj {
 463   Compile *const _compile;
 464  public:
 465   CompileWrapper(Compile* compile);
 466 
 467   ~CompileWrapper();
 468 };
 469 
 470 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 471   // the Compile* pointer is stored in the current ciEnv:
 472   ciEnv* env = compile->env();
 473   assert(env == ciEnv::current(), "must already be a ciEnv active");
 474   assert(env->compiler_data() == NULL, "compile already active?");
 475   env->set_compiler_data(compile);
 476   assert(compile == Compile::current(), "sanity");
 477 
 478   compile->set_type_dict(NULL);
 479   compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
 480   compile->clone_map().set_clone_idx(0);
 481   compile->set_type_last_size(0);
 482   compile->set_last_tf(NULL, NULL);
 483   compile->set_indexSet_arena(NULL);
 484   compile->set_indexSet_free_block_list(NULL);
 485   compile->init_type_arena();
 486   Type::Initialize(compile);
 487   _compile->begin_method();
 488   _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
 489 }
 490 CompileWrapper::~CompileWrapper() {
 491   // simulate crash during compilation
 492   assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
 493 
 494   _compile->end_method();
 495   _compile->env()->set_compiler_data(NULL);
 496 }
 497 
 498 
 499 //----------------------------print_compile_messages---------------------------
 500 void Compile::print_compile_messages() {
 501 #ifndef PRODUCT
 502   // Check if recompiling
 503   if (!subsume_loads() && PrintOpto) {
 504     // Recompiling without allowing machine instructions to subsume loads
 505     tty->print_cr("*********************************************************");
 506     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 507     tty->print_cr("*********************************************************");
 508   }
 509   if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
 510     // Recompiling without escape analysis
 511     tty->print_cr("*********************************************************");
 512     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 513     tty->print_cr("*********************************************************");
 514   }
 515   if (do_iterative_escape_analysis() != DoEscapeAnalysis && PrintOpto) {
 516     // Recompiling without iterative escape analysis
 517     tty->print_cr("*********************************************************");
 518     tty->print_cr("** Bailout: Recompile without iterative escape analysis**");
 519     tty->print_cr("*********************************************************");
 520   }
 521   if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
 522     // Recompiling without boxing elimination
 523     tty->print_cr("*********************************************************");
 524     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 525     tty->print_cr("*********************************************************");
 526   }
 527   if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
 528     // Recompiling without locks coarsening
 529     tty->print_cr("*********************************************************");
 530     tty->print_cr("** Bailout: Recompile without locks coarsening         **");
 531     tty->print_cr("*********************************************************");
 532   }
 533   if (env()->break_at_compile()) {
 534     // Open the debugger when compiling this method.
 535     tty->print("### Breaking when compiling: ");
 536     method()->print_short_name();
 537     tty->cr();
 538     BREAKPOINT;
 539   }
 540 
 541   if( PrintOpto ) {
 542     if (is_osr_compilation()) {
 543       tty->print("[OSR]%3d", _compile_id);
 544     } else {
 545       tty->print("%3d", _compile_id);
 546     }
 547   }
 548 #endif
 549 }
 550 
 551 #ifndef PRODUCT
 552 void Compile::print_ideal_ir(const char* phase_name) {
 553   ttyLocker ttyl;
 554   // keep the following output all in one block
 555   // This output goes directly to the tty, not the compiler log.
 556   // To enable tools to match it up with the compilation activity,
 557   // be sure to tag this tty output with the compile ID.
 558   if (xtty != NULL) {
 559     xtty->head("ideal compile_id='%d'%s compile_phase='%s'",
 560                compile_id(),
 561                is_osr_compilation() ? " compile_kind='osr'" : "",
 562                phase_name);
 563   }
 564   if (_output == nullptr) {
 565     tty->print_cr("AFTER: %s", phase_name);
 566     // Print out all nodes in ascending order of index.
 567     root()->dump_bfs(MaxNodeLimit, nullptr, "+S$");
 568   } else {
 569     // Dump the node blockwise if we have a scheduling
 570     _output->print_scheduling();
 571   }
 572 
 573   if (xtty != NULL) {
 574     xtty->tail("ideal");
 575   }
 576 }
 577 #endif
 578 
 579 // ============================================================================
 580 //------------------------------Compile standard-------------------------------
 581 debug_only( int Compile::_debug_idx = 100000; )
 582 
 583 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 584 // the continuation bci for on stack replacement.
 585 
 586 
 587 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
 588                   Options options, DirectiveSet* directive)
 589                 : Phase(Compiler),
 590                   _compile_id(ci_env->compile_id()),
 591                   _options(options),
 592                   _method(target),
 593                   _entry_bci(osr_bci),
 594                   _ilt(NULL),
 595                   _stub_function(NULL),
 596                   _stub_name(NULL),
 597                   _stub_entry_point(NULL),
 598                   _max_node_limit(MaxNodeLimit),
 599                   _post_loop_opts_phase(false),
 600                   _inlining_progress(false),
 601                   _inlining_incrementally(false),
 602                   _do_cleanup(false),
 603                   _has_reserved_stack_access(target->has_reserved_stack_access()),
 604 #ifndef PRODUCT
 605                   _igv_idx(0),
 606                   _trace_opto_output(directive->TraceOptoOutputOption),
 607 #endif
 608                   _has_method_handle_invokes(false),
 609                   _clinit_barrier_on_entry(false),
 610                   _stress_seed(0),
 611                   _comp_arena(mtCompiler),
 612                   _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
 613                   _env(ci_env),
 614                   _directive(directive),
 615                   _log(ci_env->log()),
 616                   _failure_reason(NULL),
 617                   _intrinsics        (comp_arena(), 0, 0, NULL),
 618                   _macro_nodes       (comp_arena(), 8, 0, NULL),
 619                   _predicate_opaqs   (comp_arena(), 8, 0, NULL),
 620                   _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL),
 621                   _expensive_nodes   (comp_arena(), 8, 0, NULL),
 622                   _for_post_loop_igvn(comp_arena(), 8, 0, NULL),
 623                   _unstable_if_traps (comp_arena(), 8, 0, NULL),
 624                   _coarsened_locks   (comp_arena(), 8, 0, NULL),
 625                   _congraph(NULL),
 626                   NOT_PRODUCT(_igv_printer(NULL) COMMA)
 627                   _dead_node_list(comp_arena()),
 628                   _dead_node_count(0),
 629                   _node_arena(mtCompiler),
 630                   _old_arena(mtCompiler),
 631                   _mach_constant_base_node(NULL),
 632                   _Compile_types(mtCompiler),
 633                   _initial_gvn(NULL),
 634                   _for_igvn(NULL),
 635                   _late_inlines(comp_arena(), 2, 0, NULL),
 636                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 637                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 638                   _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
 639                   _late_inlines_pos(0),
 640                   _number_of_mh_late_inlines(0),
 641                   _print_inlining_stream(new (mtCompiler) stringStream()),
 642                   _print_inlining_list(NULL),
 643                   _print_inlining_idx(0),
 644                   _print_inlining_output(NULL),
 645                   _replay_inline_data(NULL),
 646                   _java_calls(0),
 647                   _inner_loops(0),
 648                   _interpreter_frame_size(0),
 649                   _output(NULL)
 650 #ifndef PRODUCT
 651                   , _in_dump_cnt(0)
 652 #endif
 653 {
 654   C = this;
 655   CompileWrapper cw(this);
 656 
 657   if (CITimeVerbose) {
 658     tty->print(" ");
 659     target->holder()->name()->print();
 660     tty->print(".");
 661     target->print_short_name();
 662     tty->print("  ");
 663   }
 664   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 665   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 666 
 667 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
 668   bool print_opto_assembly = directive->PrintOptoAssemblyOption;
 669   // We can always print a disassembly, either abstract (hex dump) or
 670   // with the help of a suitable hsdis library. Thus, we should not
 671   // couple print_assembly and print_opto_assembly controls.
 672   // But: always print opto and regular assembly on compile command 'print'.
 673   bool print_assembly = directive->PrintAssemblyOption;
 674   set_print_assembly(print_opto_assembly || print_assembly);
 675 #else
 676   set_print_assembly(false); // must initialize.
 677 #endif
 678 
 679 #ifndef PRODUCT
 680   set_parsed_irreducible_loop(false);
 681 
 682   if (directive->ReplayInlineOption) {
 683     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 684   }
 685 #endif
 686   set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
 687   set_print_intrinsics(directive->PrintIntrinsicsOption);
 688   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 689 
 690   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 691     // Make sure the method being compiled gets its own MDO,
 692     // so we can at least track the decompile_count().
 693     // Need MDO to record RTM code generation state.
 694     method()->ensure_method_data();
 695   }
 696 
 697   Init(/*do_aliasing=*/ true);
 698 
 699   print_compile_messages();
 700 
 701   _ilt = InlineTree::build_inline_tree_root();
 702 
 703   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 704   assert(num_alias_types() >= AliasIdxRaw, "");
 705 
 706 #define MINIMUM_NODE_HASH  1023
 707   // Node list that Iterative GVN will start with
 708   Unique_Node_List for_igvn(comp_arena());
 709   set_for_igvn(&for_igvn);
 710 
 711   // GVN that will be run immediately on new nodes
 712   uint estimated_size = method()->code_size()*4+64;
 713   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 714   PhaseGVN gvn(node_arena(), estimated_size);
 715   set_initial_gvn(&gvn);
 716 
 717   print_inlining_init();
 718   { // Scope for timing the parser
 719     TracePhase tp("parse", &timers[_t_parser]);
 720 
 721     // Put top into the hash table ASAP.
 722     initial_gvn()->transform_no_reclaim(top());
 723 
 724     // Set up tf(), start(), and find a CallGenerator.
 725     CallGenerator* cg = NULL;
 726     if (is_osr_compilation()) {
 727       const TypeTuple *domain = StartOSRNode::osr_domain();
 728       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 729       init_tf(TypeFunc::make(domain, range));
 730       StartNode* s = new StartOSRNode(root(), domain);
 731       initial_gvn()->set_type_bottom(s);
 732       init_start(s);
 733       cg = CallGenerator::for_osr(method(), entry_bci());
 734     } else {
 735       // Normal case.
 736       init_tf(TypeFunc::make(method()));
 737       StartNode* s = new StartNode(root(), tf()->domain());
 738       initial_gvn()->set_type_bottom(s);
 739       init_start(s);
 740       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
 741         // With java.lang.ref.reference.get() we must go through the
 742         // intrinsic - even when get() is the root
 743         // method of the compile - so that, if necessary, the value in
 744         // the referent field of the reference object gets recorded by
 745         // the pre-barrier code.
 746         cg = find_intrinsic(method(), false);
 747       }
 748       if (cg == NULL) {
 749         float past_uses = method()->interpreter_invocation_count();
 750         float expected_uses = past_uses;
 751         cg = CallGenerator::for_inline(method(), expected_uses);
 752       }
 753     }
 754     if (failing())  return;
 755     if (cg == NULL) {
 756       record_method_not_compilable("cannot parse method");
 757       return;
 758     }
 759     JVMState* jvms = build_start_state(start(), tf());
 760     if ((jvms = cg->generate(jvms)) == NULL) {
 761       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 762         record_method_not_compilable("method parse failed");
 763       }
 764       return;
 765     }
 766     GraphKit kit(jvms);
 767 
 768     if (!kit.stopped()) {
 769       // Accept return values, and transfer control we know not where.
 770       // This is done by a special, unique ReturnNode bound to root.
 771       return_values(kit.jvms());
 772     }
 773 
 774     if (kit.has_exceptions()) {
 775       // Any exceptions that escape from this call must be rethrown
 776       // to whatever caller is dynamically above us on the stack.
 777       // This is done by a special, unique RethrowNode bound to root.
 778       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 779     }
 780 
 781     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 782 
 783     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 784       inline_string_calls(true);
 785     }
 786 
 787     if (failing())  return;
 788 
 789     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 790 
 791     // Remove clutter produced by parsing.
 792     if (!failing()) {
 793       ResourceMark rm;
 794       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 795     }
 796   }
 797 
 798   // Note:  Large methods are capped off in do_one_bytecode().
 799   if (failing())  return;
 800 
 801   // After parsing, node notes are no longer automagic.
 802   // They must be propagated by register_new_node_with_optimizer(),
 803   // clone(), or the like.
 804   set_default_node_notes(NULL);
 805 
 806 #ifndef PRODUCT
 807   if (should_print_igv(1)) {
 808     _igv_printer->print_inlining();
 809   }
 810 #endif
 811 
 812   if (failing())  return;
 813   NOT_PRODUCT( verify_graph_edges(); )
 814 
 815   // If any phase is randomized for stress testing, seed random number
 816   // generation and log the seed for repeatability.
 817   if (StressLCM || StressGCM || StressIGVN || StressCCP) {
 818     if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
 819       _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
 820       FLAG_SET_ERGO(StressSeed, _stress_seed);
 821     } else {
 822       _stress_seed = StressSeed;
 823     }
 824     if (_log != NULL) {
 825       _log->elem("stress_test seed='%u'", _stress_seed);
 826     }
 827   }
 828 
 829   // Now optimize
 830   Optimize();
 831   if (failing())  return;
 832   NOT_PRODUCT( verify_graph_edges(); )
 833 
 834 #ifndef PRODUCT
 835   if (should_print_ideal()) {
 836     print_ideal_ir("print_ideal");
 837   }
 838 #endif
 839 
 840 #ifdef ASSERT
 841   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 842   bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
 843 #endif
 844 
 845   // Dump compilation data to replay it.
 846   if (directive->DumpReplayOption) {
 847     env()->dump_replay_data(_compile_id);
 848   }
 849   if (directive->DumpInlineOption && (ilt() != NULL)) {
 850     env()->dump_inline_data(_compile_id);
 851   }
 852 
 853   // Now that we know the size of all the monitors we can add a fixed slot
 854   // for the original deopt pc.
 855   int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
 856   set_fixed_slots(next_slot);
 857 
 858   // Compute when to use implicit null checks. Used by matching trap based
 859   // nodes and NullCheck optimization.
 860   set_allowed_deopt_reasons();
 861 
 862   // Now generate code
 863   Code_Gen();
 864 }
 865 
 866 //------------------------------Compile----------------------------------------
 867 // Compile a runtime stub
 868 Compile::Compile( ciEnv* ci_env,
 869                   TypeFunc_generator generator,
 870                   address stub_function,
 871                   const char *stub_name,
 872                   int is_fancy_jump,
 873                   bool pass_tls,
 874                   bool return_pc,
 875                   DirectiveSet* directive)
 876   : Phase(Compiler),
 877     _compile_id(0),
 878     _options(Options::for_runtime_stub()),
 879     _method(NULL),
 880     _entry_bci(InvocationEntryBci),
 881     _stub_function(stub_function),
 882     _stub_name(stub_name),
 883     _stub_entry_point(NULL),
 884     _max_node_limit(MaxNodeLimit),
 885     _post_loop_opts_phase(false),
 886     _inlining_progress(false),
 887     _inlining_incrementally(false),
 888     _has_reserved_stack_access(false),
 889 #ifndef PRODUCT
 890     _igv_idx(0),
 891     _trace_opto_output(directive->TraceOptoOutputOption),
 892 #endif
 893     _has_method_handle_invokes(false),
 894     _clinit_barrier_on_entry(false),
 895     _stress_seed(0),
 896     _comp_arena(mtCompiler),
 897     _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
 898     _env(ci_env),
 899     _directive(directive),
 900     _log(ci_env->log()),
 901     _failure_reason(NULL),
 902     _congraph(NULL),
 903     NOT_PRODUCT(_igv_printer(NULL) COMMA)
 904     _dead_node_list(comp_arena()),
 905     _dead_node_count(0),
 906     _node_arena(mtCompiler),
 907     _old_arena(mtCompiler),
 908     _mach_constant_base_node(NULL),
 909     _Compile_types(mtCompiler),
 910     _initial_gvn(NULL),
 911     _for_igvn(NULL),
 912     _number_of_mh_late_inlines(0),
 913     _print_inlining_stream(new (mtCompiler) stringStream()),
 914     _print_inlining_list(NULL),
 915     _print_inlining_idx(0),
 916     _print_inlining_output(NULL),
 917     _replay_inline_data(NULL),
 918     _java_calls(0),
 919     _inner_loops(0),
 920     _interpreter_frame_size(0),
 921     _output(NULL),
 922 #ifndef PRODUCT
 923     _in_dump_cnt(0),
 924 #endif
 925     _allowed_reasons(0) {
 926   C = this;
 927 
 928   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
 929   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
 930 
 931 #ifndef PRODUCT
 932   set_print_assembly(PrintFrameConverterAssembly);
 933   set_parsed_irreducible_loop(false);
 934 #else
 935   set_print_assembly(false); // Must initialize.
 936 #endif
 937   set_has_irreducible_loop(false); // no loops
 938 
 939   CompileWrapper cw(this);
 940   Init(/*do_aliasing=*/ false);
 941   init_tf((*generator)());
 942 
 943   {
 944     // The following is a dummy for the sake of GraphKit::gen_stub
 945     Unique_Node_List for_igvn(comp_arena());
 946     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
 947     PhaseGVN gvn(Thread::current()->resource_area(),255);
 948     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
 949     gvn.transform_no_reclaim(top());
 950 
 951     GraphKit kit;
 952     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
 953   }
 954 
 955   NOT_PRODUCT( verify_graph_edges(); )
 956 
 957   Code_Gen();
 958 }
 959 
 960 //------------------------------Init-------------------------------------------
 961 // Prepare for a single compilation
 962 void Compile::Init(bool aliasing) {
 963   _do_aliasing = aliasing;
 964   _unique  = 0;
 965   _regalloc = NULL;
 966 
 967   _tf      = NULL;  // filled in later
 968   _top     = NULL;  // cached later
 969   _matcher = NULL;  // filled in later
 970   _cfg     = NULL;  // filled in later
 971 
 972   IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
 973 
 974   _node_note_array = NULL;
 975   _default_node_notes = NULL;
 976   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
 977 
 978   _immutable_memory = NULL; // filled in at first inquiry
 979 
 980 #ifdef ASSERT
 981   _type_verify_symmetry = true;
 982   _phase_optimize_finished = false;
 983   _exception_backedge = false;
 984 #endif
 985 
 986   // Globally visible Nodes
 987   // First set TOP to NULL to give safe behavior during creation of RootNode
 988   set_cached_top_node(NULL);
 989   set_root(new RootNode());
 990   // Now that you have a Root to point to, create the real TOP
 991   set_cached_top_node( new ConNode(Type::TOP) );
 992   set_recent_alloc(NULL, NULL);
 993 
 994   // Create Debug Information Recorder to record scopes, oopmaps, etc.
 995   env()->set_oop_recorder(new OopRecorder(env()->arena()));
 996   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
 997   env()->set_dependencies(new Dependencies(env()));
 998 
 999   _fixed_slots = 0;
1000   set_has_split_ifs(false);
1001   set_has_loops(false); // first approximation
1002   set_has_stringbuilder(false);
1003   set_has_boxed_value(false);
1004   _trap_can_recompile = false;  // no traps emitted yet
1005   _major_progress = true; // start out assuming good things will happen
1006   set_has_unsafe_access(false);
1007   set_max_vector_size(0);
1008   set_clear_upper_avx(false);  //false as default for clear upper bits of ymm registers
1009   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1010   set_decompile_count(0);
1011 
1012   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1013   _loop_opts_cnt = LoopOptsCount;
1014   set_do_inlining(Inline);
1015   set_max_inline_size(MaxInlineSize);
1016   set_freq_inline_size(FreqInlineSize);
1017   set_do_scheduling(OptoScheduling);
1018 
1019   set_do_vector_loop(false);
1020   set_has_monitors(false);
1021 
1022   if (AllowVectorizeOnDemand) {
1023     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1024       set_do_vector_loop(true);
1025       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());})
1026     } else if (has_method() && method()->name() != 0 &&
1027                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1028       set_do_vector_loop(true);
1029     }
1030   }
1031   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1032   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());})
1033 
1034   set_rtm_state(NoRTM); // No RTM lock eliding by default
1035   _max_node_limit = _directive->MaxNodeLimitOption;
1036 
1037 #if INCLUDE_RTM_OPT
1038   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1039     int rtm_state = method()->method_data()->rtm_state();
1040     if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
1041       // Don't generate RTM lock eliding code.
1042       set_rtm_state(NoRTM);
1043     } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1044       // Generate RTM lock eliding code without abort ratio calculation code.
1045       set_rtm_state(UseRTM);
1046     } else if (UseRTMDeopt) {
1047       // Generate RTM lock eliding code and include abort ratio calculation
1048       // code if UseRTMDeopt is on.
1049       set_rtm_state(ProfileRTM);
1050     }
1051   }
1052 #endif
1053   if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1054     set_clinit_barrier_on_entry(true);
1055   }
1056   if (debug_info()->recording_non_safepoints()) {
1057     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1058                         (comp_arena(), 8, 0, NULL));
1059     set_default_node_notes(Node_Notes::make(this));
1060   }
1061 
1062   const int grow_ats = 16;
1063   _max_alias_types = grow_ats;
1064   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1065   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1066   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1067   {
1068     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1069   }
1070   // Initialize the first few types.
1071   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1072   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1073   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1074   _num_alias_types = AliasIdxRaw+1;
1075   // Zero out the alias type cache.
1076   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1077   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1078   probe_alias_cache(NULL)->_index = AliasIdxTop;
1079 }
1080 
1081 //---------------------------init_start----------------------------------------
1082 // Install the StartNode on this compile object.
1083 void Compile::init_start(StartNode* s) {
1084   if (failing())
1085     return; // already failing
1086   assert(s == start(), "");
1087 }
1088 
1089 /**
1090  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1091  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1092  * the ideal graph.
1093  */
1094 StartNode* Compile::start() const {
1095   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1096   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1097     Node* start = root()->fast_out(i);
1098     if (start->is_Start()) {
1099       return start->as_Start();
1100     }
1101   }
1102   fatal("Did not find Start node!");
1103   return NULL;
1104 }
1105 
1106 //-------------------------------immutable_memory-------------------------------------
1107 // Access immutable memory
1108 Node* Compile::immutable_memory() {
1109   if (_immutable_memory != NULL) {
1110     return _immutable_memory;
1111   }
1112   StartNode* s = start();
1113   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1114     Node *p = s->fast_out(i);
1115     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1116       _immutable_memory = p;
1117       return _immutable_memory;
1118     }
1119   }
1120   ShouldNotReachHere();
1121   return NULL;
1122 }
1123 
1124 //----------------------set_cached_top_node------------------------------------
1125 // Install the cached top node, and make sure Node::is_top works correctly.
1126 void Compile::set_cached_top_node(Node* tn) {
1127   if (tn != NULL)  verify_top(tn);
1128   Node* old_top = _top;
1129   _top = tn;
1130   // Calling Node::setup_is_top allows the nodes the chance to adjust
1131   // their _out arrays.
1132   if (_top != NULL)     _top->setup_is_top();
1133   if (old_top != NULL)  old_top->setup_is_top();
1134   assert(_top == NULL || top()->is_top(), "");
1135 }
1136 
1137 #ifdef ASSERT
1138 uint Compile::count_live_nodes_by_graph_walk() {
1139   Unique_Node_List useful(comp_arena());
1140   // Get useful node list by walking the graph.
1141   identify_useful_nodes(useful);
1142   return useful.size();
1143 }
1144 
1145 void Compile::print_missing_nodes() {
1146 
1147   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1148   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1149     return;
1150   }
1151 
1152   // This is an expensive function. It is executed only when the user
1153   // specifies VerifyIdealNodeCount option or otherwise knows the
1154   // additional work that needs to be done to identify reachable nodes
1155   // by walking the flow graph and find the missing ones using
1156   // _dead_node_list.
1157 
1158   Unique_Node_List useful(comp_arena());
1159   // Get useful node list by walking the graph.
1160   identify_useful_nodes(useful);
1161 
1162   uint l_nodes = C->live_nodes();
1163   uint l_nodes_by_walk = useful.size();
1164 
1165   if (l_nodes != l_nodes_by_walk) {
1166     if (_log != NULL) {
1167       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1168       _log->stamp();
1169       _log->end_head();
1170     }
1171     VectorSet& useful_member_set = useful.member_set();
1172     int last_idx = l_nodes_by_walk;
1173     for (int i = 0; i < last_idx; i++) {
1174       if (useful_member_set.test(i)) {
1175         if (_dead_node_list.test(i)) {
1176           if (_log != NULL) {
1177             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1178           }
1179           if (PrintIdealNodeCount) {
1180             // Print the log message to tty
1181               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1182               useful.at(i)->dump();
1183           }
1184         }
1185       }
1186       else if (! _dead_node_list.test(i)) {
1187         if (_log != NULL) {
1188           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1189         }
1190         if (PrintIdealNodeCount) {
1191           // Print the log message to tty
1192           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1193         }
1194       }
1195     }
1196     if (_log != NULL) {
1197       _log->tail("mismatched_nodes");
1198     }
1199   }
1200 }
1201 void Compile::record_modified_node(Node* n) {
1202   if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) {
1203     _modified_nodes->push(n);
1204   }
1205 }
1206 
1207 void Compile::remove_modified_node(Node* n) {
1208   if (_modified_nodes != NULL) {
1209     _modified_nodes->remove(n);
1210   }
1211 }
1212 #endif
1213 
1214 #ifndef PRODUCT
1215 void Compile::verify_top(Node* tn) const {
1216   if (tn != NULL) {
1217     assert(tn->is_Con(), "top node must be a constant");
1218     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1219     assert(tn->in(0) != NULL, "must have live top node");
1220   }
1221 }
1222 #endif
1223 
1224 
1225 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1226 
1227 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1228   guarantee(arr != NULL, "");
1229   int num_blocks = arr->length();
1230   if (grow_by < num_blocks)  grow_by = num_blocks;
1231   int num_notes = grow_by * _node_notes_block_size;
1232   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1233   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1234   while (num_notes > 0) {
1235     arr->append(notes);
1236     notes     += _node_notes_block_size;
1237     num_notes -= _node_notes_block_size;
1238   }
1239   assert(num_notes == 0, "exact multiple, please");
1240 }
1241 
1242 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1243   if (source == NULL || dest == NULL)  return false;
1244 
1245   if (dest->is_Con())
1246     return false;               // Do not push debug info onto constants.
1247 
1248 #ifdef ASSERT
1249   // Leave a bread crumb trail pointing to the original node:
1250   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1251     dest->set_debug_orig(source);
1252   }
1253 #endif
1254 
1255   if (node_note_array() == NULL)
1256     return false;               // Not collecting any notes now.
1257 
1258   // This is a copy onto a pre-existing node, which may already have notes.
1259   // If both nodes have notes, do not overwrite any pre-existing notes.
1260   Node_Notes* source_notes = node_notes_at(source->_idx);
1261   if (source_notes == NULL || source_notes->is_clear())  return false;
1262   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1263   if (dest_notes == NULL || dest_notes->is_clear()) {
1264     return set_node_notes_at(dest->_idx, source_notes);
1265   }
1266 
1267   Node_Notes merged_notes = (*source_notes);
1268   // The order of operations here ensures that dest notes will win...
1269   merged_notes.update_from(dest_notes);
1270   return set_node_notes_at(dest->_idx, &merged_notes);
1271 }
1272 
1273 
1274 //--------------------------allow_range_check_smearing-------------------------
1275 // Gating condition for coalescing similar range checks.
1276 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1277 // single covering check that is at least as strong as any of them.
1278 // If the optimization succeeds, the simplified (strengthened) range check
1279 // will always succeed.  If it fails, we will deopt, and then give up
1280 // on the optimization.
1281 bool Compile::allow_range_check_smearing() const {
1282   // If this method has already thrown a range-check,
1283   // assume it was because we already tried range smearing
1284   // and it failed.
1285   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1286   return !already_trapped;
1287 }
1288 
1289 
1290 //------------------------------flatten_alias_type-----------------------------
1291 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1292   assert(do_aliasing(), "Aliasing should be enabled");
1293   int offset = tj->offset();
1294   TypePtr::PTR ptr = tj->ptr();
1295 
1296   // Known instance (scalarizable allocation) alias only with itself.
1297   bool is_known_inst = tj->isa_oopptr() != NULL &&
1298                        tj->is_oopptr()->is_known_instance();
1299 
1300   // Process weird unsafe references.
1301   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1302     assert(InlineUnsafeOps || StressReflectiveCode, "indeterminate pointers come only from unsafe ops");
1303     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1304     tj = TypeOopPtr::BOTTOM;
1305     ptr = tj->ptr();
1306     offset = tj->offset();
1307   }
1308 
1309   // Array pointers need some flattening
1310   const TypeAryPtr* ta = tj->isa_aryptr();
1311   if (ta && ta->is_stable()) {
1312     // Erase stability property for alias analysis.
1313     tj = ta = ta->cast_to_stable(false);
1314   }
1315   if( ta && is_known_inst ) {
1316     if ( offset != Type::OffsetBot &&
1317          offset > arrayOopDesc::length_offset_in_bytes() ) {
1318       offset = Type::OffsetBot; // Flatten constant access into array body only
1319       tj = ta = ta->
1320               remove_speculative()->
1321               cast_to_ptr_type(ptr)->
1322               with_offset(offset);
1323     }
1324   } else if (ta) {
1325     // For arrays indexed by constant indices, we flatten the alias
1326     // space to include all of the array body.  Only the header, klass
1327     // and array length can be accessed un-aliased.
1328     if( offset != Type::OffsetBot ) {
1329       if( ta->const_oop() ) { // MethodData* or Method*
1330         offset = Type::OffsetBot;   // Flatten constant access into array body
1331         tj = ta = ta->
1332                 remove_speculative()->
1333                 cast_to_ptr_type(ptr)->
1334                 cast_to_exactness(false)->
1335                 with_offset(offset);
1336       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1337         // range is OK as-is.
1338         tj = ta = TypeAryPtr::RANGE;
1339       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1340         tj = TypeInstPtr::KLASS; // all klass loads look alike
1341         ta = TypeAryPtr::RANGE; // generic ignored junk
1342         ptr = TypePtr::BotPTR;
1343       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1344         tj = TypeInstPtr::MARK;
1345         ta = TypeAryPtr::RANGE; // generic ignored junk
1346         ptr = TypePtr::BotPTR;
1347       } else {                  // Random constant offset into array body
1348         offset = Type::OffsetBot;   // Flatten constant access into array body
1349         tj = ta = ta->
1350                 remove_speculative()->
1351                 cast_to_ptr_type(ptr)->
1352                 cast_to_exactness(false)->
1353                 with_offset(offset);
1354       }
1355     }
1356     // Arrays of fixed size alias with arrays of unknown size.
1357     if (ta->size() != TypeInt::POS) {
1358       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1359       tj = ta = ta->
1360               remove_speculative()->
1361               cast_to_ptr_type(ptr)->
1362               with_ary(tary)->
1363               cast_to_exactness(false);
1364     }
1365     // Arrays of known objects become arrays of unknown objects.
1366     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1367       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1368       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1369     }
1370     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1371       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1372       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1373     }
1374     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1375     // cannot be distinguished by bytecode alone.
1376     if (ta->elem() == TypeInt::BOOL) {
1377       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1378       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1379       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1380     }
1381     // During the 2nd round of IterGVN, NotNull castings are removed.
1382     // Make sure the Bottom and NotNull variants alias the same.
1383     // Also, make sure exact and non-exact variants alias the same.
1384     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1385       tj = ta = ta->
1386               remove_speculative()->
1387               cast_to_ptr_type(TypePtr::BotPTR)->
1388               cast_to_exactness(false)->
1389               with_offset(offset);
1390     }
1391   }
1392 
1393   // Oop pointers need some flattening
1394   const TypeInstPtr *to = tj->isa_instptr();
1395   if (to && to != TypeOopPtr::BOTTOM) {
1396     ciInstanceKlass* ik = to->instance_klass();
1397     if( ptr == TypePtr::Constant ) {
1398       if (ik != ciEnv::current()->Class_klass() ||
1399           offset < ik->layout_helper_size_in_bytes()) {
1400         // No constant oop pointers (such as Strings); they alias with
1401         // unknown strings.
1402         assert(!is_known_inst, "not scalarizable allocation");
1403         tj = to = to->
1404                 cast_to_instance_id(TypeOopPtr::InstanceBot)->
1405                 remove_speculative()->
1406                 cast_to_ptr_type(TypePtr::BotPTR)->
1407                 cast_to_exactness(false);
1408       }
1409     } else if( is_known_inst ) {
1410       tj = to; // Keep NotNull and klass_is_exact for instance type
1411     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1412       // During the 2nd round of IterGVN, NotNull castings are removed.
1413       // Make sure the Bottom and NotNull variants alias the same.
1414       // Also, make sure exact and non-exact variants alias the same.
1415       tj = to = to->
1416               remove_speculative()->
1417               cast_to_instance_id(TypeOopPtr::InstanceBot)->
1418               cast_to_ptr_type(TypePtr::BotPTR)->
1419               cast_to_exactness(false);
1420     }
1421     if (to->speculative() != NULL) {
1422       tj = to = to->remove_speculative();
1423     }
1424     // Canonicalize the holder of this field
1425     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1426       // First handle header references such as a LoadKlassNode, even if the
1427       // object's klass is unloaded at compile time (4965979).
1428       if (!is_known_inst) { // Do it only for non-instance types
1429         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1430       }
1431     } else if (offset < 0 || offset >= ik->layout_helper_size_in_bytes()) {
1432       // Static fields are in the space above the normal instance
1433       // fields in the java.lang.Class instance.
1434       if (ik != ciEnv::current()->Class_klass()) {
1435         to = NULL;
1436         tj = TypeOopPtr::BOTTOM;
1437         offset = tj->offset();
1438       }
1439     } else {
1440       ciInstanceKlass *canonical_holder = ik->get_canonical_holder(offset);
1441       assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1442       if (!ik->equals(canonical_holder) || tj->offset() != offset) {
1443         if( is_known_inst ) {
1444           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1445         } else {
1446           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1447         }
1448       }
1449     }
1450   }
1451 
1452   // Klass pointers to object array klasses need some flattening
1453   const TypeKlassPtr *tk = tj->isa_klassptr();
1454   if( tk ) {
1455     // If we are referencing a field within a Klass, we need
1456     // to assume the worst case of an Object.  Both exact and
1457     // inexact types must flatten to the same alias class so
1458     // use NotNull as the PTR.
1459     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1460       tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull,
1461                                        env()->Object_klass(),
1462                                        offset);
1463     }
1464 
1465     if (tk->isa_aryklassptr() && tk->is_aryklassptr()->elem()->isa_klassptr()) {
1466       ciKlass* k = ciObjArrayKlass::make(env()->Object_klass());
1467       if (!k || !k->is_loaded()) {                  // Only fails for some -Xcomp runs
1468         tj = tk = TypeInstKlassPtr::make(TypePtr::NotNull, env()->Object_klass(), offset);
1469       } else {
1470         tj = tk = TypeAryKlassPtr::make(TypePtr::NotNull, tk->is_aryklassptr()->elem(), k, offset);
1471       }
1472     }
1473 
1474     // Check for precise loads from the primary supertype array and force them
1475     // to the supertype cache alias index.  Check for generic array loads from
1476     // the primary supertype array and also force them to the supertype cache
1477     // alias index.  Since the same load can reach both, we need to merge
1478     // these 2 disparate memories into the same alias class.  Since the
1479     // primary supertype array is read-only, there's no chance of confusion
1480     // where we bypass an array load and an array store.
1481     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1482     if (offset == Type::OffsetBot ||
1483         (offset >= primary_supers_offset &&
1484          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1485         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1486       offset = in_bytes(Klass::secondary_super_cache_offset());
1487       tj = tk = tk->with_offset(offset);
1488     }
1489   }
1490 
1491   // Flatten all Raw pointers together.
1492   if (tj->base() == Type::RawPtr)
1493     tj = TypeRawPtr::BOTTOM;
1494 
1495   if (tj->base() == Type::AnyPtr)
1496     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1497 
1498   offset = tj->offset();
1499   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1500 
1501   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1502           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1503           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1504           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1505           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1506           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1507           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1508           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1509   assert( tj->ptr() != TypePtr::TopPTR &&
1510           tj->ptr() != TypePtr::AnyNull &&
1511           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1512 //    assert( tj->ptr() != TypePtr::Constant ||
1513 //            tj->base() == Type::RawPtr ||
1514 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1515 
1516   return tj;
1517 }
1518 
1519 void Compile::AliasType::Init(int i, const TypePtr* at) {
1520   assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1521   _index = i;
1522   _adr_type = at;
1523   _field = NULL;
1524   _element = NULL;
1525   _is_rewritable = true; // default
1526   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1527   if (atoop != NULL && atoop->is_known_instance()) {
1528     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1529     _general_index = Compile::current()->get_alias_index(gt);
1530   } else {
1531     _general_index = 0;
1532   }
1533 }
1534 
1535 BasicType Compile::AliasType::basic_type() const {
1536   if (element() != NULL) {
1537     const Type* element = adr_type()->is_aryptr()->elem();
1538     return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1539   } if (field() != NULL) {
1540     return field()->layout_type();
1541   } else {
1542     return T_ILLEGAL; // unknown
1543   }
1544 }
1545 
1546 //---------------------------------print_on------------------------------------
1547 #ifndef PRODUCT
1548 void Compile::AliasType::print_on(outputStream* st) {
1549   if (index() < 10)
1550         st->print("@ <%d> ", index());
1551   else  st->print("@ <%d>",  index());
1552   st->print(is_rewritable() ? "   " : " RO");
1553   int offset = adr_type()->offset();
1554   if (offset == Type::OffsetBot)
1555         st->print(" +any");
1556   else  st->print(" +%-3d", offset);
1557   st->print(" in ");
1558   adr_type()->dump_on(st);
1559   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1560   if (field() != NULL && tjp) {
1561     if (tjp->is_instptr()->instance_klass()  != field()->holder() ||
1562         tjp->offset() != field()->offset_in_bytes()) {
1563       st->print(" != ");
1564       field()->print();
1565       st->print(" ***");
1566     }
1567   }
1568 }
1569 
1570 void print_alias_types() {
1571   Compile* C = Compile::current();
1572   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1573   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1574     C->alias_type(idx)->print_on(tty);
1575     tty->cr();
1576   }
1577 }
1578 #endif
1579 
1580 
1581 //----------------------------probe_alias_cache--------------------------------
1582 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1583   intptr_t key = (intptr_t) adr_type;
1584   key ^= key >> logAliasCacheSize;
1585   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1586 }
1587 
1588 
1589 //-----------------------------grow_alias_types--------------------------------
1590 void Compile::grow_alias_types() {
1591   const int old_ats  = _max_alias_types; // how many before?
1592   const int new_ats  = old_ats;          // how many more?
1593   const int grow_ats = old_ats+new_ats;  // how many now?
1594   _max_alias_types = grow_ats;
1595   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1596   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1597   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1598   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1599 }
1600 
1601 
1602 //--------------------------------find_alias_type------------------------------
1603 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1604   if (!do_aliasing()) {
1605     return alias_type(AliasIdxBot);
1606   }
1607 
1608   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1609   if (ace->_adr_type == adr_type) {
1610     return alias_type(ace->_index);
1611   }
1612 
1613   // Handle special cases.
1614   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1615   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1616 
1617   // Do it the slow way.
1618   const TypePtr* flat = flatten_alias_type(adr_type);
1619 
1620 #ifdef ASSERT
1621   {
1622     ResourceMark rm;
1623     assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1624            Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1625     assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1626            Type::str(adr_type));
1627     if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1628       const TypeOopPtr* foop = flat->is_oopptr();
1629       // Scalarizable allocations have exact klass always.
1630       bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1631       const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1632       assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1633              Type::str(foop), Type::str(xoop));
1634     }
1635   }
1636 #endif
1637 
1638   int idx = AliasIdxTop;
1639   for (int i = 0; i < num_alias_types(); i++) {
1640     if (alias_type(i)->adr_type() == flat) {
1641       idx = i;
1642       break;
1643     }
1644   }
1645 
1646   if (idx == AliasIdxTop) {
1647     if (no_create)  return NULL;
1648     // Grow the array if necessary.
1649     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1650     // Add a new alias type.
1651     idx = _num_alias_types++;
1652     _alias_types[idx]->Init(idx, flat);
1653     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1654     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1655     if (flat->isa_instptr()) {
1656       if (flat->offset() == java_lang_Class::klass_offset()
1657           && flat->is_instptr()->instance_klass() == env()->Class_klass())
1658         alias_type(idx)->set_rewritable(false);
1659     }
1660     if (flat->isa_aryptr()) {
1661 #ifdef ASSERT
1662       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1663       // (T_BYTE has the weakest alignment and size restrictions...)
1664       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1665 #endif
1666       if (flat->offset() == TypePtr::OffsetBot) {
1667         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1668       }
1669     }
1670     if (flat->isa_klassptr()) {
1671       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1672         alias_type(idx)->set_rewritable(false);
1673       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1674         alias_type(idx)->set_rewritable(false);
1675       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1676         alias_type(idx)->set_rewritable(false);
1677       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1678         alias_type(idx)->set_rewritable(false);
1679       if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1680         alias_type(idx)->set_rewritable(false);
1681     }
1682     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1683     // but the base pointer type is not distinctive enough to identify
1684     // references into JavaThread.)
1685 
1686     // Check for final fields.
1687     const TypeInstPtr* tinst = flat->isa_instptr();
1688     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1689       ciField* field;
1690       if (tinst->const_oop() != NULL &&
1691           tinst->instance_klass() == ciEnv::current()->Class_klass() &&
1692           tinst->offset() >= (tinst->instance_klass()->layout_helper_size_in_bytes())) {
1693         // static field
1694         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1695         field = k->get_field_by_offset(tinst->offset(), true);
1696       } else {
1697         ciInstanceKlass *k = tinst->instance_klass();
1698         field = k->get_field_by_offset(tinst->offset(), false);
1699       }
1700       assert(field == NULL ||
1701              original_field == NULL ||
1702              (field->holder() == original_field->holder() &&
1703               field->offset() == original_field->offset() &&
1704               field->is_static() == original_field->is_static()), "wrong field?");
1705       // Set field() and is_rewritable() attributes.
1706       if (field != NULL)  alias_type(idx)->set_field(field);
1707     }
1708   }
1709 
1710   // Fill the cache for next time.
1711   ace->_adr_type = adr_type;
1712   ace->_index    = idx;
1713   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1714 
1715   // Might as well try to fill the cache for the flattened version, too.
1716   AliasCacheEntry* face = probe_alias_cache(flat);
1717   if (face->_adr_type == NULL) {
1718     face->_adr_type = flat;
1719     face->_index    = idx;
1720     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1721   }
1722 
1723   return alias_type(idx);
1724 }
1725 
1726 
1727 Compile::AliasType* Compile::alias_type(ciField* field) {
1728   const TypeOopPtr* t;
1729   if (field->is_static())
1730     t = TypeInstPtr::make(field->holder()->java_mirror());
1731   else
1732     t = TypeOopPtr::make_from_klass_raw(field->holder());
1733   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1734   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1735   return atp;
1736 }
1737 
1738 
1739 //------------------------------have_alias_type--------------------------------
1740 bool Compile::have_alias_type(const TypePtr* adr_type) {
1741   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1742   if (ace->_adr_type == adr_type) {
1743     return true;
1744   }
1745 
1746   // Handle special cases.
1747   if (adr_type == NULL)             return true;
1748   if (adr_type == TypePtr::BOTTOM)  return true;
1749 
1750   return find_alias_type(adr_type, true, NULL) != NULL;
1751 }
1752 
1753 //-----------------------------must_alias--------------------------------------
1754 // True if all values of the given address type are in the given alias category.
1755 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1756   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1757   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1758   if (alias_idx == AliasIdxTop)         return false; // the empty category
1759   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1760 
1761   // the only remaining possible overlap is identity
1762   int adr_idx = get_alias_index(adr_type);
1763   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1764   assert(adr_idx == alias_idx ||
1765          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1766           && adr_type                       != TypeOopPtr::BOTTOM),
1767          "should not be testing for overlap with an unsafe pointer");
1768   return adr_idx == alias_idx;
1769 }
1770 
1771 //------------------------------can_alias--------------------------------------
1772 // True if any values of the given address type are in the given alias category.
1773 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1774   if (alias_idx == AliasIdxTop)         return false; // the empty category
1775   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1776   // Known instance doesn't alias with bottom memory
1777   if (alias_idx == AliasIdxBot)         return !adr_type->is_known_instance();                   // the universal category
1778   if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1779 
1780   // the only remaining possible overlap is identity
1781   int adr_idx = get_alias_index(adr_type);
1782   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1783   return adr_idx == alias_idx;
1784 }
1785 
1786 //---------------------cleanup_loop_predicates-----------------------
1787 // Remove the opaque nodes that protect the predicates so that all unused
1788 // checks and uncommon_traps will be eliminated from the ideal graph
1789 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1790   if (predicate_count()==0) return;
1791   for (int i = predicate_count(); i > 0; i--) {
1792     Node * n = predicate_opaque1_node(i-1);
1793     assert(n->Opcode() == Op_Opaque1, "must be");
1794     igvn.replace_node(n, n->in(1));
1795   }
1796   assert(predicate_count()==0, "should be clean!");
1797 }
1798 
1799 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1800   if (!n->for_post_loop_opts_igvn()) {
1801     assert(!_for_post_loop_igvn.contains(n), "duplicate");
1802     n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1803     _for_post_loop_igvn.append(n);
1804   }
1805 }
1806 
1807 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1808   n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1809   _for_post_loop_igvn.remove(n);
1810 }
1811 
1812 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1813   // Verify that all previous optimizations produced a valid graph
1814   // at least to this point, even if no loop optimizations were done.
1815   PhaseIdealLoop::verify(igvn);
1816 
1817   C->set_post_loop_opts_phase(); // no more loop opts allowed
1818 
1819   assert(!C->major_progress(), "not cleared");
1820 
1821   if (_for_post_loop_igvn.length() > 0) {
1822     while (_for_post_loop_igvn.length() > 0) {
1823       Node* n = _for_post_loop_igvn.pop();
1824       n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1825       igvn._worklist.push(n);
1826     }
1827     igvn.optimize();
1828     assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1829 
1830     // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1831     if (C->major_progress()) {
1832       C->clear_major_progress(); // ensure that major progress is now clear
1833     }
1834   }
1835 }
1836 
1837 void Compile::record_unstable_if_trap(UnstableIfTrap* trap) {
1838   if (OptimizeUnstableIf) {
1839     _unstable_if_traps.append(trap);
1840   }
1841 }
1842 
1843 void Compile::remove_useless_unstable_if_traps(Unique_Node_List& useful) {
1844   for (int i = _unstable_if_traps.length() - 1; i >= 0; i--) {
1845     UnstableIfTrap* trap = _unstable_if_traps.at(i);
1846     Node* n = trap->uncommon_trap();
1847     if (!useful.member(n)) {
1848       _unstable_if_traps.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
1849     }
1850   }
1851 }
1852 
1853 // Remove the unstable if trap associated with 'unc' from candidates. It is either dead
1854 // or fold-compares case. Return true if succeed or not found.
1855 //
1856 // In rare cases, the found trap has been processed. It is too late to delete it. Return
1857 // false and ask fold-compares to yield.
1858 //
1859 // 'fold-compares' may use the uncommon_trap of the dominating IfNode to cover the fused
1860 // IfNode. This breaks the unstable_if trap invariant: control takes the unstable path
1861 // when deoptimization does happen.
1862 bool Compile::remove_unstable_if_trap(CallStaticJavaNode* unc, bool yield) {
1863   for (int i = 0; i < _unstable_if_traps.length(); ++i) {
1864     UnstableIfTrap* trap = _unstable_if_traps.at(i);
1865     if (trap->uncommon_trap() == unc) {
1866       if (yield && trap->modified()) {
1867         return false;
1868       }
1869       _unstable_if_traps.delete_at(i);
1870       break;
1871     }
1872   }
1873   return true;
1874 }
1875 
1876 // Re-calculate unstable_if traps with the liveness of next_bci, which points to the unlikely path.
1877 // It needs to be done after igvn because fold-compares may fuse uncommon_traps and before renumbering.
1878 void Compile::process_for_unstable_if_traps(PhaseIterGVN& igvn) {
1879   for (int i = _unstable_if_traps.length() - 1; i >= 0; --i) {
1880     UnstableIfTrap* trap = _unstable_if_traps.at(i);
1881     CallStaticJavaNode* unc = trap->uncommon_trap();
1882     int next_bci = trap->next_bci();
1883     bool modified = trap->modified();
1884 
1885     if (next_bci != -1 && !modified) {
1886       assert(!_dead_node_list.test(unc->_idx), "changing a dead node!");
1887       JVMState* jvms = unc->jvms();
1888       ciMethod* method = jvms->method();
1889       ciBytecodeStream iter(method);
1890 
1891       iter.force_bci(jvms->bci());
1892       assert(next_bci == iter.next_bci() || next_bci == iter.get_dest(), "wrong next_bci at unstable_if");
1893       Bytecodes::Code c = iter.cur_bc();
1894       Node* lhs = nullptr;
1895       Node* rhs = nullptr;
1896       if (c == Bytecodes::_if_acmpeq || c == Bytecodes::_if_acmpne) {
1897         lhs = unc->peek_operand(0);
1898         rhs = unc->peek_operand(1);
1899       } else if (c == Bytecodes::_ifnull || c == Bytecodes::_ifnonnull) {
1900         lhs = unc->peek_operand(0);
1901       }
1902 
1903       ResourceMark rm;
1904       const MethodLivenessResult& live_locals = method->liveness_at_bci(next_bci);
1905       assert(live_locals.is_valid(), "broken liveness info");
1906       int len = (int)live_locals.size();
1907 
1908       for (int i = 0; i < len; i++) {
1909         Node* local = unc->local(jvms, i);
1910         // kill local using the liveness of next_bci.
1911         // give up when the local looks like an operand to secure reexecution.
1912         if (!live_locals.at(i) && !local->is_top() && local != lhs && local!= rhs) {
1913           uint idx = jvms->locoff() + i;
1914 #ifdef ASSERT
1915           if (Verbose) {
1916             tty->print("[unstable_if] kill local#%d: ", idx);
1917             local->dump();
1918             tty->cr();
1919           }
1920 #endif
1921           igvn.replace_input_of(unc, idx, top());
1922           modified = true;
1923         }
1924       }
1925     }
1926 
1927     // keep the mondified trap for late query
1928     if (modified) {
1929       trap->set_modified();
1930     } else {
1931       _unstable_if_traps.delete_at(i);
1932     }
1933   }
1934   igvn.optimize();
1935 }
1936 
1937 // StringOpts and late inlining of string methods
1938 void Compile::inline_string_calls(bool parse_time) {
1939   {
1940     // remove useless nodes to make the usage analysis simpler
1941     ResourceMark rm;
1942     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1943   }
1944 
1945   {
1946     ResourceMark rm;
1947     print_method(PHASE_BEFORE_STRINGOPTS, 3);
1948     PhaseStringOpts pso(initial_gvn());
1949     print_method(PHASE_AFTER_STRINGOPTS, 3);
1950   }
1951 
1952   // now inline anything that we skipped the first time around
1953   if (!parse_time) {
1954     _late_inlines_pos = _late_inlines.length();
1955   }
1956 
1957   while (_string_late_inlines.length() > 0) {
1958     CallGenerator* cg = _string_late_inlines.pop();
1959     cg->do_late_inline();
1960     if (failing())  return;
1961   }
1962   _string_late_inlines.trunc_to(0);
1963 }
1964 
1965 // Late inlining of boxing methods
1966 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1967   if (_boxing_late_inlines.length() > 0) {
1968     assert(has_boxed_value(), "inconsistent");
1969 
1970     PhaseGVN* gvn = initial_gvn();
1971     set_inlining_incrementally(true);
1972 
1973     assert( igvn._worklist.size() == 0, "should be done with igvn" );
1974     for_igvn()->clear();
1975     gvn->replace_with(&igvn);
1976 
1977     _late_inlines_pos = _late_inlines.length();
1978 
1979     while (_boxing_late_inlines.length() > 0) {
1980       CallGenerator* cg = _boxing_late_inlines.pop();
1981       cg->do_late_inline();
1982       if (failing())  return;
1983     }
1984     _boxing_late_inlines.trunc_to(0);
1985 
1986     inline_incrementally_cleanup(igvn);
1987 
1988     set_inlining_incrementally(false);
1989   }
1990 }
1991 
1992 bool Compile::inline_incrementally_one() {
1993   assert(IncrementalInline, "incremental inlining should be on");
1994 
1995   TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1996 
1997   set_inlining_progress(false);
1998   set_do_cleanup(false);
1999 
2000   for (int i = 0; i < _late_inlines.length(); i++) {
2001     _late_inlines_pos = i+1;
2002     CallGenerator* cg = _late_inlines.at(i);
2003     bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
2004     if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
2005       cg->do_late_inline();
2006       assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
2007       if (failing()) {
2008         return false;
2009       } else if (inlining_progress()) {
2010         _late_inlines_pos = i+1; // restore the position in case new elements were inserted
2011         print_method(PHASE_INCREMENTAL_INLINE_STEP, 3, cg->call_node());
2012         break; // process one call site at a time
2013       }
2014     } else {
2015       // Ignore late inline direct calls when inlining is not allowed.
2016       // They are left in the late inline list when node budget is exhausted until the list is fully drained.
2017     }
2018   }
2019   // Remove processed elements.
2020   _late_inlines.remove_till(_late_inlines_pos);
2021   _late_inlines_pos = 0;
2022 
2023   assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
2024 
2025   bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2026 
2027   set_inlining_progress(false);
2028   set_do_cleanup(false);
2029 
2030   bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
2031   return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
2032 }
2033 
2034 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2035   {
2036     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2037     ResourceMark rm;
2038     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2039   }
2040   {
2041     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2042     igvn = PhaseIterGVN(initial_gvn());
2043     igvn.optimize();
2044   }
2045   print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
2046 }
2047 
2048 // Perform incremental inlining until bound on number of live nodes is reached
2049 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2050   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2051 
2052   set_inlining_incrementally(true);
2053   uint low_live_nodes = 0;
2054 
2055   while (_late_inlines.length() > 0) {
2056     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2057       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2058         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2059         // PhaseIdealLoop is expensive so we only try it once we are
2060         // out of live nodes and we only try it again if the previous
2061         // helped got the number of nodes down significantly
2062         PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2063         if (failing())  return;
2064         low_live_nodes = live_nodes();
2065         _major_progress = true;
2066       }
2067 
2068       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2069         bool do_print_inlining = print_inlining() || print_intrinsics();
2070         if (do_print_inlining || log() != NULL) {
2071           // Print inlining message for candidates that we couldn't inline for lack of space.
2072           for (int i = 0; i < _late_inlines.length(); i++) {
2073             CallGenerator* cg = _late_inlines.at(i);
2074             const char* msg = "live nodes > LiveNodeCountInliningCutoff";
2075             if (do_print_inlining) {
2076               cg->print_inlining_late(msg);
2077             }
2078             log_late_inline_failure(cg, msg);
2079           }
2080         }
2081         break; // finish
2082       }
2083     }
2084 
2085     for_igvn()->clear();
2086     initial_gvn()->replace_with(&igvn);
2087 
2088     while (inline_incrementally_one()) {
2089       assert(!failing(), "inconsistent");
2090     }
2091     if (failing())  return;
2092 
2093     inline_incrementally_cleanup(igvn);
2094 
2095     print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
2096 
2097     if (failing())  return;
2098 
2099     if (_late_inlines.length() == 0) {
2100       break; // no more progress
2101     }
2102   }
2103   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2104 
2105   if (_string_late_inlines.length() > 0) {
2106     assert(has_stringbuilder(), "inconsistent");
2107     for_igvn()->clear();
2108     initial_gvn()->replace_with(&igvn);
2109 
2110     inline_string_calls(false);
2111 
2112     if (failing())  return;
2113 
2114     inline_incrementally_cleanup(igvn);
2115   }
2116 
2117   set_inlining_incrementally(false);
2118 }
2119 
2120 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
2121   // "inlining_incrementally() == false" is used to signal that no inlining is allowed
2122   // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
2123   // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL"
2124   // as if "inlining_incrementally() == true" were set.
2125   assert(inlining_incrementally() == false, "not allowed");
2126   assert(_modified_nodes == NULL, "not allowed");
2127   assert(_late_inlines.length() > 0, "sanity");
2128 
2129   while (_late_inlines.length() > 0) {
2130     for_igvn()->clear();
2131     initial_gvn()->replace_with(&igvn);
2132 
2133     while (inline_incrementally_one()) {
2134       assert(!failing(), "inconsistent");
2135     }
2136     if (failing())  return;
2137 
2138     inline_incrementally_cleanup(igvn);
2139   }
2140 }
2141 
2142 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2143   if (_loop_opts_cnt > 0) {
2144     while (major_progress() && (_loop_opts_cnt > 0)) {
2145       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2146       PhaseIdealLoop::optimize(igvn, mode);
2147       _loop_opts_cnt--;
2148       if (failing())  return false;
2149       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2150     }
2151   }
2152   return true;
2153 }
2154 
2155 // Remove edges from "root" to each SafePoint at a backward branch.
2156 // They were inserted during parsing (see add_safepoint()) to make
2157 // infinite loops without calls or exceptions visible to root, i.e.,
2158 // useful.
2159 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2160   Node *r = root();
2161   if (r != NULL) {
2162     for (uint i = r->req(); i < r->len(); ++i) {
2163       Node *n = r->in(i);
2164       if (n != NULL && n->is_SafePoint()) {
2165         r->rm_prec(i);
2166         if (n->outcnt() == 0) {
2167           igvn.remove_dead_node(n);
2168         }
2169         --i;
2170       }
2171     }
2172     // Parsing may have added top inputs to the root node (Path
2173     // leading to the Halt node proven dead). Make sure we get a
2174     // chance to clean them up.
2175     igvn._worklist.push(r);
2176     igvn.optimize();
2177   }
2178 }
2179 
2180 //------------------------------Optimize---------------------------------------
2181 // Given a graph, optimize it.
2182 void Compile::Optimize() {
2183   TracePhase tp("optimizer", &timers[_t_optimizer]);
2184 
2185 #ifndef PRODUCT
2186   if (env()->break_at_compile()) {
2187     BREAKPOINT;
2188   }
2189 
2190 #endif
2191 
2192   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2193 #ifdef ASSERT
2194   bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2195 #endif
2196 
2197   ResourceMark rm;
2198 
2199   print_inlining_reinit();
2200 
2201   NOT_PRODUCT( verify_graph_edges(); )
2202 
2203   print_method(PHASE_AFTER_PARSING, 1);
2204 
2205  {
2206   // Iterative Global Value Numbering, including ideal transforms
2207   // Initialize IterGVN with types and values from parse-time GVN
2208   PhaseIterGVN igvn(initial_gvn());
2209 #ifdef ASSERT
2210   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2211 #endif
2212   {
2213     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2214     igvn.optimize();
2215   }
2216 
2217   if (failing())  return;
2218 
2219   print_method(PHASE_ITER_GVN1, 2);
2220 
2221   process_for_unstable_if_traps(igvn);
2222 
2223   inline_incrementally(igvn);
2224 
2225   print_method(PHASE_INCREMENTAL_INLINE, 2);
2226 
2227   if (failing())  return;
2228 
2229   if (eliminate_boxing()) {
2230     // Inline valueOf() methods now.
2231     inline_boxing_calls(igvn);
2232 
2233     if (AlwaysIncrementalInline) {
2234       inline_incrementally(igvn);
2235     }
2236 
2237     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2238 
2239     if (failing())  return;
2240   }
2241 
2242   // Remove the speculative part of types and clean up the graph from
2243   // the extra CastPP nodes whose only purpose is to carry them. Do
2244   // that early so that optimizations are not disrupted by the extra
2245   // CastPP nodes.
2246   remove_speculative_types(igvn);
2247 
2248   // No more new expensive nodes will be added to the list from here
2249   // so keep only the actual candidates for optimizations.
2250   cleanup_expensive_nodes(igvn);
2251 
2252   assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2253   if (EnableVectorSupport && has_vbox_nodes()) {
2254     TracePhase tp("", &timers[_t_vector]);
2255     PhaseVector pv(igvn);
2256     pv.optimize_vector_boxes();
2257 
2258     print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2259   }
2260   assert(!has_vbox_nodes(), "sanity");
2261 
2262   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2263     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2264     initial_gvn()->replace_with(&igvn);
2265     Unique_Node_List* old_worklist = for_igvn();
2266     old_worklist->clear();
2267     Unique_Node_List new_worklist(C->comp_arena());
2268     {
2269       ResourceMark rm;
2270       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2271     }
2272     Unique_Node_List* save_for_igvn = for_igvn();
2273     set_for_igvn(&new_worklist);
2274     igvn = PhaseIterGVN(initial_gvn());
2275     igvn.optimize();
2276     set_for_igvn(old_worklist); // new_worklist is dead beyond this point
2277   }
2278 
2279   // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2280   // safepoints
2281   remove_root_to_sfpts_edges(igvn);
2282 
2283   // Perform escape analysis
2284   if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2285     if (has_loops()) {
2286       // Cleanup graph (remove dead nodes).
2287       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2288       PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2289       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2290       if (failing())  return;
2291     }
2292     bool progress;
2293     do {
2294       ConnectionGraph::do_analysis(this, &igvn);
2295 
2296       if (failing())  return;
2297 
2298       int mcount = macro_count(); // Record number of allocations and locks before IGVN
2299 
2300       // Optimize out fields loads from scalar replaceable allocations.
2301       igvn.optimize();
2302       print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2303 
2304       if (failing())  return;
2305 
2306       if (congraph() != NULL && macro_count() > 0) {
2307         TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2308         PhaseMacroExpand mexp(igvn);
2309         mexp.eliminate_macro_nodes();
2310         igvn.set_delay_transform(false);
2311 
2312         igvn.optimize();
2313         print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2314 
2315         if (failing())  return;
2316       }
2317       progress = do_iterative_escape_analysis() &&
2318                  (macro_count() < mcount) &&
2319                  ConnectionGraph::has_candidates(this);
2320       // Try again if candidates exist and made progress
2321       // by removing some allocations and/or locks.
2322     } while (progress);
2323   }
2324 
2325   // Loop transforms on the ideal graph.  Range Check Elimination,
2326   // peeling, unrolling, etc.
2327 
2328   // Set loop opts counter
2329   if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2330     {
2331       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2332       PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2333       _loop_opts_cnt--;
2334       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2335       if (failing())  return;
2336     }
2337     // Loop opts pass if partial peeling occurred in previous pass
2338     if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2339       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2340       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2341       _loop_opts_cnt--;
2342       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2343       if (failing())  return;
2344     }
2345     // Loop opts pass for loop-unrolling before CCP
2346     if(major_progress() && (_loop_opts_cnt > 0)) {
2347       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2348       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2349       _loop_opts_cnt--;
2350       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2351     }
2352     if (!failing()) {
2353       // Verify that last round of loop opts produced a valid graph
2354       PhaseIdealLoop::verify(igvn);
2355     }
2356   }
2357   if (failing())  return;
2358 
2359   // Conditional Constant Propagation;
2360   PhaseCCP ccp( &igvn );
2361   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2362   {
2363     TracePhase tp("ccp", &timers[_t_ccp]);
2364     ccp.do_transform();
2365   }
2366   print_method(PHASE_CCP1, 2);
2367 
2368   assert( true, "Break here to ccp.dump_old2new_map()");
2369 
2370   // Iterative Global Value Numbering, including ideal transforms
2371   {
2372     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2373     igvn = ccp;
2374     igvn.optimize();
2375   }
2376   print_method(PHASE_ITER_GVN2, 2);
2377 
2378   if (failing())  return;
2379 
2380   // Loop transforms on the ideal graph.  Range Check Elimination,
2381   // peeling, unrolling, etc.
2382   if (!optimize_loops(igvn, LoopOptsDefault)) {
2383     return;
2384   }
2385 
2386   if (failing())  return;
2387 
2388   C->clear_major_progress(); // ensure that major progress is now clear
2389 
2390   process_for_post_loop_opts_igvn(igvn);
2391 
2392 #ifdef ASSERT
2393   bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2394 #endif
2395 
2396   {
2397     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2398     PhaseMacroExpand  mex(igvn);
2399     if (mex.expand_macro_nodes()) {
2400       assert(failing(), "must bail out w/ explicit message");
2401       return;
2402     }
2403     print_method(PHASE_MACRO_EXPANSION, 2);
2404   }
2405 
2406   {
2407     TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2408     if (bs->expand_barriers(this, igvn)) {
2409       assert(failing(), "must bail out w/ explicit message");
2410       return;
2411     }
2412     print_method(PHASE_BARRIER_EXPANSION, 2);
2413   }
2414 
2415   if (C->max_vector_size() > 0) {
2416     C->optimize_logic_cones(igvn);
2417     igvn.optimize();
2418   }
2419 
2420   DEBUG_ONLY( _modified_nodes = NULL; )
2421 
2422   assert(igvn._worklist.size() == 0, "not empty");
2423 
2424   assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2425 
2426   if (_late_inlines.length() > 0) {
2427     // More opportunities to optimize virtual and MH calls.
2428     // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2429     process_late_inline_calls_no_inline(igvn);
2430   }
2431  } // (End scope of igvn; run destructor if necessary for asserts.)
2432 
2433  check_no_dead_use();
2434 
2435  process_print_inlining();
2436 
2437  // A method with only infinite loops has no edges entering loops from root
2438  {
2439    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2440    if (final_graph_reshaping()) {
2441      assert(failing(), "must bail out w/ explicit message");
2442      return;
2443    }
2444  }
2445 
2446  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2447  DEBUG_ONLY(set_phase_optimize_finished();)
2448 }
2449 
2450 #ifdef ASSERT
2451 void Compile::check_no_dead_use() const {
2452   ResourceMark rm;
2453   Unique_Node_List wq;
2454   wq.push(root());
2455   for (uint i = 0; i < wq.size(); ++i) {
2456     Node* n = wq.at(i);
2457     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2458       Node* u = n->fast_out(j);
2459       if (u->outcnt() == 0 && !u->is_Con()) {
2460         u->dump();
2461         fatal("no reachable node should have no use");
2462       }
2463       wq.push(u);
2464     }
2465   }
2466 }
2467 #endif
2468 
2469 void Compile::inline_vector_reboxing_calls() {
2470   if (C->_vector_reboxing_late_inlines.length() > 0) {
2471     _late_inlines_pos = C->_late_inlines.length();
2472     while (_vector_reboxing_late_inlines.length() > 0) {
2473       CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2474       cg->do_late_inline();
2475       if (failing())  return;
2476       print_method(PHASE_INLINE_VECTOR_REBOX, 3, cg->call_node());
2477     }
2478     _vector_reboxing_late_inlines.trunc_to(0);
2479   }
2480 }
2481 
2482 bool Compile::has_vbox_nodes() {
2483   if (C->_vector_reboxing_late_inlines.length() > 0) {
2484     return true;
2485   }
2486   for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2487     Node * n = C->macro_node(macro_idx);
2488     assert(n->is_macro(), "only macro nodes expected here");
2489     if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2490       return true;
2491     }
2492   }
2493   return false;
2494 }
2495 
2496 //---------------------------- Bitwise operation packing optimization ---------------------------
2497 
2498 static bool is_vector_unary_bitwise_op(Node* n) {
2499   return n->Opcode() == Op_XorV &&
2500          VectorNode::is_vector_bitwise_not_pattern(n);
2501 }
2502 
2503 static bool is_vector_binary_bitwise_op(Node* n) {
2504   switch (n->Opcode()) {
2505     case Op_AndV:
2506     case Op_OrV:
2507       return true;
2508 
2509     case Op_XorV:
2510       return !is_vector_unary_bitwise_op(n);
2511 
2512     default:
2513       return false;
2514   }
2515 }
2516 
2517 static bool is_vector_ternary_bitwise_op(Node* n) {
2518   return n->Opcode() == Op_MacroLogicV;
2519 }
2520 
2521 static bool is_vector_bitwise_op(Node* n) {
2522   return is_vector_unary_bitwise_op(n)  ||
2523          is_vector_binary_bitwise_op(n) ||
2524          is_vector_ternary_bitwise_op(n);
2525 }
2526 
2527 static bool is_vector_bitwise_cone_root(Node* n) {
2528   if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2529     return false;
2530   }
2531   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2532     if (is_vector_bitwise_op(n->fast_out(i))) {
2533       return false;
2534     }
2535   }
2536   return true;
2537 }
2538 
2539 static uint collect_unique_inputs(Node* n, Unique_Node_List& inputs) {
2540   uint cnt = 0;
2541   if (is_vector_bitwise_op(n)) {
2542     uint inp_cnt = n->is_predicated_vector() ? n->req()-1 : n->req();
2543     if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2544       for (uint i = 1; i < inp_cnt; i++) {
2545         Node* in = n->in(i);
2546         bool skip = VectorNode::is_all_ones_vector(in);
2547         if (!skip && !inputs.member(in)) {
2548           inputs.push(in);
2549           cnt++;
2550         }
2551       }
2552       assert(cnt <= 1, "not unary");
2553     } else {
2554       uint last_req = inp_cnt;
2555       if (is_vector_ternary_bitwise_op(n)) {
2556         last_req = inp_cnt - 1; // skip last input
2557       }
2558       for (uint i = 1; i < last_req; i++) {
2559         Node* def = n->in(i);
2560         if (!inputs.member(def)) {
2561           inputs.push(def);
2562           cnt++;
2563         }
2564       }
2565     }
2566   } else { // not a bitwise operations
2567     if (!inputs.member(n)) {
2568       inputs.push(n);
2569       cnt++;
2570     }
2571   }
2572   return cnt;
2573 }
2574 
2575 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2576   Unique_Node_List useful_nodes;
2577   C->identify_useful_nodes(useful_nodes);
2578 
2579   for (uint i = 0; i < useful_nodes.size(); i++) {
2580     Node* n = useful_nodes.at(i);
2581     if (is_vector_bitwise_cone_root(n)) {
2582       list.push(n);
2583     }
2584   }
2585 }
2586 
2587 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2588                                     const TypeVect* vt,
2589                                     Unique_Node_List& partition,
2590                                     Unique_Node_List& inputs) {
2591   assert(partition.size() == 2 || partition.size() == 3, "not supported");
2592   assert(inputs.size()    == 2 || inputs.size()    == 3, "not supported");
2593   assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2594 
2595   Node* in1 = inputs.at(0);
2596   Node* in2 = inputs.at(1);
2597   Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2598 
2599   uint func = compute_truth_table(partition, inputs);
2600 
2601   Node* pn = partition.at(partition.size() - 1);
2602   Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
2603   return igvn.transform(MacroLogicVNode::make(igvn, in1, in2, in3, mask, func, vt));
2604 }
2605 
2606 static uint extract_bit(uint func, uint pos) {
2607   return (func & (1 << pos)) >> pos;
2608 }
2609 
2610 //
2611 //  A macro logic node represents a truth table. It has 4 inputs,
2612 //  First three inputs corresponds to 3 columns of a truth table
2613 //  and fourth input captures the logic function.
2614 //
2615 //  eg.  fn = (in1 AND in2) OR in3;
2616 //
2617 //      MacroNode(in1,in2,in3,fn)
2618 //
2619 //  -----------------
2620 //  in1 in2 in3  fn
2621 //  -----------------
2622 //  0    0   0    0
2623 //  0    0   1    1
2624 //  0    1   0    0
2625 //  0    1   1    1
2626 //  1    0   0    0
2627 //  1    0   1    1
2628 //  1    1   0    1
2629 //  1    1   1    1
2630 //
2631 
2632 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2633   int res = 0;
2634   for (int i = 0; i < 8; i++) {
2635     int bit1 = extract_bit(in1, i);
2636     int bit2 = extract_bit(in2, i);
2637     int bit3 = extract_bit(in3, i);
2638 
2639     int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2640     int func_bit = extract_bit(func, func_bit_pos);
2641 
2642     res |= func_bit << i;
2643   }
2644   return res;
2645 }
2646 
2647 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2648   assert(n != NULL, "");
2649   assert(eval_map.contains(n), "absent");
2650   return *(eval_map.get(n));
2651 }
2652 
2653 static void eval_operands(Node* n,
2654                           uint& func1, uint& func2, uint& func3,
2655                           ResourceHashtable<Node*,uint>& eval_map) {
2656   assert(is_vector_bitwise_op(n), "");
2657 
2658   if (is_vector_unary_bitwise_op(n)) {
2659     Node* opnd = n->in(1);
2660     if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2661       opnd = n->in(2);
2662     }
2663     func1 = eval_operand(opnd, eval_map);
2664   } else if (is_vector_binary_bitwise_op(n)) {
2665     func1 = eval_operand(n->in(1), eval_map);
2666     func2 = eval_operand(n->in(2), eval_map);
2667   } else {
2668     assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2669     func1 = eval_operand(n->in(1), eval_map);
2670     func2 = eval_operand(n->in(2), eval_map);
2671     func3 = eval_operand(n->in(3), eval_map);
2672   }
2673 }
2674 
2675 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2676   assert(inputs.size() <= 3, "sanity");
2677   ResourceMark rm;
2678   uint res = 0;
2679   ResourceHashtable<Node*,uint> eval_map;
2680 
2681   // Populate precomputed functions for inputs.
2682   // Each input corresponds to one column of 3 input truth-table.
2683   uint input_funcs[] = { 0xAA,   // (_, _, c) -> c
2684                          0xCC,   // (_, b, _) -> b
2685                          0xF0 }; // (a, _, _) -> a
2686   for (uint i = 0; i < inputs.size(); i++) {
2687     eval_map.put(inputs.at(i), input_funcs[2-i]);
2688   }
2689 
2690   for (uint i = 0; i < partition.size(); i++) {
2691     Node* n = partition.at(i);
2692 
2693     uint func1 = 0, func2 = 0, func3 = 0;
2694     eval_operands(n, func1, func2, func3, eval_map);
2695 
2696     switch (n->Opcode()) {
2697       case Op_OrV:
2698         assert(func3 == 0, "not binary");
2699         res = func1 | func2;
2700         break;
2701       case Op_AndV:
2702         assert(func3 == 0, "not binary");
2703         res = func1 & func2;
2704         break;
2705       case Op_XorV:
2706         if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2707           assert(func2 == 0 && func3 == 0, "not unary");
2708           res = (~func1) & 0xFF;
2709         } else {
2710           assert(func3 == 0, "not binary");
2711           res = func1 ^ func2;
2712         }
2713         break;
2714       case Op_MacroLogicV:
2715         // Ordering of inputs may change during evaluation of sub-tree
2716         // containing MacroLogic node as a child node, thus a re-evaluation
2717         // makes sure that function is evaluated in context of current
2718         // inputs.
2719         res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2720         break;
2721 
2722       default: assert(false, "not supported: %s", n->Name());
2723     }
2724     assert(res <= 0xFF, "invalid");
2725     eval_map.put(n, res);
2726   }
2727   return res;
2728 }
2729 
2730 // Criteria under which nodes gets packed into a macro logic node:-
2731 //  1) Parent and both child nodes are all unmasked or masked with
2732 //     same predicates.
2733 //  2) Masked parent can be packed with left child if it is predicated
2734 //     and both have same predicates.
2735 //  3) Masked parent can be packed with right child if its un-predicated
2736 //     or has matching predication condition.
2737 //  4) An unmasked parent can be packed with an unmasked child.
2738 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2739   assert(partition.size() == 0, "not empty");
2740   assert(inputs.size() == 0, "not empty");
2741   if (is_vector_ternary_bitwise_op(n)) {
2742     return false;
2743   }
2744 
2745   bool is_unary_op = is_vector_unary_bitwise_op(n);
2746   if (is_unary_op) {
2747     assert(collect_unique_inputs(n, inputs) == 1, "not unary");
2748     return false; // too few inputs
2749   }
2750 
2751   bool pack_left_child = true;
2752   bool pack_right_child = true;
2753 
2754   bool left_child_LOP = is_vector_bitwise_op(n->in(1));
2755   bool right_child_LOP = is_vector_bitwise_op(n->in(2));
2756 
2757   int left_child_input_cnt = 0;
2758   int right_child_input_cnt = 0;
2759 
2760   bool parent_is_predicated = n->is_predicated_vector();
2761   bool left_child_predicated = n->in(1)->is_predicated_vector();
2762   bool right_child_predicated = n->in(2)->is_predicated_vector();
2763 
2764   Node* parent_pred = parent_is_predicated ? n->in(n->req()-1) : NULL;
2765   Node* left_child_pred = left_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
2766   Node* right_child_pred = right_child_predicated ? n->in(1)->in(n->in(1)->req()-1) : NULL;
2767 
2768   do {
2769     if (pack_left_child && left_child_LOP &&
2770         ((!parent_is_predicated && !left_child_predicated) ||
2771         ((parent_is_predicated && left_child_predicated &&
2772           parent_pred == left_child_pred)))) {
2773        partition.push(n->in(1));
2774        left_child_input_cnt = collect_unique_inputs(n->in(1), inputs);
2775     } else {
2776        inputs.push(n->in(1));
2777        left_child_input_cnt = 1;
2778     }
2779 
2780     if (pack_right_child && right_child_LOP &&
2781         (!right_child_predicated ||
2782          (right_child_predicated && parent_is_predicated &&
2783           parent_pred == right_child_pred))) {
2784        partition.push(n->in(2));
2785        right_child_input_cnt = collect_unique_inputs(n->in(2), inputs);
2786     } else {
2787        inputs.push(n->in(2));
2788        right_child_input_cnt = 1;
2789     }
2790 
2791     if (inputs.size() > 3) {
2792       assert(partition.size() > 0, "");
2793       inputs.clear();
2794       partition.clear();
2795       if (left_child_input_cnt > right_child_input_cnt) {
2796         pack_left_child = false;
2797       } else {
2798         pack_right_child = false;
2799       }
2800     } else {
2801       break;
2802     }
2803   } while(true);
2804 
2805   if(partition.size()) {
2806     partition.push(n);
2807   }
2808 
2809   return (partition.size() == 2 || partition.size() == 3) &&
2810          (inputs.size()    == 2 || inputs.size()    == 3);
2811 }
2812 
2813 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2814   assert(is_vector_bitwise_op(n), "not a root");
2815 
2816   visited.set(n->_idx);
2817 
2818   // 1) Do a DFS walk over the logic cone.
2819   for (uint i = 1; i < n->req(); i++) {
2820     Node* in = n->in(i);
2821     if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2822       process_logic_cone_root(igvn, in, visited);
2823     }
2824   }
2825 
2826   // 2) Bottom up traversal: Merge node[s] with
2827   // the parent to form macro logic node.
2828   Unique_Node_List partition;
2829   Unique_Node_List inputs;
2830   if (compute_logic_cone(n, partition, inputs)) {
2831     const TypeVect* vt = n->bottom_type()->is_vect();
2832     Node* pn = partition.at(partition.size() - 1);
2833     Node* mask = pn->is_predicated_vector() ? pn->in(pn->req()-1) : NULL;
2834     if (mask == NULL ||
2835         Matcher::match_rule_supported_vector_masked(Op_MacroLogicV, vt->length(), vt->element_basic_type())) {
2836       Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2837 #ifdef ASSERT
2838       if (TraceNewVectors) {
2839         tty->print("new Vector node: ");
2840         macro_logic->dump();
2841       }
2842 #endif
2843       igvn.replace_node(n, macro_logic);
2844     }
2845   }
2846 }
2847 
2848 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2849   ResourceMark rm;
2850   if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2851     Unique_Node_List list;
2852     collect_logic_cone_roots(list);
2853 
2854     while (list.size() > 0) {
2855       Node* n = list.pop();
2856       const TypeVect* vt = n->bottom_type()->is_vect();
2857       bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2858       if (supported) {
2859         VectorSet visited(comp_arena());
2860         process_logic_cone_root(igvn, n, visited);
2861       }
2862     }
2863   }
2864 }
2865 
2866 //------------------------------Code_Gen---------------------------------------
2867 // Given a graph, generate code for it
2868 void Compile::Code_Gen() {
2869   if (failing()) {
2870     return;
2871   }
2872 
2873   // Perform instruction selection.  You might think we could reclaim Matcher
2874   // memory PDQ, but actually the Matcher is used in generating spill code.
2875   // Internals of the Matcher (including some VectorSets) must remain live
2876   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2877   // set a bit in reclaimed memory.
2878 
2879   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2880   // nodes.  Mapping is only valid at the root of each matched subtree.
2881   NOT_PRODUCT( verify_graph_edges(); )
2882 
2883   Matcher matcher;
2884   _matcher = &matcher;
2885   {
2886     TracePhase tp("matcher", &timers[_t_matcher]);
2887     matcher.match();
2888     if (failing()) {
2889       return;
2890     }
2891   }
2892   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2893   // nodes.  Mapping is only valid at the root of each matched subtree.
2894   NOT_PRODUCT( verify_graph_edges(); )
2895 
2896   // If you have too many nodes, or if matching has failed, bail out
2897   check_node_count(0, "out of nodes matching instructions");
2898   if (failing()) {
2899     return;
2900   }
2901 
2902   print_method(PHASE_MATCHING, 2);
2903 
2904   // Build a proper-looking CFG
2905   PhaseCFG cfg(node_arena(), root(), matcher);
2906   _cfg = &cfg;
2907   {
2908     TracePhase tp("scheduler", &timers[_t_scheduler]);
2909     bool success = cfg.do_global_code_motion();
2910     if (!success) {
2911       return;
2912     }
2913 
2914     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2915     NOT_PRODUCT( verify_graph_edges(); )
2916     cfg.verify();
2917   }
2918 
2919   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2920   _regalloc = &regalloc;
2921   {
2922     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2923     // Perform register allocation.  After Chaitin, use-def chains are
2924     // no longer accurate (at spill code) and so must be ignored.
2925     // Node->LRG->reg mappings are still accurate.
2926     _regalloc->Register_Allocate();
2927 
2928     // Bail out if the allocator builds too many nodes
2929     if (failing()) {
2930       return;
2931     }
2932   }
2933 
2934   // Prior to register allocation we kept empty basic blocks in case the
2935   // the allocator needed a place to spill.  After register allocation we
2936   // are not adding any new instructions.  If any basic block is empty, we
2937   // can now safely remove it.
2938   {
2939     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2940     cfg.remove_empty_blocks();
2941     if (do_freq_based_layout()) {
2942       PhaseBlockLayout layout(cfg);
2943     } else {
2944       cfg.set_loop_alignment();
2945     }
2946     cfg.fixup_flow();
2947     cfg.remove_unreachable_blocks();
2948     cfg.verify_dominator_tree();
2949   }
2950 
2951   // Apply peephole optimizations
2952   if( OptoPeephole ) {
2953     TracePhase tp("peephole", &timers[_t_peephole]);
2954     PhasePeephole peep( _regalloc, cfg);
2955     peep.do_transform();
2956   }
2957 
2958   // Do late expand if CPU requires this.
2959   if (Matcher::require_postalloc_expand) {
2960     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2961     cfg.postalloc_expand(_regalloc);
2962   }
2963 
2964   // Convert Nodes to instruction bits in a buffer
2965   {
2966     TracePhase tp("output", &timers[_t_output]);
2967     PhaseOutput output;
2968     output.Output();
2969     if (failing())  return;
2970     output.install();
2971   }
2972 
2973   print_method(PHASE_FINAL_CODE, 1);
2974 
2975   // He's dead, Jim.
2976   _cfg     = (PhaseCFG*)((intptr_t)0xdeadbeef);
2977   _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2978 }
2979 
2980 //------------------------------Final_Reshape_Counts---------------------------
2981 // This class defines counters to help identify when a method
2982 // may/must be executed using hardware with only 24-bit precision.
2983 struct Final_Reshape_Counts : public StackObj {
2984   int  _call_count;             // count non-inlined 'common' calls
2985   int  _float_count;            // count float ops requiring 24-bit precision
2986   int  _double_count;           // count double ops requiring more precision
2987   int  _java_call_count;        // count non-inlined 'java' calls
2988   int  _inner_loop_count;       // count loops which need alignment
2989   VectorSet _visited;           // Visitation flags
2990   Node_List _tests;             // Set of IfNodes & PCTableNodes
2991 
2992   Final_Reshape_Counts() :
2993     _call_count(0), _float_count(0), _double_count(0),
2994     _java_call_count(0), _inner_loop_count(0) { }
2995 
2996   void inc_call_count  () { _call_count  ++; }
2997   void inc_float_count () { _float_count ++; }
2998   void inc_double_count() { _double_count++; }
2999   void inc_java_call_count() { _java_call_count++; }
3000   void inc_inner_loop_count() { _inner_loop_count++; }
3001 
3002   int  get_call_count  () const { return _call_count  ; }
3003   int  get_float_count () const { return _float_count ; }
3004   int  get_double_count() const { return _double_count; }
3005   int  get_java_call_count() const { return _java_call_count; }
3006   int  get_inner_loop_count() const { return _inner_loop_count; }
3007 };
3008 
3009 // Eliminate trivially redundant StoreCMs and accumulate their
3010 // precedence edges.
3011 void Compile::eliminate_redundant_card_marks(Node* n) {
3012   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
3013   if (n->in(MemNode::Address)->outcnt() > 1) {
3014     // There are multiple users of the same address so it might be
3015     // possible to eliminate some of the StoreCMs
3016     Node* mem = n->in(MemNode::Memory);
3017     Node* adr = n->in(MemNode::Address);
3018     Node* val = n->in(MemNode::ValueIn);
3019     Node* prev = n;
3020     bool done = false;
3021     // Walk the chain of StoreCMs eliminating ones that match.  As
3022     // long as it's a chain of single users then the optimization is
3023     // safe.  Eliminating partially redundant StoreCMs would require
3024     // cloning copies down the other paths.
3025     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
3026       if (adr == mem->in(MemNode::Address) &&
3027           val == mem->in(MemNode::ValueIn)) {
3028         // redundant StoreCM
3029         if (mem->req() > MemNode::OopStore) {
3030           // Hasn't been processed by this code yet.
3031           n->add_prec(mem->in(MemNode::OopStore));
3032         } else {
3033           // Already converted to precedence edge
3034           for (uint i = mem->req(); i < mem->len(); i++) {
3035             // Accumulate any precedence edges
3036             if (mem->in(i) != NULL) {
3037               n->add_prec(mem->in(i));
3038             }
3039           }
3040           // Everything above this point has been processed.
3041           done = true;
3042         }
3043         // Eliminate the previous StoreCM
3044         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
3045         assert(mem->outcnt() == 0, "should be dead");
3046         mem->disconnect_inputs(this);
3047       } else {
3048         prev = mem;
3049       }
3050       mem = prev->in(MemNode::Memory);
3051     }
3052   }
3053 }
3054 
3055 //------------------------------final_graph_reshaping_impl----------------------
3056 // Implement items 1-5 from final_graph_reshaping below.
3057 void Compile::final_graph_reshaping_impl(Node *n, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3058 
3059   if ( n->outcnt() == 0 ) return; // dead node
3060   uint nop = n->Opcode();
3061 
3062   // Check for 2-input instruction with "last use" on right input.
3063   // Swap to left input.  Implements item (2).
3064   if( n->req() == 3 &&          // two-input instruction
3065       n->in(1)->outcnt() > 1 && // left use is NOT a last use
3066       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
3067       n->in(2)->outcnt() == 1 &&// right use IS a last use
3068       !n->in(2)->is_Con() ) {   // right use is not a constant
3069     // Check for commutative opcode
3070     switch( nop ) {
3071     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
3072     case Op_MaxI:  case Op_MaxL:  case Op_MaxF:  case Op_MaxD:
3073     case Op_MinI:  case Op_MinL:  case Op_MinF:  case Op_MinD:
3074     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
3075     case Op_AndL:  case Op_XorL:  case Op_OrL:
3076     case Op_AndI:  case Op_XorI:  case Op_OrI: {
3077       // Move "last use" input to left by swapping inputs
3078       n->swap_edges(1, 2);
3079       break;
3080     }
3081     default:
3082       break;
3083     }
3084   }
3085 
3086 #ifdef ASSERT
3087   if( n->is_Mem() ) {
3088     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
3089     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
3090             // oop will be recorded in oop map if load crosses safepoint
3091             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
3092                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
3093             "raw memory operations should have control edge");
3094   }
3095   if (n->is_MemBar()) {
3096     MemBarNode* mb = n->as_MemBar();
3097     if (mb->trailing_store() || mb->trailing_load_store()) {
3098       assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
3099       Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
3100       assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
3101              (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
3102     } else if (mb->leading()) {
3103       assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
3104     }
3105   }
3106 #endif
3107   // Count FPU ops and common calls, implements item (3)
3108   bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop, dead_nodes);
3109   if (!gc_handled) {
3110     final_graph_reshaping_main_switch(n, frc, nop, dead_nodes);
3111   }
3112 
3113   // Collect CFG split points
3114   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3115     frc._tests.push(n);
3116   }
3117 }
3118 
3119 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop, Unique_Node_List& dead_nodes) {
3120   switch( nop ) {
3121   // Count all float operations that may use FPU
3122   case Op_AddF:
3123   case Op_SubF:
3124   case Op_MulF:
3125   case Op_DivF:
3126   case Op_NegF:
3127   case Op_ModF:
3128   case Op_ConvI2F:
3129   case Op_ConF:
3130   case Op_CmpF:
3131   case Op_CmpF3:
3132   case Op_StoreF:
3133   case Op_LoadF:
3134   // case Op_ConvL2F: // longs are split into 32-bit halves
3135     frc.inc_float_count();
3136     break;
3137 
3138   case Op_ConvF2D:
3139   case Op_ConvD2F:
3140     frc.inc_float_count();
3141     frc.inc_double_count();
3142     break;
3143 
3144   // Count all double operations that may use FPU
3145   case Op_AddD:
3146   case Op_SubD:
3147   case Op_MulD:
3148   case Op_DivD:
3149   case Op_NegD:
3150   case Op_ModD:
3151   case Op_ConvI2D:
3152   case Op_ConvD2I:
3153   // case Op_ConvL2D: // handled by leaf call
3154   // case Op_ConvD2L: // handled by leaf call
3155   case Op_ConD:
3156   case Op_CmpD:
3157   case Op_CmpD3:
3158   case Op_StoreD:
3159   case Op_LoadD:
3160   case Op_LoadD_unaligned:
3161     frc.inc_double_count();
3162     break;
3163   case Op_Opaque1:              // Remove Opaque Nodes before matching
3164   case Op_Opaque3:
3165     n->subsume_by(n->in(1), this);
3166     break;
3167   case Op_CallStaticJava:
3168   case Op_CallJava:
3169   case Op_CallDynamicJava:
3170     frc.inc_java_call_count(); // Count java call site;
3171   case Op_CallRuntime:
3172   case Op_CallLeaf:
3173   case Op_CallLeafVector:
3174   case Op_CallLeafNoFP: {
3175     assert (n->is_Call(), "");
3176     CallNode *call = n->as_Call();
3177     // Count call sites where the FP mode bit would have to be flipped.
3178     // Do not count uncommon runtime calls:
3179     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3180     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3181     if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3182       frc.inc_call_count();   // Count the call site
3183     } else {                  // See if uncommon argument is shared
3184       Node *n = call->in(TypeFunc::Parms);
3185       int nop = n->Opcode();
3186       // Clone shared simple arguments to uncommon calls, item (1).
3187       if (n->outcnt() > 1 &&
3188           !n->is_Proj() &&
3189           nop != Op_CreateEx &&
3190           nop != Op_CheckCastPP &&
3191           nop != Op_DecodeN &&
3192           nop != Op_DecodeNKlass &&
3193           !n->is_Mem() &&
3194           !n->is_Phi()) {
3195         Node *x = n->clone();
3196         call->set_req(TypeFunc::Parms, x);
3197       }
3198     }
3199     break;
3200   }
3201 
3202   case Op_StoreCM:
3203     {
3204       // Convert OopStore dependence into precedence edge
3205       Node* prec = n->in(MemNode::OopStore);
3206       n->del_req(MemNode::OopStore);
3207       n->add_prec(prec);
3208       eliminate_redundant_card_marks(n);
3209     }
3210 
3211     // fall through
3212 
3213   case Op_StoreB:
3214   case Op_StoreC:
3215   case Op_StoreI:
3216   case Op_StoreL:
3217   case Op_CompareAndSwapB:
3218   case Op_CompareAndSwapS:
3219   case Op_CompareAndSwapI:
3220   case Op_CompareAndSwapL:
3221   case Op_CompareAndSwapP:
3222   case Op_CompareAndSwapN:
3223   case Op_WeakCompareAndSwapB:
3224   case Op_WeakCompareAndSwapS:
3225   case Op_WeakCompareAndSwapI:
3226   case Op_WeakCompareAndSwapL:
3227   case Op_WeakCompareAndSwapP:
3228   case Op_WeakCompareAndSwapN:
3229   case Op_CompareAndExchangeB:
3230   case Op_CompareAndExchangeS:
3231   case Op_CompareAndExchangeI:
3232   case Op_CompareAndExchangeL:
3233   case Op_CompareAndExchangeP:
3234   case Op_CompareAndExchangeN:
3235   case Op_GetAndAddS:
3236   case Op_GetAndAddB:
3237   case Op_GetAndAddI:
3238   case Op_GetAndAddL:
3239   case Op_GetAndSetS:
3240   case Op_GetAndSetB:
3241   case Op_GetAndSetI:
3242   case Op_GetAndSetL:
3243   case Op_GetAndSetP:
3244   case Op_GetAndSetN:
3245   case Op_StoreP:
3246   case Op_StoreN:
3247   case Op_StoreNKlass:
3248   case Op_LoadB:
3249   case Op_LoadUB:
3250   case Op_LoadUS:
3251   case Op_LoadI:
3252   case Op_LoadKlass:
3253   case Op_LoadNKlass:
3254   case Op_LoadL:
3255   case Op_LoadL_unaligned:
3256   case Op_LoadP:
3257   case Op_LoadN:
3258   case Op_LoadRange:
3259   case Op_LoadS:
3260     break;
3261 
3262   case Op_AddP: {               // Assert sane base pointers
3263     Node *addp = n->in(AddPNode::Address);
3264     assert( !addp->is_AddP() ||
3265             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3266             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3267             "Base pointers must match (addp %u)", addp->_idx );
3268 #ifdef _LP64
3269     if ((UseCompressedOops || UseCompressedClassPointers) &&
3270         addp->Opcode() == Op_ConP &&
3271         addp == n->in(AddPNode::Base) &&
3272         n->in(AddPNode::Offset)->is_Con()) {
3273       // If the transformation of ConP to ConN+DecodeN is beneficial depends
3274       // on the platform and on the compressed oops mode.
3275       // Use addressing with narrow klass to load with offset on x86.
3276       // Some platforms can use the constant pool to load ConP.
3277       // Do this transformation here since IGVN will convert ConN back to ConP.
3278       const Type* t = addp->bottom_type();
3279       bool is_oop   = t->isa_oopptr() != NULL;
3280       bool is_klass = t->isa_klassptr() != NULL;
3281 
3282       if ((is_oop   && Matcher::const_oop_prefer_decode()  ) ||
3283           (is_klass && Matcher::const_klass_prefer_decode())) {
3284         Node* nn = NULL;
3285 
3286         int op = is_oop ? Op_ConN : Op_ConNKlass;
3287 
3288         // Look for existing ConN node of the same exact type.
3289         Node* r  = root();
3290         uint cnt = r->outcnt();
3291         for (uint i = 0; i < cnt; i++) {
3292           Node* m = r->raw_out(i);
3293           if (m!= NULL && m->Opcode() == op &&
3294               m->bottom_type()->make_ptr() == t) {
3295             nn = m;
3296             break;
3297           }
3298         }
3299         if (nn != NULL) {
3300           // Decode a narrow oop to match address
3301           // [R12 + narrow_oop_reg<<3 + offset]
3302           if (is_oop) {
3303             nn = new DecodeNNode(nn, t);
3304           } else {
3305             nn = new DecodeNKlassNode(nn, t);
3306           }
3307           // Check for succeeding AddP which uses the same Base.
3308           // Otherwise we will run into the assertion above when visiting that guy.
3309           for (uint i = 0; i < n->outcnt(); ++i) {
3310             Node *out_i = n->raw_out(i);
3311             if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3312               out_i->set_req(AddPNode::Base, nn);
3313 #ifdef ASSERT
3314               for (uint j = 0; j < out_i->outcnt(); ++j) {
3315                 Node *out_j = out_i->raw_out(j);
3316                 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3317                        "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3318               }
3319 #endif
3320             }
3321           }
3322           n->set_req(AddPNode::Base, nn);
3323           n->set_req(AddPNode::Address, nn);
3324           if (addp->outcnt() == 0) {
3325             addp->disconnect_inputs(this);
3326           }
3327         }
3328       }
3329     }
3330 #endif
3331     break;
3332   }
3333 
3334   case Op_CastPP: {
3335     // Remove CastPP nodes to gain more freedom during scheduling but
3336     // keep the dependency they encode as control or precedence edges
3337     // (if control is set already) on memory operations. Some CastPP
3338     // nodes don't have a control (don't carry a dependency): skip
3339     // those.
3340     if (n->in(0) != NULL) {
3341       ResourceMark rm;
3342       Unique_Node_List wq;
3343       wq.push(n);
3344       for (uint next = 0; next < wq.size(); ++next) {
3345         Node *m = wq.at(next);
3346         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3347           Node* use = m->fast_out(i);
3348           if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3349             use->ensure_control_or_add_prec(n->in(0));
3350           } else {
3351             switch(use->Opcode()) {
3352             case Op_AddP:
3353             case Op_DecodeN:
3354             case Op_DecodeNKlass:
3355             case Op_CheckCastPP:
3356             case Op_CastPP:
3357               wq.push(use);
3358               break;
3359             }
3360           }
3361         }
3362       }
3363     }
3364     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3365     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3366       Node* in1 = n->in(1);
3367       const Type* t = n->bottom_type();
3368       Node* new_in1 = in1->clone();
3369       new_in1->as_DecodeN()->set_type(t);
3370 
3371       if (!Matcher::narrow_oop_use_complex_address()) {
3372         //
3373         // x86, ARM and friends can handle 2 adds in addressing mode
3374         // and Matcher can fold a DecodeN node into address by using
3375         // a narrow oop directly and do implicit NULL check in address:
3376         //
3377         // [R12 + narrow_oop_reg<<3 + offset]
3378         // NullCheck narrow_oop_reg
3379         //
3380         // On other platforms (Sparc) we have to keep new DecodeN node and
3381         // use it to do implicit NULL check in address:
3382         //
3383         // decode_not_null narrow_oop_reg, base_reg
3384         // [base_reg + offset]
3385         // NullCheck base_reg
3386         //
3387         // Pin the new DecodeN node to non-null path on these platform (Sparc)
3388         // to keep the information to which NULL check the new DecodeN node
3389         // corresponds to use it as value in implicit_null_check().
3390         //
3391         new_in1->set_req(0, n->in(0));
3392       }
3393 
3394       n->subsume_by(new_in1, this);
3395       if (in1->outcnt() == 0) {
3396         in1->disconnect_inputs(this);
3397       }
3398     } else {
3399       n->subsume_by(n->in(1), this);
3400       if (n->outcnt() == 0) {
3401         n->disconnect_inputs(this);
3402       }
3403     }
3404     break;
3405   }
3406 #ifdef _LP64
3407   case Op_CmpP:
3408     // Do this transformation here to preserve CmpPNode::sub() and
3409     // other TypePtr related Ideal optimizations (for example, ptr nullness).
3410     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3411       Node* in1 = n->in(1);
3412       Node* in2 = n->in(2);
3413       if (!in1->is_DecodeNarrowPtr()) {
3414         in2 = in1;
3415         in1 = n->in(2);
3416       }
3417       assert(in1->is_DecodeNarrowPtr(), "sanity");
3418 
3419       Node* new_in2 = NULL;
3420       if (in2->is_DecodeNarrowPtr()) {
3421         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3422         new_in2 = in2->in(1);
3423       } else if (in2->Opcode() == Op_ConP) {
3424         const Type* t = in2->bottom_type();
3425         if (t == TypePtr::NULL_PTR) {
3426           assert(in1->is_DecodeN(), "compare klass to null?");
3427           // Don't convert CmpP null check into CmpN if compressed
3428           // oops implicit null check is not generated.
3429           // This will allow to generate normal oop implicit null check.
3430           if (Matcher::gen_narrow_oop_implicit_null_checks())
3431             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3432           //
3433           // This transformation together with CastPP transformation above
3434           // will generated code for implicit NULL checks for compressed oops.
3435           //
3436           // The original code after Optimize()
3437           //
3438           //    LoadN memory, narrow_oop_reg
3439           //    decode narrow_oop_reg, base_reg
3440           //    CmpP base_reg, NULL
3441           //    CastPP base_reg // NotNull
3442           //    Load [base_reg + offset], val_reg
3443           //
3444           // after these transformations will be
3445           //
3446           //    LoadN memory, narrow_oop_reg
3447           //    CmpN narrow_oop_reg, NULL
3448           //    decode_not_null narrow_oop_reg, base_reg
3449           //    Load [base_reg + offset], val_reg
3450           //
3451           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3452           // since narrow oops can be used in debug info now (see the code in
3453           // final_graph_reshaping_walk()).
3454           //
3455           // At the end the code will be matched to
3456           // on x86:
3457           //
3458           //    Load_narrow_oop memory, narrow_oop_reg
3459           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3460           //    NullCheck narrow_oop_reg
3461           //
3462           // and on sparc:
3463           //
3464           //    Load_narrow_oop memory, narrow_oop_reg
3465           //    decode_not_null narrow_oop_reg, base_reg
3466           //    Load [base_reg + offset], val_reg
3467           //    NullCheck base_reg
3468           //
3469         } else if (t->isa_oopptr()) {
3470           new_in2 = ConNode::make(t->make_narrowoop());
3471         } else if (t->isa_klassptr()) {
3472           new_in2 = ConNode::make(t->make_narrowklass());
3473         }
3474       }
3475       if (new_in2 != NULL) {
3476         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3477         n->subsume_by(cmpN, this);
3478         if (in1->outcnt() == 0) {
3479           in1->disconnect_inputs(this);
3480         }
3481         if (in2->outcnt() == 0) {
3482           in2->disconnect_inputs(this);
3483         }
3484       }
3485     }
3486     break;
3487 
3488   case Op_DecodeN:
3489   case Op_DecodeNKlass:
3490     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3491     // DecodeN could be pinned when it can't be fold into
3492     // an address expression, see the code for Op_CastPP above.
3493     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3494     break;
3495 
3496   case Op_EncodeP:
3497   case Op_EncodePKlass: {
3498     Node* in1 = n->in(1);
3499     if (in1->is_DecodeNarrowPtr()) {
3500       n->subsume_by(in1->in(1), this);
3501     } else if (in1->Opcode() == Op_ConP) {
3502       const Type* t = in1->bottom_type();
3503       if (t == TypePtr::NULL_PTR) {
3504         assert(t->isa_oopptr(), "null klass?");
3505         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3506       } else if (t->isa_oopptr()) {
3507         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3508       } else if (t->isa_klassptr()) {
3509         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3510       }
3511     }
3512     if (in1->outcnt() == 0) {
3513       in1->disconnect_inputs(this);
3514     }
3515     break;
3516   }
3517 
3518   case Op_Proj: {
3519     if (OptimizeStringConcat || IncrementalInline) {
3520       ProjNode* proj = n->as_Proj();
3521       if (proj->_is_io_use) {
3522         assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3523         // Separate projections were used for the exception path which
3524         // are normally removed by a late inline.  If it wasn't inlined
3525         // then they will hang around and should just be replaced with
3526         // the original one. Merge them.
3527         Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3528         if (non_io_proj  != NULL) {
3529           proj->subsume_by(non_io_proj , this);
3530         }
3531       }
3532     }
3533     break;
3534   }
3535 
3536   case Op_Phi:
3537     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3538       // The EncodeP optimization may create Phi with the same edges
3539       // for all paths. It is not handled well by Register Allocator.
3540       Node* unique_in = n->in(1);
3541       assert(unique_in != NULL, "");
3542       uint cnt = n->req();
3543       for (uint i = 2; i < cnt; i++) {
3544         Node* m = n->in(i);
3545         assert(m != NULL, "");
3546         if (unique_in != m)
3547           unique_in = NULL;
3548       }
3549       if (unique_in != NULL) {
3550         n->subsume_by(unique_in, this);
3551       }
3552     }
3553     break;
3554 
3555 #endif
3556 
3557 #ifdef ASSERT
3558   case Op_CastII:
3559     // Verify that all range check dependent CastII nodes were removed.
3560     if (n->isa_CastII()->has_range_check()) {
3561       n->dump(3);
3562       assert(false, "Range check dependent CastII node was not removed");
3563     }
3564     break;
3565 #endif
3566 
3567   case Op_ModI:
3568     if (UseDivMod) {
3569       // Check if a%b and a/b both exist
3570       Node* d = n->find_similar(Op_DivI);
3571       if (d) {
3572         // Replace them with a fused divmod if supported
3573         if (Matcher::has_match_rule(Op_DivModI)) {
3574           DivModINode* divmod = DivModINode::make(n);
3575           d->subsume_by(divmod->div_proj(), this);
3576           n->subsume_by(divmod->mod_proj(), this);
3577         } else {
3578           // replace a%b with a-((a/b)*b)
3579           Node* mult = new MulINode(d, d->in(2));
3580           Node* sub  = new SubINode(d->in(1), mult);
3581           n->subsume_by(sub, this);
3582         }
3583       }
3584     }
3585     break;
3586 
3587   case Op_ModL:
3588     if (UseDivMod) {
3589       // Check if a%b and a/b both exist
3590       Node* d = n->find_similar(Op_DivL);
3591       if (d) {
3592         // Replace them with a fused divmod if supported
3593         if (Matcher::has_match_rule(Op_DivModL)) {
3594           DivModLNode* divmod = DivModLNode::make(n);
3595           d->subsume_by(divmod->div_proj(), this);
3596           n->subsume_by(divmod->mod_proj(), this);
3597         } else {
3598           // replace a%b with a-((a/b)*b)
3599           Node* mult = new MulLNode(d, d->in(2));
3600           Node* sub  = new SubLNode(d->in(1), mult);
3601           n->subsume_by(sub, this);
3602         }
3603       }
3604     }
3605     break;
3606 
3607   case Op_UModI:
3608     if (UseDivMod) {
3609       // Check if a%b and a/b both exist
3610       Node* d = n->find_similar(Op_UDivI);
3611       if (d) {
3612         // Replace them with a fused unsigned divmod if supported
3613         if (Matcher::has_match_rule(Op_UDivModI)) {
3614           UDivModINode* divmod = UDivModINode::make(n);
3615           d->subsume_by(divmod->div_proj(), this);
3616           n->subsume_by(divmod->mod_proj(), this);
3617         } else {
3618           // replace a%b with a-((a/b)*b)
3619           Node* mult = new MulINode(d, d->in(2));
3620           Node* sub  = new SubINode(d->in(1), mult);
3621           n->subsume_by(sub, this);
3622         }
3623       }
3624     }
3625     break;
3626 
3627   case Op_UModL:
3628     if (UseDivMod) {
3629       // Check if a%b and a/b both exist
3630       Node* d = n->find_similar(Op_UDivL);
3631       if (d) {
3632         // Replace them with a fused unsigned divmod if supported
3633         if (Matcher::has_match_rule(Op_UDivModL)) {
3634           UDivModLNode* divmod = UDivModLNode::make(n);
3635           d->subsume_by(divmod->div_proj(), this);
3636           n->subsume_by(divmod->mod_proj(), this);
3637         } else {
3638           // replace a%b with a-((a/b)*b)
3639           Node* mult = new MulLNode(d, d->in(2));
3640           Node* sub  = new SubLNode(d->in(1), mult);
3641           n->subsume_by(sub, this);
3642         }
3643       }
3644     }
3645     break;
3646 
3647   case Op_LoadVector:
3648   case Op_StoreVector:
3649   case Op_LoadVectorGather:
3650   case Op_StoreVectorScatter:
3651   case Op_LoadVectorGatherMasked:
3652   case Op_StoreVectorScatterMasked:
3653   case Op_VectorCmpMasked:
3654   case Op_VectorMaskGen:
3655   case Op_LoadVectorMasked:
3656   case Op_StoreVectorMasked:
3657     break;
3658 
3659   case Op_AddReductionVI:
3660   case Op_AddReductionVL:
3661   case Op_AddReductionVF:
3662   case Op_AddReductionVD:
3663   case Op_MulReductionVI:
3664   case Op_MulReductionVL:
3665   case Op_MulReductionVF:
3666   case Op_MulReductionVD:
3667   case Op_MinReductionV:
3668   case Op_MaxReductionV:
3669   case Op_AndReductionV:
3670   case Op_OrReductionV:
3671   case Op_XorReductionV:
3672     break;
3673 
3674   case Op_PackB:
3675   case Op_PackS:
3676   case Op_PackI:
3677   case Op_PackF:
3678   case Op_PackL:
3679   case Op_PackD:
3680     if (n->req()-1 > 2) {
3681       // Replace many operand PackNodes with a binary tree for matching
3682       PackNode* p = (PackNode*) n;
3683       Node* btp = p->binary_tree_pack(1, n->req());
3684       n->subsume_by(btp, this);
3685     }
3686     break;
3687   case Op_Loop:
3688     assert(!n->as_Loop()->is_loop_nest_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3689   case Op_CountedLoop:
3690   case Op_LongCountedLoop:
3691   case Op_OuterStripMinedLoop:
3692     if (n->as_Loop()->is_inner_loop()) {
3693       frc.inc_inner_loop_count();
3694     }
3695     n->as_Loop()->verify_strip_mined(0);
3696     break;
3697   case Op_LShiftI:
3698   case Op_RShiftI:
3699   case Op_URShiftI:
3700   case Op_LShiftL:
3701   case Op_RShiftL:
3702   case Op_URShiftL:
3703     if (Matcher::need_masked_shift_count) {
3704       // The cpu's shift instructions don't restrict the count to the
3705       // lower 5/6 bits. We need to do the masking ourselves.
3706       Node* in2 = n->in(2);
3707       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3708       const TypeInt* t = in2->find_int_type();
3709       if (t != NULL && t->is_con()) {
3710         juint shift = t->get_con();
3711         if (shift > mask) { // Unsigned cmp
3712           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3713         }
3714       } else {
3715         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3716           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3717           n->set_req(2, shift);
3718         }
3719       }
3720       if (in2->outcnt() == 0) { // Remove dead node
3721         in2->disconnect_inputs(this);
3722       }
3723     }
3724     break;
3725   case Op_MemBarStoreStore:
3726   case Op_MemBarRelease:
3727     // Break the link with AllocateNode: it is no longer useful and
3728     // confuses register allocation.
3729     if (n->req() > MemBarNode::Precedent) {
3730       n->set_req(MemBarNode::Precedent, top());
3731     }
3732     break;
3733   case Op_MemBarAcquire: {
3734     if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3735       // At parse time, the trailing MemBarAcquire for a volatile load
3736       // is created with an edge to the load. After optimizations,
3737       // that input may be a chain of Phis. If those phis have no
3738       // other use, then the MemBarAcquire keeps them alive and
3739       // register allocation can be confused.
3740       dead_nodes.push(n->in(MemBarNode::Precedent));
3741       n->set_req(MemBarNode::Precedent, top());
3742     }
3743     break;
3744   }
3745   case Op_Blackhole:
3746     break;
3747   case Op_RangeCheck: {
3748     RangeCheckNode* rc = n->as_RangeCheck();
3749     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3750     n->subsume_by(iff, this);
3751     frc._tests.push(iff);
3752     break;
3753   }
3754   case Op_ConvI2L: {
3755     if (!Matcher::convi2l_type_required) {
3756       // Code generation on some platforms doesn't need accurate
3757       // ConvI2L types. Widening the type can help remove redundant
3758       // address computations.
3759       n->as_Type()->set_type(TypeLong::INT);
3760       ResourceMark rm;
3761       Unique_Node_List wq;
3762       wq.push(n);
3763       for (uint next = 0; next < wq.size(); next++) {
3764         Node *m = wq.at(next);
3765 
3766         for(;;) {
3767           // Loop over all nodes with identical inputs edges as m
3768           Node* k = m->find_similar(m->Opcode());
3769           if (k == NULL) {
3770             break;
3771           }
3772           // Push their uses so we get a chance to remove node made
3773           // redundant
3774           for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3775             Node* u = k->fast_out(i);
3776             if (u->Opcode() == Op_LShiftL ||
3777                 u->Opcode() == Op_AddL ||
3778                 u->Opcode() == Op_SubL ||
3779                 u->Opcode() == Op_AddP) {
3780               wq.push(u);
3781             }
3782           }
3783           // Replace all nodes with identical edges as m with m
3784           k->subsume_by(m, this);
3785         }
3786       }
3787     }
3788     break;
3789   }
3790   case Op_CmpUL: {
3791     if (!Matcher::has_match_rule(Op_CmpUL)) {
3792       // No support for unsigned long comparisons
3793       ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3794       Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3795       Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3796       ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3797       Node* andl = new AndLNode(orl, remove_sign_mask);
3798       Node* cmp = new CmpLNode(andl, n->in(2));
3799       n->subsume_by(cmp, this);
3800     }
3801     break;
3802   }
3803   default:
3804     assert(!n->is_Call(), "");
3805     assert(!n->is_Mem(), "");
3806     assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3807     break;
3808   }
3809 }
3810 
3811 //------------------------------final_graph_reshaping_walk---------------------
3812 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3813 // requires that the walk visits a node's inputs before visiting the node.
3814 void Compile::final_graph_reshaping_walk(Node_Stack& nstack, Node* root, Final_Reshape_Counts& frc, Unique_Node_List& dead_nodes) {
3815   Unique_Node_List sfpt;
3816 
3817   frc._visited.set(root->_idx); // first, mark node as visited
3818   uint cnt = root->req();
3819   Node *n = root;
3820   uint  i = 0;
3821   while (true) {
3822     if (i < cnt) {
3823       // Place all non-visited non-null inputs onto stack
3824       Node* m = n->in(i);
3825       ++i;
3826       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3827         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3828           // compute worst case interpreter size in case of a deoptimization
3829           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3830 
3831           sfpt.push(m);
3832         }
3833         cnt = m->req();
3834         nstack.push(n, i); // put on stack parent and next input's index
3835         n = m;
3836         i = 0;
3837       }
3838     } else {
3839       // Now do post-visit work
3840       final_graph_reshaping_impl(n, frc, dead_nodes);
3841       if (nstack.is_empty())
3842         break;             // finished
3843       n = nstack.node();   // Get node from stack
3844       cnt = n->req();
3845       i = nstack.index();
3846       nstack.pop();        // Shift to the next node on stack
3847     }
3848   }
3849 
3850   // Skip next transformation if compressed oops are not used.
3851   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3852       (!UseCompressedOops && !UseCompressedClassPointers))
3853     return;
3854 
3855   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3856   // It could be done for an uncommon traps or any safepoints/calls
3857   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3858   while (sfpt.size() > 0) {
3859     n = sfpt.pop();
3860     JVMState *jvms = n->as_SafePoint()->jvms();
3861     assert(jvms != NULL, "sanity");
3862     int start = jvms->debug_start();
3863     int end   = n->req();
3864     bool is_uncommon = (n->is_CallStaticJava() &&
3865                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3866     for (int j = start; j < end; j++) {
3867       Node* in = n->in(j);
3868       if (in->is_DecodeNarrowPtr()) {
3869         bool safe_to_skip = true;
3870         if (!is_uncommon ) {
3871           // Is it safe to skip?
3872           for (uint i = 0; i < in->outcnt(); i++) {
3873             Node* u = in->raw_out(i);
3874             if (!u->is_SafePoint() ||
3875                 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3876               safe_to_skip = false;
3877             }
3878           }
3879         }
3880         if (safe_to_skip) {
3881           n->set_req(j, in->in(1));
3882         }
3883         if (in->outcnt() == 0) {
3884           in->disconnect_inputs(this);
3885         }
3886       }
3887     }
3888   }
3889 }
3890 
3891 //------------------------------final_graph_reshaping--------------------------
3892 // Final Graph Reshaping.
3893 //
3894 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3895 //     and not commoned up and forced early.  Must come after regular
3896 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3897 //     inputs to Loop Phis; these will be split by the allocator anyways.
3898 //     Remove Opaque nodes.
3899 // (2) Move last-uses by commutative operations to the left input to encourage
3900 //     Intel update-in-place two-address operations and better register usage
3901 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3902 //     calls canonicalizing them back.
3903 // (3) Count the number of double-precision FP ops, single-precision FP ops
3904 //     and call sites.  On Intel, we can get correct rounding either by
3905 //     forcing singles to memory (requires extra stores and loads after each
3906 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3907 //     clearing the mode bit around call sites).  The mode bit is only used
3908 //     if the relative frequency of single FP ops to calls is low enough.
3909 //     This is a key transform for SPEC mpeg_audio.
3910 // (4) Detect infinite loops; blobs of code reachable from above but not
3911 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3912 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3913 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3914 //     Detection is by looking for IfNodes where only 1 projection is
3915 //     reachable from below or CatchNodes missing some targets.
3916 // (5) Assert for insane oop offsets in debug mode.
3917 
3918 bool Compile::final_graph_reshaping() {
3919   // an infinite loop may have been eliminated by the optimizer,
3920   // in which case the graph will be empty.
3921   if (root()->req() == 1) {
3922     record_method_not_compilable("trivial infinite loop");
3923     return true;
3924   }
3925 
3926   // Expensive nodes have their control input set to prevent the GVN
3927   // from freely commoning them. There's no GVN beyond this point so
3928   // no need to keep the control input. We want the expensive nodes to
3929   // be freely moved to the least frequent code path by gcm.
3930   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3931   for (int i = 0; i < expensive_count(); i++) {
3932     _expensive_nodes.at(i)->set_req(0, NULL);
3933   }
3934 
3935   Final_Reshape_Counts frc;
3936 
3937   // Visit everybody reachable!
3938   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3939   Node_Stack nstack(live_nodes() >> 1);
3940   Unique_Node_List dead_nodes;
3941   final_graph_reshaping_walk(nstack, root(), frc, dead_nodes);
3942 
3943   // Check for unreachable (from below) code (i.e., infinite loops).
3944   for( uint i = 0; i < frc._tests.size(); i++ ) {
3945     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3946     // Get number of CFG targets.
3947     // Note that PCTables include exception targets after calls.
3948     uint required_outcnt = n->required_outcnt();
3949     if (n->outcnt() != required_outcnt) {
3950       // Check for a few special cases.  Rethrow Nodes never take the
3951       // 'fall-thru' path, so expected kids is 1 less.
3952       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3953         if (n->in(0)->in(0)->is_Call()) {
3954           CallNode* call = n->in(0)->in(0)->as_Call();
3955           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3956             required_outcnt--;      // Rethrow always has 1 less kid
3957           } else if (call->req() > TypeFunc::Parms &&
3958                      call->is_CallDynamicJava()) {
3959             // Check for null receiver. In such case, the optimizer has
3960             // detected that the virtual call will always result in a null
3961             // pointer exception. The fall-through projection of this CatchNode
3962             // will not be populated.
3963             Node* arg0 = call->in(TypeFunc::Parms);
3964             if (arg0->is_Type() &&
3965                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3966               required_outcnt--;
3967             }
3968           } else if (call->entry_point() == OptoRuntime::new_array_Java() ||
3969                      call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3970             // Check for illegal array length. In such case, the optimizer has
3971             // detected that the allocation attempt will always result in an
3972             // exception. There is no fall-through projection of this CatchNode .
3973             assert(call->is_CallStaticJava(), "static call expected");
3974             assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3975             uint valid_length_test_input = call->req() - 1;
3976             Node* valid_length_test = call->in(valid_length_test_input);
3977             call->del_req(valid_length_test_input);
3978             if (valid_length_test->find_int_con(1) == 0) {
3979               required_outcnt--;
3980             }
3981             dead_nodes.push(valid_length_test);
3982             assert(n->outcnt() == required_outcnt, "malformed control flow");
3983             continue;
3984           }
3985         }
3986       }
3987       // Recheck with a better notion of 'required_outcnt'
3988       if (n->outcnt() != required_outcnt) {
3989         record_method_not_compilable("malformed control flow");
3990         return true;            // Not all targets reachable!
3991       }
3992     } else if (n->is_PCTable() && n->in(0) && n->in(0)->in(0) && n->in(0)->in(0)->is_Call()) {
3993       CallNode* call = n->in(0)->in(0)->as_Call();
3994       if (call->entry_point() == OptoRuntime::new_array_Java() ||
3995           call->entry_point() == OptoRuntime::new_array_nozero_Java()) {
3996         assert(call->is_CallStaticJava(), "static call expected");
3997         assert(call->req() == call->jvms()->endoff() + 1, "missing extra input");
3998         uint valid_length_test_input = call->req() - 1;
3999         dead_nodes.push(call->in(valid_length_test_input));
4000         call->del_req(valid_length_test_input); // valid length test useless now
4001       }
4002     }
4003     // Check that I actually visited all kids.  Unreached kids
4004     // must be infinite loops.
4005     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
4006       if (!frc._visited.test(n->fast_out(j)->_idx)) {
4007         record_method_not_compilable("infinite loop");
4008         return true;            // Found unvisited kid; must be unreach
4009       }
4010 
4011     // Here so verification code in final_graph_reshaping_walk()
4012     // always see an OuterStripMinedLoopEnd
4013     if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
4014       IfNode* init_iff = n->as_If();
4015       Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
4016       n->subsume_by(iff, this);
4017     }
4018   }
4019 
4020   while (dead_nodes.size() > 0) {
4021     Node* m = dead_nodes.pop();
4022     if (m->outcnt() == 0 && m != top()) {
4023       for (uint j = 0; j < m->req(); j++) {
4024         Node* in = m->in(j);
4025         if (in != NULL) {
4026           dead_nodes.push(in);
4027         }
4028       }
4029       m->disconnect_inputs(this);
4030     }
4031   }
4032 
4033 #ifdef IA32
4034   // If original bytecodes contained a mixture of floats and doubles
4035   // check if the optimizer has made it homogeneous, item (3).
4036   if (UseSSE == 0 &&
4037       frc.get_float_count() > 32 &&
4038       frc.get_double_count() == 0 &&
4039       (10 * frc.get_call_count() < frc.get_float_count()) ) {
4040     set_24_bit_selection_and_mode(false, true);
4041   }
4042 #endif // IA32
4043 
4044   set_java_calls(frc.get_java_call_count());
4045   set_inner_loops(frc.get_inner_loop_count());
4046 
4047   // No infinite loops, no reason to bail out.
4048   return false;
4049 }
4050 
4051 //-----------------------------too_many_traps----------------------------------
4052 // Report if there are too many traps at the current method and bci.
4053 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
4054 bool Compile::too_many_traps(ciMethod* method,
4055                              int bci,
4056                              Deoptimization::DeoptReason reason) {
4057   ciMethodData* md = method->method_data();
4058   if (md->is_empty()) {
4059     // Assume the trap has not occurred, or that it occurred only
4060     // because of a transient condition during start-up in the interpreter.
4061     return false;
4062   }
4063   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4064   if (md->has_trap_at(bci, m, reason) != 0) {
4065     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
4066     // Also, if there are multiple reasons, or if there is no per-BCI record,
4067     // assume the worst.
4068     if (log())
4069       log()->elem("observe trap='%s' count='%d'",
4070                   Deoptimization::trap_reason_name(reason),
4071                   md->trap_count(reason));
4072     return true;
4073   } else {
4074     // Ignore method/bci and see if there have been too many globally.
4075     return too_many_traps(reason, md);
4076   }
4077 }
4078 
4079 // Less-accurate variant which does not require a method and bci.
4080 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
4081                              ciMethodData* logmd) {
4082   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
4083     // Too many traps globally.
4084     // Note that we use cumulative trap_count, not just md->trap_count.
4085     if (log()) {
4086       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
4087       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
4088                   Deoptimization::trap_reason_name(reason),
4089                   mcount, trap_count(reason));
4090     }
4091     return true;
4092   } else {
4093     // The coast is clear.
4094     return false;
4095   }
4096 }
4097 
4098 //--------------------------too_many_recompiles--------------------------------
4099 // Report if there are too many recompiles at the current method and bci.
4100 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
4101 // Is not eager to return true, since this will cause the compiler to use
4102 // Action_none for a trap point, to avoid too many recompilations.
4103 bool Compile::too_many_recompiles(ciMethod* method,
4104                                   int bci,
4105                                   Deoptimization::DeoptReason reason) {
4106   ciMethodData* md = method->method_data();
4107   if (md->is_empty()) {
4108     // Assume the trap has not occurred, or that it occurred only
4109     // because of a transient condition during start-up in the interpreter.
4110     return false;
4111   }
4112   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
4113   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
4114   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
4115   Deoptimization::DeoptReason per_bc_reason
4116     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
4117   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
4118   if ((per_bc_reason == Deoptimization::Reason_none
4119        || md->has_trap_at(bci, m, reason) != 0)
4120       // The trap frequency measure we care about is the recompile count:
4121       && md->trap_recompiled_at(bci, m)
4122       && md->overflow_recompile_count() >= bc_cutoff) {
4123     // Do not emit a trap here if it has already caused recompilations.
4124     // Also, if there are multiple reasons, or if there is no per-BCI record,
4125     // assume the worst.
4126     if (log())
4127       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
4128                   Deoptimization::trap_reason_name(reason),
4129                   md->trap_count(reason),
4130                   md->overflow_recompile_count());
4131     return true;
4132   } else if (trap_count(reason) != 0
4133              && decompile_count() >= m_cutoff) {
4134     // Too many recompiles globally, and we have seen this sort of trap.
4135     // Use cumulative decompile_count, not just md->decompile_count.
4136     if (log())
4137       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
4138                   Deoptimization::trap_reason_name(reason),
4139                   md->trap_count(reason), trap_count(reason),
4140                   md->decompile_count(), decompile_count());
4141     return true;
4142   } else {
4143     // The coast is clear.
4144     return false;
4145   }
4146 }
4147 
4148 // Compute when not to trap. Used by matching trap based nodes and
4149 // NullCheck optimization.
4150 void Compile::set_allowed_deopt_reasons() {
4151   _allowed_reasons = 0;
4152   if (is_method_compilation()) {
4153     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
4154       assert(rs < BitsPerInt, "recode bit map");
4155       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
4156         _allowed_reasons |= nth_bit(rs);
4157       }
4158     }
4159   }
4160 }
4161 
4162 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
4163   return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
4164 }
4165 
4166 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
4167   return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
4168 }
4169 
4170 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
4171   if (holder->is_initialized()) {
4172     return false;
4173   }
4174   if (holder->is_being_initialized()) {
4175     if (accessing_method->holder() == holder) {
4176       // Access inside a class. The barrier can be elided when access happens in <clinit>,
4177       // <init>, or a static method. In all those cases, there was an initialization
4178       // barrier on the holder klass passed.
4179       if (accessing_method->is_static_initializer() ||
4180           accessing_method->is_object_initializer() ||
4181           accessing_method->is_static()) {
4182         return false;
4183       }
4184     } else if (accessing_method->holder()->is_subclass_of(holder)) {
4185       // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
4186       // In case of <init> or a static method, the barrier is on the subclass is not enough:
4187       // child class can become fully initialized while its parent class is still being initialized.
4188       if (accessing_method->is_static_initializer()) {
4189         return false;
4190       }
4191     }
4192     ciMethod* root = method(); // the root method of compilation
4193     if (root != accessing_method) {
4194       return needs_clinit_barrier(holder, root); // check access in the context of compilation root
4195     }
4196   }
4197   return true;
4198 }
4199 
4200 #ifndef PRODUCT
4201 //------------------------------verify_bidirectional_edges---------------------
4202 // For each input edge to a node (ie - for each Use-Def edge), verify that
4203 // there is a corresponding Def-Use edge.
4204 void Compile::verify_bidirectional_edges(Unique_Node_List &visited) {
4205   // Allocate stack of size C->live_nodes()/16 to avoid frequent realloc
4206   uint stack_size = live_nodes() >> 4;
4207   Node_List nstack(MAX2(stack_size, (uint)OptoNodeListSize));
4208   nstack.push(_root);
4209 
4210   while (nstack.size() > 0) {
4211     Node* n = nstack.pop();
4212     if (visited.member(n)) {
4213       continue;
4214     }
4215     visited.push(n);
4216 
4217     // Walk over all input edges, checking for correspondence
4218     uint length = n->len();
4219     for (uint i = 0; i < length; i++) {
4220       Node* in = n->in(i);
4221       if (in != NULL && !visited.member(in)) {
4222         nstack.push(in); // Put it on stack
4223       }
4224       if (in != NULL && !in->is_top()) {
4225         // Count instances of `next`
4226         int cnt = 0;
4227         for (uint idx = 0; idx < in->_outcnt; idx++) {
4228           if (in->_out[idx] == n) {
4229             cnt++;
4230           }
4231         }
4232         assert(cnt > 0, "Failed to find Def-Use edge.");
4233         // Check for duplicate edges
4234         // walk the input array downcounting the input edges to n
4235         for (uint j = 0; j < length; j++) {
4236           if (n->in(j) == in) {
4237             cnt--;
4238           }
4239         }
4240         assert(cnt == 0, "Mismatched edge count.");
4241       } else if (in == NULL) {
4242         assert(i == 0 || i >= n->req() ||
4243                n->is_Region() || n->is_Phi() || n->is_ArrayCopy() ||
4244                (n->is_Unlock() && i == (n->req() - 1)) ||
4245                (n->is_MemBar() && i == 5), // the precedence edge to a membar can be removed during macro node expansion
4246               "only region, phi, arraycopy, unlock or membar nodes have null data edges");
4247       } else {
4248         assert(in->is_top(), "sanity");
4249         // Nothing to check.
4250       }
4251     }
4252   }
4253 }
4254 
4255 //------------------------------verify_graph_edges---------------------------
4256 // Walk the Graph and verify that there is a one-to-one correspondence
4257 // between Use-Def edges and Def-Use edges in the graph.
4258 void Compile::verify_graph_edges(bool no_dead_code) {
4259   if (VerifyGraphEdges) {
4260     Unique_Node_List visited;
4261 
4262     // Call graph walk to check edges
4263     verify_bidirectional_edges(visited);
4264     if (no_dead_code) {
4265       // Now make sure that no visited node is used by an unvisited node.
4266       bool dead_nodes = false;
4267       Unique_Node_List checked;
4268       while (visited.size() > 0) {
4269         Node* n = visited.pop();
4270         checked.push(n);
4271         for (uint i = 0; i < n->outcnt(); i++) {
4272           Node* use = n->raw_out(i);
4273           if (checked.member(use))  continue;  // already checked
4274           if (visited.member(use))  continue;  // already in the graph
4275           if (use->is_Con())        continue;  // a dead ConNode is OK
4276           // At this point, we have found a dead node which is DU-reachable.
4277           if (!dead_nodes) {
4278             tty->print_cr("*** Dead nodes reachable via DU edges:");
4279             dead_nodes = true;
4280           }
4281           use->dump(2);
4282           tty->print_cr("---");
4283           checked.push(use);  // No repeats; pretend it is now checked.
4284         }
4285       }
4286       assert(!dead_nodes, "using nodes must be reachable from root");
4287     }
4288   }
4289 }
4290 #endif
4291 
4292 // The Compile object keeps track of failure reasons separately from the ciEnv.
4293 // This is required because there is not quite a 1-1 relation between the
4294 // ciEnv and its compilation task and the Compile object.  Note that one
4295 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4296 // to backtrack and retry without subsuming loads.  Other than this backtracking
4297 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4298 // by the logic in C2Compiler.
4299 void Compile::record_failure(const char* reason) {
4300   if (log() != NULL) {
4301     log()->elem("failure reason='%s' phase='compile'", reason);
4302   }
4303   if (_failure_reason == NULL) {
4304     // Record the first failure reason.
4305     _failure_reason = reason;
4306   }
4307 
4308   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4309     C->print_method(PHASE_FAILURE, 1);
4310   }
4311   _root = NULL;  // flush the graph, too
4312 }
4313 
4314 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4315   : TraceTime(name, accumulator, CITime, CITimeVerbose),
4316     _phase_name(name), _dolog(CITimeVerbose)
4317 {
4318   if (_dolog) {
4319     C = Compile::current();
4320     _log = C->log();
4321   } else {
4322     C = NULL;
4323     _log = NULL;
4324   }
4325   if (_log != NULL) {
4326     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4327     _log->stamp();
4328     _log->end_head();
4329   }
4330 }
4331 
4332 Compile::TracePhase::~TracePhase() {
4333 
4334   C = Compile::current();
4335   if (_dolog) {
4336     _log = C->log();
4337   } else {
4338     _log = NULL;
4339   }
4340 
4341 #ifdef ASSERT
4342   if (PrintIdealNodeCount) {
4343     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4344                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4345   }
4346 
4347   if (VerifyIdealNodeCount) {
4348     Compile::current()->print_missing_nodes();
4349   }
4350 #endif
4351 
4352   if (_log != NULL) {
4353     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4354   }
4355 }
4356 
4357 //----------------------------static_subtype_check-----------------------------
4358 // Shortcut important common cases when superklass is exact:
4359 // (0) superklass is java.lang.Object (can occur in reflective code)
4360 // (1) subklass is already limited to a subtype of superklass => always ok
4361 // (2) subklass does not overlap with superklass => always fail
4362 // (3) superklass has NO subtypes and we can check with a simple compare.
4363 Compile::SubTypeCheckResult Compile::static_subtype_check(const TypeKlassPtr* superk, const TypeKlassPtr* subk, bool skip) {
4364   if (skip) {
4365     return SSC_full_test;       // Let caller generate the general case.
4366   }
4367 
4368   if (subk->is_java_subtype_of(superk)) {
4369     return SSC_always_true; // (0) and (1)  this test cannot fail
4370   }
4371 
4372   if (!subk->maybe_java_subtype_of(superk)) {
4373     return SSC_always_false; // (2) true path dead; no dynamic test needed
4374   }
4375 
4376   const Type* superelem = superk;
4377   if (superk->isa_aryklassptr()) {
4378     int ignored;
4379     superelem = superk->is_aryklassptr()->base_element_type(ignored);
4380   }
4381 
4382   if (superelem->isa_instklassptr()) {
4383     ciInstanceKlass* ik = superelem->is_instklassptr()->instance_klass();
4384     if (!ik->has_subklass()) {
4385       if (!ik->is_final()) {
4386         // Add a dependency if there is a chance of a later subclass.
4387         dependencies()->assert_leaf_type(ik);
4388       }
4389       if (!superk->maybe_java_subtype_of(subk)) {
4390         return SSC_always_false;
4391       }
4392       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
4393     }
4394   } else {
4395     // A primitive array type has no subtypes.
4396     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
4397   }
4398 
4399   return SSC_full_test;
4400 }
4401 
4402 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4403 #ifdef _LP64
4404   // The scaled index operand to AddP must be a clean 64-bit value.
4405   // Java allows a 32-bit int to be incremented to a negative
4406   // value, which appears in a 64-bit register as a large
4407   // positive number.  Using that large positive number as an
4408   // operand in pointer arithmetic has bad consequences.
4409   // On the other hand, 32-bit overflow is rare, and the possibility
4410   // can often be excluded, if we annotate the ConvI2L node with
4411   // a type assertion that its value is known to be a small positive
4412   // number.  (The prior range check has ensured this.)
4413   // This assertion is used by ConvI2LNode::Ideal.
4414   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4415   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4416   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4417   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4418 #endif
4419   return idx;
4420 }
4421 
4422 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4423 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4424   if (ctrl != NULL) {
4425     // Express control dependency by a CastII node with a narrow type.
4426     value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4427     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4428     // node from floating above the range check during loop optimizations. Otherwise, the
4429     // ConvI2L node may be eliminated independently of the range check, causing the data path
4430     // to become TOP while the control path is still there (although it's unreachable).
4431     value->set_req(0, ctrl);
4432     value = phase->transform(value);
4433   }
4434   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4435   return phase->transform(new ConvI2LNode(value, ltype));
4436 }
4437 
4438 // The message about the current inlining is accumulated in
4439 // _print_inlining_stream and transferred into the _print_inlining_list
4440 // once we know whether inlining succeeds or not. For regular
4441 // inlining, messages are appended to the buffer pointed by
4442 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4443 // a new buffer is added after _print_inlining_idx in the list. This
4444 // way we can update the inlining message for late inlining call site
4445 // when the inlining is attempted again.
4446 void Compile::print_inlining_init() {
4447   if (print_inlining() || print_intrinsics()) {
4448     // print_inlining_init is actually called several times.
4449     print_inlining_reset();
4450     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4451   }
4452 }
4453 
4454 void Compile::print_inlining_reinit() {
4455   if (print_inlining() || print_intrinsics()) {
4456     print_inlining_reset();
4457   }
4458 }
4459 
4460 void Compile::print_inlining_reset() {
4461   _print_inlining_stream->reset();
4462 }
4463 
4464 void Compile::print_inlining_commit() {
4465   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4466   // Transfer the message from _print_inlining_stream to the current
4467   // _print_inlining_list buffer and clear _print_inlining_stream.
4468   _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4469   print_inlining_reset();
4470 }
4471 
4472 void Compile::print_inlining_push() {
4473   // Add new buffer to the _print_inlining_list at current position
4474   _print_inlining_idx++;
4475   _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4476 }
4477 
4478 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4479   return _print_inlining_list->at(_print_inlining_idx);
4480 }
4481 
4482 void Compile::print_inlining_update(CallGenerator* cg) {
4483   if (print_inlining() || print_intrinsics()) {
4484     if (cg->is_late_inline()) {
4485       if (print_inlining_current()->cg() != cg &&
4486           (print_inlining_current()->cg() != NULL ||
4487            print_inlining_current()->ss()->size() != 0)) {
4488         print_inlining_push();
4489       }
4490       print_inlining_commit();
4491       print_inlining_current()->set_cg(cg);
4492     } else {
4493       if (print_inlining_current()->cg() != NULL) {
4494         print_inlining_push();
4495       }
4496       print_inlining_commit();
4497     }
4498   }
4499 }
4500 
4501 void Compile::print_inlining_move_to(CallGenerator* cg) {
4502   // We resume inlining at a late inlining call site. Locate the
4503   // corresponding inlining buffer so that we can update it.
4504   if (print_inlining() || print_intrinsics()) {
4505     for (int i = 0; i < _print_inlining_list->length(); i++) {
4506       if (_print_inlining_list->at(i)->cg() == cg) {
4507         _print_inlining_idx = i;
4508         return;
4509       }
4510     }
4511     ShouldNotReachHere();
4512   }
4513 }
4514 
4515 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4516   if (print_inlining() || print_intrinsics()) {
4517     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4518     assert(print_inlining_current()->cg() == cg, "wrong entry");
4519     // replace message with new message
4520     _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4521     print_inlining_commit();
4522     print_inlining_current()->set_cg(cg);
4523   }
4524 }
4525 
4526 void Compile::print_inlining_assert_ready() {
4527   assert(!_print_inlining || _print_inlining_stream->size() == 0, "losing data");
4528 }
4529 
4530 void Compile::process_print_inlining() {
4531   assert(_late_inlines.length() == 0, "not drained yet");
4532   if (print_inlining() || print_intrinsics()) {
4533     ResourceMark rm;
4534     stringStream ss;
4535     assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
4536     for (int i = 0; i < _print_inlining_list->length(); i++) {
4537       PrintInliningBuffer* pib = _print_inlining_list->at(i);
4538       ss.print("%s", pib->ss()->freeze());
4539       delete pib;
4540       DEBUG_ONLY(_print_inlining_list->at_put(i, NULL));
4541     }
4542     // Reset _print_inlining_list, it only contains destructed objects.
4543     // It is on the arena, so it will be freed when the arena is reset.
4544     _print_inlining_list = NULL;
4545     // _print_inlining_stream won't be used anymore, either.
4546     print_inlining_reset();
4547     size_t end = ss.size();
4548     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4549     strncpy(_print_inlining_output, ss.freeze(), end+1);
4550     _print_inlining_output[end] = 0;
4551   }
4552 }
4553 
4554 void Compile::dump_print_inlining() {
4555   if (_print_inlining_output != NULL) {
4556     tty->print_raw(_print_inlining_output);
4557   }
4558 }
4559 
4560 void Compile::log_late_inline(CallGenerator* cg) {
4561   if (log() != NULL) {
4562     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4563                 cg->unique_id());
4564     JVMState* p = cg->call_node()->jvms();
4565     while (p != NULL) {
4566       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4567       p = p->caller();
4568     }
4569     log()->tail("late_inline");
4570   }
4571 }
4572 
4573 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4574   log_late_inline(cg);
4575   if (log() != NULL) {
4576     log()->inline_fail(msg);
4577   }
4578 }
4579 
4580 void Compile::log_inline_id(CallGenerator* cg) {
4581   if (log() != NULL) {
4582     // The LogCompilation tool needs a unique way to identify late
4583     // inline call sites. This id must be unique for this call site in
4584     // this compilation. Try to have it unique across compilations as
4585     // well because it can be convenient when grepping through the log
4586     // file.
4587     // Distinguish OSR compilations from others in case CICountOSR is
4588     // on.
4589     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4590     cg->set_unique_id(id);
4591     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4592   }
4593 }
4594 
4595 void Compile::log_inline_failure(const char* msg) {
4596   if (C->log() != NULL) {
4597     C->log()->inline_fail(msg);
4598   }
4599 }
4600 
4601 
4602 // Dump inlining replay data to the stream.
4603 // Don't change thread state and acquire any locks.
4604 void Compile::dump_inline_data(outputStream* out) {
4605   InlineTree* inl_tree = ilt();
4606   if (inl_tree != NULL) {
4607     out->print(" inline %d", inl_tree->count());
4608     inl_tree->dump_replay_data(out);
4609   }
4610 }
4611 
4612 void Compile::dump_inline_data_reduced(outputStream* out) {
4613   assert(ReplayReduce, "");
4614 
4615   InlineTree* inl_tree = ilt();
4616   if (inl_tree == NULL) {
4617     return;
4618   }
4619   // Enable iterative replay file reduction
4620   // Output "compile" lines for depth 1 subtrees,
4621   // simulating that those trees were compiled
4622   // instead of inlined.
4623   for (int i = 0; i < inl_tree->subtrees().length(); ++i) {
4624     InlineTree* sub = inl_tree->subtrees().at(i);
4625     if (sub->inline_level() != 1) {
4626       continue;
4627     }
4628 
4629     ciMethod* method = sub->method();
4630     int entry_bci = -1;
4631     int comp_level = env()->task()->comp_level();
4632     out->print("compile ");
4633     method->dump_name_as_ascii(out);
4634     out->print(" %d %d", entry_bci, comp_level);
4635     out->print(" inline %d", sub->count());
4636     sub->dump_replay_data(out, -1);
4637     out->cr();
4638   }
4639 }
4640 
4641 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4642   if (n1->Opcode() < n2->Opcode())      return -1;
4643   else if (n1->Opcode() > n2->Opcode()) return 1;
4644 
4645   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4646   for (uint i = 1; i < n1->req(); i++) {
4647     if (n1->in(i) < n2->in(i))      return -1;
4648     else if (n1->in(i) > n2->in(i)) return 1;
4649   }
4650 
4651   return 0;
4652 }
4653 
4654 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4655   Node* n1 = *n1p;
4656   Node* n2 = *n2p;
4657 
4658   return cmp_expensive_nodes(n1, n2);
4659 }
4660 
4661 void Compile::sort_expensive_nodes() {
4662   if (!expensive_nodes_sorted()) {
4663     _expensive_nodes.sort(cmp_expensive_nodes);
4664   }
4665 }
4666 
4667 bool Compile::expensive_nodes_sorted() const {
4668   for (int i = 1; i < _expensive_nodes.length(); i++) {
4669     if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4670       return false;
4671     }
4672   }
4673   return true;
4674 }
4675 
4676 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4677   if (_expensive_nodes.length() == 0) {
4678     return false;
4679   }
4680 
4681   assert(OptimizeExpensiveOps, "optimization off?");
4682 
4683   // Take this opportunity to remove dead nodes from the list
4684   int j = 0;
4685   for (int i = 0; i < _expensive_nodes.length(); i++) {
4686     Node* n = _expensive_nodes.at(i);
4687     if (!n->is_unreachable(igvn)) {
4688       assert(n->is_expensive(), "should be expensive");
4689       _expensive_nodes.at_put(j, n);
4690       j++;
4691     }
4692   }
4693   _expensive_nodes.trunc_to(j);
4694 
4695   // Then sort the list so that similar nodes are next to each other
4696   // and check for at least two nodes of identical kind with same data
4697   // inputs.
4698   sort_expensive_nodes();
4699 
4700   for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4701     if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4702       return true;
4703     }
4704   }
4705 
4706   return false;
4707 }
4708 
4709 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4710   if (_expensive_nodes.length() == 0) {
4711     return;
4712   }
4713 
4714   assert(OptimizeExpensiveOps, "optimization off?");
4715 
4716   // Sort to bring similar nodes next to each other and clear the
4717   // control input of nodes for which there's only a single copy.
4718   sort_expensive_nodes();
4719 
4720   int j = 0;
4721   int identical = 0;
4722   int i = 0;
4723   bool modified = false;
4724   for (; i < _expensive_nodes.length()-1; i++) {
4725     assert(j <= i, "can't write beyond current index");
4726     if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4727       identical++;
4728       _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4729       continue;
4730     }
4731     if (identical > 0) {
4732       _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4733       identical = 0;
4734     } else {
4735       Node* n = _expensive_nodes.at(i);
4736       igvn.replace_input_of(n, 0, NULL);
4737       igvn.hash_insert(n);
4738       modified = true;
4739     }
4740   }
4741   if (identical > 0) {
4742     _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4743   } else if (_expensive_nodes.length() >= 1) {
4744     Node* n = _expensive_nodes.at(i);
4745     igvn.replace_input_of(n, 0, NULL);
4746     igvn.hash_insert(n);
4747     modified = true;
4748   }
4749   _expensive_nodes.trunc_to(j);
4750   if (modified) {
4751     igvn.optimize();
4752   }
4753 }
4754 
4755 void Compile::add_expensive_node(Node * n) {
4756   assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4757   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4758   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4759   if (OptimizeExpensiveOps) {
4760     _expensive_nodes.append(n);
4761   } else {
4762     // Clear control input and let IGVN optimize expensive nodes if
4763     // OptimizeExpensiveOps is off.
4764     n->set_req(0, NULL);
4765   }
4766 }
4767 
4768 /**
4769  * Track coarsened Lock and Unlock nodes.
4770  */
4771 
4772 class Lock_List : public Node_List {
4773   uint _origin_cnt;
4774 public:
4775   Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4776   uint origin_cnt() const { return _origin_cnt; }
4777 };
4778 
4779 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4780   int length = locks.length();
4781   if (length > 0) {
4782     // Have to keep this list until locks elimination during Macro nodes elimination.
4783     Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4784     for (int i = 0; i < length; i++) {
4785       AbstractLockNode* lock = locks.at(i);
4786       assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4787       locks_list->push(lock);
4788     }
4789     _coarsened_locks.append(locks_list);
4790   }
4791 }
4792 
4793 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4794   int count = coarsened_count();
4795   for (int i = 0; i < count; i++) {
4796     Node_List* locks_list = _coarsened_locks.at(i);
4797     for (uint j = 0; j < locks_list->size(); j++) {
4798       Node* lock = locks_list->at(j);
4799       assert(lock->is_AbstractLock(), "sanity");
4800       if (!useful.member(lock)) {
4801         locks_list->yank(lock);
4802       }
4803     }
4804   }
4805 }
4806 
4807 void Compile::remove_coarsened_lock(Node* n) {
4808   if (n->is_AbstractLock()) {
4809     int count = coarsened_count();
4810     for (int i = 0; i < count; i++) {
4811       Node_List* locks_list = _coarsened_locks.at(i);
4812       locks_list->yank(n);
4813     }
4814   }
4815 }
4816 
4817 bool Compile::coarsened_locks_consistent() {
4818   int count = coarsened_count();
4819   for (int i = 0; i < count; i++) {
4820     bool unbalanced = false;
4821     bool modified = false; // track locks kind modifications
4822     Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4823     uint size = locks_list->size();
4824     if (size == 0) {
4825       unbalanced = false; // All locks were eliminated - good
4826     } else if (size != locks_list->origin_cnt()) {
4827       unbalanced = true; // Some locks were removed from list
4828     } else {
4829       for (uint j = 0; j < size; j++) {
4830         Node* lock = locks_list->at(j);
4831         // All nodes in group should have the same state (modified or not)
4832         if (!lock->as_AbstractLock()->is_coarsened()) {
4833           if (j == 0) {
4834             // first on list was modified, the rest should be too for consistency
4835             modified = true;
4836           } else if (!modified) {
4837             // this lock was modified but previous locks on the list were not
4838             unbalanced = true;
4839             break;
4840           }
4841         } else if (modified) {
4842           // previous locks on list were modified but not this lock
4843           unbalanced = true;
4844           break;
4845         }
4846       }
4847     }
4848     if (unbalanced) {
4849       // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4850 #ifdef ASSERT
4851       if (PrintEliminateLocks) {
4852         tty->print_cr("=== unbalanced coarsened locks ===");
4853         for (uint l = 0; l < size; l++) {
4854           locks_list->at(l)->dump();
4855         }
4856       }
4857 #endif
4858       record_failure(C2Compiler::retry_no_locks_coarsening());
4859       return false;
4860     }
4861   }
4862   return true;
4863 }
4864 
4865 /**
4866  * Remove the speculative part of types and clean up the graph
4867  */
4868 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4869   if (UseTypeSpeculation) {
4870     Unique_Node_List worklist;
4871     worklist.push(root());
4872     int modified = 0;
4873     // Go over all type nodes that carry a speculative type, drop the
4874     // speculative part of the type and enqueue the node for an igvn
4875     // which may optimize it out.
4876     for (uint next = 0; next < worklist.size(); ++next) {
4877       Node *n  = worklist.at(next);
4878       if (n->is_Type()) {
4879         TypeNode* tn = n->as_Type();
4880         const Type* t = tn->type();
4881         const Type* t_no_spec = t->remove_speculative();
4882         if (t_no_spec != t) {
4883           bool in_hash = igvn.hash_delete(n);
4884           assert(in_hash, "node should be in igvn hash table");
4885           tn->set_type(t_no_spec);
4886           igvn.hash_insert(n);
4887           igvn._worklist.push(n); // give it a chance to go away
4888           modified++;
4889         }
4890       }
4891       // Iterate over outs - endless loops is unreachable from below
4892       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4893         Node *m = n->fast_out(i);
4894         if (not_a_node(m)) {
4895           continue;
4896         }
4897         worklist.push(m);
4898       }
4899     }
4900     // Drop the speculative part of all types in the igvn's type table
4901     igvn.remove_speculative_types();
4902     if (modified > 0) {
4903       igvn.optimize();
4904     }
4905 #ifdef ASSERT
4906     // Verify that after the IGVN is over no speculative type has resurfaced
4907     worklist.clear();
4908     worklist.push(root());
4909     for (uint next = 0; next < worklist.size(); ++next) {
4910       Node *n  = worklist.at(next);
4911       const Type* t = igvn.type_or_null(n);
4912       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4913       if (n->is_Type()) {
4914         t = n->as_Type()->type();
4915         assert(t == t->remove_speculative(), "no more speculative types");
4916       }
4917       // Iterate over outs - endless loops is unreachable from below
4918       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4919         Node *m = n->fast_out(i);
4920         if (not_a_node(m)) {
4921           continue;
4922         }
4923         worklist.push(m);
4924       }
4925     }
4926     igvn.check_no_speculative_types();
4927 #endif
4928   }
4929 }
4930 
4931 // Auxiliary methods to support randomized stressing/fuzzing.
4932 
4933 int Compile::random() {
4934   _stress_seed = os::next_random(_stress_seed);
4935   return static_cast<int>(_stress_seed);
4936 }
4937 
4938 // This method can be called the arbitrary number of times, with current count
4939 // as the argument. The logic allows selecting a single candidate from the
4940 // running list of candidates as follows:
4941 //    int count = 0;
4942 //    Cand* selected = null;
4943 //    while(cand = cand->next()) {
4944 //      if (randomized_select(++count)) {
4945 //        selected = cand;
4946 //      }
4947 //    }
4948 //
4949 // Including count equalizes the chances any candidate is "selected".
4950 // This is useful when we don't have the complete list of candidates to choose
4951 // from uniformly. In this case, we need to adjust the randomicity of the
4952 // selection, or else we will end up biasing the selection towards the latter
4953 // candidates.
4954 //
4955 // Quick back-envelope calculation shows that for the list of n candidates
4956 // the equal probability for the candidate to persist as "best" can be
4957 // achieved by replacing it with "next" k-th candidate with the probability
4958 // of 1/k. It can be easily shown that by the end of the run, the
4959 // probability for any candidate is converged to 1/n, thus giving the
4960 // uniform distribution among all the candidates.
4961 //
4962 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4963 #define RANDOMIZED_DOMAIN_POW 29
4964 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4965 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4966 bool Compile::randomized_select(int count) {
4967   assert(count > 0, "only positive");
4968   return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4969 }
4970 
4971 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4972 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4973 
4974 void NodeCloneInfo::dump() const {
4975   tty->print(" {%d:%d} ", idx(), gen());
4976 }
4977 
4978 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4979   uint64_t val = value(old->_idx);
4980   NodeCloneInfo cio(val);
4981   assert(val != 0, "old node should be in the map");
4982   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4983   insert(nnn->_idx, cin.get());
4984 #ifndef PRODUCT
4985   if (is_debug()) {
4986     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4987   }
4988 #endif
4989 }
4990 
4991 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4992   NodeCloneInfo cio(value(old->_idx));
4993   if (cio.get() == 0) {
4994     cio.set(old->_idx, 0);
4995     insert(old->_idx, cio.get());
4996 #ifndef PRODUCT
4997     if (is_debug()) {
4998       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4999     }
5000 #endif
5001   }
5002   clone(old, nnn, gen);
5003 }
5004 
5005 int CloneMap::max_gen() const {
5006   int g = 0;
5007   DictI di(_dict);
5008   for(; di.test(); ++di) {
5009     int t = gen(di._key);
5010     if (g < t) {
5011       g = t;
5012 #ifndef PRODUCT
5013       if (is_debug()) {
5014         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
5015       }
5016 #endif
5017     }
5018   }
5019   return g;
5020 }
5021 
5022 void CloneMap::dump(node_idx_t key) const {
5023   uint64_t val = value(key);
5024   if (val != 0) {
5025     NodeCloneInfo ni(val);
5026     ni.dump();
5027   }
5028 }
5029 
5030 // Move Allocate nodes to the start of the list
5031 void Compile::sort_macro_nodes() {
5032   int count = macro_count();
5033   int allocates = 0;
5034   for (int i = 0; i < count; i++) {
5035     Node* n = macro_node(i);
5036     if (n->is_Allocate()) {
5037       if (i != allocates) {
5038         Node* tmp = macro_node(allocates);
5039         _macro_nodes.at_put(allocates, n);
5040         _macro_nodes.at_put(i, tmp);
5041       }
5042       allocates++;
5043     }
5044   }
5045 }
5046 
5047 void Compile::print_method(CompilerPhaseType cpt, int level, Node* n) {
5048   EventCompilerPhase event;
5049   if (event.should_commit()) {
5050     CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
5051   }
5052 #ifndef PRODUCT
5053   ResourceMark rm;
5054   stringStream ss;
5055   ss.print_raw(CompilerPhaseTypeHelper::to_description(cpt));
5056   if (n != nullptr) {
5057     ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
5058   }
5059 
5060   const char* name = ss.as_string();
5061   if (should_print_igv(level)) {
5062     _igv_printer->print_method(name, level);
5063   }
5064   if (should_print_phase(cpt)) {
5065     print_ideal_ir(CompilerPhaseTypeHelper::to_name(cpt));
5066   }
5067 #endif
5068   C->_latest_stage_start_counter.stamp();
5069 }
5070 
5071 // Only used from CompileWrapper
5072 void Compile::begin_method() {
5073 #ifndef PRODUCT
5074   if (_method != NULL && should_print_igv(1)) {
5075     _igv_printer->begin_method();
5076   }
5077 #endif
5078   C->_latest_stage_start_counter.stamp();
5079 }
5080 
5081 // Only used from CompileWrapper
5082 void Compile::end_method() {
5083   EventCompilerPhase event;
5084   if (event.should_commit()) {
5085     CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, 1);
5086   }
5087 
5088 #ifndef PRODUCT
5089   if (_method != NULL && should_print_igv(1)) {
5090     _igv_printer->end_method();
5091   }
5092 #endif
5093 }
5094 
5095 bool Compile::should_print_phase(CompilerPhaseType cpt) {
5096 #ifndef PRODUCT
5097   if ((_directive->ideal_phase_mask() & CompilerPhaseTypeHelper::to_bitmask(cpt)) != 0) {
5098     return true;
5099   }
5100 #endif
5101   return false;
5102 }
5103 
5104 bool Compile::should_print_igv(int level) {
5105 #ifndef PRODUCT
5106   if (PrintIdealGraphLevel < 0) { // disabled by the user
5107     return false;
5108   }
5109 
5110   bool need = directive()->IGVPrintLevelOption >= level;
5111   if (need && !_igv_printer) {
5112     _igv_printer = IdealGraphPrinter::printer();
5113     _igv_printer->set_compile(this);
5114   }
5115   return need;
5116 #else
5117   return false;
5118 #endif
5119 }
5120 
5121 #ifndef PRODUCT
5122 IdealGraphPrinter* Compile::_debug_file_printer = NULL;
5123 IdealGraphPrinter* Compile::_debug_network_printer = NULL;
5124 
5125 // Called from debugger. Prints method to the default file with the default phase name.
5126 // This works regardless of any Ideal Graph Visualizer flags set or not.
5127 void igv_print() {
5128   Compile::current()->igv_print_method_to_file();
5129 }
5130 
5131 // Same as igv_print() above but with a specified phase name.
5132 void igv_print(const char* phase_name) {
5133   Compile::current()->igv_print_method_to_file(phase_name);
5134 }
5135 
5136 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
5137 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
5138 // This works regardless of any Ideal Graph Visualizer flags set or not.
5139 void igv_print(bool network) {
5140   if (network) {
5141     Compile::current()->igv_print_method_to_network();
5142   } else {
5143     Compile::current()->igv_print_method_to_file();
5144   }
5145 }
5146 
5147 // Same as igv_print(bool network) above but with a specified phase name.
5148 void igv_print(bool network, const char* phase_name) {
5149   if (network) {
5150     Compile::current()->igv_print_method_to_network(phase_name);
5151   } else {
5152     Compile::current()->igv_print_method_to_file(phase_name);
5153   }
5154 }
5155 
5156 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
5157 void igv_print_default() {
5158   Compile::current()->print_method(PHASE_DEBUG, 0);
5159 }
5160 
5161 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
5162 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
5163 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
5164 void igv_append() {
5165   Compile::current()->igv_print_method_to_file("Debug", true);
5166 }
5167 
5168 // Same as igv_append() above but with a specified phase name.
5169 void igv_append(const char* phase_name) {
5170   Compile::current()->igv_print_method_to_file(phase_name, true);
5171 }
5172 
5173 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
5174   const char* file_name = "custom_debug.xml";
5175   if (_debug_file_printer == NULL) {
5176     _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
5177   } else {
5178     _debug_file_printer->update_compiled_method(C->method());
5179   }
5180   tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
5181   _debug_file_printer->print(phase_name, (Node*)C->root());
5182 }
5183 
5184 void Compile::igv_print_method_to_network(const char* phase_name) {
5185   if (_debug_network_printer == NULL) {
5186     _debug_network_printer = new IdealGraphPrinter(C);
5187   } else {
5188     _debug_network_printer->update_compiled_method(C->method());
5189   }
5190   tty->print_cr("Method printed over network stream to IGV");
5191   _debug_network_printer->print(phase_name, (Node*)C->root());
5192 }
5193 #endif
5194 
5195 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
5196   if (type != NULL && phase->type(value)->higher_equal(type)) {
5197     return value;
5198   }
5199   Node* result = NULL;
5200   if (bt == T_BYTE) {
5201     result = phase->transform(new LShiftINode(value, phase->intcon(24)));
5202     result = new RShiftINode(result, phase->intcon(24));
5203   } else if (bt == T_BOOLEAN) {
5204     result = new AndINode(value, phase->intcon(0xFF));
5205   } else if (bt == T_CHAR) {
5206     result = new AndINode(value,phase->intcon(0xFFFF));
5207   } else {
5208     assert(bt == T_SHORT, "unexpected narrow type");
5209     result = phase->transform(new LShiftINode(value, phase->intcon(16)));
5210     result = new RShiftINode(result, phase->intcon(16));
5211   }
5212   if (transform_res) {
5213     result = phase->transform(result);
5214   }
5215   return result;
5216 }
5217