1 /*
   2  * Copyright (c) 1997, 2021, 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 "jvm_io.h"
  27 #include "asm/macroAssembler.hpp"
  28 #include "asm/macroAssembler.inline.hpp"
  29 #include "ci/ciReplay.hpp"
  30 #include "classfile/javaClasses.hpp"
  31 #include "code/exceptionHandlerTable.hpp"
  32 #include "code/nmethod.hpp"
  33 #include "compiler/compileBroker.hpp"
  34 #include "compiler/compileLog.hpp"
  35 #include "compiler/disassembler.hpp"
  36 #include "compiler/oopMap.hpp"
  37 #include "gc/shared/barrierSet.hpp"
  38 #include "gc/shared/c2/barrierSetC2.hpp"
  39 #include "jfr/jfrEvents.hpp"
  40 #include "memory/resourceArea.hpp"
  41 #include "opto/addnode.hpp"
  42 #include "opto/block.hpp"
  43 #include "opto/c2compiler.hpp"
  44 #include "opto/callGenerator.hpp"
  45 #include "opto/callnode.hpp"
  46 #include "opto/castnode.hpp"
  47 #include "opto/cfgnode.hpp"
  48 #include "opto/chaitin.hpp"
  49 #include "opto/compile.hpp"
  50 #include "opto/connode.hpp"
  51 #include "opto/convertnode.hpp"
  52 #include "opto/divnode.hpp"
  53 #include "opto/escape.hpp"
  54 #include "opto/idealGraphPrinter.hpp"
  55 #include "opto/loopnode.hpp"
  56 #include "opto/machnode.hpp"
  57 #include "opto/macro.hpp"
  58 #include "opto/matcher.hpp"
  59 #include "opto/mathexactnode.hpp"
  60 #include "opto/memnode.hpp"
  61 #include "opto/mulnode.hpp"
  62 #include "opto/narrowptrnode.hpp"
  63 #include "opto/node.hpp"
  64 #include "opto/opcodes.hpp"
  65 #include "opto/output.hpp"
  66 #include "opto/parse.hpp"
  67 #include "opto/phaseX.hpp"
  68 #include "opto/rootnode.hpp"
  69 #include "opto/runtime.hpp"
  70 #include "opto/stringopts.hpp"
  71 #include "opto/type.hpp"
  72 #include "opto/vector.hpp"
  73 #include "opto/vectornode.hpp"
  74 #include "runtime/globals_extension.hpp"
  75 #include "runtime/sharedRuntime.hpp"
  76 #include "runtime/signature.hpp"
  77 #include "runtime/stubRoutines.hpp"
  78 #include "runtime/timer.hpp"
  79 #include "utilities/align.hpp"
  80 #include "utilities/copy.hpp"
  81 #include "utilities/macros.hpp"
  82 #include "utilities/resourceHash.hpp"
  83 
  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     PhaseCCP::print_statistics();
 253     PhaseRegAlloc::print_statistics();
 254     PhaseOutput::print_statistics();
 255     PhasePeephole::print_statistics();
 256     PhaseIdealLoop::print_statistics();
 257     if (xtty != NULL)  xtty->tail("statistics");
 258   }
 259   if (_intrinsic_hist_flags[as_int(vmIntrinsics::_none)] != 0) {
 260     // put this under its own <statistics> element.
 261     print_intrinsic_statistics();
 262   }
 263 }
 264 #endif //PRODUCT
 265 
 266 void Compile::gvn_replace_by(Node* n, Node* nn) {
 267   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 268     Node* use = n->last_out(i);
 269     bool is_in_table = initial_gvn()->hash_delete(use);
 270     uint uses_found = 0;
 271     for (uint j = 0; j < use->len(); j++) {
 272       if (use->in(j) == n) {
 273         if (j < use->req())
 274           use->set_req(j, nn);
 275         else
 276           use->set_prec(j, nn);
 277         uses_found++;
 278       }
 279     }
 280     if (is_in_table) {
 281       // reinsert into table
 282       initial_gvn()->hash_find_insert(use);
 283     }
 284     record_for_igvn(use);
 285     i -= uses_found;    // we deleted 1 or more copies of this edge
 286   }
 287 }
 288 
 289 
 290 // Identify all nodes that are reachable from below, useful.
 291 // Use breadth-first pass that records state in a Unique_Node_List,
 292 // recursive traversal is slower.
 293 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 294   int estimated_worklist_size = live_nodes();
 295   useful.map( estimated_worklist_size, NULL );  // preallocate space
 296 
 297   // Initialize worklist
 298   if (root() != NULL)     { useful.push(root()); }
 299   // If 'top' is cached, declare it useful to preserve cached node
 300   if( cached_top_node() ) { useful.push(cached_top_node()); }
 301 
 302   // Push all useful nodes onto the list, breadthfirst
 303   for( uint next = 0; next < useful.size(); ++next ) {
 304     assert( next < unique(), "Unique useful nodes < total nodes");
 305     Node *n  = useful.at(next);
 306     uint max = n->len();
 307     for( uint i = 0; i < max; ++i ) {
 308       Node *m = n->in(i);
 309       if (not_a_node(m))  continue;
 310       useful.push(m);
 311     }
 312   }
 313 }
 314 
 315 // Update dead_node_list with any missing dead nodes using useful
 316 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 317 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 318   uint max_idx = unique();
 319   VectorSet& useful_node_set = useful.member_set();
 320 
 321   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 322     // If node with index node_idx is not in useful set,
 323     // mark it as dead in dead node list.
 324     if (!useful_node_set.test(node_idx)) {
 325       record_dead_node(node_idx);
 326     }
 327   }
 328 }
 329 
 330 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 331   int shift = 0;
 332   for (int i = 0; i < inlines->length(); i++) {
 333     CallGenerator* cg = inlines->at(i);
 334     if (useful.member(cg->call_node())) {
 335       if (shift > 0) {
 336         inlines->at_put(i - shift, cg);
 337       }
 338     } else {
 339       shift++; // skip over the dead element
 340     }
 341   }
 342   if (shift > 0) {
 343     inlines->trunc_to(inlines->length() - shift); // remove last elements from compacted array
 344   }
 345 }
 346 
 347 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Node* dead) {
 348   assert(dead != NULL && dead->is_Call(), "sanity");
 349   int found = 0;
 350   for (int i = 0; i < inlines->length(); i++) {
 351     if (inlines->at(i)->call_node() == dead) {
 352       inlines->remove_at(i);
 353       found++;
 354       NOT_DEBUG( break; ) // elements are unique, so exit early
 355     }
 356   }
 357   assert(found <= 1, "not unique");
 358 }
 359 
 360 void Compile::remove_useless_nodes(GrowableArray<Node*>& node_list, Unique_Node_List& useful) {
 361   for (int i = node_list.length() - 1; i >= 0; i--) {
 362     Node* n = node_list.at(i);
 363     if (!useful.member(n)) {
 364       node_list.delete_at(i); // replaces i-th with last element which is known to be useful (already processed)
 365     }
 366   }
 367 }
 368 
 369 void Compile::remove_useless_node(Node* dead) {
 370   remove_modified_node(dead);
 371 
 372   // Constant node that has no out-edges and has only one in-edge from
 373   // root is usually dead. However, sometimes reshaping walk makes
 374   // it reachable by adding use edges. So, we will NOT count Con nodes
 375   // as dead to be conservative about the dead node count at any
 376   // given time.
 377   if (!dead->is_Con()) {
 378     record_dead_node(dead->_idx);
 379   }
 380   if (dead->is_macro()) {
 381     remove_macro_node(dead);
 382   }
 383   if (dead->is_expensive()) {
 384     remove_expensive_node(dead);
 385   }
 386   if (dead->Opcode() == Op_Opaque4) {
 387     remove_skeleton_predicate_opaq(dead);
 388   }
 389   if (dead->for_post_loop_opts_igvn()) {
 390     remove_from_post_loop_opts_igvn(dead);
 391   }
 392   if (dead->is_Call()) {
 393     remove_useless_late_inlines(                &_late_inlines, dead);
 394     remove_useless_late_inlines(         &_string_late_inlines, dead);
 395     remove_useless_late_inlines(         &_boxing_late_inlines, dead);
 396     remove_useless_late_inlines(&_vector_reboxing_late_inlines, dead);
 397   }
 398   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 399   bs->unregister_potential_barrier_node(dead);
 400 }
 401 
 402 // Disconnect all useless nodes by disconnecting those at the boundary.
 403 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 404   uint next = 0;
 405   while (next < useful.size()) {
 406     Node *n = useful.at(next++);
 407     if (n->is_SafePoint()) {
 408       // We're done with a parsing phase. Replaced nodes are not valid
 409       // beyond that point.
 410       n->as_SafePoint()->delete_replaced_nodes();
 411     }
 412     // Use raw traversal of out edges since this code removes out edges
 413     int max = n->outcnt();
 414     for (int j = 0; j < max; ++j) {
 415       Node* child = n->raw_out(j);
 416       if (!useful.member(child)) {
 417         assert(!child->is_top() || child != top(),
 418                "If top is cached in Compile object it is in useful list");
 419         // Only need to remove this out-edge to the useless node
 420         n->raw_del_out(j);
 421         --j;
 422         --max;
 423       }
 424     }
 425     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 426       record_for_igvn(n->unique_out());
 427     }
 428   }
 429 
 430   remove_useless_nodes(_macro_nodes,        useful); // remove useless macro nodes
 431   remove_useless_nodes(_predicate_opaqs,    useful); // remove useless predicate opaque nodes
 432   remove_useless_nodes(_skeleton_predicate_opaqs, useful);
 433   remove_useless_nodes(_expensive_nodes,    useful); // remove useless expensive nodes
 434   remove_useless_nodes(_for_post_loop_igvn, useful); // remove useless node recorded for post loop opts IGVN pass
 435   remove_useless_coarsened_locks(useful);            // remove useless coarsened locks nodes
 436 
 437   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 438   bs->eliminate_useless_gc_barriers(useful, this);
 439   // clean up the late inline lists
 440   remove_useless_late_inlines(                &_late_inlines, useful);
 441   remove_useless_late_inlines(         &_string_late_inlines, useful);
 442   remove_useless_late_inlines(         &_boxing_late_inlines, useful);
 443   remove_useless_late_inlines(&_vector_reboxing_late_inlines, useful);
 444   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 445 }
 446 
 447 // ============================================================================
 448 //------------------------------CompileWrapper---------------------------------
 449 class CompileWrapper : public StackObj {
 450   Compile *const _compile;
 451  public:
 452   CompileWrapper(Compile* compile);
 453 
 454   ~CompileWrapper();
 455 };
 456 
 457 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 458   // the Compile* pointer is stored in the current ciEnv:
 459   ciEnv* env = compile->env();
 460   assert(env == ciEnv::current(), "must already be a ciEnv active");
 461   assert(env->compiler_data() == NULL, "compile already active?");
 462   env->set_compiler_data(compile);
 463   assert(compile == Compile::current(), "sanity");
 464 
 465   compile->set_type_dict(NULL);
 466   compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
 467   compile->clone_map().set_clone_idx(0);
 468   compile->set_type_last_size(0);
 469   compile->set_last_tf(NULL, NULL);
 470   compile->set_indexSet_arena(NULL);
 471   compile->set_indexSet_free_block_list(NULL);
 472   compile->init_type_arena();
 473   Type::Initialize(compile);
 474   _compile->begin_method();
 475   _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
 476 }
 477 CompileWrapper::~CompileWrapper() {
 478   // simulate crash during compilation
 479   assert(CICrashAt < 0 || _compile->compile_id() != CICrashAt, "just as planned");
 480 
 481   _compile->end_method();
 482   _compile->env()->set_compiler_data(NULL);
 483 }
 484 
 485 
 486 //----------------------------print_compile_messages---------------------------
 487 void Compile::print_compile_messages() {
 488 #ifndef PRODUCT
 489   // Check if recompiling
 490   if (!subsume_loads() && PrintOpto) {
 491     // Recompiling without allowing machine instructions to subsume loads
 492     tty->print_cr("*********************************************************");
 493     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 494     tty->print_cr("*********************************************************");
 495   }
 496   if ((do_escape_analysis() != DoEscapeAnalysis) && PrintOpto) {
 497     // Recompiling without escape analysis
 498     tty->print_cr("*********************************************************");
 499     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 500     tty->print_cr("*********************************************************");
 501   }
 502   if ((eliminate_boxing() != EliminateAutoBox) && PrintOpto) {
 503     // Recompiling without boxing elimination
 504     tty->print_cr("*********************************************************");
 505     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 506     tty->print_cr("*********************************************************");
 507   }
 508   if ((do_locks_coarsening() != EliminateLocks) && PrintOpto) {
 509     // Recompiling without locks coarsening
 510     tty->print_cr("*********************************************************");
 511     tty->print_cr("** Bailout: Recompile without locks coarsening         **");
 512     tty->print_cr("*********************************************************");
 513   }
 514   if (env()->break_at_compile()) {
 515     // Open the debugger when compiling this method.
 516     tty->print("### Breaking when compiling: ");
 517     method()->print_short_name();
 518     tty->cr();
 519     BREAKPOINT;
 520   }
 521 
 522   if( PrintOpto ) {
 523     if (is_osr_compilation()) {
 524       tty->print("[OSR]%3d", _compile_id);
 525     } else {
 526       tty->print("%3d", _compile_id);
 527     }
 528   }
 529 #endif
 530 }
 531 
 532 // ============================================================================
 533 //------------------------------Compile standard-------------------------------
 534 debug_only( int Compile::_debug_idx = 100000; )
 535 
 536 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 537 // the continuation bci for on stack replacement.
 538 
 539 
 540 Compile::Compile( ciEnv* ci_env, ciMethod* target, int osr_bci,
 541                   Options options, DirectiveSet* directive)
 542                 : Phase(Compiler),
 543                   _compile_id(ci_env->compile_id()),
 544                   _options(options),
 545                   _method(target),
 546                   _entry_bci(osr_bci),
 547                   _ilt(NULL),
 548                   _stub_function(NULL),
 549                   _stub_name(NULL),
 550                   _stub_entry_point(NULL),
 551                   _max_node_limit(MaxNodeLimit),
 552                   _post_loop_opts_phase(false),
 553                   _inlining_progress(false),
 554                   _inlining_incrementally(false),
 555                   _do_cleanup(false),
 556                   _has_reserved_stack_access(target->has_reserved_stack_access()),
 557 #ifndef PRODUCT
 558                   _igv_idx(0),
 559                   _trace_opto_output(directive->TraceOptoOutputOption),
 560                   _print_ideal(directive->PrintIdealOption),
 561 #endif
 562                   _has_method_handle_invokes(false),
 563                   _clinit_barrier_on_entry(false),
 564                   _stress_seed(0),
 565                   _comp_arena(mtCompiler),
 566                   _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
 567                   _env(ci_env),
 568                   _directive(directive),
 569                   _log(ci_env->log()),
 570                   _failure_reason(NULL),
 571                   _intrinsics        (comp_arena(), 0, 0, NULL),
 572                   _macro_nodes       (comp_arena(), 8, 0, NULL),
 573                   _predicate_opaqs   (comp_arena(), 8, 0, NULL),
 574                   _skeleton_predicate_opaqs (comp_arena(), 8, 0, NULL),
 575                   _expensive_nodes   (comp_arena(), 8, 0, NULL),
 576                   _for_post_loop_igvn(comp_arena(), 8, 0, NULL),
 577                   _coarsened_locks   (comp_arena(), 8, 0, NULL),
 578                   _congraph(NULL),
 579                   NOT_PRODUCT(_printer(NULL) COMMA)
 580                   _dead_node_list(comp_arena()),
 581                   _dead_node_count(0),
 582                   _node_arena(mtCompiler),
 583                   _old_arena(mtCompiler),
 584                   _mach_constant_base_node(NULL),
 585                   _Compile_types(mtCompiler),
 586                   _initial_gvn(NULL),
 587                   _for_igvn(NULL),
 588                   _late_inlines(comp_arena(), 2, 0, NULL),
 589                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 590                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 591                   _vector_reboxing_late_inlines(comp_arena(), 2, 0, NULL),
 592                   _late_inlines_pos(0),
 593                   _number_of_mh_late_inlines(0),
 594                   _print_inlining_stream(NULL),
 595                   _print_inlining_list(NULL),
 596                   _print_inlining_idx(0),
 597                   _print_inlining_output(NULL),
 598                   _replay_inline_data(NULL),
 599                   _java_calls(0),
 600                   _inner_loops(0),
 601                   _interpreter_frame_size(0)
 602 #ifndef PRODUCT
 603                   , _in_dump_cnt(0)
 604 #endif
 605 {
 606   C = this;
 607   CompileWrapper cw(this);
 608 
 609   if (CITimeVerbose) {
 610     tty->print(" ");
 611     target->holder()->name()->print();
 612     tty->print(".");
 613     target->print_short_name();
 614     tty->print("  ");
 615   }
 616   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 617   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 618 
 619 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
 620   bool print_opto_assembly = directive->PrintOptoAssemblyOption;
 621   // We can always print a disassembly, either abstract (hex dump) or
 622   // with the help of a suitable hsdis library. Thus, we should not
 623   // couple print_assembly and print_opto_assembly controls.
 624   // But: always print opto and regular assembly on compile command 'print'.
 625   bool print_assembly = directive->PrintAssemblyOption;
 626   set_print_assembly(print_opto_assembly || print_assembly);
 627 #else
 628   set_print_assembly(false); // must initialize.
 629 #endif
 630 
 631 #ifndef PRODUCT
 632   set_parsed_irreducible_loop(false);
 633 
 634   if (directive->ReplayInlineOption) {
 635     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 636   }
 637 #endif
 638   set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
 639   set_print_intrinsics(directive->PrintIntrinsicsOption);
 640   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 641 
 642   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 643     // Make sure the method being compiled gets its own MDO,
 644     // so we can at least track the decompile_count().
 645     // Need MDO to record RTM code generation state.
 646     method()->ensure_method_data();
 647   }
 648 
 649   Init(::AliasLevel);
 650 
 651 
 652   print_compile_messages();
 653 
 654   _ilt = InlineTree::build_inline_tree_root();
 655 
 656   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 657   assert(num_alias_types() >= AliasIdxRaw, "");
 658 
 659 #define MINIMUM_NODE_HASH  1023
 660   // Node list that Iterative GVN will start with
 661   Unique_Node_List for_igvn(comp_arena());
 662   set_for_igvn(&for_igvn);
 663 
 664   // GVN that will be run immediately on new nodes
 665   uint estimated_size = method()->code_size()*4+64;
 666   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 667   PhaseGVN gvn(node_arena(), estimated_size);
 668   set_initial_gvn(&gvn);
 669 
 670   print_inlining_init();
 671   { // Scope for timing the parser
 672     TracePhase tp("parse", &timers[_t_parser]);
 673 
 674     // Put top into the hash table ASAP.
 675     initial_gvn()->transform_no_reclaim(top());
 676 
 677     // Set up tf(), start(), and find a CallGenerator.
 678     CallGenerator* cg = NULL;
 679     if (is_osr_compilation()) {
 680       const TypeTuple *domain = StartOSRNode::osr_domain();
 681       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 682       init_tf(TypeFunc::make(domain, range));
 683       StartNode* s = new StartOSRNode(root(), domain);
 684       initial_gvn()->set_type_bottom(s);
 685       init_start(s);
 686       cg = CallGenerator::for_osr(method(), entry_bci());
 687     } else {
 688       // Normal case.
 689       init_tf(TypeFunc::make(method()));
 690       StartNode* s = new StartNode(root(), tf()->domain());
 691       initial_gvn()->set_type_bottom(s);
 692       init_start(s);
 693       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
 694         // With java.lang.ref.reference.get() we must go through the
 695         // intrinsic - even when get() is the root
 696         // method of the compile - so that, if necessary, the value in
 697         // the referent field of the reference object gets recorded by
 698         // the pre-barrier code.
 699         cg = find_intrinsic(method(), false);
 700       }
 701       if (cg == NULL) {
 702         float past_uses = method()->interpreter_invocation_count();
 703         float expected_uses = past_uses;
 704         cg = CallGenerator::for_inline(method(), expected_uses);
 705       }
 706     }
 707     if (failing())  return;
 708     if (cg == NULL) {
 709       record_method_not_compilable("cannot parse method");
 710       return;
 711     }
 712     JVMState* jvms = build_start_state(start(), tf());
 713     if ((jvms = cg->generate(jvms)) == NULL) {
 714       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 715         record_method_not_compilable("method parse failed");
 716       }
 717       return;
 718     }
 719     GraphKit kit(jvms);
 720 
 721     if (!kit.stopped()) {
 722       // Accept return values, and transfer control we know not where.
 723       // This is done by a special, unique ReturnNode bound to root.
 724       return_values(kit.jvms());
 725     }
 726 
 727     if (kit.has_exceptions()) {
 728       // Any exceptions that escape from this call must be rethrown
 729       // to whatever caller is dynamically above us on the stack.
 730       // This is done by a special, unique RethrowNode bound to root.
 731       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 732     }
 733 
 734     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 735 
 736     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 737       inline_string_calls(true);
 738     }
 739 
 740     if (failing())  return;
 741 
 742     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 743 
 744     // Remove clutter produced by parsing.
 745     if (!failing()) {
 746       ResourceMark rm;
 747       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 748     }
 749   }
 750 
 751   // Note:  Large methods are capped off in do_one_bytecode().
 752   if (failing())  return;
 753 
 754   // After parsing, node notes are no longer automagic.
 755   // They must be propagated by register_new_node_with_optimizer(),
 756   // clone(), or the like.
 757   set_default_node_notes(NULL);
 758 
 759 #ifndef PRODUCT
 760   if (should_print(1)) {
 761     _printer->print_inlining();
 762   }
 763 #endif
 764 
 765   if (failing())  return;
 766   NOT_PRODUCT( verify_graph_edges(); )
 767 
 768   // If any phase is randomized for stress testing, seed random number
 769   // generation and log the seed for repeatability.
 770   if (StressLCM || StressGCM || StressIGVN || StressCCP) {
 771     if (FLAG_IS_DEFAULT(StressSeed) || (FLAG_IS_ERGO(StressSeed) && RepeatCompilation)) {
 772       _stress_seed = static_cast<uint>(Ticks::now().nanoseconds());
 773       FLAG_SET_ERGO(StressSeed, _stress_seed);
 774     } else {
 775       _stress_seed = StressSeed;
 776     }
 777     if (_log != NULL) {
 778       _log->elem("stress_test seed='%u'", _stress_seed);
 779     }
 780   }
 781 
 782   // Now optimize
 783   Optimize();
 784   if (failing())  return;
 785   NOT_PRODUCT( verify_graph_edges(); )
 786 
 787 #ifndef PRODUCT
 788   if (print_ideal()) {
 789     ttyLocker ttyl;  // keep the following output all in one block
 790     // This output goes directly to the tty, not the compiler log.
 791     // To enable tools to match it up with the compilation activity,
 792     // be sure to tag this tty output with the compile ID.
 793     if (xtty != NULL) {
 794       xtty->head("ideal compile_id='%d'%s", compile_id(),
 795                  is_osr_compilation()    ? " compile_kind='osr'" :
 796                  "");
 797     }
 798     root()->dump(9999);
 799     if (xtty != NULL) {
 800       xtty->tail("ideal");
 801     }
 802   }
 803 #endif
 804 
 805 #ifdef ASSERT
 806   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 807   bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
 808 #endif
 809 
 810   // Dump compilation data to replay it.
 811   if (directive->DumpReplayOption) {
 812     env()->dump_replay_data(_compile_id);
 813   }
 814   if (directive->DumpInlineOption && (ilt() != NULL)) {
 815     env()->dump_inline_data(_compile_id);
 816   }
 817 
 818   // Now that we know the size of all the monitors we can add a fixed slot
 819   // for the original deopt pc.
 820   int next_slot = fixed_slots() + (sizeof(address) / VMRegImpl::stack_slot_size);
 821   set_fixed_slots(next_slot);
 822 
 823   // Compute when to use implicit null checks. Used by matching trap based
 824   // nodes and NullCheck optimization.
 825   set_allowed_deopt_reasons();
 826 
 827   // Now generate code
 828   Code_Gen();
 829 }
 830 
 831 //------------------------------Compile----------------------------------------
 832 // Compile a runtime stub
 833 Compile::Compile( ciEnv* ci_env,
 834                   TypeFunc_generator generator,
 835                   address stub_function,
 836                   const char *stub_name,
 837                   int is_fancy_jump,
 838                   bool pass_tls,
 839                   bool return_pc,
 840                   DirectiveSet* directive)
 841   : Phase(Compiler),
 842     _compile_id(0),
 843     _options(Options::for_runtime_stub()),
 844     _method(NULL),
 845     _entry_bci(InvocationEntryBci),
 846     _stub_function(stub_function),
 847     _stub_name(stub_name),
 848     _stub_entry_point(NULL),
 849     _max_node_limit(MaxNodeLimit),
 850     _post_loop_opts_phase(false),
 851     _inlining_progress(false),
 852     _inlining_incrementally(false),
 853     _has_reserved_stack_access(false),
 854 #ifndef PRODUCT
 855     _igv_idx(0),
 856     _trace_opto_output(directive->TraceOptoOutputOption),
 857     _print_ideal(directive->PrintIdealOption),
 858 #endif
 859     _has_method_handle_invokes(false),
 860     _clinit_barrier_on_entry(false),
 861     _stress_seed(0),
 862     _comp_arena(mtCompiler),
 863     _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
 864     _env(ci_env),
 865     _directive(directive),
 866     _log(ci_env->log()),
 867     _failure_reason(NULL),
 868     _congraph(NULL),
 869     NOT_PRODUCT(_printer(NULL) COMMA)
 870     _dead_node_list(comp_arena()),
 871     _dead_node_count(0),
 872     _node_arena(mtCompiler),
 873     _old_arena(mtCompiler),
 874     _mach_constant_base_node(NULL),
 875     _Compile_types(mtCompiler),
 876     _initial_gvn(NULL),
 877     _for_igvn(NULL),
 878     _number_of_mh_late_inlines(0),
 879     _print_inlining_stream(NULL),
 880     _print_inlining_list(NULL),
 881     _print_inlining_idx(0),
 882     _print_inlining_output(NULL),
 883     _replay_inline_data(NULL),
 884     _java_calls(0),
 885     _inner_loops(0),
 886     _interpreter_frame_size(0),
 887 #ifndef PRODUCT
 888     _in_dump_cnt(0),
 889 #endif
 890     _allowed_reasons(0) {
 891   C = this;
 892 
 893   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
 894   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
 895 
 896 #ifndef PRODUCT
 897   set_print_assembly(PrintFrameConverterAssembly);
 898   set_parsed_irreducible_loop(false);
 899 #else
 900   set_print_assembly(false); // Must initialize.
 901 #endif
 902   set_has_irreducible_loop(false); // no loops
 903 
 904   CompileWrapper cw(this);
 905   Init(/*AliasLevel=*/ 0);
 906   init_tf((*generator)());
 907 
 908   {
 909     // The following is a dummy for the sake of GraphKit::gen_stub
 910     Unique_Node_List for_igvn(comp_arena());
 911     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
 912     PhaseGVN gvn(Thread::current()->resource_area(),255);
 913     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
 914     gvn.transform_no_reclaim(top());
 915 
 916     GraphKit kit;
 917     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
 918   }
 919 
 920   NOT_PRODUCT( verify_graph_edges(); )
 921 
 922   Code_Gen();
 923 }
 924 
 925 //------------------------------Init-------------------------------------------
 926 // Prepare for a single compilation
 927 void Compile::Init(int aliaslevel) {
 928   _unique  = 0;
 929   _regalloc = NULL;
 930 
 931   _tf      = NULL;  // filled in later
 932   _top     = NULL;  // cached later
 933   _matcher = NULL;  // filled in later
 934   _cfg     = NULL;  // filled in later
 935 
 936   IA32_ONLY( set_24_bit_selection_and_mode(true, false); )
 937 
 938   _node_note_array = NULL;
 939   _default_node_notes = NULL;
 940   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
 941 
 942   _immutable_memory = NULL; // filled in at first inquiry
 943 
 944   // Globally visible Nodes
 945   // First set TOP to NULL to give safe behavior during creation of RootNode
 946   set_cached_top_node(NULL);
 947   set_root(new RootNode());
 948   // Now that you have a Root to point to, create the real TOP
 949   set_cached_top_node( new ConNode(Type::TOP) );
 950   set_recent_alloc(NULL, NULL);
 951 
 952   // Create Debug Information Recorder to record scopes, oopmaps, etc.
 953   env()->set_oop_recorder(new OopRecorder(env()->arena()));
 954   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
 955   env()->set_dependencies(new Dependencies(env()));
 956 
 957   _fixed_slots = 0;
 958   set_has_split_ifs(false);
 959   set_has_loops(false); // first approximation
 960   set_has_stringbuilder(false);
 961   set_has_boxed_value(false);
 962   _trap_can_recompile = false;  // no traps emitted yet
 963   _major_progress = true; // start out assuming good things will happen
 964   set_has_unsafe_access(false);
 965   set_max_vector_size(0);
 966   set_clear_upper_avx(false);  //false as default for clear upper bits of ymm registers
 967   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
 968   set_decompile_count(0);
 969 
 970   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
 971   _loop_opts_cnt = LoopOptsCount;
 972   set_do_inlining(Inline);
 973   set_max_inline_size(MaxInlineSize);
 974   set_freq_inline_size(FreqInlineSize);
 975   set_do_scheduling(OptoScheduling);
 976 
 977   set_do_vector_loop(false);
 978 
 979   if (AllowVectorizeOnDemand) {
 980     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
 981       set_do_vector_loop(true);
 982       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());})
 983     } else if (has_method() && method()->name() != 0 &&
 984                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
 985       set_do_vector_loop(true);
 986     }
 987   }
 988   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
 989   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());})
 990 
 991   set_age_code(has_method() && method()->profile_aging());
 992   set_rtm_state(NoRTM); // No RTM lock eliding by default
 993   _max_node_limit = _directive->MaxNodeLimitOption;
 994 
 995 #if INCLUDE_RTM_OPT
 996   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
 997     int rtm_state = method()->method_data()->rtm_state();
 998     if (method_has_option(CompileCommand::NoRTMLockEliding) || ((rtm_state & NoRTM) != 0)) {
 999       // Don't generate RTM lock eliding code.
1000       set_rtm_state(NoRTM);
1001     } else if (method_has_option(CompileCommand::UseRTMLockEliding) || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1002       // Generate RTM lock eliding code without abort ratio calculation code.
1003       set_rtm_state(UseRTM);
1004     } else if (UseRTMDeopt) {
1005       // Generate RTM lock eliding code and include abort ratio calculation
1006       // code if UseRTMDeopt is on.
1007       set_rtm_state(ProfileRTM);
1008     }
1009   }
1010 #endif
1011   if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1012     set_clinit_barrier_on_entry(true);
1013   }
1014   if (debug_info()->recording_non_safepoints()) {
1015     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1016                         (comp_arena(), 8, 0, NULL));
1017     set_default_node_notes(Node_Notes::make(this));
1018   }
1019 
1020   // // -- Initialize types before each compile --
1021   // // Update cached type information
1022   // if( _method && _method->constants() )
1023   //   Type::update_loaded_types(_method, _method->constants());
1024 
1025   // Init alias_type map.
1026   if (!do_escape_analysis() && aliaslevel == 3) {
1027     aliaslevel = 2;  // No unique types without escape analysis
1028   }
1029   _AliasLevel = aliaslevel;
1030   const int grow_ats = 16;
1031   _max_alias_types = grow_ats;
1032   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1033   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1034   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1035   {
1036     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1037   }
1038   // Initialize the first few types.
1039   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1040   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1041   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1042   _num_alias_types = AliasIdxRaw+1;
1043   // Zero out the alias type cache.
1044   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1045   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1046   probe_alias_cache(NULL)->_index = AliasIdxTop;
1047 
1048 #ifdef ASSERT
1049   _type_verify_symmetry = true;
1050   _phase_optimize_finished = false;
1051   _exception_backedge = false;
1052 #endif
1053 }
1054 
1055 //---------------------------init_start----------------------------------------
1056 // Install the StartNode on this compile object.
1057 void Compile::init_start(StartNode* s) {
1058   if (failing())
1059     return; // already failing
1060   assert(s == start(), "");
1061 }
1062 
1063 /**
1064  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1065  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1066  * the ideal graph.
1067  */
1068 StartNode* Compile::start() const {
1069   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1070   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1071     Node* start = root()->fast_out(i);
1072     if (start->is_Start()) {
1073       return start->as_Start();
1074     }
1075   }
1076   fatal("Did not find Start node!");
1077   return NULL;
1078 }
1079 
1080 //-------------------------------immutable_memory-------------------------------------
1081 // Access immutable memory
1082 Node* Compile::immutable_memory() {
1083   if (_immutable_memory != NULL) {
1084     return _immutable_memory;
1085   }
1086   StartNode* s = start();
1087   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1088     Node *p = s->fast_out(i);
1089     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1090       _immutable_memory = p;
1091       return _immutable_memory;
1092     }
1093   }
1094   ShouldNotReachHere();
1095   return NULL;
1096 }
1097 
1098 //----------------------set_cached_top_node------------------------------------
1099 // Install the cached top node, and make sure Node::is_top works correctly.
1100 void Compile::set_cached_top_node(Node* tn) {
1101   if (tn != NULL)  verify_top(tn);
1102   Node* old_top = _top;
1103   _top = tn;
1104   // Calling Node::setup_is_top allows the nodes the chance to adjust
1105   // their _out arrays.
1106   if (_top != NULL)     _top->setup_is_top();
1107   if (old_top != NULL)  old_top->setup_is_top();
1108   assert(_top == NULL || top()->is_top(), "");
1109 }
1110 
1111 #ifdef ASSERT
1112 uint Compile::count_live_nodes_by_graph_walk() {
1113   Unique_Node_List useful(comp_arena());
1114   // Get useful node list by walking the graph.
1115   identify_useful_nodes(useful);
1116   return useful.size();
1117 }
1118 
1119 void Compile::print_missing_nodes() {
1120 
1121   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1122   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1123     return;
1124   }
1125 
1126   // This is an expensive function. It is executed only when the user
1127   // specifies VerifyIdealNodeCount option or otherwise knows the
1128   // additional work that needs to be done to identify reachable nodes
1129   // by walking the flow graph and find the missing ones using
1130   // _dead_node_list.
1131 
1132   Unique_Node_List useful(comp_arena());
1133   // Get useful node list by walking the graph.
1134   identify_useful_nodes(useful);
1135 
1136   uint l_nodes = C->live_nodes();
1137   uint l_nodes_by_walk = useful.size();
1138 
1139   if (l_nodes != l_nodes_by_walk) {
1140     if (_log != NULL) {
1141       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1142       _log->stamp();
1143       _log->end_head();
1144     }
1145     VectorSet& useful_member_set = useful.member_set();
1146     int last_idx = l_nodes_by_walk;
1147     for (int i = 0; i < last_idx; i++) {
1148       if (useful_member_set.test(i)) {
1149         if (_dead_node_list.test(i)) {
1150           if (_log != NULL) {
1151             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1152           }
1153           if (PrintIdealNodeCount) {
1154             // Print the log message to tty
1155               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1156               useful.at(i)->dump();
1157           }
1158         }
1159       }
1160       else if (! _dead_node_list.test(i)) {
1161         if (_log != NULL) {
1162           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1163         }
1164         if (PrintIdealNodeCount) {
1165           // Print the log message to tty
1166           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1167         }
1168       }
1169     }
1170     if (_log != NULL) {
1171       _log->tail("mismatched_nodes");
1172     }
1173   }
1174 }
1175 void Compile::record_modified_node(Node* n) {
1176   if (_modified_nodes != NULL && !_inlining_incrementally && !n->is_Con()) {
1177     _modified_nodes->push(n);
1178   }
1179 }
1180 
1181 void Compile::remove_modified_node(Node* n) {
1182   if (_modified_nodes != NULL) {
1183     _modified_nodes->remove(n);
1184   }
1185 }
1186 #endif
1187 
1188 #ifndef PRODUCT
1189 void Compile::verify_top(Node* tn) const {
1190   if (tn != NULL) {
1191     assert(tn->is_Con(), "top node must be a constant");
1192     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1193     assert(tn->in(0) != NULL, "must have live top node");
1194   }
1195 }
1196 #endif
1197 
1198 
1199 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1200 
1201 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1202   guarantee(arr != NULL, "");
1203   int num_blocks = arr->length();
1204   if (grow_by < num_blocks)  grow_by = num_blocks;
1205   int num_notes = grow_by * _node_notes_block_size;
1206   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1207   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1208   while (num_notes > 0) {
1209     arr->append(notes);
1210     notes     += _node_notes_block_size;
1211     num_notes -= _node_notes_block_size;
1212   }
1213   assert(num_notes == 0, "exact multiple, please");
1214 }
1215 
1216 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1217   if (source == NULL || dest == NULL)  return false;
1218 
1219   if (dest->is_Con())
1220     return false;               // Do not push debug info onto constants.
1221 
1222 #ifdef ASSERT
1223   // Leave a bread crumb trail pointing to the original node:
1224   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1225     dest->set_debug_orig(source);
1226   }
1227 #endif
1228 
1229   if (node_note_array() == NULL)
1230     return false;               // Not collecting any notes now.
1231 
1232   // This is a copy onto a pre-existing node, which may already have notes.
1233   // If both nodes have notes, do not overwrite any pre-existing notes.
1234   Node_Notes* source_notes = node_notes_at(source->_idx);
1235   if (source_notes == NULL || source_notes->is_clear())  return false;
1236   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1237   if (dest_notes == NULL || dest_notes->is_clear()) {
1238     return set_node_notes_at(dest->_idx, source_notes);
1239   }
1240 
1241   Node_Notes merged_notes = (*source_notes);
1242   // The order of operations here ensures that dest notes will win...
1243   merged_notes.update_from(dest_notes);
1244   return set_node_notes_at(dest->_idx, &merged_notes);
1245 }
1246 
1247 
1248 //--------------------------allow_range_check_smearing-------------------------
1249 // Gating condition for coalescing similar range checks.
1250 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1251 // single covering check that is at least as strong as any of them.
1252 // If the optimization succeeds, the simplified (strengthened) range check
1253 // will always succeed.  If it fails, we will deopt, and then give up
1254 // on the optimization.
1255 bool Compile::allow_range_check_smearing() const {
1256   // If this method has already thrown a range-check,
1257   // assume it was because we already tried range smearing
1258   // and it failed.
1259   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1260   return !already_trapped;
1261 }
1262 
1263 
1264 //------------------------------flatten_alias_type-----------------------------
1265 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1266   int offset = tj->offset();
1267   TypePtr::PTR ptr = tj->ptr();
1268 
1269   // Known instance (scalarizable allocation) alias only with itself.
1270   bool is_known_inst = tj->isa_oopptr() != NULL &&
1271                        tj->is_oopptr()->is_known_instance();
1272 
1273   // Process weird unsafe references.
1274   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1275     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1276     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1277     tj = TypeOopPtr::BOTTOM;
1278     ptr = tj->ptr();
1279     offset = tj->offset();
1280   }
1281 
1282   // Array pointers need some flattening
1283   const TypeAryPtr *ta = tj->isa_aryptr();
1284   if (ta && ta->is_stable()) {
1285     // Erase stability property for alias analysis.
1286     tj = ta = ta->cast_to_stable(false);
1287   }
1288   if( ta && is_known_inst ) {
1289     if ( offset != Type::OffsetBot &&
1290          offset > arrayOopDesc::length_offset_in_bytes() ) {
1291       offset = Type::OffsetBot; // Flatten constant access into array body only
1292       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1293     }
1294   } else if( ta && _AliasLevel >= 2 ) {
1295     // For arrays indexed by constant indices, we flatten the alias
1296     // space to include all of the array body.  Only the header, klass
1297     // and array length can be accessed un-aliased.
1298     if( offset != Type::OffsetBot ) {
1299       if( ta->const_oop() ) { // MethodData* or Method*
1300         offset = Type::OffsetBot;   // Flatten constant access into array body
1301         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1302       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1303         // range is OK as-is.
1304         tj = ta = TypeAryPtr::RANGE;
1305       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1306         tj = TypeInstPtr::KLASS; // all klass loads look alike
1307         ta = TypeAryPtr::RANGE; // generic ignored junk
1308         ptr = TypePtr::BotPTR;
1309       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1310         tj = TypeInstPtr::MARK;
1311         ta = TypeAryPtr::RANGE; // generic ignored junk
1312         ptr = TypePtr::BotPTR;
1313       } else {                  // Random constant offset into array body
1314         offset = Type::OffsetBot;   // Flatten constant access into array body
1315         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1316       }
1317     }
1318     // Arrays of fixed size alias with arrays of unknown size.
1319     if (ta->size() != TypeInt::POS) {
1320       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1321       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1322     }
1323     // Arrays of known objects become arrays of unknown objects.
1324     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1325       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1326       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1327     }
1328     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1329       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1330       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1331     }
1332     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1333     // cannot be distinguished by bytecode alone.
1334     if (ta->elem() == TypeInt::BOOL) {
1335       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1336       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1337       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1338     }
1339     // During the 2nd round of IterGVN, NotNull castings are removed.
1340     // Make sure the Bottom and NotNull variants alias the same.
1341     // Also, make sure exact and non-exact variants alias the same.
1342     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1343       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1344     }
1345   }
1346 
1347   // Oop pointers need some flattening
1348   const TypeInstPtr *to = tj->isa_instptr();
1349   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1350     ciInstanceKlass *k = to->klass()->as_instance_klass();
1351     if( ptr == TypePtr::Constant ) {
1352       if (to->klass() != ciEnv::current()->Class_klass() ||
1353           offset < k->layout_helper_size_in_bytes()) {
1354         // No constant oop pointers (such as Strings); they alias with
1355         // unknown strings.
1356         assert(!is_known_inst, "not scalarizable allocation");
1357         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1358       }
1359     } else if( is_known_inst ) {
1360       tj = to; // Keep NotNull and klass_is_exact for instance type
1361     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1362       // During the 2nd round of IterGVN, NotNull castings are removed.
1363       // Make sure the Bottom and NotNull variants alias the same.
1364       // Also, make sure exact and non-exact variants alias the same.
1365       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1366     }
1367     if (to->speculative() != NULL) {
1368       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1369     }
1370     // Canonicalize the holder of this field
1371     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1372       // First handle header references such as a LoadKlassNode, even if the
1373       // object's klass is unloaded at compile time (4965979).
1374       if (!is_known_inst) { // Do it only for non-instance types
1375         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1376       }
1377     } else if (offset < 0 || offset >= k->layout_helper_size_in_bytes()) {
1378       // Static fields are in the space above the normal instance
1379       // fields in the java.lang.Class instance.
1380       if (to->klass() != ciEnv::current()->Class_klass()) {
1381         to = NULL;
1382         tj = TypeOopPtr::BOTTOM;
1383         offset = tj->offset();
1384       }
1385     } else {
1386       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1387       assert(offset < canonical_holder->layout_helper_size_in_bytes(), "");
1388       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1389         if( is_known_inst ) {
1390           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1391         } else {
1392           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1393         }
1394       }
1395     }
1396   }
1397 
1398   // Klass pointers to object array klasses need some flattening
1399   const TypeKlassPtr *tk = tj->isa_klassptr();
1400   if( tk ) {
1401     // If we are referencing a field within a Klass, we need
1402     // to assume the worst case of an Object.  Both exact and
1403     // inexact types must flatten to the same alias class so
1404     // use NotNull as the PTR.
1405     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1406 
1407       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1408                                    TypeInstKlassPtr::OBJECT->klass(),
1409                                    offset);
1410     }
1411 
1412     ciKlass* klass = tk->klass();
1413     if( klass->is_obj_array_klass() ) {
1414       ciKlass* k = TypeAryPtr::OOPS->klass();
1415       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1416         k = TypeInstPtr::BOTTOM->klass();
1417       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1418     }
1419 
1420     // Check for precise loads from the primary supertype array and force them
1421     // to the supertype cache alias index.  Check for generic array loads from
1422     // the primary supertype array and also force them to the supertype cache
1423     // alias index.  Since the same load can reach both, we need to merge
1424     // these 2 disparate memories into the same alias class.  Since the
1425     // primary supertype array is read-only, there's no chance of confusion
1426     // where we bypass an array load and an array store.
1427     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1428     if (offset == Type::OffsetBot ||
1429         (offset >= primary_supers_offset &&
1430          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1431         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1432       offset = in_bytes(Klass::secondary_super_cache_offset());
1433       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1434     }
1435   }
1436 
1437   // Flatten all Raw pointers together.
1438   if (tj->base() == Type::RawPtr)
1439     tj = TypeRawPtr::BOTTOM;
1440 
1441   if (tj->base() == Type::AnyPtr)
1442     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1443 
1444   // Flatten all to bottom for now
1445   switch( _AliasLevel ) {
1446   case 0:
1447     tj = TypePtr::BOTTOM;
1448     break;
1449   case 1:                       // Flatten to: oop, static, field or array
1450     switch (tj->base()) {
1451     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1452     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1453     case Type::AryPtr:   // do not distinguish arrays at all
1454     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1455     case Type::KlassPtr:
1456     case Type::AryKlassPtr:
1457     case Type::InstKlassPtr: tj = TypeInstKlassPtr::OBJECT; break;
1458     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1459     default: ShouldNotReachHere();
1460     }
1461     break;
1462   case 2:                       // No collapsing at level 2; keep all splits
1463   case 3:                       // No collapsing at level 3; keep all splits
1464     break;
1465   default:
1466     Unimplemented();
1467   }
1468 
1469   offset = tj->offset();
1470   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1471 
1472   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1473           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1474           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1475           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1476           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1477           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1478           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr),
1479           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1480   assert( tj->ptr() != TypePtr::TopPTR &&
1481           tj->ptr() != TypePtr::AnyNull &&
1482           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1483 //    assert( tj->ptr() != TypePtr::Constant ||
1484 //            tj->base() == Type::RawPtr ||
1485 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1486 
1487   return tj;
1488 }
1489 
1490 void Compile::AliasType::Init(int i, const TypePtr* at) {
1491   assert(AliasIdxTop <= i && i < Compile::current()->_max_alias_types, "Invalid alias index");
1492   _index = i;
1493   _adr_type = at;
1494   _field = NULL;
1495   _element = NULL;
1496   _is_rewritable = true; // default
1497   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1498   if (atoop != NULL && atoop->is_known_instance()) {
1499     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1500     _general_index = Compile::current()->get_alias_index(gt);
1501   } else {
1502     _general_index = 0;
1503   }
1504 }
1505 
1506 BasicType Compile::AliasType::basic_type() const {
1507   if (element() != NULL) {
1508     const Type* element = adr_type()->is_aryptr()->elem();
1509     return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1510   } if (field() != NULL) {
1511     return field()->layout_type();
1512   } else {
1513     return T_ILLEGAL; // unknown
1514   }
1515 }
1516 
1517 //---------------------------------print_on------------------------------------
1518 #ifndef PRODUCT
1519 void Compile::AliasType::print_on(outputStream* st) {
1520   if (index() < 10)
1521         st->print("@ <%d> ", index());
1522   else  st->print("@ <%d>",  index());
1523   st->print(is_rewritable() ? "   " : " RO");
1524   int offset = adr_type()->offset();
1525   if (offset == Type::OffsetBot)
1526         st->print(" +any");
1527   else  st->print(" +%-3d", offset);
1528   st->print(" in ");
1529   adr_type()->dump_on(st);
1530   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1531   if (field() != NULL && tjp) {
1532     if (tjp->klass()  != field()->holder() ||
1533         tjp->offset() != field()->offset_in_bytes()) {
1534       st->print(" != ");
1535       field()->print();
1536       st->print(" ***");
1537     }
1538   }
1539 }
1540 
1541 void print_alias_types() {
1542   Compile* C = Compile::current();
1543   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1544   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1545     C->alias_type(idx)->print_on(tty);
1546     tty->cr();
1547   }
1548 }
1549 #endif
1550 
1551 
1552 //----------------------------probe_alias_cache--------------------------------
1553 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1554   intptr_t key = (intptr_t) adr_type;
1555   key ^= key >> logAliasCacheSize;
1556   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1557 }
1558 
1559 
1560 //-----------------------------grow_alias_types--------------------------------
1561 void Compile::grow_alias_types() {
1562   const int old_ats  = _max_alias_types; // how many before?
1563   const int new_ats  = old_ats;          // how many more?
1564   const int grow_ats = old_ats+new_ats;  // how many now?
1565   _max_alias_types = grow_ats;
1566   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1567   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1568   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1569   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1570 }
1571 
1572 
1573 //--------------------------------find_alias_type------------------------------
1574 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1575   if (_AliasLevel == 0)
1576     return alias_type(AliasIdxBot);
1577 
1578   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1579   if (ace->_adr_type == adr_type) {
1580     return alias_type(ace->_index);
1581   }
1582 
1583   // Handle special cases.
1584   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1585   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1586 
1587   // Do it the slow way.
1588   const TypePtr* flat = flatten_alias_type(adr_type);
1589 
1590 #ifdef ASSERT
1591   {
1592     ResourceMark rm;
1593     assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1594            Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1595     assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1596            Type::str(adr_type));
1597     if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1598       const TypeOopPtr* foop = flat->is_oopptr();
1599       // Scalarizable allocations have exact klass always.
1600       bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1601       const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1602       assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1603              Type::str(foop), Type::str(xoop));
1604     }
1605   }
1606 #endif
1607 
1608   int idx = AliasIdxTop;
1609   for (int i = 0; i < num_alias_types(); i++) {
1610     if (alias_type(i)->adr_type() == flat) {
1611       idx = i;
1612       break;
1613     }
1614   }
1615 
1616   if (idx == AliasIdxTop) {
1617     if (no_create)  return NULL;
1618     // Grow the array if necessary.
1619     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1620     // Add a new alias type.
1621     idx = _num_alias_types++;
1622     _alias_types[idx]->Init(idx, flat);
1623     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1624     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1625     if (flat->isa_instptr()) {
1626       if (flat->offset() == java_lang_Class::klass_offset()
1627           && flat->is_instptr()->klass() == env()->Class_klass())
1628         alias_type(idx)->set_rewritable(false);
1629     }
1630     if (flat->isa_aryptr()) {
1631 #ifdef ASSERT
1632       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1633       // (T_BYTE has the weakest alignment and size restrictions...)
1634       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1635 #endif
1636       if (flat->offset() == TypePtr::OffsetBot) {
1637         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1638       }
1639     }
1640     if (flat->isa_klassptr()) {
1641       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1642         alias_type(idx)->set_rewritable(false);
1643       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1644         alias_type(idx)->set_rewritable(false);
1645       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1646         alias_type(idx)->set_rewritable(false);
1647       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1648         alias_type(idx)->set_rewritable(false);
1649       if (flat->offset() == in_bytes(Klass::secondary_super_cache_offset()))
1650         alias_type(idx)->set_rewritable(false);
1651     }
1652     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1653     // but the base pointer type is not distinctive enough to identify
1654     // references into JavaThread.)
1655 
1656     // Check for final fields.
1657     const TypeInstPtr* tinst = flat->isa_instptr();
1658     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1659       ciField* field;
1660       if (tinst->const_oop() != NULL &&
1661           tinst->klass() == ciEnv::current()->Class_klass() &&
1662           tinst->offset() >= (tinst->klass()->as_instance_klass()->layout_helper_size_in_bytes())) {
1663         // static field
1664         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1665         field = k->get_field_by_offset(tinst->offset(), true);
1666       } else {
1667         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1668         field = k->get_field_by_offset(tinst->offset(), false);
1669       }
1670       assert(field == NULL ||
1671              original_field == NULL ||
1672              (field->holder() == original_field->holder() &&
1673               field->offset() == original_field->offset() &&
1674               field->is_static() == original_field->is_static()), "wrong field?");
1675       // Set field() and is_rewritable() attributes.
1676       if (field != NULL)  alias_type(idx)->set_field(field);
1677     }
1678   }
1679 
1680   // Fill the cache for next time.
1681   ace->_adr_type = adr_type;
1682   ace->_index    = idx;
1683   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1684 
1685   // Might as well try to fill the cache for the flattened version, too.
1686   AliasCacheEntry* face = probe_alias_cache(flat);
1687   if (face->_adr_type == NULL) {
1688     face->_adr_type = flat;
1689     face->_index    = idx;
1690     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1691   }
1692 
1693   return alias_type(idx);
1694 }
1695 
1696 
1697 Compile::AliasType* Compile::alias_type(ciField* field) {
1698   const TypeOopPtr* t;
1699   if (field->is_static())
1700     t = TypeInstPtr::make(field->holder()->java_mirror());
1701   else
1702     t = TypeOopPtr::make_from_klass_raw(field->holder());
1703   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1704   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1705   return atp;
1706 }
1707 
1708 
1709 //------------------------------have_alias_type--------------------------------
1710 bool Compile::have_alias_type(const TypePtr* adr_type) {
1711   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1712   if (ace->_adr_type == adr_type) {
1713     return true;
1714   }
1715 
1716   // Handle special cases.
1717   if (adr_type == NULL)             return true;
1718   if (adr_type == TypePtr::BOTTOM)  return true;
1719 
1720   return find_alias_type(adr_type, true, NULL) != NULL;
1721 }
1722 
1723 //-----------------------------must_alias--------------------------------------
1724 // True if all values of the given address type are in the given alias category.
1725 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1726   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1727   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1728   if (alias_idx == AliasIdxTop)         return false; // the empty category
1729   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1730 
1731   // the only remaining possible overlap is identity
1732   int adr_idx = get_alias_index(adr_type);
1733   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1734   assert(adr_idx == alias_idx ||
1735          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1736           && adr_type                       != TypeOopPtr::BOTTOM),
1737          "should not be testing for overlap with an unsafe pointer");
1738   return adr_idx == alias_idx;
1739 }
1740 
1741 //------------------------------can_alias--------------------------------------
1742 // True if any values of the given address type are in the given alias category.
1743 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1744   if (alias_idx == AliasIdxTop)         return false; // the empty category
1745   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1746   // Known instance doesn't alias with bottom memory
1747   if (alias_idx == AliasIdxBot)         return !adr_type->is_known_instance();                   // the universal category
1748   if (adr_type->base() == Type::AnyPtr) return !C->get_adr_type(alias_idx)->is_known_instance(); // TypePtr::BOTTOM or its twins
1749 
1750   // the only remaining possible overlap is identity
1751   int adr_idx = get_alias_index(adr_type);
1752   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1753   return adr_idx == alias_idx;
1754 }
1755 
1756 //---------------------cleanup_loop_predicates-----------------------
1757 // Remove the opaque nodes that protect the predicates so that all unused
1758 // checks and uncommon_traps will be eliminated from the ideal graph
1759 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1760   if (predicate_count()==0) return;
1761   for (int i = predicate_count(); i > 0; i--) {
1762     Node * n = predicate_opaque1_node(i-1);
1763     assert(n->Opcode() == Op_Opaque1, "must be");
1764     igvn.replace_node(n, n->in(1));
1765   }
1766   assert(predicate_count()==0, "should be clean!");
1767 }
1768 
1769 void Compile::record_for_post_loop_opts_igvn(Node* n) {
1770   if (!n->for_post_loop_opts_igvn()) {
1771     assert(!_for_post_loop_igvn.contains(n), "duplicate");
1772     n->add_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1773     _for_post_loop_igvn.append(n);
1774   }
1775 }
1776 
1777 void Compile::remove_from_post_loop_opts_igvn(Node* n) {
1778   n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1779   _for_post_loop_igvn.remove(n);
1780 }
1781 
1782 void Compile::process_for_post_loop_opts_igvn(PhaseIterGVN& igvn) {
1783   // Verify that all previous optimizations produced a valid graph
1784   // at least to this point, even if no loop optimizations were done.
1785   PhaseIdealLoop::verify(igvn);
1786 
1787   C->set_post_loop_opts_phase(); // no more loop opts allowed
1788 
1789   assert(!C->major_progress(), "not cleared");
1790 
1791   if (_for_post_loop_igvn.length() > 0) {
1792     while (_for_post_loop_igvn.length() > 0) {
1793       Node* n = _for_post_loop_igvn.pop();
1794       n->remove_flag(Node::NodeFlags::Flag_for_post_loop_opts_igvn);
1795       igvn._worklist.push(n);
1796     }
1797     igvn.optimize();
1798     assert(_for_post_loop_igvn.length() == 0, "no more delayed nodes allowed");
1799 
1800     // Sometimes IGVN sets major progress (e.g., when processing loop nodes).
1801     if (C->major_progress()) {
1802       C->clear_major_progress(); // ensure that major progress is now clear
1803     }
1804   }
1805 }
1806 
1807 // StringOpts and late inlining of string methods
1808 void Compile::inline_string_calls(bool parse_time) {
1809   {
1810     // remove useless nodes to make the usage analysis simpler
1811     ResourceMark rm;
1812     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1813   }
1814 
1815   {
1816     ResourceMark rm;
1817     print_method(PHASE_BEFORE_STRINGOPTS, 3);
1818     PhaseStringOpts pso(initial_gvn(), for_igvn());
1819     print_method(PHASE_AFTER_STRINGOPTS, 3);
1820   }
1821 
1822   // now inline anything that we skipped the first time around
1823   if (!parse_time) {
1824     _late_inlines_pos = _late_inlines.length();
1825   }
1826 
1827   while (_string_late_inlines.length() > 0) {
1828     CallGenerator* cg = _string_late_inlines.pop();
1829     cg->do_late_inline();
1830     if (failing())  return;
1831   }
1832   _string_late_inlines.trunc_to(0);
1833 }
1834 
1835 // Late inlining of boxing methods
1836 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1837   if (_boxing_late_inlines.length() > 0) {
1838     assert(has_boxed_value(), "inconsistent");
1839 
1840     PhaseGVN* gvn = initial_gvn();
1841     set_inlining_incrementally(true);
1842 
1843     assert( igvn._worklist.size() == 0, "should be done with igvn" );
1844     for_igvn()->clear();
1845     gvn->replace_with(&igvn);
1846 
1847     _late_inlines_pos = _late_inlines.length();
1848 
1849     while (_boxing_late_inlines.length() > 0) {
1850       CallGenerator* cg = _boxing_late_inlines.pop();
1851       cg->do_late_inline();
1852       if (failing())  return;
1853     }
1854     _boxing_late_inlines.trunc_to(0);
1855 
1856     inline_incrementally_cleanup(igvn);
1857 
1858     set_inlining_incrementally(false);
1859   }
1860 }
1861 
1862 bool Compile::inline_incrementally_one() {
1863   assert(IncrementalInline, "incremental inlining should be on");
1864 
1865   TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
1866 
1867   set_inlining_progress(false);
1868   set_do_cleanup(false);
1869 
1870   for (int i = 0; i < _late_inlines.length(); i++) {
1871     _late_inlines_pos = i+1;
1872     CallGenerator* cg = _late_inlines.at(i);
1873     bool does_dispatch = cg->is_virtual_late_inline() || cg->is_mh_late_inline();
1874     if (inlining_incrementally() || does_dispatch) { // a call can be either inlined or strength-reduced to a direct call
1875       cg->do_late_inline();
1876       assert(_late_inlines.at(i) == cg, "no insertions before current position allowed");
1877       if (failing()) {
1878         return false;
1879       } else if (inlining_progress()) {
1880         _late_inlines_pos = i+1; // restore the position in case new elements were inserted
1881         print_method(PHASE_INCREMENTAL_INLINE_STEP, cg->call_node(), 3);
1882         break; // process one call site at a time
1883       }
1884     } else {
1885       // Ignore late inline direct calls when inlining is not allowed.
1886       // They are left in the late inline list when node budget is exhausted until the list is fully drained.
1887     }
1888   }
1889   // Remove processed elements.
1890   _late_inlines.remove_till(_late_inlines_pos);
1891   _late_inlines_pos = 0;
1892 
1893   assert(inlining_progress() || _late_inlines.length() == 0, "no progress");
1894 
1895   bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
1896 
1897   set_inlining_progress(false);
1898   set_do_cleanup(false);
1899 
1900   bool force_cleanup = directive()->IncrementalInlineForceCleanupOption;
1901   return (_late_inlines.length() > 0) && !needs_cleanup && !force_cleanup;
1902 }
1903 
1904 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
1905   {
1906     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
1907     ResourceMark rm;
1908     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1909   }
1910   {
1911     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
1912     igvn = PhaseIterGVN(initial_gvn());
1913     igvn.optimize();
1914   }
1915   print_method(PHASE_INCREMENTAL_INLINE_CLEANUP, 3);
1916 }
1917 
1918 // Perform incremental inlining until bound on number of live nodes is reached
1919 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1920   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
1921 
1922   set_inlining_incrementally(true);
1923   uint low_live_nodes = 0;
1924 
1925   while (_late_inlines.length() > 0) {
1926     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1927       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1928         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
1929         // PhaseIdealLoop is expensive so we only try it once we are
1930         // out of live nodes and we only try it again if the previous
1931         // helped got the number of nodes down significantly
1932         PhaseIdealLoop::optimize(igvn, LoopOptsNone);
1933         if (failing())  return;
1934         low_live_nodes = live_nodes();
1935         _major_progress = true;
1936       }
1937 
1938       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1939         bool do_print_inlining = print_inlining() || print_intrinsics();
1940         if (do_print_inlining || log() != NULL) {
1941           // Print inlining message for candidates that we couldn't inline for lack of space.
1942           for (int i = 0; i < _late_inlines.length(); i++) {
1943             CallGenerator* cg = _late_inlines.at(i);
1944             const char* msg = "live nodes > LiveNodeCountInliningCutoff";
1945             if (do_print_inlining) {
1946               cg->print_inlining_late(msg);
1947             }
1948             log_late_inline_failure(cg, msg);
1949           }
1950         }
1951         break; // finish
1952       }
1953     }
1954 
1955     for_igvn()->clear();
1956     initial_gvn()->replace_with(&igvn);
1957 
1958     while (inline_incrementally_one()) {
1959       assert(!failing(), "inconsistent");
1960     }
1961     if (failing())  return;
1962 
1963     inline_incrementally_cleanup(igvn);
1964 
1965     print_method(PHASE_INCREMENTAL_INLINE_STEP, 3);
1966 
1967     if (failing())  return;
1968 
1969     if (_late_inlines.length() == 0) {
1970       break; // no more progress
1971     }
1972   }
1973   assert( igvn._worklist.size() == 0, "should be done with igvn" );
1974 
1975   if (_string_late_inlines.length() > 0) {
1976     assert(has_stringbuilder(), "inconsistent");
1977     for_igvn()->clear();
1978     initial_gvn()->replace_with(&igvn);
1979 
1980     inline_string_calls(false);
1981 
1982     if (failing())  return;
1983 
1984     inline_incrementally_cleanup(igvn);
1985   }
1986 
1987   set_inlining_incrementally(false);
1988 }
1989 
1990 void Compile::process_late_inline_calls_no_inline(PhaseIterGVN& igvn) {
1991   // "inlining_incrementally() == false" is used to signal that no inlining is allowed
1992   // (see LateInlineVirtualCallGenerator::do_late_inline_check() for details).
1993   // Tracking and verification of modified nodes is disabled by setting "_modified_nodes == NULL"
1994   // as if "inlining_incrementally() == true" were set.
1995   assert(inlining_incrementally() == false, "not allowed");
1996   assert(_modified_nodes == NULL, "not allowed");
1997   assert(_late_inlines.length() > 0, "sanity");
1998 
1999   while (_late_inlines.length() > 0) {
2000     for_igvn()->clear();
2001     initial_gvn()->replace_with(&igvn);
2002 
2003     while (inline_incrementally_one()) {
2004       assert(!failing(), "inconsistent");
2005     }
2006     if (failing())  return;
2007 
2008     inline_incrementally_cleanup(igvn);
2009   }
2010 }
2011 
2012 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2013   if (_loop_opts_cnt > 0) {
2014     debug_only( int cnt = 0; );
2015     while (major_progress() && (_loop_opts_cnt > 0)) {
2016       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2017       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2018       PhaseIdealLoop::optimize(igvn, mode);
2019       _loop_opts_cnt--;
2020       if (failing())  return false;
2021       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2022     }
2023   }
2024   return true;
2025 }
2026 
2027 // Remove edges from "root" to each SafePoint at a backward branch.
2028 // They were inserted during parsing (see add_safepoint()) to make
2029 // infinite loops without calls or exceptions visible to root, i.e.,
2030 // useful.
2031 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2032   Node *r = root();
2033   if (r != NULL) {
2034     for (uint i = r->req(); i < r->len(); ++i) {
2035       Node *n = r->in(i);
2036       if (n != NULL && n->is_SafePoint()) {
2037         r->rm_prec(i);
2038         if (n->outcnt() == 0) {
2039           igvn.remove_dead_node(n);
2040         }
2041         --i;
2042       }
2043     }
2044     // Parsing may have added top inputs to the root node (Path
2045     // leading to the Halt node proven dead). Make sure we get a
2046     // chance to clean them up.
2047     igvn._worklist.push(r);
2048     igvn.optimize();
2049   }
2050 }
2051 
2052 //------------------------------Optimize---------------------------------------
2053 // Given a graph, optimize it.
2054 void Compile::Optimize() {
2055   TracePhase tp("optimizer", &timers[_t_optimizer]);
2056 
2057 #ifndef PRODUCT
2058   if (env()->break_at_compile()) {
2059     BREAKPOINT;
2060   }
2061 
2062 #endif
2063 
2064   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2065 #ifdef ASSERT
2066   bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2067 #endif
2068 
2069   ResourceMark rm;
2070 
2071   print_inlining_reinit();
2072 
2073   NOT_PRODUCT( verify_graph_edges(); )
2074 
2075   print_method(PHASE_AFTER_PARSING);
2076 
2077  {
2078   // Iterative Global Value Numbering, including ideal transforms
2079   // Initialize IterGVN with types and values from parse-time GVN
2080   PhaseIterGVN igvn(initial_gvn());
2081 #ifdef ASSERT
2082   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2083 #endif
2084   {
2085     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2086     igvn.optimize();
2087   }
2088 
2089   if (failing())  return;
2090 
2091   print_method(PHASE_ITER_GVN1, 2);
2092 
2093   inline_incrementally(igvn);
2094 
2095   print_method(PHASE_INCREMENTAL_INLINE, 2);
2096 
2097   if (failing())  return;
2098 
2099   if (eliminate_boxing()) {
2100     // Inline valueOf() methods now.
2101     inline_boxing_calls(igvn);
2102 
2103     if (AlwaysIncrementalInline) {
2104       inline_incrementally(igvn);
2105     }
2106 
2107     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2108 
2109     if (failing())  return;
2110   }
2111 
2112   // Remove the speculative part of types and clean up the graph from
2113   // the extra CastPP nodes whose only purpose is to carry them. Do
2114   // that early so that optimizations are not disrupted by the extra
2115   // CastPP nodes.
2116   remove_speculative_types(igvn);
2117 
2118   // No more new expensive nodes will be added to the list from here
2119   // so keep only the actual candidates for optimizations.
2120   cleanup_expensive_nodes(igvn);
2121 
2122   assert(EnableVectorSupport || !has_vbox_nodes(), "sanity");
2123   if (EnableVectorSupport && has_vbox_nodes()) {
2124     TracePhase tp("", &timers[_t_vector]);
2125     PhaseVector pv(igvn);
2126     pv.optimize_vector_boxes();
2127 
2128     print_method(PHASE_ITER_GVN_AFTER_VECTOR, 2);
2129   }
2130   assert(!has_vbox_nodes(), "sanity");
2131 
2132   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2133     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2134     initial_gvn()->replace_with(&igvn);
2135     Unique_Node_List* old_worklist = for_igvn();
2136     old_worklist->clear();
2137     Unique_Node_List new_worklist(C->comp_arena());
2138     {
2139       ResourceMark rm;
2140       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2141     }
2142     Unique_Node_List* save_for_igvn = for_igvn();
2143     set_for_igvn(&new_worklist);
2144     igvn = PhaseIterGVN(initial_gvn());
2145     igvn.optimize();
2146     set_for_igvn(old_worklist); // new_worklist is dead beyond this point
2147   }
2148 
2149   // Now that all inlining is over and no PhaseRemoveUseless will run, cut edge from root to loop
2150   // safepoints
2151   remove_root_to_sfpts_edges(igvn);
2152 
2153   // Perform escape analysis
2154   if (do_escape_analysis() && ConnectionGraph::has_candidates(this)) {
2155     if (has_loops()) {
2156       // Cleanup graph (remove dead nodes).
2157       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2158       PhaseIdealLoop::optimize(igvn, LoopOptsMaxUnroll);
2159       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2160       if (failing())  return;
2161     }
2162     ConnectionGraph::do_analysis(this, &igvn);
2163 
2164     if (failing())  return;
2165 
2166     // Optimize out fields loads from scalar replaceable allocations.
2167     igvn.optimize();
2168     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2169 
2170     if (failing())  return;
2171 
2172     if (congraph() != NULL && macro_count() > 0) {
2173       TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2174       PhaseMacroExpand mexp(igvn);
2175       mexp.eliminate_macro_nodes();
2176       igvn.set_delay_transform(false);
2177 
2178       igvn.optimize();
2179       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2180 
2181       if (failing())  return;
2182     }
2183   }
2184 
2185   // Loop transforms on the ideal graph.  Range Check Elimination,
2186   // peeling, unrolling, etc.
2187 
2188   // Set loop opts counter
2189   if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2190     {
2191       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2192       PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2193       _loop_opts_cnt--;
2194       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2195       if (failing())  return;
2196     }
2197     // Loop opts pass if partial peeling occurred in previous pass
2198     if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2199       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2200       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2201       _loop_opts_cnt--;
2202       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2203       if (failing())  return;
2204     }
2205     // Loop opts pass for loop-unrolling before CCP
2206     if(major_progress() && (_loop_opts_cnt > 0)) {
2207       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2208       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2209       _loop_opts_cnt--;
2210       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2211     }
2212     if (!failing()) {
2213       // Verify that last round of loop opts produced a valid graph
2214       PhaseIdealLoop::verify(igvn);
2215     }
2216   }
2217   if (failing())  return;
2218 
2219   // Conditional Constant Propagation;
2220   PhaseCCP ccp( &igvn );
2221   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2222   {
2223     TracePhase tp("ccp", &timers[_t_ccp]);
2224     ccp.do_transform();
2225   }
2226   print_method(PHASE_CCP1, 2);
2227 
2228   assert( true, "Break here to ccp.dump_old2new_map()");
2229 
2230   // Iterative Global Value Numbering, including ideal transforms
2231   {
2232     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2233     igvn = ccp;
2234     igvn.optimize();
2235   }
2236   print_method(PHASE_ITER_GVN2, 2);
2237 
2238   if (failing())  return;
2239 
2240   // Loop transforms on the ideal graph.  Range Check Elimination,
2241   // peeling, unrolling, etc.
2242   if (!optimize_loops(igvn, LoopOptsDefault)) {
2243     return;
2244   }
2245 
2246   if (failing())  return;
2247 
2248   C->clear_major_progress(); // ensure that major progress is now clear
2249 
2250   process_for_post_loop_opts_igvn(igvn);
2251 
2252 #ifdef ASSERT
2253   bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2254 #endif
2255 
2256   {
2257     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2258     PhaseMacroExpand  mex(igvn);
2259     if (mex.expand_macro_nodes()) {
2260       assert(failing(), "must bail out w/ explicit message");
2261       return;
2262     }
2263     print_method(PHASE_MACRO_EXPANSION, 2);
2264   }
2265 
2266   {
2267     TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2268     if (bs->expand_barriers(this, igvn)) {
2269       assert(failing(), "must bail out w/ explicit message");
2270       return;
2271     }
2272     print_method(PHASE_BARRIER_EXPANSION, 2);
2273   }
2274 
2275   if (C->max_vector_size() > 0) {
2276     C->optimize_logic_cones(igvn);
2277     igvn.optimize();
2278   }
2279 
2280   DEBUG_ONLY( _modified_nodes = NULL; )
2281 
2282   assert(igvn._worklist.size() == 0, "not empty");
2283 
2284   assert(_late_inlines.length() == 0 || IncrementalInlineMH || IncrementalInlineVirtual, "not empty");
2285 
2286   if (_late_inlines.length() > 0) {
2287     // More opportunities to optimize virtual and MH calls.
2288     // Though it's maybe too late to perform inlining, strength-reducing them to direct calls is still an option.
2289     process_late_inline_calls_no_inline(igvn);
2290   }
2291  } // (End scope of igvn; run destructor if necessary for asserts.)
2292 
2293  check_no_dead_use();
2294 
2295  process_print_inlining();
2296 
2297  // A method with only infinite loops has no edges entering loops from root
2298  {
2299    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2300    if (final_graph_reshaping()) {
2301      assert(failing(), "must bail out w/ explicit message");
2302      return;
2303    }
2304  }
2305 
2306  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2307  DEBUG_ONLY(set_phase_optimize_finished();)
2308 }
2309 
2310 #ifdef ASSERT
2311 void Compile::check_no_dead_use() const {
2312   ResourceMark rm;
2313   Unique_Node_List wq;
2314   wq.push(root());
2315   for (uint i = 0; i < wq.size(); ++i) {
2316     Node* n = wq.at(i);
2317     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
2318       Node* u = n->fast_out(j);
2319       if (u->outcnt() == 0 && !u->is_Con()) {
2320         u->dump();
2321         fatal("no reachable node should have no use");
2322       }
2323       wq.push(u);
2324     }
2325   }
2326 }
2327 #endif
2328 
2329 void Compile::inline_vector_reboxing_calls() {
2330   if (C->_vector_reboxing_late_inlines.length() > 0) {
2331     _late_inlines_pos = C->_late_inlines.length();
2332     while (_vector_reboxing_late_inlines.length() > 0) {
2333       CallGenerator* cg = _vector_reboxing_late_inlines.pop();
2334       cg->do_late_inline();
2335       if (failing())  return;
2336       print_method(PHASE_INLINE_VECTOR_REBOX, cg->call_node());
2337     }
2338     _vector_reboxing_late_inlines.trunc_to(0);
2339   }
2340 }
2341 
2342 bool Compile::has_vbox_nodes() {
2343   if (C->_vector_reboxing_late_inlines.length() > 0) {
2344     return true;
2345   }
2346   for (int macro_idx = C->macro_count() - 1; macro_idx >= 0; macro_idx--) {
2347     Node * n = C->macro_node(macro_idx);
2348     assert(n->is_macro(), "only macro nodes expected here");
2349     if (n->Opcode() == Op_VectorUnbox || n->Opcode() == Op_VectorBox || n->Opcode() == Op_VectorBoxAllocate) {
2350       return true;
2351     }
2352   }
2353   return false;
2354 }
2355 
2356 //---------------------------- Bitwise operation packing optimization ---------------------------
2357 
2358 static bool is_vector_unary_bitwise_op(Node* n) {
2359   return n->Opcode() == Op_XorV &&
2360          n->req() == 2 &&
2361          VectorNode::is_vector_bitwise_not_pattern(n);
2362 }
2363 
2364 static bool is_vector_binary_bitwise_op(Node* n) {
2365   switch (n->Opcode()) {
2366     case Op_AndV:
2367     case Op_OrV:
2368       return n->req() == 2;
2369 
2370     case Op_XorV:
2371       return !is_vector_unary_bitwise_op(n);
2372 
2373     default:
2374       return false;
2375   }
2376 }
2377 
2378 static bool is_vector_ternary_bitwise_op(Node* n) {
2379   return n->Opcode() == Op_MacroLogicV;
2380 }
2381 
2382 static bool is_vector_bitwise_op(Node* n) {
2383   return is_vector_unary_bitwise_op(n)  ||
2384          is_vector_binary_bitwise_op(n) ||
2385          is_vector_ternary_bitwise_op(n);
2386 }
2387 
2388 static bool is_vector_bitwise_cone_root(Node* n) {
2389   if (n->bottom_type()->isa_vectmask() || !is_vector_bitwise_op(n)) {
2390     return false;
2391   }
2392   for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2393     if (is_vector_bitwise_op(n->fast_out(i))) {
2394       return false;
2395     }
2396   }
2397   return true;
2398 }
2399 
2400 static uint collect_unique_inputs(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2401   uint cnt = 0;
2402   if (is_vector_bitwise_op(n)) {
2403     if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2404       for (uint i = 1; i < n->req(); i++) {
2405         Node* in = n->in(i);
2406         bool skip = VectorNode::is_all_ones_vector(in);
2407         if (!skip && !inputs.member(in)) {
2408           inputs.push(in);
2409           cnt++;
2410         }
2411       }
2412       assert(cnt <= 1, "not unary");
2413     } else {
2414       uint last_req = n->req();
2415       if (is_vector_ternary_bitwise_op(n)) {
2416         last_req = n->req() - 1; // skip last input
2417       }
2418       for (uint i = 1; i < last_req; i++) {
2419         Node* def = n->in(i);
2420         if (!inputs.member(def)) {
2421           inputs.push(def);
2422           cnt++;
2423         }
2424       }
2425     }
2426     partition.push(n);
2427   } else { // not a bitwise operations
2428     if (!inputs.member(n)) {
2429       inputs.push(n);
2430       cnt++;
2431     }
2432   }
2433   return cnt;
2434 }
2435 
2436 void Compile::collect_logic_cone_roots(Unique_Node_List& list) {
2437   Unique_Node_List useful_nodes;
2438   C->identify_useful_nodes(useful_nodes);
2439 
2440   for (uint i = 0; i < useful_nodes.size(); i++) {
2441     Node* n = useful_nodes.at(i);
2442     if (is_vector_bitwise_cone_root(n)) {
2443       list.push(n);
2444     }
2445   }
2446 }
2447 
2448 Node* Compile::xform_to_MacroLogicV(PhaseIterGVN& igvn,
2449                                     const TypeVect* vt,
2450                                     Unique_Node_List& partition,
2451                                     Unique_Node_List& inputs) {
2452   assert(partition.size() == 2 || partition.size() == 3, "not supported");
2453   assert(inputs.size()    == 2 || inputs.size()    == 3, "not supported");
2454   assert(Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type()), "not supported");
2455 
2456   Node* in1 = inputs.at(0);
2457   Node* in2 = inputs.at(1);
2458   Node* in3 = (inputs.size() == 3 ? inputs.at(2) : in2);
2459 
2460   uint func = compute_truth_table(partition, inputs);
2461   return igvn.transform(MacroLogicVNode::make(igvn, in3, in2, in1, func, vt));
2462 }
2463 
2464 static uint extract_bit(uint func, uint pos) {
2465   return (func & (1 << pos)) >> pos;
2466 }
2467 
2468 //
2469 //  A macro logic node represents a truth table. It has 4 inputs,
2470 //  First three inputs corresponds to 3 columns of a truth table
2471 //  and fourth input captures the logic function.
2472 //
2473 //  eg.  fn = (in1 AND in2) OR in3;
2474 //
2475 //      MacroNode(in1,in2,in3,fn)
2476 //
2477 //  -----------------
2478 //  in1 in2 in3  fn
2479 //  -----------------
2480 //  0    0   0    0
2481 //  0    0   1    1
2482 //  0    1   0    0
2483 //  0    1   1    1
2484 //  1    0   0    0
2485 //  1    0   1    1
2486 //  1    1   0    1
2487 //  1    1   1    1
2488 //
2489 
2490 uint Compile::eval_macro_logic_op(uint func, uint in1 , uint in2, uint in3) {
2491   int res = 0;
2492   for (int i = 0; i < 8; i++) {
2493     int bit1 = extract_bit(in1, i);
2494     int bit2 = extract_bit(in2, i);
2495     int bit3 = extract_bit(in3, i);
2496 
2497     int func_bit_pos = (bit1 << 2 | bit2 << 1 | bit3);
2498     int func_bit = extract_bit(func, func_bit_pos);
2499 
2500     res |= func_bit << i;
2501   }
2502   return res;
2503 }
2504 
2505 static uint eval_operand(Node* n, ResourceHashtable<Node*,uint>& eval_map) {
2506   assert(n != NULL, "");
2507   assert(eval_map.contains(n), "absent");
2508   return *(eval_map.get(n));
2509 }
2510 
2511 static void eval_operands(Node* n,
2512                           uint& func1, uint& func2, uint& func3,
2513                           ResourceHashtable<Node*,uint>& eval_map) {
2514   assert(is_vector_bitwise_op(n), "");
2515 
2516   if (is_vector_unary_bitwise_op(n)) {
2517     Node* opnd = n->in(1);
2518     if (VectorNode::is_vector_bitwise_not_pattern(n) && VectorNode::is_all_ones_vector(opnd)) {
2519       opnd = n->in(2);
2520     }
2521     func1 = eval_operand(opnd, eval_map);
2522   } else if (is_vector_binary_bitwise_op(n)) {
2523     func1 = eval_operand(n->in(1), eval_map);
2524     func2 = eval_operand(n->in(2), eval_map);
2525   } else {
2526     assert(is_vector_ternary_bitwise_op(n), "unknown operation");
2527     func1 = eval_operand(n->in(1), eval_map);
2528     func2 = eval_operand(n->in(2), eval_map);
2529     func3 = eval_operand(n->in(3), eval_map);
2530   }
2531 }
2532 
2533 uint Compile::compute_truth_table(Unique_Node_List& partition, Unique_Node_List& inputs) {
2534   assert(inputs.size() <= 3, "sanity");
2535   ResourceMark rm;
2536   uint res = 0;
2537   ResourceHashtable<Node*,uint> eval_map;
2538 
2539   // Populate precomputed functions for inputs.
2540   // Each input corresponds to one column of 3 input truth-table.
2541   uint input_funcs[] = { 0xAA,   // (_, _, a) -> a
2542                          0xCC,   // (_, b, _) -> b
2543                          0xF0 }; // (c, _, _) -> c
2544   for (uint i = 0; i < inputs.size(); i++) {
2545     eval_map.put(inputs.at(i), input_funcs[i]);
2546   }
2547 
2548   for (uint i = 0; i < partition.size(); i++) {
2549     Node* n = partition.at(i);
2550 
2551     uint func1 = 0, func2 = 0, func3 = 0;
2552     eval_operands(n, func1, func2, func3, eval_map);
2553 
2554     switch (n->Opcode()) {
2555       case Op_OrV:
2556         assert(func3 == 0, "not binary");
2557         res = func1 | func2;
2558         break;
2559       case Op_AndV:
2560         assert(func3 == 0, "not binary");
2561         res = func1 & func2;
2562         break;
2563       case Op_XorV:
2564         if (VectorNode::is_vector_bitwise_not_pattern(n)) {
2565           assert(func2 == 0 && func3 == 0, "not unary");
2566           res = (~func1) & 0xFF;
2567         } else {
2568           assert(func3 == 0, "not binary");
2569           res = func1 ^ func2;
2570         }
2571         break;
2572       case Op_MacroLogicV:
2573         // Ordering of inputs may change during evaluation of sub-tree
2574         // containing MacroLogic node as a child node, thus a re-evaluation
2575         // makes sure that function is evaluated in context of current
2576         // inputs.
2577         res = eval_macro_logic_op(n->in(4)->get_int(), func1, func2, func3);
2578         break;
2579 
2580       default: assert(false, "not supported: %s", n->Name());
2581     }
2582     assert(res <= 0xFF, "invalid");
2583     eval_map.put(n, res);
2584   }
2585   return res;
2586 }
2587 
2588 bool Compile::compute_logic_cone(Node* n, Unique_Node_List& partition, Unique_Node_List& inputs) {
2589   assert(partition.size() == 0, "not empty");
2590   assert(inputs.size() == 0, "not empty");
2591   if (is_vector_ternary_bitwise_op(n)) {
2592     return false;
2593   }
2594 
2595   bool is_unary_op = is_vector_unary_bitwise_op(n);
2596   if (is_unary_op) {
2597     assert(collect_unique_inputs(n, partition, inputs) == 1, "not unary");
2598     return false; // too few inputs
2599   }
2600 
2601   assert(is_vector_binary_bitwise_op(n), "not binary");
2602   Node* in1 = n->in(1);
2603   Node* in2 = n->in(2);
2604 
2605   int in1_unique_inputs_cnt = collect_unique_inputs(in1, partition, inputs);
2606   int in2_unique_inputs_cnt = collect_unique_inputs(in2, partition, inputs);
2607   partition.push(n);
2608 
2609   // Too many inputs?
2610   if (inputs.size() > 3) {
2611     partition.clear();
2612     inputs.clear();
2613     { // Recompute in2 inputs
2614       Unique_Node_List not_used;
2615       in2_unique_inputs_cnt = collect_unique_inputs(in2, not_used, not_used);
2616     }
2617     // Pick the node with minimum number of inputs.
2618     if (in1_unique_inputs_cnt >= 3 && in2_unique_inputs_cnt >= 3) {
2619       return false; // still too many inputs
2620     }
2621     // Recompute partition & inputs.
2622     Node* child       = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in1 : in2);
2623     collect_unique_inputs(child, partition, inputs);
2624 
2625     Node* other_input = (in1_unique_inputs_cnt < in2_unique_inputs_cnt ? in2 : in1);
2626     inputs.push(other_input);
2627 
2628     partition.push(n);
2629   }
2630 
2631   return (partition.size() == 2 || partition.size() == 3) &&
2632          (inputs.size()    == 2 || inputs.size()    == 3);
2633 }
2634 
2635 
2636 void Compile::process_logic_cone_root(PhaseIterGVN &igvn, Node *n, VectorSet &visited) {
2637   assert(is_vector_bitwise_op(n), "not a root");
2638 
2639   visited.set(n->_idx);
2640 
2641   // 1) Do a DFS walk over the logic cone.
2642   for (uint i = 1; i < n->req(); i++) {
2643     Node* in = n->in(i);
2644     if (!visited.test(in->_idx) && is_vector_bitwise_op(in)) {
2645       process_logic_cone_root(igvn, in, visited);
2646     }
2647   }
2648 
2649   // 2) Bottom up traversal: Merge node[s] with
2650   // the parent to form macro logic node.
2651   Unique_Node_List partition;
2652   Unique_Node_List inputs;
2653   if (compute_logic_cone(n, partition, inputs)) {
2654     const TypeVect* vt = n->bottom_type()->is_vect();
2655     Node* macro_logic = xform_to_MacroLogicV(igvn, vt, partition, inputs);
2656     igvn.replace_node(n, macro_logic);
2657   }
2658 }
2659 
2660 void Compile::optimize_logic_cones(PhaseIterGVN &igvn) {
2661   ResourceMark rm;
2662   if (Matcher::match_rule_supported(Op_MacroLogicV)) {
2663     Unique_Node_List list;
2664     collect_logic_cone_roots(list);
2665 
2666     while (list.size() > 0) {
2667       Node* n = list.pop();
2668       const TypeVect* vt = n->bottom_type()->is_vect();
2669       bool supported = Matcher::match_rule_supported_vector(Op_MacroLogicV, vt->length(), vt->element_basic_type());
2670       if (supported) {
2671         VectorSet visited(comp_arena());
2672         process_logic_cone_root(igvn, n, visited);
2673       }
2674     }
2675   }
2676 }
2677 
2678 //------------------------------Code_Gen---------------------------------------
2679 // Given a graph, generate code for it
2680 void Compile::Code_Gen() {
2681   if (failing()) {
2682     return;
2683   }
2684 
2685   // Perform instruction selection.  You might think we could reclaim Matcher
2686   // memory PDQ, but actually the Matcher is used in generating spill code.
2687   // Internals of the Matcher (including some VectorSets) must remain live
2688   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2689   // set a bit in reclaimed memory.
2690 
2691   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2692   // nodes.  Mapping is only valid at the root of each matched subtree.
2693   NOT_PRODUCT( verify_graph_edges(); )
2694 
2695   Matcher matcher;
2696   _matcher = &matcher;
2697   {
2698     TracePhase tp("matcher", &timers[_t_matcher]);
2699     matcher.match();
2700     if (failing()) {
2701       return;
2702     }
2703   }
2704   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2705   // nodes.  Mapping is only valid at the root of each matched subtree.
2706   NOT_PRODUCT( verify_graph_edges(); )
2707 
2708   // If you have too many nodes, or if matching has failed, bail out
2709   check_node_count(0, "out of nodes matching instructions");
2710   if (failing()) {
2711     return;
2712   }
2713 
2714   print_method(PHASE_MATCHING, 2);
2715 
2716   // Build a proper-looking CFG
2717   PhaseCFG cfg(node_arena(), root(), matcher);
2718   _cfg = &cfg;
2719   {
2720     TracePhase tp("scheduler", &timers[_t_scheduler]);
2721     bool success = cfg.do_global_code_motion();
2722     if (!success) {
2723       return;
2724     }
2725 
2726     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2727     NOT_PRODUCT( verify_graph_edges(); )
2728     cfg.verify();
2729   }
2730 
2731   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2732   _regalloc = &regalloc;
2733   {
2734     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2735     // Perform register allocation.  After Chaitin, use-def chains are
2736     // no longer accurate (at spill code) and so must be ignored.
2737     // Node->LRG->reg mappings are still accurate.
2738     _regalloc->Register_Allocate();
2739 
2740     // Bail out if the allocator builds too many nodes
2741     if (failing()) {
2742       return;
2743     }
2744   }
2745 
2746   // Prior to register allocation we kept empty basic blocks in case the
2747   // the allocator needed a place to spill.  After register allocation we
2748   // are not adding any new instructions.  If any basic block is empty, we
2749   // can now safely remove it.
2750   {
2751     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2752     cfg.remove_empty_blocks();
2753     if (do_freq_based_layout()) {
2754       PhaseBlockLayout layout(cfg);
2755     } else {
2756       cfg.set_loop_alignment();
2757     }
2758     cfg.fixup_flow();
2759   }
2760 
2761   // Apply peephole optimizations
2762   if( OptoPeephole ) {
2763     TracePhase tp("peephole", &timers[_t_peephole]);
2764     PhasePeephole peep( _regalloc, cfg);
2765     peep.do_transform();
2766   }
2767 
2768   // Do late expand if CPU requires this.
2769   if (Matcher::require_postalloc_expand) {
2770     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2771     cfg.postalloc_expand(_regalloc);
2772   }
2773 
2774   // Convert Nodes to instruction bits in a buffer
2775   {
2776     TracePhase tp("output", &timers[_t_output]);
2777     PhaseOutput output;
2778     output.Output();
2779     if (failing())  return;
2780     output.install();
2781   }
2782 
2783   print_method(PHASE_FINAL_CODE);
2784 
2785   // He's dead, Jim.
2786   _cfg     = (PhaseCFG*)((intptr_t)0xdeadbeef);
2787   _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2788 }
2789 
2790 //------------------------------Final_Reshape_Counts---------------------------
2791 // This class defines counters to help identify when a method
2792 // may/must be executed using hardware with only 24-bit precision.
2793 struct Final_Reshape_Counts : public StackObj {
2794   int  _call_count;             // count non-inlined 'common' calls
2795   int  _float_count;            // count float ops requiring 24-bit precision
2796   int  _double_count;           // count double ops requiring more precision
2797   int  _java_call_count;        // count non-inlined 'java' calls
2798   int  _inner_loop_count;       // count loops which need alignment
2799   VectorSet _visited;           // Visitation flags
2800   Node_List _tests;             // Set of IfNodes & PCTableNodes
2801 
2802   Final_Reshape_Counts() :
2803     _call_count(0), _float_count(0), _double_count(0),
2804     _java_call_count(0), _inner_loop_count(0) { }
2805 
2806   void inc_call_count  () { _call_count  ++; }
2807   void inc_float_count () { _float_count ++; }
2808   void inc_double_count() { _double_count++; }
2809   void inc_java_call_count() { _java_call_count++; }
2810   void inc_inner_loop_count() { _inner_loop_count++; }
2811 
2812   int  get_call_count  () const { return _call_count  ; }
2813   int  get_float_count () const { return _float_count ; }
2814   int  get_double_count() const { return _double_count; }
2815   int  get_java_call_count() const { return _java_call_count; }
2816   int  get_inner_loop_count() const { return _inner_loop_count; }
2817 };
2818 
2819 // Eliminate trivially redundant StoreCMs and accumulate their
2820 // precedence edges.
2821 void Compile::eliminate_redundant_card_marks(Node* n) {
2822   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2823   if (n->in(MemNode::Address)->outcnt() > 1) {
2824     // There are multiple users of the same address so it might be
2825     // possible to eliminate some of the StoreCMs
2826     Node* mem = n->in(MemNode::Memory);
2827     Node* adr = n->in(MemNode::Address);
2828     Node* val = n->in(MemNode::ValueIn);
2829     Node* prev = n;
2830     bool done = false;
2831     // Walk the chain of StoreCMs eliminating ones that match.  As
2832     // long as it's a chain of single users then the optimization is
2833     // safe.  Eliminating partially redundant StoreCMs would require
2834     // cloning copies down the other paths.
2835     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2836       if (adr == mem->in(MemNode::Address) &&
2837           val == mem->in(MemNode::ValueIn)) {
2838         // redundant StoreCM
2839         if (mem->req() > MemNode::OopStore) {
2840           // Hasn't been processed by this code yet.
2841           n->add_prec(mem->in(MemNode::OopStore));
2842         } else {
2843           // Already converted to precedence edge
2844           for (uint i = mem->req(); i < mem->len(); i++) {
2845             // Accumulate any precedence edges
2846             if (mem->in(i) != NULL) {
2847               n->add_prec(mem->in(i));
2848             }
2849           }
2850           // Everything above this point has been processed.
2851           done = true;
2852         }
2853         // Eliminate the previous StoreCM
2854         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2855         assert(mem->outcnt() == 0, "should be dead");
2856         mem->disconnect_inputs(this);
2857       } else {
2858         prev = mem;
2859       }
2860       mem = prev->in(MemNode::Memory);
2861     }
2862   }
2863 }
2864 
2865 //------------------------------final_graph_reshaping_impl----------------------
2866 // Implement items 1-5 from final_graph_reshaping below.
2867 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2868 
2869   if ( n->outcnt() == 0 ) return; // dead node
2870   uint nop = n->Opcode();
2871 
2872   // Check for 2-input instruction with "last use" on right input.
2873   // Swap to left input.  Implements item (2).
2874   if( n->req() == 3 &&          // two-input instruction
2875       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2876       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2877       n->in(2)->outcnt() == 1 &&// right use IS a last use
2878       !n->in(2)->is_Con() ) {   // right use is not a constant
2879     // Check for commutative opcode
2880     switch( nop ) {
2881     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2882     case Op_MaxI:  case Op_MaxL:  case Op_MaxF:  case Op_MaxD:
2883     case Op_MinI:  case Op_MinL:  case Op_MinF:  case Op_MinD:
2884     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2885     case Op_AndL:  case Op_XorL:  case Op_OrL:
2886     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2887       // Move "last use" input to left by swapping inputs
2888       n->swap_edges(1, 2);
2889       break;
2890     }
2891     default:
2892       break;
2893     }
2894   }
2895 
2896 #ifdef ASSERT
2897   if( n->is_Mem() ) {
2898     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2899     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2900             // oop will be recorded in oop map if load crosses safepoint
2901             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2902                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2903             "raw memory operations should have control edge");
2904   }
2905   if (n->is_MemBar()) {
2906     MemBarNode* mb = n->as_MemBar();
2907     if (mb->trailing_store() || mb->trailing_load_store()) {
2908       assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2909       Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2910       assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2911              (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2912     } else if (mb->leading()) {
2913       assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2914     }
2915   }
2916 #endif
2917   // Count FPU ops and common calls, implements item (3)
2918   bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2919   if (!gc_handled) {
2920     final_graph_reshaping_main_switch(n, frc, nop);
2921   }
2922 
2923   // Collect CFG split points
2924   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2925     frc._tests.push(n);
2926   }
2927 }
2928 
2929 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2930   switch( nop ) {
2931   // Count all float operations that may use FPU
2932   case Op_AddF:
2933   case Op_SubF:
2934   case Op_MulF:
2935   case Op_DivF:
2936   case Op_NegF:
2937   case Op_ModF:
2938   case Op_ConvI2F:
2939   case Op_ConF:
2940   case Op_CmpF:
2941   case Op_CmpF3:
2942   case Op_StoreF:
2943   case Op_LoadF:
2944   // case Op_ConvL2F: // longs are split into 32-bit halves
2945     frc.inc_float_count();
2946     break;
2947 
2948   case Op_ConvF2D:
2949   case Op_ConvD2F:
2950     frc.inc_float_count();
2951     frc.inc_double_count();
2952     break;
2953 
2954   // Count all double operations that may use FPU
2955   case Op_AddD:
2956   case Op_SubD:
2957   case Op_MulD:
2958   case Op_DivD:
2959   case Op_NegD:
2960   case Op_ModD:
2961   case Op_ConvI2D:
2962   case Op_ConvD2I:
2963   // case Op_ConvL2D: // handled by leaf call
2964   // case Op_ConvD2L: // handled by leaf call
2965   case Op_ConD:
2966   case Op_CmpD:
2967   case Op_CmpD3:
2968   case Op_StoreD:
2969   case Op_LoadD:
2970   case Op_LoadD_unaligned:
2971     frc.inc_double_count();
2972     break;
2973   case Op_Opaque1:              // Remove Opaque Nodes before matching
2974   case Op_Opaque2:              // Remove Opaque Nodes before matching
2975   case Op_Opaque3:
2976     n->subsume_by(n->in(1), this);
2977     break;
2978   case Op_CallStaticJava:
2979   case Op_CallJava:
2980   case Op_CallDynamicJava:
2981     frc.inc_java_call_count(); // Count java call site;
2982   case Op_CallRuntime:
2983   case Op_CallLeaf:
2984   case Op_CallLeafVector:
2985   case Op_CallNative:
2986   case Op_CallLeafNoFP: {
2987     assert (n->is_Call(), "");
2988     CallNode *call = n->as_Call();
2989     // Count call sites where the FP mode bit would have to be flipped.
2990     // Do not count uncommon runtime calls:
2991     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2992     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2993     if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2994       frc.inc_call_count();   // Count the call site
2995     } else {                  // See if uncommon argument is shared
2996       Node *n = call->in(TypeFunc::Parms);
2997       int nop = n->Opcode();
2998       // Clone shared simple arguments to uncommon calls, item (1).
2999       if (n->outcnt() > 1 &&
3000           !n->is_Proj() &&
3001           nop != Op_CreateEx &&
3002           nop != Op_CheckCastPP &&
3003           nop != Op_DecodeN &&
3004           nop != Op_DecodeNKlass &&
3005           !n->is_Mem() &&
3006           !n->is_Phi()) {
3007         Node *x = n->clone();
3008         call->set_req(TypeFunc::Parms, x);
3009       }
3010     }
3011     break;
3012   }
3013 
3014   case Op_StoreCM:
3015     {
3016       // Convert OopStore dependence into precedence edge
3017       Node* prec = n->in(MemNode::OopStore);
3018       n->del_req(MemNode::OopStore);
3019       n->add_prec(prec);
3020       eliminate_redundant_card_marks(n);
3021     }
3022 
3023     // fall through
3024 
3025   case Op_StoreB:
3026   case Op_StoreC:
3027   case Op_StorePConditional:
3028   case Op_StoreI:
3029   case Op_StoreL:
3030   case Op_StoreIConditional:
3031   case Op_StoreLConditional:
3032   case Op_CompareAndSwapB:
3033   case Op_CompareAndSwapS:
3034   case Op_CompareAndSwapI:
3035   case Op_CompareAndSwapL:
3036   case Op_CompareAndSwapP:
3037   case Op_CompareAndSwapN:
3038   case Op_WeakCompareAndSwapB:
3039   case Op_WeakCompareAndSwapS:
3040   case Op_WeakCompareAndSwapI:
3041   case Op_WeakCompareAndSwapL:
3042   case Op_WeakCompareAndSwapP:
3043   case Op_WeakCompareAndSwapN:
3044   case Op_CompareAndExchangeB:
3045   case Op_CompareAndExchangeS:
3046   case Op_CompareAndExchangeI:
3047   case Op_CompareAndExchangeL:
3048   case Op_CompareAndExchangeP:
3049   case Op_CompareAndExchangeN:
3050   case Op_GetAndAddS:
3051   case Op_GetAndAddB:
3052   case Op_GetAndAddI:
3053   case Op_GetAndAddL:
3054   case Op_GetAndSetS:
3055   case Op_GetAndSetB:
3056   case Op_GetAndSetI:
3057   case Op_GetAndSetL:
3058   case Op_GetAndSetP:
3059   case Op_GetAndSetN:
3060   case Op_StoreP:
3061   case Op_StoreN:
3062   case Op_StoreNKlass:
3063   case Op_LoadB:
3064   case Op_LoadUB:
3065   case Op_LoadUS:
3066   case Op_LoadI:
3067   case Op_LoadKlass:
3068   case Op_LoadNKlass:
3069   case Op_LoadL:
3070   case Op_LoadL_unaligned:
3071   case Op_LoadPLocked:
3072   case Op_LoadP:
3073   case Op_LoadN:
3074   case Op_LoadRange:
3075   case Op_LoadS:
3076     break;
3077 
3078   case Op_AddP: {               // Assert sane base pointers
3079     Node *addp = n->in(AddPNode::Address);
3080     assert( !addp->is_AddP() ||
3081             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3082             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3083             "Base pointers must match (addp %u)", addp->_idx );
3084 #ifdef _LP64
3085     if ((UseCompressedOops || UseCompressedClassPointers) &&
3086         addp->Opcode() == Op_ConP &&
3087         addp == n->in(AddPNode::Base) &&
3088         n->in(AddPNode::Offset)->is_Con()) {
3089       // If the transformation of ConP to ConN+DecodeN is beneficial depends
3090       // on the platform and on the compressed oops mode.
3091       // Use addressing with narrow klass to load with offset on x86.
3092       // Some platforms can use the constant pool to load ConP.
3093       // Do this transformation here since IGVN will convert ConN back to ConP.
3094       const Type* t = addp->bottom_type();
3095       bool is_oop   = t->isa_oopptr() != NULL;
3096       bool is_klass = t->isa_klassptr() != NULL;
3097 
3098       if ((is_oop   && Matcher::const_oop_prefer_decode()  ) ||
3099           (is_klass && Matcher::const_klass_prefer_decode())) {
3100         Node* nn = NULL;
3101 
3102         int op = is_oop ? Op_ConN : Op_ConNKlass;
3103 
3104         // Look for existing ConN node of the same exact type.
3105         Node* r  = root();
3106         uint cnt = r->outcnt();
3107         for (uint i = 0; i < cnt; i++) {
3108           Node* m = r->raw_out(i);
3109           if (m!= NULL && m->Opcode() == op &&
3110               m->bottom_type()->make_ptr() == t) {
3111             nn = m;
3112             break;
3113           }
3114         }
3115         if (nn != NULL) {
3116           // Decode a narrow oop to match address
3117           // [R12 + narrow_oop_reg<<3 + offset]
3118           if (is_oop) {
3119             nn = new DecodeNNode(nn, t);
3120           } else {
3121             nn = new DecodeNKlassNode(nn, t);
3122           }
3123           // Check for succeeding AddP which uses the same Base.
3124           // Otherwise we will run into the assertion above when visiting that guy.
3125           for (uint i = 0; i < n->outcnt(); ++i) {
3126             Node *out_i = n->raw_out(i);
3127             if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3128               out_i->set_req(AddPNode::Base, nn);
3129 #ifdef ASSERT
3130               for (uint j = 0; j < out_i->outcnt(); ++j) {
3131                 Node *out_j = out_i->raw_out(j);
3132                 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3133                        "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3134               }
3135 #endif
3136             }
3137           }
3138           n->set_req(AddPNode::Base, nn);
3139           n->set_req(AddPNode::Address, nn);
3140           if (addp->outcnt() == 0) {
3141             addp->disconnect_inputs(this);
3142           }
3143         }
3144       }
3145     }
3146 #endif
3147     break;
3148   }
3149 
3150   case Op_CastPP: {
3151     // Remove CastPP nodes to gain more freedom during scheduling but
3152     // keep the dependency they encode as control or precedence edges
3153     // (if control is set already) on memory operations. Some CastPP
3154     // nodes don't have a control (don't carry a dependency): skip
3155     // those.
3156     if (n->in(0) != NULL) {
3157       ResourceMark rm;
3158       Unique_Node_List wq;
3159       wq.push(n);
3160       for (uint next = 0; next < wq.size(); ++next) {
3161         Node *m = wq.at(next);
3162         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3163           Node* use = m->fast_out(i);
3164           if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3165             use->ensure_control_or_add_prec(n->in(0));
3166           } else {
3167             switch(use->Opcode()) {
3168             case Op_AddP:
3169             case Op_DecodeN:
3170             case Op_DecodeNKlass:
3171             case Op_CheckCastPP:
3172             case Op_CastPP:
3173               wq.push(use);
3174               break;
3175             }
3176           }
3177         }
3178       }
3179     }
3180     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3181     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3182       Node* in1 = n->in(1);
3183       const Type* t = n->bottom_type();
3184       Node* new_in1 = in1->clone();
3185       new_in1->as_DecodeN()->set_type(t);
3186 
3187       if (!Matcher::narrow_oop_use_complex_address()) {
3188         //
3189         // x86, ARM and friends can handle 2 adds in addressing mode
3190         // and Matcher can fold a DecodeN node into address by using
3191         // a narrow oop directly and do implicit NULL check in address:
3192         //
3193         // [R12 + narrow_oop_reg<<3 + offset]
3194         // NullCheck narrow_oop_reg
3195         //
3196         // On other platforms (Sparc) we have to keep new DecodeN node and
3197         // use it to do implicit NULL check in address:
3198         //
3199         // decode_not_null narrow_oop_reg, base_reg
3200         // [base_reg + offset]
3201         // NullCheck base_reg
3202         //
3203         // Pin the new DecodeN node to non-null path on these platform (Sparc)
3204         // to keep the information to which NULL check the new DecodeN node
3205         // corresponds to use it as value in implicit_null_check().
3206         //
3207         new_in1->set_req(0, n->in(0));
3208       }
3209 
3210       n->subsume_by(new_in1, this);
3211       if (in1->outcnt() == 0) {
3212         in1->disconnect_inputs(this);
3213       }
3214     } else {
3215       n->subsume_by(n->in(1), this);
3216       if (n->outcnt() == 0) {
3217         n->disconnect_inputs(this);
3218       }
3219     }
3220     break;
3221   }
3222 #ifdef _LP64
3223   case Op_CmpP:
3224     // Do this transformation here to preserve CmpPNode::sub() and
3225     // other TypePtr related Ideal optimizations (for example, ptr nullness).
3226     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3227       Node* in1 = n->in(1);
3228       Node* in2 = n->in(2);
3229       if (!in1->is_DecodeNarrowPtr()) {
3230         in2 = in1;
3231         in1 = n->in(2);
3232       }
3233       assert(in1->is_DecodeNarrowPtr(), "sanity");
3234 
3235       Node* new_in2 = NULL;
3236       if (in2->is_DecodeNarrowPtr()) {
3237         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3238         new_in2 = in2->in(1);
3239       } else if (in2->Opcode() == Op_ConP) {
3240         const Type* t = in2->bottom_type();
3241         if (t == TypePtr::NULL_PTR) {
3242           assert(in1->is_DecodeN(), "compare klass to null?");
3243           // Don't convert CmpP null check into CmpN if compressed
3244           // oops implicit null check is not generated.
3245           // This will allow to generate normal oop implicit null check.
3246           if (Matcher::gen_narrow_oop_implicit_null_checks())
3247             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3248           //
3249           // This transformation together with CastPP transformation above
3250           // will generated code for implicit NULL checks for compressed oops.
3251           //
3252           // The original code after Optimize()
3253           //
3254           //    LoadN memory, narrow_oop_reg
3255           //    decode narrow_oop_reg, base_reg
3256           //    CmpP base_reg, NULL
3257           //    CastPP base_reg // NotNull
3258           //    Load [base_reg + offset], val_reg
3259           //
3260           // after these transformations will be
3261           //
3262           //    LoadN memory, narrow_oop_reg
3263           //    CmpN narrow_oop_reg, NULL
3264           //    decode_not_null narrow_oop_reg, base_reg
3265           //    Load [base_reg + offset], val_reg
3266           //
3267           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3268           // since narrow oops can be used in debug info now (see the code in
3269           // final_graph_reshaping_walk()).
3270           //
3271           // At the end the code will be matched to
3272           // on x86:
3273           //
3274           //    Load_narrow_oop memory, narrow_oop_reg
3275           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3276           //    NullCheck narrow_oop_reg
3277           //
3278           // and on sparc:
3279           //
3280           //    Load_narrow_oop memory, narrow_oop_reg
3281           //    decode_not_null narrow_oop_reg, base_reg
3282           //    Load [base_reg + offset], val_reg
3283           //    NullCheck base_reg
3284           //
3285         } else if (t->isa_oopptr()) {
3286           new_in2 = ConNode::make(t->make_narrowoop());
3287         } else if (t->isa_klassptr()) {
3288           new_in2 = ConNode::make(t->make_narrowklass());
3289         }
3290       }
3291       if (new_in2 != NULL) {
3292         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3293         n->subsume_by(cmpN, this);
3294         if (in1->outcnt() == 0) {
3295           in1->disconnect_inputs(this);
3296         }
3297         if (in2->outcnt() == 0) {
3298           in2->disconnect_inputs(this);
3299         }
3300       }
3301     }
3302     break;
3303 
3304   case Op_DecodeN:
3305   case Op_DecodeNKlass:
3306     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3307     // DecodeN could be pinned when it can't be fold into
3308     // an address expression, see the code for Op_CastPP above.
3309     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3310     break;
3311 
3312   case Op_EncodeP:
3313   case Op_EncodePKlass: {
3314     Node* in1 = n->in(1);
3315     if (in1->is_DecodeNarrowPtr()) {
3316       n->subsume_by(in1->in(1), this);
3317     } else if (in1->Opcode() == Op_ConP) {
3318       const Type* t = in1->bottom_type();
3319       if (t == TypePtr::NULL_PTR) {
3320         assert(t->isa_oopptr(), "null klass?");
3321         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3322       } else if (t->isa_oopptr()) {
3323         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3324       } else if (t->isa_klassptr()) {
3325         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3326       }
3327     }
3328     if (in1->outcnt() == 0) {
3329       in1->disconnect_inputs(this);
3330     }
3331     break;
3332   }
3333 
3334   case Op_Proj: {
3335     if (OptimizeStringConcat || IncrementalInline) {
3336       ProjNode* proj = n->as_Proj();
3337       if (proj->_is_io_use) {
3338         assert(proj->_con == TypeFunc::I_O || proj->_con == TypeFunc::Memory, "");
3339         // Separate projections were used for the exception path which
3340         // are normally removed by a late inline.  If it wasn't inlined
3341         // then they will hang around and should just be replaced with
3342         // the original one. Merge them.
3343         Node* non_io_proj = proj->in(0)->as_Multi()->proj_out_or_null(proj->_con, false /*is_io_use*/);
3344         if (non_io_proj  != NULL) {
3345           proj->subsume_by(non_io_proj , this);
3346         }
3347       }
3348     }
3349     break;
3350   }
3351 
3352   case Op_Phi:
3353     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3354       // The EncodeP optimization may create Phi with the same edges
3355       // for all paths. It is not handled well by Register Allocator.
3356       Node* unique_in = n->in(1);
3357       assert(unique_in != NULL, "");
3358       uint cnt = n->req();
3359       for (uint i = 2; i < cnt; i++) {
3360         Node* m = n->in(i);
3361         assert(m != NULL, "");
3362         if (unique_in != m)
3363           unique_in = NULL;
3364       }
3365       if (unique_in != NULL) {
3366         n->subsume_by(unique_in, this);
3367       }
3368     }
3369     break;
3370 
3371 #endif
3372 
3373 #ifdef ASSERT
3374   case Op_CastII:
3375     // Verify that all range check dependent CastII nodes were removed.
3376     if (n->isa_CastII()->has_range_check()) {
3377       n->dump(3);
3378       assert(false, "Range check dependent CastII node was not removed");
3379     }
3380     break;
3381 #endif
3382 
3383   case Op_ModI:
3384     if (UseDivMod) {
3385       // Check if a%b and a/b both exist
3386       Node* d = n->find_similar(Op_DivI);
3387       if (d) {
3388         // Replace them with a fused divmod if supported
3389         if (Matcher::has_match_rule(Op_DivModI)) {
3390           DivModINode* divmod = DivModINode::make(n);
3391           d->subsume_by(divmod->div_proj(), this);
3392           n->subsume_by(divmod->mod_proj(), this);
3393         } else {
3394           // replace a%b with a-((a/b)*b)
3395           Node* mult = new MulINode(d, d->in(2));
3396           Node* sub  = new SubINode(d->in(1), mult);
3397           n->subsume_by(sub, this);
3398         }
3399       }
3400     }
3401     break;
3402 
3403   case Op_ModL:
3404     if (UseDivMod) {
3405       // Check if a%b and a/b both exist
3406       Node* d = n->find_similar(Op_DivL);
3407       if (d) {
3408         // Replace them with a fused divmod if supported
3409         if (Matcher::has_match_rule(Op_DivModL)) {
3410           DivModLNode* divmod = DivModLNode::make(n);
3411           d->subsume_by(divmod->div_proj(), this);
3412           n->subsume_by(divmod->mod_proj(), this);
3413         } else {
3414           // replace a%b with a-((a/b)*b)
3415           Node* mult = new MulLNode(d, d->in(2));
3416           Node* sub  = new SubLNode(d->in(1), mult);
3417           n->subsume_by(sub, this);
3418         }
3419       }
3420     }
3421     break;
3422 
3423   case Op_LoadVector:
3424   case Op_StoreVector:
3425   case Op_LoadVectorGather:
3426   case Op_StoreVectorScatter:
3427   case Op_LoadVectorGatherMasked:
3428   case Op_StoreVectorScatterMasked:
3429   case Op_VectorCmpMasked:
3430   case Op_VectorMaskGen:
3431   case Op_LoadVectorMasked:
3432   case Op_StoreVectorMasked:
3433     break;
3434 
3435   case Op_AddReductionVI:
3436   case Op_AddReductionVL:
3437   case Op_AddReductionVF:
3438   case Op_AddReductionVD:
3439   case Op_MulReductionVI:
3440   case Op_MulReductionVL:
3441   case Op_MulReductionVF:
3442   case Op_MulReductionVD:
3443   case Op_MinReductionV:
3444   case Op_MaxReductionV:
3445   case Op_AndReductionV:
3446   case Op_OrReductionV:
3447   case Op_XorReductionV:
3448     break;
3449 
3450   case Op_PackB:
3451   case Op_PackS:
3452   case Op_PackI:
3453   case Op_PackF:
3454   case Op_PackL:
3455   case Op_PackD:
3456     if (n->req()-1 > 2) {
3457       // Replace many operand PackNodes with a binary tree for matching
3458       PackNode* p = (PackNode*) n;
3459       Node* btp = p->binary_tree_pack(1, n->req());
3460       n->subsume_by(btp, this);
3461     }
3462     break;
3463   case Op_Loop:
3464     assert(!n->as_Loop()->is_transformed_long_inner_loop() || _loop_opts_cnt == 0, "should have been turned into a counted loop");
3465   case Op_CountedLoop:
3466   case Op_LongCountedLoop:
3467   case Op_OuterStripMinedLoop:
3468     if (n->as_Loop()->is_inner_loop()) {
3469       frc.inc_inner_loop_count();
3470     }
3471     n->as_Loop()->verify_strip_mined(0);
3472     break;
3473   case Op_LShiftI:
3474   case Op_RShiftI:
3475   case Op_URShiftI:
3476   case Op_LShiftL:
3477   case Op_RShiftL:
3478   case Op_URShiftL:
3479     if (Matcher::need_masked_shift_count) {
3480       // The cpu's shift instructions don't restrict the count to the
3481       // lower 5/6 bits. We need to do the masking ourselves.
3482       Node* in2 = n->in(2);
3483       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3484       const TypeInt* t = in2->find_int_type();
3485       if (t != NULL && t->is_con()) {
3486         juint shift = t->get_con();
3487         if (shift > mask) { // Unsigned cmp
3488           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3489         }
3490       } else {
3491         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3492           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3493           n->set_req(2, shift);
3494         }
3495       }
3496       if (in2->outcnt() == 0) { // Remove dead node
3497         in2->disconnect_inputs(this);
3498       }
3499     }
3500     break;
3501   case Op_MemBarStoreStore:
3502   case Op_MemBarRelease:
3503     // Break the link with AllocateNode: it is no longer useful and
3504     // confuses register allocation.
3505     if (n->req() > MemBarNode::Precedent) {
3506       n->set_req(MemBarNode::Precedent, top());
3507     }
3508     break;
3509   case Op_MemBarAcquire: {
3510     if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3511       // At parse time, the trailing MemBarAcquire for a volatile load
3512       // is created with an edge to the load. After optimizations,
3513       // that input may be a chain of Phis. If those phis have no
3514       // other use, then the MemBarAcquire keeps them alive and
3515       // register allocation can be confused.
3516       ResourceMark rm;
3517       Unique_Node_List wq;
3518       wq.push(n->in(MemBarNode::Precedent));
3519       n->set_req(MemBarNode::Precedent, top());
3520       while (wq.size() > 0) {
3521         Node* m = wq.pop();
3522         if (m->outcnt() == 0) {
3523           for (uint j = 0; j < m->req(); j++) {
3524             Node* in = m->in(j);
3525             if (in != NULL) {
3526               wq.push(in);
3527             }
3528           }
3529           m->disconnect_inputs(this);
3530         }
3531       }
3532     }
3533     break;
3534   }
3535   case Op_Blackhole:
3536     break;
3537   case Op_RangeCheck: {
3538     RangeCheckNode* rc = n->as_RangeCheck();
3539     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3540     n->subsume_by(iff, this);
3541     frc._tests.push(iff);
3542     break;
3543   }
3544   case Op_ConvI2L: {
3545     if (!Matcher::convi2l_type_required) {
3546       // Code generation on some platforms doesn't need accurate
3547       // ConvI2L types. Widening the type can help remove redundant
3548       // address computations.
3549       n->as_Type()->set_type(TypeLong::INT);
3550       ResourceMark rm;
3551       Unique_Node_List wq;
3552       wq.push(n);
3553       for (uint next = 0; next < wq.size(); next++) {
3554         Node *m = wq.at(next);
3555 
3556         for(;;) {
3557           // Loop over all nodes with identical inputs edges as m
3558           Node* k = m->find_similar(m->Opcode());
3559           if (k == NULL) {
3560             break;
3561           }
3562           // Push their uses so we get a chance to remove node made
3563           // redundant
3564           for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3565             Node* u = k->fast_out(i);
3566             if (u->Opcode() == Op_LShiftL ||
3567                 u->Opcode() == Op_AddL ||
3568                 u->Opcode() == Op_SubL ||
3569                 u->Opcode() == Op_AddP) {
3570               wq.push(u);
3571             }
3572           }
3573           // Replace all nodes with identical edges as m with m
3574           k->subsume_by(m, this);
3575         }
3576       }
3577     }
3578     break;
3579   }
3580   case Op_CmpUL: {
3581     if (!Matcher::has_match_rule(Op_CmpUL)) {
3582       // No support for unsigned long comparisons
3583       ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3584       Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3585       Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3586       ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3587       Node* andl = new AndLNode(orl, remove_sign_mask);
3588       Node* cmp = new CmpLNode(andl, n->in(2));
3589       n->subsume_by(cmp, this);
3590     }
3591     break;
3592   }
3593   default:
3594     assert(!n->is_Call(), "");
3595     assert(!n->is_Mem(), "");
3596     assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3597     break;
3598   }
3599 }
3600 
3601 //------------------------------final_graph_reshaping_walk---------------------
3602 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3603 // requires that the walk visits a node's inputs before visiting the node.
3604 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3605   Unique_Node_List sfpt;
3606 
3607   frc._visited.set(root->_idx); // first, mark node as visited
3608   uint cnt = root->req();
3609   Node *n = root;
3610   uint  i = 0;
3611   while (true) {
3612     if (i < cnt) {
3613       // Place all non-visited non-null inputs onto stack
3614       Node* m = n->in(i);
3615       ++i;
3616       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3617         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3618           // compute worst case interpreter size in case of a deoptimization
3619           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3620 
3621           sfpt.push(m);
3622         }
3623         cnt = m->req();
3624         nstack.push(n, i); // put on stack parent and next input's index
3625         n = m;
3626         i = 0;
3627       }
3628     } else {
3629       // Now do post-visit work
3630       final_graph_reshaping_impl( n, frc );
3631       if (nstack.is_empty())
3632         break;             // finished
3633       n = nstack.node();   // Get node from stack
3634       cnt = n->req();
3635       i = nstack.index();
3636       nstack.pop();        // Shift to the next node on stack
3637     }
3638   }
3639 
3640   // Skip next transformation if compressed oops are not used.
3641   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3642       (!UseCompressedOops && !UseCompressedClassPointers))
3643     return;
3644 
3645   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3646   // It could be done for an uncommon traps or any safepoints/calls
3647   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3648   while (sfpt.size() > 0) {
3649     n = sfpt.pop();
3650     JVMState *jvms = n->as_SafePoint()->jvms();
3651     assert(jvms != NULL, "sanity");
3652     int start = jvms->debug_start();
3653     int end   = n->req();
3654     bool is_uncommon = (n->is_CallStaticJava() &&
3655                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3656     for (int j = start; j < end; j++) {
3657       Node* in = n->in(j);
3658       if (in->is_DecodeNarrowPtr()) {
3659         bool safe_to_skip = true;
3660         if (!is_uncommon ) {
3661           // Is it safe to skip?
3662           for (uint i = 0; i < in->outcnt(); i++) {
3663             Node* u = in->raw_out(i);
3664             if (!u->is_SafePoint() ||
3665                 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3666               safe_to_skip = false;
3667             }
3668           }
3669         }
3670         if (safe_to_skip) {
3671           n->set_req(j, in->in(1));
3672         }
3673         if (in->outcnt() == 0) {
3674           in->disconnect_inputs(this);
3675         }
3676       }
3677     }
3678   }
3679 }
3680 
3681 //------------------------------final_graph_reshaping--------------------------
3682 // Final Graph Reshaping.
3683 //
3684 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3685 //     and not commoned up and forced early.  Must come after regular
3686 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3687 //     inputs to Loop Phis; these will be split by the allocator anyways.
3688 //     Remove Opaque nodes.
3689 // (2) Move last-uses by commutative operations to the left input to encourage
3690 //     Intel update-in-place two-address operations and better register usage
3691 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3692 //     calls canonicalizing them back.
3693 // (3) Count the number of double-precision FP ops, single-precision FP ops
3694 //     and call sites.  On Intel, we can get correct rounding either by
3695 //     forcing singles to memory (requires extra stores and loads after each
3696 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3697 //     clearing the mode bit around call sites).  The mode bit is only used
3698 //     if the relative frequency of single FP ops to calls is low enough.
3699 //     This is a key transform for SPEC mpeg_audio.
3700 // (4) Detect infinite loops; blobs of code reachable from above but not
3701 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3702 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3703 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3704 //     Detection is by looking for IfNodes where only 1 projection is
3705 //     reachable from below or CatchNodes missing some targets.
3706 // (5) Assert for insane oop offsets in debug mode.
3707 
3708 bool Compile::final_graph_reshaping() {
3709   // an infinite loop may have been eliminated by the optimizer,
3710   // in which case the graph will be empty.
3711   if (root()->req() == 1) {
3712     record_method_not_compilable("trivial infinite loop");
3713     return true;
3714   }
3715 
3716   // Expensive nodes have their control input set to prevent the GVN
3717   // from freely commoning them. There's no GVN beyond this point so
3718   // no need to keep the control input. We want the expensive nodes to
3719   // be freely moved to the least frequent code path by gcm.
3720   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3721   for (int i = 0; i < expensive_count(); i++) {
3722     _expensive_nodes.at(i)->set_req(0, NULL);
3723   }
3724 
3725   Final_Reshape_Counts frc;
3726 
3727   // Visit everybody reachable!
3728   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3729   Node_Stack nstack(live_nodes() >> 1);
3730   final_graph_reshaping_walk(nstack, root(), frc);
3731 
3732   // Check for unreachable (from below) code (i.e., infinite loops).
3733   for( uint i = 0; i < frc._tests.size(); i++ ) {
3734     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3735     // Get number of CFG targets.
3736     // Note that PCTables include exception targets after calls.
3737     uint required_outcnt = n->required_outcnt();
3738     if (n->outcnt() != required_outcnt) {
3739       // Check for a few special cases.  Rethrow Nodes never take the
3740       // 'fall-thru' path, so expected kids is 1 less.
3741       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3742         if (n->in(0)->in(0)->is_Call()) {
3743           CallNode *call = n->in(0)->in(0)->as_Call();
3744           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3745             required_outcnt--;      // Rethrow always has 1 less kid
3746           } else if (call->req() > TypeFunc::Parms &&
3747                      call->is_CallDynamicJava()) {
3748             // Check for null receiver. In such case, the optimizer has
3749             // detected that the virtual call will always result in a null
3750             // pointer exception. The fall-through projection of this CatchNode
3751             // will not be populated.
3752             Node *arg0 = call->in(TypeFunc::Parms);
3753             if (arg0->is_Type() &&
3754                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3755               required_outcnt--;
3756             }
3757           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3758                      call->req() > TypeFunc::Parms+1 &&
3759                      call->is_CallStaticJava()) {
3760             // Check for negative array length. In such case, the optimizer has
3761             // detected that the allocation attempt will always result in an
3762             // exception. There is no fall-through projection of this CatchNode .
3763             Node *arg1 = call->in(TypeFunc::Parms+1);
3764             if (arg1->is_Type() &&
3765                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3766               required_outcnt--;
3767             }
3768           }
3769         }
3770       }
3771       // Recheck with a better notion of 'required_outcnt'
3772       if (n->outcnt() != required_outcnt) {
3773         record_method_not_compilable("malformed control flow");
3774         return true;            // Not all targets reachable!
3775       }
3776     }
3777     // Check that I actually visited all kids.  Unreached kids
3778     // must be infinite loops.
3779     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3780       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3781         record_method_not_compilable("infinite loop");
3782         return true;            // Found unvisited kid; must be unreach
3783       }
3784 
3785     // Here so verification code in final_graph_reshaping_walk()
3786     // always see an OuterStripMinedLoopEnd
3787     if (n->is_OuterStripMinedLoopEnd() || n->is_LongCountedLoopEnd()) {
3788       IfNode* init_iff = n->as_If();
3789       Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3790       n->subsume_by(iff, this);
3791     }
3792   }
3793 
3794 #ifdef IA32
3795   // If original bytecodes contained a mixture of floats and doubles
3796   // check if the optimizer has made it homogenous, item (3).
3797   if (UseSSE == 0 &&
3798       frc.get_float_count() > 32 &&
3799       frc.get_double_count() == 0 &&
3800       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3801     set_24_bit_selection_and_mode(false, true);
3802   }
3803 #endif // IA32
3804 
3805   set_java_calls(frc.get_java_call_count());
3806   set_inner_loops(frc.get_inner_loop_count());
3807 
3808   // No infinite loops, no reason to bail out.
3809   return false;
3810 }
3811 
3812 //-----------------------------too_many_traps----------------------------------
3813 // Report if there are too many traps at the current method and bci.
3814 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3815 bool Compile::too_many_traps(ciMethod* method,
3816                              int bci,
3817                              Deoptimization::DeoptReason reason) {
3818   ciMethodData* md = method->method_data();
3819   if (md->is_empty()) {
3820     // Assume the trap has not occurred, or that it occurred only
3821     // because of a transient condition during start-up in the interpreter.
3822     return false;
3823   }
3824   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3825   if (md->has_trap_at(bci, m, reason) != 0) {
3826     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3827     // Also, if there are multiple reasons, or if there is no per-BCI record,
3828     // assume the worst.
3829     if (log())
3830       log()->elem("observe trap='%s' count='%d'",
3831                   Deoptimization::trap_reason_name(reason),
3832                   md->trap_count(reason));
3833     return true;
3834   } else {
3835     // Ignore method/bci and see if there have been too many globally.
3836     return too_many_traps(reason, md);
3837   }
3838 }
3839 
3840 // Less-accurate variant which does not require a method and bci.
3841 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3842                              ciMethodData* logmd) {
3843   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3844     // Too many traps globally.
3845     // Note that we use cumulative trap_count, not just md->trap_count.
3846     if (log()) {
3847       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3848       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3849                   Deoptimization::trap_reason_name(reason),
3850                   mcount, trap_count(reason));
3851     }
3852     return true;
3853   } else {
3854     // The coast is clear.
3855     return false;
3856   }
3857 }
3858 
3859 //--------------------------too_many_recompiles--------------------------------
3860 // Report if there are too many recompiles at the current method and bci.
3861 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3862 // Is not eager to return true, since this will cause the compiler to use
3863 // Action_none for a trap point, to avoid too many recompilations.
3864 bool Compile::too_many_recompiles(ciMethod* method,
3865                                   int bci,
3866                                   Deoptimization::DeoptReason reason) {
3867   ciMethodData* md = method->method_data();
3868   if (md->is_empty()) {
3869     // Assume the trap has not occurred, or that it occurred only
3870     // because of a transient condition during start-up in the interpreter.
3871     return false;
3872   }
3873   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3874   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3875   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3876   Deoptimization::DeoptReason per_bc_reason
3877     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3878   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3879   if ((per_bc_reason == Deoptimization::Reason_none
3880        || md->has_trap_at(bci, m, reason) != 0)
3881       // The trap frequency measure we care about is the recompile count:
3882       && md->trap_recompiled_at(bci, m)
3883       && md->overflow_recompile_count() >= bc_cutoff) {
3884     // Do not emit a trap here if it has already caused recompilations.
3885     // Also, if there are multiple reasons, or if there is no per-BCI record,
3886     // assume the worst.
3887     if (log())
3888       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3889                   Deoptimization::trap_reason_name(reason),
3890                   md->trap_count(reason),
3891                   md->overflow_recompile_count());
3892     return true;
3893   } else if (trap_count(reason) != 0
3894              && decompile_count() >= m_cutoff) {
3895     // Too many recompiles globally, and we have seen this sort of trap.
3896     // Use cumulative decompile_count, not just md->decompile_count.
3897     if (log())
3898       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3899                   Deoptimization::trap_reason_name(reason),
3900                   md->trap_count(reason), trap_count(reason),
3901                   md->decompile_count(), decompile_count());
3902     return true;
3903   } else {
3904     // The coast is clear.
3905     return false;
3906   }
3907 }
3908 
3909 // Compute when not to trap. Used by matching trap based nodes and
3910 // NullCheck optimization.
3911 void Compile::set_allowed_deopt_reasons() {
3912   _allowed_reasons = 0;
3913   if (is_method_compilation()) {
3914     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3915       assert(rs < BitsPerInt, "recode bit map");
3916       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3917         _allowed_reasons |= nth_bit(rs);
3918       }
3919     }
3920   }
3921 }
3922 
3923 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
3924   return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
3925 }
3926 
3927 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
3928   return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
3929 }
3930 
3931 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
3932   if (holder->is_initialized()) {
3933     return false;
3934   }
3935   if (holder->is_being_initialized()) {
3936     if (accessing_method->holder() == holder) {
3937       // Access inside a class. The barrier can be elided when access happens in <clinit>,
3938       // <init>, or a static method. In all those cases, there was an initialization
3939       // barrier on the holder klass passed.
3940       if (accessing_method->is_static_initializer() ||
3941           accessing_method->is_object_initializer() ||
3942           accessing_method->is_static()) {
3943         return false;
3944       }
3945     } else if (accessing_method->holder()->is_subclass_of(holder)) {
3946       // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
3947       // In case of <init> or a static method, the barrier is on the subclass is not enough:
3948       // child class can become fully initialized while its parent class is still being initialized.
3949       if (accessing_method->is_static_initializer()) {
3950         return false;
3951       }
3952     }
3953     ciMethod* root = method(); // the root method of compilation
3954     if (root != accessing_method) {
3955       return needs_clinit_barrier(holder, root); // check access in the context of compilation root
3956     }
3957   }
3958   return true;
3959 }
3960 
3961 #ifndef PRODUCT
3962 //------------------------------verify_graph_edges---------------------------
3963 // Walk the Graph and verify that there is a one-to-one correspondence
3964 // between Use-Def edges and Def-Use edges in the graph.
3965 void Compile::verify_graph_edges(bool no_dead_code) {
3966   if (VerifyGraphEdges) {
3967     Unique_Node_List visited;
3968     // Call recursive graph walk to check edges
3969     _root->verify_edges(visited);
3970     if (no_dead_code) {
3971       // Now make sure that no visited node is used by an unvisited node.
3972       bool dead_nodes = false;
3973       Unique_Node_List checked;
3974       while (visited.size() > 0) {
3975         Node* n = visited.pop();
3976         checked.push(n);
3977         for (uint i = 0; i < n->outcnt(); i++) {
3978           Node* use = n->raw_out(i);
3979           if (checked.member(use))  continue;  // already checked
3980           if (visited.member(use))  continue;  // already in the graph
3981           if (use->is_Con())        continue;  // a dead ConNode is OK
3982           // At this point, we have found a dead node which is DU-reachable.
3983           if (!dead_nodes) {
3984             tty->print_cr("*** Dead nodes reachable via DU edges:");
3985             dead_nodes = true;
3986           }
3987           use->dump(2);
3988           tty->print_cr("---");
3989           checked.push(use);  // No repeats; pretend it is now checked.
3990         }
3991       }
3992       assert(!dead_nodes, "using nodes must be reachable from root");
3993     }
3994   }
3995 }
3996 #endif
3997 
3998 // The Compile object keeps track of failure reasons separately from the ciEnv.
3999 // This is required because there is not quite a 1-1 relation between the
4000 // ciEnv and its compilation task and the Compile object.  Note that one
4001 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4002 // to backtrack and retry without subsuming loads.  Other than this backtracking
4003 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4004 // by the logic in C2Compiler.
4005 void Compile::record_failure(const char* reason) {
4006   if (log() != NULL) {
4007     log()->elem("failure reason='%s' phase='compile'", reason);
4008   }
4009   if (_failure_reason == NULL) {
4010     // Record the first failure reason.
4011     _failure_reason = reason;
4012   }
4013 
4014   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4015     C->print_method(PHASE_FAILURE);
4016   }
4017   _root = NULL;  // flush the graph, too
4018 }
4019 
4020 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4021   : TraceTime(name, accumulator, CITime, CITimeVerbose),
4022     _phase_name(name), _dolog(CITimeVerbose)
4023 {
4024   if (_dolog) {
4025     C = Compile::current();
4026     _log = C->log();
4027   } else {
4028     C = NULL;
4029     _log = NULL;
4030   }
4031   if (_log != NULL) {
4032     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4033     _log->stamp();
4034     _log->end_head();
4035   }
4036 }
4037 
4038 Compile::TracePhase::~TracePhase() {
4039 
4040   C = Compile::current();
4041   if (_dolog) {
4042     _log = C->log();
4043   } else {
4044     _log = NULL;
4045   }
4046 
4047 #ifdef ASSERT
4048   if (PrintIdealNodeCount) {
4049     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4050                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4051   }
4052 
4053   if (VerifyIdealNodeCount) {
4054     Compile::current()->print_missing_nodes();
4055   }
4056 #endif
4057 
4058   if (_log != NULL) {
4059     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4060   }
4061 }
4062 
4063 //----------------------------static_subtype_check-----------------------------
4064 // Shortcut important common cases when superklass is exact:
4065 // (0) superklass is java.lang.Object (can occur in reflective code)
4066 // (1) subklass is already limited to a subtype of superklass => always ok
4067 // (2) subklass does not overlap with superklass => always fail
4068 // (3) superklass has NO subtypes and we can check with a simple compare.
4069 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4070   if (StressReflectiveCode) {
4071     return SSC_full_test;       // Let caller generate the general case.
4072   }
4073 
4074   if (superk == env()->Object_klass()) {
4075     return SSC_always_true;     // (0) this test cannot fail
4076   }
4077 
4078   ciType* superelem = superk;
4079   ciType* subelem = subk;
4080   if (superelem->is_array_klass()) {
4081     superelem = superelem->as_array_klass()->base_element_type();
4082   }
4083   if (subelem->is_array_klass()) {
4084     subelem = subelem->as_array_klass()->base_element_type();
4085   }
4086 
4087   if (!subk->is_interface()) {  // cannot trust static interface types yet
4088     if (subk->is_subtype_of(superk)) {
4089       return SSC_always_true;   // (1) false path dead; no dynamic test needed
4090     }
4091     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4092         !(subelem->is_klass() && subelem->as_klass()->is_interface()) &&
4093         !superk->is_subtype_of(subk)) {
4094       return SSC_always_false;  // (2) true path dead; no dynamic test needed
4095     }
4096   }
4097 
4098   // If casting to an instance klass, it must have no subtypes
4099   if (superk->is_interface()) {
4100     // Cannot trust interfaces yet.
4101     // %%% S.B. superk->nof_implementors() == 1
4102   } else if (superelem->is_instance_klass()) {
4103     ciInstanceKlass* ik = superelem->as_instance_klass();
4104     if (!ik->has_subklass() && !ik->is_interface()) {
4105       if (!ik->is_final()) {
4106         // Add a dependency if there is a chance of a later subclass.
4107         dependencies()->assert_leaf_type(ik);
4108       }
4109       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
4110     }
4111   } else {
4112     // A primitive array type has no subtypes.
4113     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
4114   }
4115 
4116   return SSC_full_test;
4117 }
4118 
4119 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4120 #ifdef _LP64
4121   // The scaled index operand to AddP must be a clean 64-bit value.
4122   // Java allows a 32-bit int to be incremented to a negative
4123   // value, which appears in a 64-bit register as a large
4124   // positive number.  Using that large positive number as an
4125   // operand in pointer arithmetic has bad consequences.
4126   // On the other hand, 32-bit overflow is rare, and the possibility
4127   // can often be excluded, if we annotate the ConvI2L node with
4128   // a type assertion that its value is known to be a small positive
4129   // number.  (The prior range check has ensured this.)
4130   // This assertion is used by ConvI2LNode::Ideal.
4131   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4132   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4133   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4134   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4135 #endif
4136   return idx;
4137 }
4138 
4139 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4140 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl, bool carry_dependency) {
4141   if (ctrl != NULL) {
4142     // Express control dependency by a CastII node with a narrow type.
4143     value = new CastIINode(value, itype, carry_dependency ? ConstraintCastNode::StrongDependency : ConstraintCastNode::RegularDependency, true /* range check dependency */);
4144     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4145     // node from floating above the range check during loop optimizations. Otherwise, the
4146     // ConvI2L node may be eliminated independently of the range check, causing the data path
4147     // to become TOP while the control path is still there (although it's unreachable).
4148     value->set_req(0, ctrl);
4149     value = phase->transform(value);
4150   }
4151   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4152   return phase->transform(new ConvI2LNode(value, ltype));
4153 }
4154 
4155 void Compile::print_inlining_stream_free() {
4156   if (_print_inlining_stream != NULL) {
4157     _print_inlining_stream->~stringStream();
4158     _print_inlining_stream = NULL;
4159   }
4160 }
4161 
4162 // The message about the current inlining is accumulated in
4163 // _print_inlining_stream and transfered into the _print_inlining_list
4164 // once we know whether inlining succeeds or not. For regular
4165 // inlining, messages are appended to the buffer pointed by
4166 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4167 // a new buffer is added after _print_inlining_idx in the list. This
4168 // way we can update the inlining message for late inlining call site
4169 // when the inlining is attempted again.
4170 void Compile::print_inlining_init() {
4171   if (print_inlining() || print_intrinsics()) {
4172     // print_inlining_init is actually called several times.
4173     print_inlining_stream_free();
4174     _print_inlining_stream = new stringStream();
4175     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer*>(comp_arena(), 1, 1, new PrintInliningBuffer());
4176   }
4177 }
4178 
4179 void Compile::print_inlining_reinit() {
4180   if (print_inlining() || print_intrinsics()) {
4181     print_inlining_stream_free();
4182     // Re allocate buffer when we change ResourceMark
4183     _print_inlining_stream = new stringStream();
4184   }
4185 }
4186 
4187 void Compile::print_inlining_reset() {
4188   _print_inlining_stream->reset();
4189 }
4190 
4191 void Compile::print_inlining_commit() {
4192   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4193   // Transfer the message from _print_inlining_stream to the current
4194   // _print_inlining_list buffer and clear _print_inlining_stream.
4195   _print_inlining_list->at(_print_inlining_idx)->ss()->write(_print_inlining_stream->base(), _print_inlining_stream->size());
4196   print_inlining_reset();
4197 }
4198 
4199 void Compile::print_inlining_push() {
4200   // Add new buffer to the _print_inlining_list at current position
4201   _print_inlining_idx++;
4202   _print_inlining_list->insert_before(_print_inlining_idx, new PrintInliningBuffer());
4203 }
4204 
4205 Compile::PrintInliningBuffer* Compile::print_inlining_current() {
4206   return _print_inlining_list->at(_print_inlining_idx);
4207 }
4208 
4209 void Compile::print_inlining_update(CallGenerator* cg) {
4210   if (print_inlining() || print_intrinsics()) {
4211     if (cg->is_late_inline()) {
4212       if (print_inlining_current()->cg() != cg &&
4213           (print_inlining_current()->cg() != NULL ||
4214            print_inlining_current()->ss()->size() != 0)) {
4215         print_inlining_push();
4216       }
4217       print_inlining_commit();
4218       print_inlining_current()->set_cg(cg);
4219     } else {
4220       if (print_inlining_current()->cg() != NULL) {
4221         print_inlining_push();
4222       }
4223       print_inlining_commit();
4224     }
4225   }
4226 }
4227 
4228 void Compile::print_inlining_move_to(CallGenerator* cg) {
4229   // We resume inlining at a late inlining call site. Locate the
4230   // corresponding inlining buffer so that we can update it.
4231   if (print_inlining() || print_intrinsics()) {
4232     for (int i = 0; i < _print_inlining_list->length(); i++) {
4233       if (_print_inlining_list->at(i)->cg() == cg) {
4234         _print_inlining_idx = i;
4235         return;
4236       }
4237     }
4238     ShouldNotReachHere();
4239   }
4240 }
4241 
4242 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4243   if (print_inlining() || print_intrinsics()) {
4244     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4245     assert(print_inlining_current()->cg() == cg, "wrong entry");
4246     // replace message with new message
4247     _print_inlining_list->at_put(_print_inlining_idx, new PrintInliningBuffer());
4248     print_inlining_commit();
4249     print_inlining_current()->set_cg(cg);
4250   }
4251 }
4252 
4253 void Compile::print_inlining_assert_ready() {
4254   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4255 }
4256 
4257 void Compile::process_print_inlining() {
4258   assert(_late_inlines.length() == 0, "not drained yet");
4259   if (print_inlining() || print_intrinsics()) {
4260     ResourceMark rm;
4261     stringStream ss;
4262     assert(_print_inlining_list != NULL, "process_print_inlining should be called only once.");
4263     for (int i = 0; i < _print_inlining_list->length(); i++) {
4264       PrintInliningBuffer* pib = _print_inlining_list->at(i);
4265       ss.print("%s", pib->ss()->as_string());
4266       delete pib;
4267       DEBUG_ONLY(_print_inlining_list->at_put(i, NULL));
4268     }
4269     // Reset _print_inlining_list, it only contains destructed objects.
4270     // It is on the arena, so it will be freed when the arena is reset.
4271     _print_inlining_list = NULL;
4272     // _print_inlining_stream won't be used anymore, either.
4273     print_inlining_stream_free();
4274     size_t end = ss.size();
4275     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4276     strncpy(_print_inlining_output, ss.base(), end+1);
4277     _print_inlining_output[end] = 0;
4278   }
4279 }
4280 
4281 void Compile::dump_print_inlining() {
4282   if (_print_inlining_output != NULL) {
4283     tty->print_raw(_print_inlining_output);
4284   }
4285 }
4286 
4287 void Compile::log_late_inline(CallGenerator* cg) {
4288   if (log() != NULL) {
4289     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4290                 cg->unique_id());
4291     JVMState* p = cg->call_node()->jvms();
4292     while (p != NULL) {
4293       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4294       p = p->caller();
4295     }
4296     log()->tail("late_inline");
4297   }
4298 }
4299 
4300 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4301   log_late_inline(cg);
4302   if (log() != NULL) {
4303     log()->inline_fail(msg);
4304   }
4305 }
4306 
4307 void Compile::log_inline_id(CallGenerator* cg) {
4308   if (log() != NULL) {
4309     // The LogCompilation tool needs a unique way to identify late
4310     // inline call sites. This id must be unique for this call site in
4311     // this compilation. Try to have it unique across compilations as
4312     // well because it can be convenient when grepping through the log
4313     // file.
4314     // Distinguish OSR compilations from others in case CICountOSR is
4315     // on.
4316     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4317     cg->set_unique_id(id);
4318     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4319   }
4320 }
4321 
4322 void Compile::log_inline_failure(const char* msg) {
4323   if (C->log() != NULL) {
4324     C->log()->inline_fail(msg);
4325   }
4326 }
4327 
4328 
4329 // Dump inlining replay data to the stream.
4330 // Don't change thread state and acquire any locks.
4331 void Compile::dump_inline_data(outputStream* out) {
4332   InlineTree* inl_tree = ilt();
4333   if (inl_tree != NULL) {
4334     out->print(" inline %d", inl_tree->count());
4335     inl_tree->dump_replay_data(out);
4336   }
4337 }
4338 
4339 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4340   if (n1->Opcode() < n2->Opcode())      return -1;
4341   else if (n1->Opcode() > n2->Opcode()) return 1;
4342 
4343   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4344   for (uint i = 1; i < n1->req(); i++) {
4345     if (n1->in(i) < n2->in(i))      return -1;
4346     else if (n1->in(i) > n2->in(i)) return 1;
4347   }
4348 
4349   return 0;
4350 }
4351 
4352 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4353   Node* n1 = *n1p;
4354   Node* n2 = *n2p;
4355 
4356   return cmp_expensive_nodes(n1, n2);
4357 }
4358 
4359 void Compile::sort_expensive_nodes() {
4360   if (!expensive_nodes_sorted()) {
4361     _expensive_nodes.sort(cmp_expensive_nodes);
4362   }
4363 }
4364 
4365 bool Compile::expensive_nodes_sorted() const {
4366   for (int i = 1; i < _expensive_nodes.length(); i++) {
4367     if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i-1)) < 0) {
4368       return false;
4369     }
4370   }
4371   return true;
4372 }
4373 
4374 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4375   if (_expensive_nodes.length() == 0) {
4376     return false;
4377   }
4378 
4379   assert(OptimizeExpensiveOps, "optimization off?");
4380 
4381   // Take this opportunity to remove dead nodes from the list
4382   int j = 0;
4383   for (int i = 0; i < _expensive_nodes.length(); i++) {
4384     Node* n = _expensive_nodes.at(i);
4385     if (!n->is_unreachable(igvn)) {
4386       assert(n->is_expensive(), "should be expensive");
4387       _expensive_nodes.at_put(j, n);
4388       j++;
4389     }
4390   }
4391   _expensive_nodes.trunc_to(j);
4392 
4393   // Then sort the list so that similar nodes are next to each other
4394   // and check for at least two nodes of identical kind with same data
4395   // inputs.
4396   sort_expensive_nodes();
4397 
4398   for (int i = 0; i < _expensive_nodes.length()-1; i++) {
4399     if (cmp_expensive_nodes(_expensive_nodes.adr_at(i), _expensive_nodes.adr_at(i+1)) == 0) {
4400       return true;
4401     }
4402   }
4403 
4404   return false;
4405 }
4406 
4407 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4408   if (_expensive_nodes.length() == 0) {
4409     return;
4410   }
4411 
4412   assert(OptimizeExpensiveOps, "optimization off?");
4413 
4414   // Sort to bring similar nodes next to each other and clear the
4415   // control input of nodes for which there's only a single copy.
4416   sort_expensive_nodes();
4417 
4418   int j = 0;
4419   int identical = 0;
4420   int i = 0;
4421   bool modified = false;
4422   for (; i < _expensive_nodes.length()-1; i++) {
4423     assert(j <= i, "can't write beyond current index");
4424     if (_expensive_nodes.at(i)->Opcode() == _expensive_nodes.at(i+1)->Opcode()) {
4425       identical++;
4426       _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4427       continue;
4428     }
4429     if (identical > 0) {
4430       _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4431       identical = 0;
4432     } else {
4433       Node* n = _expensive_nodes.at(i);
4434       igvn.replace_input_of(n, 0, NULL);
4435       igvn.hash_insert(n);
4436       modified = true;
4437     }
4438   }
4439   if (identical > 0) {
4440     _expensive_nodes.at_put(j++, _expensive_nodes.at(i));
4441   } else if (_expensive_nodes.length() >= 1) {
4442     Node* n = _expensive_nodes.at(i);
4443     igvn.replace_input_of(n, 0, NULL);
4444     igvn.hash_insert(n);
4445     modified = true;
4446   }
4447   _expensive_nodes.trunc_to(j);
4448   if (modified) {
4449     igvn.optimize();
4450   }
4451 }
4452 
4453 void Compile::add_expensive_node(Node * n) {
4454   assert(!_expensive_nodes.contains(n), "duplicate entry in expensive list");
4455   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4456   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4457   if (OptimizeExpensiveOps) {
4458     _expensive_nodes.append(n);
4459   } else {
4460     // Clear control input and let IGVN optimize expensive nodes if
4461     // OptimizeExpensiveOps is off.
4462     n->set_req(0, NULL);
4463   }
4464 }
4465 
4466 /**
4467  * Track coarsened Lock and Unlock nodes.
4468  */
4469 
4470 class Lock_List : public Node_List {
4471   uint _origin_cnt;
4472 public:
4473   Lock_List(Arena *a, uint cnt) : Node_List(a), _origin_cnt(cnt) {}
4474   uint origin_cnt() const { return _origin_cnt; }
4475 };
4476 
4477 void Compile::add_coarsened_locks(GrowableArray<AbstractLockNode*>& locks) {
4478   int length = locks.length();
4479   if (length > 0) {
4480     // Have to keep this list until locks elimination during Macro nodes elimination.
4481     Lock_List* locks_list = new (comp_arena()) Lock_List(comp_arena(), length);
4482     for (int i = 0; i < length; i++) {
4483       AbstractLockNode* lock = locks.at(i);
4484       assert(lock->is_coarsened(), "expecting only coarsened AbstractLock nodes, but got '%s'[%d] node", lock->Name(), lock->_idx);
4485       locks_list->push(lock);
4486     }
4487     _coarsened_locks.append(locks_list);
4488   }
4489 }
4490 
4491 void Compile::remove_useless_coarsened_locks(Unique_Node_List& useful) {
4492   int count = coarsened_count();
4493   for (int i = 0; i < count; i++) {
4494     Node_List* locks_list = _coarsened_locks.at(i);
4495     for (uint j = 0; j < locks_list->size(); j++) {
4496       Node* lock = locks_list->at(j);
4497       assert(lock->is_AbstractLock(), "sanity");
4498       if (!useful.member(lock)) {
4499         locks_list->yank(lock);
4500       }
4501     }
4502   }
4503 }
4504 
4505 void Compile::remove_coarsened_lock(Node* n) {
4506   if (n->is_AbstractLock()) {
4507     int count = coarsened_count();
4508     for (int i = 0; i < count; i++) {
4509       Node_List* locks_list = _coarsened_locks.at(i);
4510       locks_list->yank(n);
4511     }
4512   }
4513 }
4514 
4515 bool Compile::coarsened_locks_consistent() {
4516   int count = coarsened_count();
4517   for (int i = 0; i < count; i++) {
4518     bool unbalanced = false;
4519     bool modified = false; // track locks kind modifications
4520     Lock_List* locks_list = (Lock_List*)_coarsened_locks.at(i);
4521     uint size = locks_list->size();
4522     if (size == 0) {
4523       unbalanced = false; // All locks were eliminated - good
4524     } else if (size != locks_list->origin_cnt()) {
4525       unbalanced = true; // Some locks were removed from list
4526     } else {
4527       for (uint j = 0; j < size; j++) {
4528         Node* lock = locks_list->at(j);
4529         // All nodes in group should have the same state (modified or not)
4530         if (!lock->as_AbstractLock()->is_coarsened()) {
4531           if (j == 0) {
4532             // first on list was modified, the rest should be too for consistency
4533             modified = true;
4534           } else if (!modified) {
4535             // this lock was modified but previous locks on the list were not
4536             unbalanced = true;
4537             break;
4538           }
4539         } else if (modified) {
4540           // previous locks on list were modified but not this lock
4541           unbalanced = true;
4542           break;
4543         }
4544       }
4545     }
4546     if (unbalanced) {
4547       // unbalanced monitor enter/exit - only some [un]lock nodes were removed or modified
4548 #ifdef ASSERT
4549       if (PrintEliminateLocks) {
4550         tty->print_cr("=== unbalanced coarsened locks ===");
4551         for (uint l = 0; l < size; l++) {
4552           locks_list->at(l)->dump();
4553         }
4554       }
4555 #endif
4556       record_failure(C2Compiler::retry_no_locks_coarsening());
4557       return false;
4558     }
4559   }
4560   return true;
4561 }
4562 
4563 /**
4564  * Remove the speculative part of types and clean up the graph
4565  */
4566 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4567   if (UseTypeSpeculation) {
4568     Unique_Node_List worklist;
4569     worklist.push(root());
4570     int modified = 0;
4571     // Go over all type nodes that carry a speculative type, drop the
4572     // speculative part of the type and enqueue the node for an igvn
4573     // which may optimize it out.
4574     for (uint next = 0; next < worklist.size(); ++next) {
4575       Node *n  = worklist.at(next);
4576       if (n->is_Type()) {
4577         TypeNode* tn = n->as_Type();
4578         const Type* t = tn->type();
4579         const Type* t_no_spec = t->remove_speculative();
4580         if (t_no_spec != t) {
4581           bool in_hash = igvn.hash_delete(n);
4582           assert(in_hash, "node should be in igvn hash table");
4583           tn->set_type(t_no_spec);
4584           igvn.hash_insert(n);
4585           igvn._worklist.push(n); // give it a chance to go away
4586           modified++;
4587         }
4588       }
4589       // Iterate over outs - endless loops is unreachable from below
4590       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4591         Node *m = n->fast_out(i);
4592         if (not_a_node(m)) {
4593           continue;
4594         }
4595         worklist.push(m);
4596       }
4597     }
4598     // Drop the speculative part of all types in the igvn's type table
4599     igvn.remove_speculative_types();
4600     if (modified > 0) {
4601       igvn.optimize();
4602     }
4603 #ifdef ASSERT
4604     // Verify that after the IGVN is over no speculative type has resurfaced
4605     worklist.clear();
4606     worklist.push(root());
4607     for (uint next = 0; next < worklist.size(); ++next) {
4608       Node *n  = worklist.at(next);
4609       const Type* t = igvn.type_or_null(n);
4610       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4611       if (n->is_Type()) {
4612         t = n->as_Type()->type();
4613         assert(t == t->remove_speculative(), "no more speculative types");
4614       }
4615       // Iterate over outs - endless loops is unreachable from below
4616       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4617         Node *m = n->fast_out(i);
4618         if (not_a_node(m)) {
4619           continue;
4620         }
4621         worklist.push(m);
4622       }
4623     }
4624     igvn.check_no_speculative_types();
4625 #endif
4626   }
4627 }
4628 
4629 // Auxiliary methods to support randomized stressing/fuzzing.
4630 
4631 int Compile::random() {
4632   _stress_seed = os::next_random(_stress_seed);
4633   return static_cast<int>(_stress_seed);
4634 }
4635 
4636 // This method can be called the arbitrary number of times, with current count
4637 // as the argument. The logic allows selecting a single candidate from the
4638 // running list of candidates as follows:
4639 //    int count = 0;
4640 //    Cand* selected = null;
4641 //    while(cand = cand->next()) {
4642 //      if (randomized_select(++count)) {
4643 //        selected = cand;
4644 //      }
4645 //    }
4646 //
4647 // Including count equalizes the chances any candidate is "selected".
4648 // This is useful when we don't have the complete list of candidates to choose
4649 // from uniformly. In this case, we need to adjust the randomicity of the
4650 // selection, or else we will end up biasing the selection towards the latter
4651 // candidates.
4652 //
4653 // Quick back-envelope calculation shows that for the list of n candidates
4654 // the equal probability for the candidate to persist as "best" can be
4655 // achieved by replacing it with "next" k-th candidate with the probability
4656 // of 1/k. It can be easily shown that by the end of the run, the
4657 // probability for any candidate is converged to 1/n, thus giving the
4658 // uniform distribution among all the candidates.
4659 //
4660 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4661 #define RANDOMIZED_DOMAIN_POW 29
4662 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4663 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4664 bool Compile::randomized_select(int count) {
4665   assert(count > 0, "only positive");
4666   return (random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4667 }
4668 
4669 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4670 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4671 
4672 void NodeCloneInfo::dump() const {
4673   tty->print(" {%d:%d} ", idx(), gen());
4674 }
4675 
4676 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4677   uint64_t val = value(old->_idx);
4678   NodeCloneInfo cio(val);
4679   assert(val != 0, "old node should be in the map");
4680   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4681   insert(nnn->_idx, cin.get());
4682 #ifndef PRODUCT
4683   if (is_debug()) {
4684     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4685   }
4686 #endif
4687 }
4688 
4689 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4690   NodeCloneInfo cio(value(old->_idx));
4691   if (cio.get() == 0) {
4692     cio.set(old->_idx, 0);
4693     insert(old->_idx, cio.get());
4694 #ifndef PRODUCT
4695     if (is_debug()) {
4696       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4697     }
4698 #endif
4699   }
4700   clone(old, nnn, gen);
4701 }
4702 
4703 int CloneMap::max_gen() const {
4704   int g = 0;
4705   DictI di(_dict);
4706   for(; di.test(); ++di) {
4707     int t = gen(di._key);
4708     if (g < t) {
4709       g = t;
4710 #ifndef PRODUCT
4711       if (is_debug()) {
4712         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4713       }
4714 #endif
4715     }
4716   }
4717   return g;
4718 }
4719 
4720 void CloneMap::dump(node_idx_t key) const {
4721   uint64_t val = value(key);
4722   if (val != 0) {
4723     NodeCloneInfo ni(val);
4724     ni.dump();
4725   }
4726 }
4727 
4728 // Move Allocate nodes to the start of the list
4729 void Compile::sort_macro_nodes() {
4730   int count = macro_count();
4731   int allocates = 0;
4732   for (int i = 0; i < count; i++) {
4733     Node* n = macro_node(i);
4734     if (n->is_Allocate()) {
4735       if (i != allocates) {
4736         Node* tmp = macro_node(allocates);
4737         _macro_nodes.at_put(allocates, n);
4738         _macro_nodes.at_put(i, tmp);
4739       }
4740       allocates++;
4741     }
4742   }
4743 }
4744 
4745 void Compile::print_method(CompilerPhaseType cpt, const char *name, int level) {
4746   EventCompilerPhase event;
4747   if (event.should_commit()) {
4748     CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, cpt, C->_compile_id, level);
4749   }
4750 #ifndef PRODUCT
4751   if (should_print(level)) {
4752     _printer->print_method(name, level);
4753   }
4754 #endif
4755   C->_latest_stage_start_counter.stamp();
4756 }
4757 
4758 void Compile::print_method(CompilerPhaseType cpt, int level, int idx) {
4759   char output[1024];
4760 #ifndef PRODUCT
4761   if (idx != 0) {
4762     jio_snprintf(output, sizeof(output), "%s:%d", CompilerPhaseTypeHelper::to_string(cpt), idx);
4763   } else {
4764     jio_snprintf(output, sizeof(output), "%s", CompilerPhaseTypeHelper::to_string(cpt));
4765   }
4766 #endif
4767   print_method(cpt, output, level);
4768 }
4769 
4770 void Compile::print_method(CompilerPhaseType cpt, Node* n, int level) {
4771   ResourceMark rm;
4772   stringStream ss;
4773   ss.print_raw(CompilerPhaseTypeHelper::to_string(cpt));
4774   if (n != NULL) {
4775     ss.print(": %d %s ", n->_idx, NodeClassNames[n->Opcode()]);
4776   } else {
4777     ss.print_raw(": NULL");
4778   }
4779   C->print_method(cpt, ss.as_string(), level);
4780 }
4781 
4782 void Compile::end_method(int level) {
4783   EventCompilerPhase event;
4784   if (event.should_commit()) {
4785     CompilerEvent::PhaseEvent::post(event, C->_latest_stage_start_counter, PHASE_END, C->_compile_id, level);
4786   }
4787 
4788 #ifndef PRODUCT
4789   if (_method != NULL && should_print(level)) {
4790     _printer->end_method();
4791   }
4792 #endif
4793 }
4794 
4795 
4796 #ifndef PRODUCT
4797 IdealGraphPrinter* Compile::_debug_file_printer = NULL;
4798 IdealGraphPrinter* Compile::_debug_network_printer = NULL;
4799 
4800 // Called from debugger. Prints method to the default file with the default phase name.
4801 // This works regardless of any Ideal Graph Visualizer flags set or not.
4802 void igv_print() {
4803   Compile::current()->igv_print_method_to_file();
4804 }
4805 
4806 // Same as igv_print() above but with a specified phase name.
4807 void igv_print(const char* phase_name) {
4808   Compile::current()->igv_print_method_to_file(phase_name);
4809 }
4810 
4811 // Called from debugger. Prints method with the default phase name to the default network or the one specified with
4812 // the network flags for the Ideal Graph Visualizer, or to the default file depending on the 'network' argument.
4813 // This works regardless of any Ideal Graph Visualizer flags set or not.
4814 void igv_print(bool network) {
4815   if (network) {
4816     Compile::current()->igv_print_method_to_network();
4817   } else {
4818     Compile::current()->igv_print_method_to_file();
4819   }
4820 }
4821 
4822 // Same as igv_print(bool network) above but with a specified phase name.
4823 void igv_print(bool network, const char* phase_name) {
4824   if (network) {
4825     Compile::current()->igv_print_method_to_network(phase_name);
4826   } else {
4827     Compile::current()->igv_print_method_to_file(phase_name);
4828   }
4829 }
4830 
4831 // Called from debugger. Normal write to the default _printer. Only works if Ideal Graph Visualizer printing flags are set.
4832 void igv_print_default() {
4833   Compile::current()->print_method(PHASE_DEBUG, 0);
4834 }
4835 
4836 // Called from debugger, especially when replaying a trace in which the program state cannot be altered like with rr replay.
4837 // A method is appended to an existing default file with the default phase name. This means that igv_append() must follow
4838 // an earlier igv_print(*) call which sets up the file. This works regardless of any Ideal Graph Visualizer flags set or not.
4839 void igv_append() {
4840   Compile::current()->igv_print_method_to_file("Debug", true);
4841 }
4842 
4843 // Same as igv_append() above but with a specified phase name.
4844 void igv_append(const char* phase_name) {
4845   Compile::current()->igv_print_method_to_file(phase_name, true);
4846 }
4847 
4848 void Compile::igv_print_method_to_file(const char* phase_name, bool append) {
4849   const char* file_name = "custom_debug.xml";
4850   if (_debug_file_printer == NULL) {
4851     _debug_file_printer = new IdealGraphPrinter(C, file_name, append);
4852   } else {
4853     _debug_file_printer->update_compiled_method(C->method());
4854   }
4855   tty->print_cr("Method %s to %s", append ? "appended" : "printed", file_name);
4856   _debug_file_printer->print(phase_name, (Node*)C->root());
4857 }
4858 
4859 void Compile::igv_print_method_to_network(const char* phase_name) {
4860   if (_debug_network_printer == NULL) {
4861     _debug_network_printer = new IdealGraphPrinter(C);
4862   } else {
4863     _debug_network_printer->update_compiled_method(C->method());
4864   }
4865   tty->print_cr("Method printed over network stream to IGV");
4866   _debug_network_printer->print(phase_name, (Node*)C->root());
4867 }
4868 #endif
4869 
4870 Node* Compile::narrow_value(BasicType bt, Node* value, const Type* type, PhaseGVN* phase, bool transform_res) {
4871   if (type != NULL && phase->type(value)->higher_equal(type)) {
4872     return value;
4873   }
4874   Node* result = NULL;
4875   if (bt == T_BYTE) {
4876     result = phase->transform(new LShiftINode(value, phase->intcon(24)));
4877     result = new RShiftINode(result, phase->intcon(24));
4878   } else if (bt == T_BOOLEAN) {
4879     result = new AndINode(value, phase->intcon(0xFF));
4880   } else if (bt == T_CHAR) {
4881     result = new AndINode(value,phase->intcon(0xFFFF));
4882   } else {
4883     assert(bt == T_SHORT, "unexpected narrow type");
4884     result = phase->transform(new LShiftINode(value, phase->intcon(16)));
4885     result = new RShiftINode(result, phase->intcon(16));
4886   }
4887   if (transform_res) {
4888     result = phase->transform(result);
4889   }
4890   return result;
4891 }
4892