rev 55090 : secret-sfac

   1 /*
   2  * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.hpp"
  27 #include "asm/macroAssembler.inline.hpp"
  28 #include "ci/ciReplay.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/exceptionHandlerTable.hpp"
  31 #include "code/nmethod.hpp"
  32 #include "compiler/compileBroker.hpp"
  33 #include "compiler/compileLog.hpp"
  34 #include "compiler/disassembler.hpp"
  35 #include "compiler/oopMap.hpp"
  36 #include "gc/shared/barrierSet.hpp"
  37 #include "gc/shared/c2/barrierSetC2.hpp"
  38 #include "memory/resourceArea.hpp"
  39 #include "opto/addnode.hpp"
  40 #include "opto/block.hpp"
  41 #include "opto/c2compiler.hpp"
  42 #include "opto/callGenerator.hpp"
  43 #include "opto/callnode.hpp"
  44 #include "opto/castnode.hpp"
  45 #include "opto/cfgnode.hpp"
  46 #include "opto/chaitin.hpp"
  47 #include "opto/compile.hpp"
  48 #include "opto/connode.hpp"
  49 #include "opto/convertnode.hpp"
  50 #include "opto/divnode.hpp"
  51 #include "opto/escape.hpp"
  52 #include "opto/idealGraphPrinter.hpp"
  53 #include "opto/loopnode.hpp"
  54 #include "opto/machnode.hpp"
  55 #include "opto/macro.hpp"
  56 #include "opto/matcher.hpp"
  57 #include "opto/mathexactnode.hpp"
  58 #include "opto/memnode.hpp"
  59 #include "opto/mulnode.hpp"
  60 #include "opto/narrowptrnode.hpp"
  61 #include "opto/node.hpp"
  62 #include "opto/opcodes.hpp"
  63 #include "opto/output.hpp"
  64 #include "opto/parse.hpp"
  65 #include "opto/phaseX.hpp"
  66 #include "opto/rootnode.hpp"
  67 #include "opto/runtime.hpp"
  68 #include "opto/stringopts.hpp"
  69 #include "opto/type.hpp"
  70 #include "opto/valuetypenode.hpp"
  71 #include "opto/vectornode.hpp"
  72 #include "runtime/arguments.hpp"
  73 #include "runtime/sharedRuntime.hpp"
  74 #include "runtime/signature.hpp"
  75 #include "runtime/stubRoutines.hpp"
  76 #include "runtime/timer.hpp"
  77 #include "utilities/align.hpp"
  78 #include "utilities/copy.hpp"
  79 #include "utilities/macros.hpp"
  80 #if INCLUDE_ZGC
  81 #include "gc/z/c2/zBarrierSetC2.hpp"
  82 #endif
  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   if (_intrinsics == NULL) {
 137     _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 138   }
 139   int len = _intrinsics->length();
 140   bool found = false;
 141   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
 142   assert(!found, "registering twice");
 143   _intrinsics->insert_before(index, cg);
 144   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 145 }
 146 
 147 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 148   assert(m->is_loaded(), "don't try this on unloaded methods");
 149   if (_intrinsics != NULL) {
 150     bool found = false;
 151     int index = intrinsic_insertion_index(m, is_virtual, found);
 152      if (found) {
 153       return _intrinsics->at(index);
 154     }
 155   }
 156   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 157   if (m->intrinsic_id() != vmIntrinsics::_none &&
 158       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 159     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 160     if (cg != NULL) {
 161       // Save it for next time:
 162       register_intrinsic(cg);
 163       return cg;
 164     } else {
 165       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 166     }
 167   }
 168   return NULL;
 169 }
 170 
 171 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 172 // in library_call.cpp.
 173 
 174 
 175 #ifndef PRODUCT
 176 // statistics gathering...
 177 
 178 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 179 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 180 
 181 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 182   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 183   int oflags = _intrinsic_hist_flags[id];
 184   assert(flags != 0, "what happened?");
 185   if (is_virtual) {
 186     flags |= _intrinsic_virtual;
 187   }
 188   bool changed = (flags != oflags);
 189   if ((flags & _intrinsic_worked) != 0) {
 190     juint count = (_intrinsic_hist_count[id] += 1);
 191     if (count == 1) {
 192       changed = true;           // first time
 193     }
 194     // increment the overall count also:
 195     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 196   }
 197   if (changed) {
 198     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 199       // Something changed about the intrinsic's virtuality.
 200       if ((flags & _intrinsic_virtual) != 0) {
 201         // This is the first use of this intrinsic as a virtual call.
 202         if (oflags != 0) {
 203           // We already saw it as a non-virtual, so note both cases.
 204           flags |= _intrinsic_both;
 205         }
 206       } else if ((oflags & _intrinsic_both) == 0) {
 207         // This is the first use of this intrinsic as a non-virtual
 208         flags |= _intrinsic_both;
 209       }
 210     }
 211     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 212   }
 213   // update the overall flags also:
 214   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 215   return changed;
 216 }
 217 
 218 static char* format_flags(int flags, char* buf) {
 219   buf[0] = 0;
 220   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 221   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 222   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 223   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 224   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 225   if (buf[0] == 0)  strcat(buf, ",");
 226   assert(buf[0] == ',', "must be");
 227   return &buf[1];
 228 }
 229 
 230 void Compile::print_intrinsic_statistics() {
 231   char flagsbuf[100];
 232   ttyLocker ttyl;
 233   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 234   tty->print_cr("Compiler intrinsic usage:");
 235   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 236   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 237   #define PRINT_STAT_LINE(name, c, f) \
 238     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 239   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 240     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 241     int   flags = _intrinsic_hist_flags[id];
 242     juint count = _intrinsic_hist_count[id];
 243     if ((flags | count) != 0) {
 244       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 245     }
 246   }
 247   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 248   if (xtty != NULL)  xtty->tail("statistics");
 249 }
 250 
 251 void Compile::print_statistics() {
 252   { ttyLocker ttyl;
 253     if (xtty != NULL)  xtty->head("statistics type='opto'");
 254     Parse::print_statistics();
 255     PhaseCCP::print_statistics();
 256     PhaseRegAlloc::print_statistics();
 257     Scheduling::print_statistics();
 258     PhasePeephole::print_statistics();
 259     PhaseIdealLoop::print_statistics();
 260     if (xtty != NULL)  xtty->tail("statistics");
 261   }
 262   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 263     // put this under its own <statistics> element.
 264     print_intrinsic_statistics();
 265   }
 266 }
 267 #endif //PRODUCT
 268 
 269 // Support for bundling info
 270 Bundle* Compile::node_bundling(const Node *n) {
 271   assert(valid_bundle_info(n), "oob");
 272   return &_node_bundling_base[n->_idx];
 273 }
 274 
 275 bool Compile::valid_bundle_info(const Node *n) {
 276   return (_node_bundling_limit > n->_idx);
 277 }
 278 
 279 
 280 void Compile::gvn_replace_by(Node* n, Node* nn) {
 281   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 282     Node* use = n->last_out(i);
 283     bool is_in_table = initial_gvn()->hash_delete(use);
 284     uint uses_found = 0;
 285     for (uint j = 0; j < use->len(); j++) {
 286       if (use->in(j) == n) {
 287         if (j < use->req())
 288           use->set_req(j, nn);
 289         else
 290           use->set_prec(j, nn);
 291         uses_found++;
 292       }
 293     }
 294     if (is_in_table) {
 295       // reinsert into table
 296       initial_gvn()->hash_find_insert(use);
 297     }
 298     record_for_igvn(use);
 299     i -= uses_found;    // we deleted 1 or more copies of this edge
 300   }
 301 }
 302 
 303 
 304 static inline bool not_a_node(const Node* n) {
 305   if (n == NULL)                   return true;
 306   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 307   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 308   return false;
 309 }
 310 
 311 // Identify all nodes that are reachable from below, useful.
 312 // Use breadth-first pass that records state in a Unique_Node_List,
 313 // recursive traversal is slower.
 314 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 315   int estimated_worklist_size = live_nodes();
 316   useful.map( estimated_worklist_size, NULL );  // preallocate space
 317 
 318   // Initialize worklist
 319   if (root() != NULL)     { useful.push(root()); }
 320   // If 'top' is cached, declare it useful to preserve cached node
 321   if( cached_top_node() ) { useful.push(cached_top_node()); }
 322 
 323   // Push all useful nodes onto the list, breadthfirst
 324   for( uint next = 0; next < useful.size(); ++next ) {
 325     assert( next < unique(), "Unique useful nodes < total nodes");
 326     Node *n  = useful.at(next);
 327     uint max = n->len();
 328     for( uint i = 0; i < max; ++i ) {
 329       Node *m = n->in(i);
 330       if (not_a_node(m))  continue;
 331       useful.push(m);
 332     }
 333   }
 334 }
 335 
 336 // Update dead_node_list with any missing dead nodes using useful
 337 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 338 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 339   uint max_idx = unique();
 340   VectorSet& useful_node_set = useful.member_set();
 341 
 342   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 343     // If node with index node_idx is not in useful set,
 344     // mark it as dead in dead node list.
 345     if (! useful_node_set.test(node_idx) ) {
 346       record_dead_node(node_idx);
 347     }
 348   }
 349 }
 350 
 351 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 352   int shift = 0;
 353   for (int i = 0; i < inlines->length(); i++) {
 354     CallGenerator* cg = inlines->at(i);
 355     CallNode* call = cg->call_node();
 356     if (shift > 0) {
 357       inlines->at_put(i-shift, cg);
 358     }
 359     if (!useful.member(call)) {
 360       shift++;
 361     }
 362   }
 363   inlines->trunc_to(inlines->length()-shift);
 364 }
 365 
 366 // Disconnect all useless nodes by disconnecting those at the boundary.
 367 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 368   uint next = 0;
 369   while (next < useful.size()) {
 370     Node *n = useful.at(next++);
 371     if (n->is_SafePoint()) {
 372       // We're done with a parsing phase. Replaced nodes are not valid
 373       // beyond that point.
 374       n->as_SafePoint()->delete_replaced_nodes();
 375     }
 376     // Use raw traversal of out edges since this code removes out edges
 377     int max = n->outcnt();
 378     for (int j = 0; j < max; ++j) {
 379       Node* child = n->raw_out(j);
 380       if (! useful.member(child)) {
 381         assert(!child->is_top() || child != top(),
 382                "If top is cached in Compile object it is in useful list");
 383         // Only need to remove this out-edge to the useless node
 384         n->raw_del_out(j);
 385         --j;
 386         --max;
 387       }
 388     }
 389     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 390       record_for_igvn(n->unique_out());
 391     }
 392   }
 393   // Remove useless macro and predicate opaq nodes
 394   for (int i = C->macro_count()-1; i >= 0; i--) {
 395     Node* n = C->macro_node(i);
 396     if (!useful.member(n)) {
 397       remove_macro_node(n);
 398     }
 399   }
 400   // Remove useless CastII nodes with range check dependency
 401   for (int i = range_check_cast_count() - 1; i >= 0; i--) {
 402     Node* cast = range_check_cast_node(i);
 403     if (!useful.member(cast)) {
 404       remove_range_check_cast(cast);
 405     }
 406   }
 407   // Remove useless expensive nodes
 408   for (int i = C->expensive_count()-1; i >= 0; i--) {
 409     Node* n = C->expensive_node(i);
 410     if (!useful.member(n)) {
 411       remove_expensive_node(n);
 412     }
 413   }
 414   // Remove useless Opaque4 nodes
 415   for (int i = opaque4_count() - 1; i >= 0; i--) {
 416     Node* opaq = opaque4_node(i);
 417     if (!useful.member(opaq)) {
 418       remove_opaque4_node(opaq);
 419     }
 420   }
 421   // Remove useless value type nodes
 422   if (_value_type_nodes != NULL) {
 423     _value_type_nodes->remove_useless_nodes(useful.member_set());
 424   }
 425   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 426   bs->eliminate_useless_gc_barriers(useful, this);
 427   // clean up the late inline lists
 428   remove_useless_late_inlines(&_string_late_inlines, useful);
 429   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 430   remove_useless_late_inlines(&_late_inlines, useful);
 431   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 432 }
 433 
 434 //------------------------------frame_size_in_words-----------------------------
 435 // frame_slots in units of words
 436 int Compile::frame_size_in_words() const {
 437   // shift is 0 in LP32 and 1 in LP64
 438   const int shift = (LogBytesPerWord - LogBytesPerInt);
 439   int words = _frame_slots >> shift;
 440   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 441   return words;
 442 }
 443 
 444 // To bang the stack of this compiled method we use the stack size
 445 // that the interpreter would need in case of a deoptimization. This
 446 // removes the need to bang the stack in the deoptimization blob which
 447 // in turn simplifies stack overflow handling.
 448 int Compile::bang_size_in_bytes() const {
 449   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
 450 }
 451 
 452 // ============================================================================
 453 //------------------------------CompileWrapper---------------------------------
 454 class CompileWrapper : public StackObj {
 455   Compile *const _compile;
 456  public:
 457   CompileWrapper(Compile* compile);
 458 
 459   ~CompileWrapper();
 460 };
 461 
 462 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 463   // the Compile* pointer is stored in the current ciEnv:
 464   ciEnv* env = compile->env();
 465   assert(env == ciEnv::current(), "must already be a ciEnv active");
 466   assert(env->compiler_data() == NULL, "compile already active?");
 467   env->set_compiler_data(compile);
 468   assert(compile == Compile::current(), "sanity");
 469 
 470   compile->set_type_dict(NULL);
 471   compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
 472   compile->clone_map().set_clone_idx(0);
 473   compile->set_type_hwm(NULL);
 474   compile->set_type_last_size(0);
 475   compile->set_last_tf(NULL, NULL);
 476   compile->set_indexSet_arena(NULL);
 477   compile->set_indexSet_free_block_list(NULL);
 478   compile->init_type_arena();
 479   Type::Initialize(compile);
 480   _compile->set_scratch_buffer_blob(NULL);
 481   _compile->begin_method();
 482   _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
 483 }
 484 CompileWrapper::~CompileWrapper() {
 485   _compile->end_method();
 486   if (_compile->scratch_buffer_blob() != NULL)
 487     BufferBlob::free(_compile->scratch_buffer_blob());
 488   _compile->env()->set_compiler_data(NULL);
 489 }
 490 
 491 
 492 //----------------------------print_compile_messages---------------------------
 493 void Compile::print_compile_messages() {
 494 #ifndef PRODUCT
 495   // Check if recompiling
 496   if (_subsume_loads == false && PrintOpto) {
 497     // Recompiling without allowing machine instructions to subsume loads
 498     tty->print_cr("*********************************************************");
 499     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 500     tty->print_cr("*********************************************************");
 501   }
 502   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 503     // Recompiling without escape analysis
 504     tty->print_cr("*********************************************************");
 505     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 506     tty->print_cr("*********************************************************");
 507   }
 508   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 509     // Recompiling without boxing elimination
 510     tty->print_cr("*********************************************************");
 511     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 512     tty->print_cr("*********************************************************");
 513   }
 514   if (C->directive()->BreakAtCompileOption) {
 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 //-----------------------init_scratch_buffer_blob------------------------------
 534 // Construct a temporary BufferBlob and cache it for this compile.
 535 void Compile::init_scratch_buffer_blob(int const_size) {
 536   // If there is already a scratch buffer blob allocated and the
 537   // constant section is big enough, use it.  Otherwise free the
 538   // current and allocate a new one.
 539   BufferBlob* blob = scratch_buffer_blob();
 540   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 541     // Use the current blob.
 542   } else {
 543     if (blob != NULL) {
 544       BufferBlob::free(blob);
 545     }
 546 
 547     ResourceMark rm;
 548     _scratch_const_size = const_size;
 549     int size = C2Compiler::initial_code_buffer_size(const_size);
 550 #ifdef ASSERT
 551     if (C->has_scalarized_args()) {
 552       // Oop verification for loading object fields from scalarized value types in the new entry point requires lots of space
 553       size += 5120;
 554     }
 555 #endif
 556     blob = BufferBlob::create("Compile::scratch_buffer", size);
 557     // Record the buffer blob for next time.
 558     set_scratch_buffer_blob(blob);
 559     // Have we run out of code space?
 560     if (scratch_buffer_blob() == NULL) {
 561       // Let CompilerBroker disable further compilations.
 562       record_failure("Not enough space for scratch buffer in CodeCache");
 563       return;
 564     }
 565   }
 566 
 567   // Initialize the relocation buffers
 568   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 569   set_scratch_locs_memory(locs_buf);
 570 }
 571 
 572 
 573 //-----------------------scratch_emit_size-------------------------------------
 574 // Helper function that computes size by emitting code
 575 uint Compile::scratch_emit_size(const Node* n) {
 576   // Start scratch_emit_size section.
 577   set_in_scratch_emit_size(true);
 578 
 579   // Emit into a trash buffer and count bytes emitted.
 580   // This is a pretty expensive way to compute a size,
 581   // but it works well enough if seldom used.
 582   // All common fixed-size instructions are given a size
 583   // method by the AD file.
 584   // Note that the scratch buffer blob and locs memory are
 585   // allocated at the beginning of the compile task, and
 586   // may be shared by several calls to scratch_emit_size.
 587   // The allocation of the scratch buffer blob is particularly
 588   // expensive, since it has to grab the code cache lock.
 589   BufferBlob* blob = this->scratch_buffer_blob();
 590   assert(blob != NULL, "Initialize BufferBlob at start");
 591   assert(blob->size() > MAX_inst_size, "sanity");
 592   relocInfo* locs_buf = scratch_locs_memory();
 593   address blob_begin = blob->content_begin();
 594   address blob_end   = (address)locs_buf;
 595   assert(blob->contains(blob_end), "sanity");
 596   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 597   buf.initialize_consts_size(_scratch_const_size);
 598   buf.initialize_stubs_size(MAX_stubs_size);
 599   assert(locs_buf != NULL, "sanity");
 600   int lsize = MAX_locs_size / 3;
 601   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 602   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 603   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 604   // Mark as scratch buffer.
 605   buf.consts()->set_scratch_emit();
 606   buf.insts()->set_scratch_emit();
 607   buf.stubs()->set_scratch_emit();
 608 
 609   // Do the emission.
 610 
 611   Label fakeL; // Fake label for branch instructions.
 612   Label*   saveL = NULL;
 613   uint save_bnum = 0;
 614   bool is_branch = n->is_MachBranch();
 615   if (is_branch) {
 616     MacroAssembler masm(&buf);
 617     masm.bind(fakeL);
 618     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 619     n->as_MachBranch()->label_set(&fakeL, 0);
 620   } else if (n->is_MachProlog()) {
 621     saveL = ((MachPrologNode*)n)->_verified_entry;
 622     ((MachPrologNode*)n)->_verified_entry = &fakeL;
 623   }
 624   n->emit(buf, this->regalloc());
 625 
 626   // Emitting into the scratch buffer should not fail
 627   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
 628 
 629   // Restore label.
 630   if (is_branch) {
 631     n->as_MachBranch()->label_set(saveL, save_bnum);
 632   } else if (n->is_MachProlog()) {
 633     ((MachPrologNode*)n)->_verified_entry = saveL;
 634   }
 635 
 636   // End scratch_emit_size section.
 637   set_in_scratch_emit_size(false);
 638 
 639   return buf.insts_size();
 640 }
 641 
 642 
 643 // ============================================================================
 644 //------------------------------Compile standard-------------------------------
 645 debug_only( int Compile::_debug_idx = 100000; )
 646 
 647 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 648 // the continuation bci for on stack replacement.
 649 
 650 
 651 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 652                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
 653                 : Phase(Compiler),
 654                   _compile_id(ci_env->compile_id()),
 655                   _save_argument_registers(false),
 656                   _subsume_loads(subsume_loads),
 657                   _do_escape_analysis(do_escape_analysis),
 658                   _eliminate_boxing(eliminate_boxing),
 659                   _method(target),
 660                   _entry_bci(osr_bci),
 661                   _stub_function(NULL),
 662                   _stub_name(NULL),
 663                   _stub_entry_point(NULL),
 664                   _max_node_limit(MaxNodeLimit),
 665                   _orig_pc_slot(0),
 666                   _orig_pc_slot_offset_in_bytes(0),
 667                   _sp_inc_slot(0),
 668                   _sp_inc_slot_offset_in_bytes(0),
 669                   _inlining_progress(false),
 670                   _inlining_incrementally(false),
 671                   _do_cleanup(false),
 672                   _has_reserved_stack_access(target->has_reserved_stack_access()),
 673 #ifndef PRODUCT
 674                   _trace_opto_output(directive->TraceOptoOutputOption),
 675 #endif
 676                   _has_method_handle_invokes(false),
 677                   _comp_arena(mtCompiler),
 678                   _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
 679                   _env(ci_env),
 680                   _directive(directive),
 681                   _log(ci_env->log()),
 682                   _failure_reason(NULL),
 683                   _congraph(NULL),
 684 #ifndef PRODUCT
 685                   _printer(IdealGraphPrinter::printer()),
 686 #endif
 687                   _dead_node_list(comp_arena()),
 688                   _dead_node_count(0),
 689                   _node_arena(mtCompiler),
 690                   _old_arena(mtCompiler),
 691                   _mach_constant_base_node(NULL),
 692                   _Compile_types(mtCompiler),
 693                   _initial_gvn(NULL),
 694                   _for_igvn(NULL),
 695                   _warm_calls(NULL),
 696                   _late_inlines(comp_arena(), 2, 0, NULL),
 697                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 698                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 699                   _late_inlines_pos(0),
 700                   _number_of_mh_late_inlines(0),
 701                   _print_inlining_stream(NULL),
 702                   _print_inlining_list(NULL),
 703                   _print_inlining_idx(0),
 704                   _print_inlining_output(NULL),
 705                   _replay_inline_data(NULL),
 706                   _java_calls(0),
 707                   _inner_loops(0),
 708                   _interpreter_frame_size(0),
 709                   _node_bundling_limit(0),
 710                   _node_bundling_base(NULL),
 711                   _code_buffer("Compile::Fill_buffer"),
 712                   _scratch_const_size(-1),
 713                   _in_scratch_emit_size(false)
 714 #ifndef PRODUCT
 715                   , _in_dump_cnt(0)
 716 #endif
 717 {
 718   C = this;
 719 #ifndef PRODUCT
 720   if (_printer != NULL) {
 721     _printer->set_compile(this);
 722   }
 723 #endif
 724   CompileWrapper cw(this);
 725 
 726   if (CITimeVerbose) {
 727     tty->print(" ");
 728     target->holder()->name()->print();
 729     tty->print(".");
 730     target->print_short_name();
 731     tty->print("  ");
 732   }
 733   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 734   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 735 
 736 #ifndef PRODUCT
 737   bool print_opto_assembly = directive->PrintOptoAssemblyOption;
 738   if (!print_opto_assembly) {
 739     bool print_assembly = directive->PrintAssemblyOption;
 740     if (print_assembly && !Disassembler::can_decode()) {
 741       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 742       print_opto_assembly = true;
 743     }
 744   }
 745   set_print_assembly(print_opto_assembly);
 746   set_parsed_irreducible_loop(false);
 747 
 748   if (directive->ReplayInlineOption) {
 749     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 750   }
 751 #endif
 752   set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
 753   set_print_intrinsics(directive->PrintIntrinsicsOption);
 754   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 755 
 756   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 757     // Make sure the method being compiled gets its own MDO,
 758     // so we can at least track the decompile_count().
 759     // Need MDO to record RTM code generation state.
 760     method()->ensure_method_data();
 761   }
 762 
 763   Init(::AliasLevel);
 764 
 765 
 766   print_compile_messages();
 767 
 768   _ilt = InlineTree::build_inline_tree_root();
 769 
 770   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 771   assert(num_alias_types() >= AliasIdxRaw, "");
 772 
 773 #define MINIMUM_NODE_HASH  1023
 774   // Node list that Iterative GVN will start with
 775   Unique_Node_List for_igvn(comp_arena());
 776   set_for_igvn(&for_igvn);
 777 
 778   // GVN that will be run immediately on new nodes
 779   uint estimated_size = method()->code_size()*4+64;
 780   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 781   PhaseGVN gvn(node_arena(), estimated_size);
 782   set_initial_gvn(&gvn);
 783 
 784   print_inlining_init();
 785   { // Scope for timing the parser
 786     TracePhase tp("parse", &timers[_t_parser]);
 787 
 788     // Put top into the hash table ASAP.
 789     initial_gvn()->transform_no_reclaim(top());
 790 
 791     // Set up tf(), start(), and find a CallGenerator.
 792     CallGenerator* cg = NULL;
 793     if (is_osr_compilation()) {
 794       const TypeTuple *domain = StartOSRNode::osr_domain();
 795       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 796       init_tf(TypeFunc::make(domain, range));
 797       StartNode* s = new StartOSRNode(root(), domain);
 798       initial_gvn()->set_type_bottom(s);
 799       init_start(s);
 800       cg = CallGenerator::for_osr(method(), entry_bci());
 801     } else {
 802       // Normal case.
 803       init_tf(TypeFunc::make(method()));
 804       StartNode* s = new StartNode(root(), tf()->domain_cc());
 805       initial_gvn()->set_type_bottom(s);
 806       init_start(s);
 807       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
 808         // With java.lang.ref.reference.get() we must go through the
 809         // intrinsic - even when get() is the root
 810         // method of the compile - so that, if necessary, the value in
 811         // the referent field of the reference object gets recorded by
 812         // the pre-barrier code.
 813         cg = find_intrinsic(method(), false);
 814       }
 815       if (cg == NULL) {
 816         float past_uses = method()->interpreter_invocation_count();
 817         float expected_uses = past_uses;
 818         cg = CallGenerator::for_inline(method(), expected_uses);
 819       }
 820     }
 821     if (failing())  return;
 822     if (cg == NULL) {
 823       record_method_not_compilable("cannot parse method");
 824       return;
 825     }
 826     JVMState* jvms = build_start_state(start(), tf());
 827     if ((jvms = cg->generate(jvms)) == NULL) {
 828       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 829         record_method_not_compilable("method parse failed");
 830       }
 831       return;
 832     }
 833     GraphKit kit(jvms);
 834 
 835     if (!kit.stopped()) {
 836       // Accept return values, and transfer control we know not where.
 837       // This is done by a special, unique ReturnNode bound to root.
 838       return_values(kit.jvms());
 839     }
 840 
 841     if (kit.has_exceptions()) {
 842       // Any exceptions that escape from this call must be rethrown
 843       // to whatever caller is dynamically above us on the stack.
 844       // This is done by a special, unique RethrowNode bound to root.
 845       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 846     }
 847 
 848     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 849 
 850     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 851       inline_string_calls(true);
 852     }
 853 
 854     if (failing())  return;
 855 
 856     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 857 
 858     // Remove clutter produced by parsing.
 859     if (!failing()) {
 860       ResourceMark rm;
 861       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 862     }
 863   }
 864 
 865   // Note:  Large methods are capped off in do_one_bytecode().
 866   if (failing())  return;
 867 
 868   // After parsing, node notes are no longer automagic.
 869   // They must be propagated by register_new_node_with_optimizer(),
 870   // clone(), or the like.
 871   set_default_node_notes(NULL);
 872 
 873   for (;;) {
 874     int successes = Inline_Warm();
 875     if (failing())  return;
 876     if (successes == 0)  break;
 877   }
 878 
 879   // Drain the list.
 880   Finish_Warm();
 881 #ifndef PRODUCT
 882   if (_printer && _printer->should_print(1)) {
 883     _printer->print_inlining();
 884   }
 885 #endif
 886 
 887   if (failing())  return;
 888   NOT_PRODUCT( verify_graph_edges(); )
 889 
 890   // Now optimize
 891   Optimize();
 892   if (failing())  return;
 893   NOT_PRODUCT( verify_graph_edges(); )
 894 
 895 #ifndef PRODUCT
 896   if (PrintIdeal) {
 897     ttyLocker ttyl;  // keep the following output all in one block
 898     // This output goes directly to the tty, not the compiler log.
 899     // To enable tools to match it up with the compilation activity,
 900     // be sure to tag this tty output with the compile ID.
 901     if (xtty != NULL) {
 902       xtty->head("ideal compile_id='%d'%s", compile_id(),
 903                  is_osr_compilation()    ? " compile_kind='osr'" :
 904                  "");
 905     }
 906     root()->dump(9999);
 907     if (xtty != NULL) {
 908       xtty->tail("ideal");
 909     }
 910   }
 911 #endif
 912 
 913 #ifdef ASSERT
 914   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 915   bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
 916 #endif
 917 
 918   // Dump compilation data to replay it.
 919   if (directive->DumpReplayOption) {
 920     env()->dump_replay_data(_compile_id);
 921   }
 922   if (directive->DumpInlineOption && (ilt() != NULL)) {
 923     env()->dump_inline_data(_compile_id);
 924   }
 925 
 926   // Now that we know the size of all the monitors we can add a fixed slot
 927   // for the original deopt pc.
 928 
 929   _orig_pc_slot = fixed_slots();
 930   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 931 
 932   if (needs_stack_repair()) {
 933     // One extra slot for the special stack increment value
 934     _sp_inc_slot = next_slot;
 935     next_slot += 2;
 936   }
 937 
 938   set_fixed_slots(next_slot);
 939 
 940   // Compute when to use implicit null checks. Used by matching trap based
 941   // nodes and NullCheck optimization.
 942   set_allowed_deopt_reasons();
 943 
 944   // Now generate code
 945   Code_Gen();
 946   if (failing())  return;
 947 
 948   // Check if we want to skip execution of all compiled code.
 949   {
 950 #ifndef PRODUCT
 951     if (OptoNoExecute) {
 952       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 953       return;
 954     }
 955 #endif
 956     TracePhase tp("install_code", &timers[_t_registerMethod]);
 957 
 958     if (is_osr_compilation()) {
 959       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 960       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 961     } else {
 962       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 963       if (_code_offsets.value(CodeOffsets::Verified_Value_Entry) == -1) {
 964         _code_offsets.set_value(CodeOffsets::Verified_Value_Entry, _first_block_size);
 965       }
 966       if (_code_offsets.value(CodeOffsets::Verified_Value_Entry_RO) == -1) {
 967         _code_offsets.set_value(CodeOffsets::Verified_Value_Entry_RO, _first_block_size);
 968       }
 969       if (_code_offsets.value(CodeOffsets::Entry) == -1) {
 970         _code_offsets.set_value(CodeOffsets::Entry, _first_block_size);
 971       }
 972       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 973     }
 974 
 975     env()->register_method(_method, _entry_bci,
 976                            &_code_offsets,
 977                            _orig_pc_slot_offset_in_bytes,
 978                            code_buffer(),
 979                            frame_size_in_words(), _oop_map_set,
 980                            &_handler_table, &_inc_table,
 981                            compiler,
 982                            has_unsafe_access(),
 983                            SharedRuntime::is_wide_vector(max_vector_size()),
 984                            rtm_state()
 985                            );
 986 
 987     if (log() != NULL) // Print code cache state into compiler log
 988       log()->code_cache_state();
 989   }
 990 }
 991 
 992 //------------------------------Compile----------------------------------------
 993 // Compile a runtime stub
 994 Compile::Compile( ciEnv* ci_env,
 995                   TypeFunc_generator generator,
 996                   address stub_function,
 997                   const char *stub_name,
 998                   int is_fancy_jump,
 999                   bool pass_tls,
1000                   bool save_arg_registers,
1001                   bool return_pc,
1002                   DirectiveSet* directive)
1003   : Phase(Compiler),
1004     _compile_id(0),
1005     _save_argument_registers(save_arg_registers),
1006     _subsume_loads(true),
1007     _do_escape_analysis(false),
1008     _eliminate_boxing(false),
1009     _method(NULL),
1010     _entry_bci(InvocationEntryBci),
1011     _stub_function(stub_function),
1012     _stub_name(stub_name),
1013     _stub_entry_point(NULL),
1014     _max_node_limit(MaxNodeLimit),
1015     _orig_pc_slot(0),
1016     _orig_pc_slot_offset_in_bytes(0),
1017     _sp_inc_slot(0),
1018     _sp_inc_slot_offset_in_bytes(0),
1019     _inlining_progress(false),
1020     _inlining_incrementally(false),
1021     _has_reserved_stack_access(false),
1022 #ifndef PRODUCT
1023     _trace_opto_output(directive->TraceOptoOutputOption),
1024 #endif
1025     _has_method_handle_invokes(false),
1026     _comp_arena(mtCompiler),
1027     _env(ci_env),
1028     _directive(directive),
1029     _log(ci_env->log()),
1030     _failure_reason(NULL),
1031     _congraph(NULL),
1032 #ifndef PRODUCT
1033     _printer(NULL),
1034 #endif
1035     _dead_node_list(comp_arena()),
1036     _dead_node_count(0),
1037     _node_arena(mtCompiler),
1038     _old_arena(mtCompiler),
1039     _mach_constant_base_node(NULL),
1040     _Compile_types(mtCompiler),
1041     _initial_gvn(NULL),
1042     _for_igvn(NULL),
1043     _warm_calls(NULL),
1044     _number_of_mh_late_inlines(0),
1045     _print_inlining_stream(NULL),
1046     _print_inlining_list(NULL),
1047     _print_inlining_idx(0),
1048     _print_inlining_output(NULL),
1049     _replay_inline_data(NULL),
1050     _java_calls(0),
1051     _inner_loops(0),
1052     _interpreter_frame_size(0),
1053     _node_bundling_limit(0),
1054     _node_bundling_base(NULL),
1055     _code_buffer("Compile::Fill_buffer"),
1056 #ifndef PRODUCT
1057     _in_dump_cnt(0),
1058 #endif
1059     _allowed_reasons(0) {
1060   C = this;
1061 
1062   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1063   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1064 
1065 #ifndef PRODUCT
1066   set_print_assembly(PrintFrameConverterAssembly);
1067   set_parsed_irreducible_loop(false);
1068 #endif
1069   set_has_irreducible_loop(false); // no loops
1070 
1071   CompileWrapper cw(this);
1072   Init(/*AliasLevel=*/ 0);
1073   init_tf((*generator)());
1074 
1075   {
1076     // The following is a dummy for the sake of GraphKit::gen_stub
1077     Unique_Node_List for_igvn(comp_arena());
1078     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1079     PhaseGVN gvn(Thread::current()->resource_area(),255);
1080     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1081     gvn.transform_no_reclaim(top());
1082 
1083     GraphKit kit;
1084     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1085   }
1086 
1087   NOT_PRODUCT( verify_graph_edges(); )
1088   Code_Gen();
1089   if (failing())  return;
1090 
1091 
1092   // Entry point will be accessed using compile->stub_entry_point();
1093   if (code_buffer() == NULL) {
1094     Matcher::soft_match_failure();
1095   } else {
1096     if (PrintAssembly && (WizardMode || Verbose))
1097       tty->print_cr("### Stub::%s", stub_name);
1098 
1099     if (!failing()) {
1100       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1101 
1102       // Make the NMethod
1103       // For now we mark the frame as never safe for profile stackwalking
1104       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1105                                                       code_buffer(),
1106                                                       CodeOffsets::frame_never_safe,
1107                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1108                                                       frame_size_in_words(),
1109                                                       _oop_map_set,
1110                                                       save_arg_registers);
1111       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1112 
1113       _stub_entry_point = rs->entry_point();
1114     }
1115   }
1116 }
1117 
1118 //------------------------------Init-------------------------------------------
1119 // Prepare for a single compilation
1120 void Compile::Init(int aliaslevel) {
1121   _unique  = 0;
1122   _regalloc = NULL;
1123 
1124   _tf      = NULL;  // filled in later
1125   _top     = NULL;  // cached later
1126   _matcher = NULL;  // filled in later
1127   _cfg     = NULL;  // filled in later
1128 
1129   set_24_bit_selection_and_mode(Use24BitFP, false);
1130 
1131   _node_note_array = NULL;
1132   _default_node_notes = NULL;
1133   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1134 
1135   _immutable_memory = NULL; // filled in at first inquiry
1136 
1137   // Globally visible Nodes
1138   // First set TOP to NULL to give safe behavior during creation of RootNode
1139   set_cached_top_node(NULL);
1140   set_root(new RootNode());
1141   // Now that you have a Root to point to, create the real TOP
1142   set_cached_top_node( new ConNode(Type::TOP) );
1143   set_recent_alloc(NULL, NULL);
1144 
1145   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1146   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1147   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1148   env()->set_dependencies(new Dependencies(env()));
1149 
1150   _fixed_slots = 0;
1151   set_has_split_ifs(false);
1152   set_has_loops(has_method() && method()->has_loops()); // first approximation
1153   set_has_stringbuilder(false);
1154   set_has_boxed_value(false);
1155   _trap_can_recompile = false;  // no traps emitted yet
1156   _major_progress = true; // start out assuming good things will happen
1157   set_has_unsafe_access(false);
1158   set_max_vector_size(0);
1159   set_clear_upper_avx(false);  //false as default for clear upper bits of ymm registers
1160   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1161   set_decompile_count(0);
1162 
1163   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1164   _loop_opts_cnt = LoopOptsCount;
1165   set_do_inlining(Inline);
1166   set_max_inline_size(MaxInlineSize);
1167   set_freq_inline_size(FreqInlineSize);
1168   set_do_scheduling(OptoScheduling);
1169   set_do_count_invocations(false);
1170   set_do_method_data_update(false);
1171 
1172   set_do_vector_loop(false);
1173 
1174   if (AllowVectorizeOnDemand) {
1175     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1176       set_do_vector_loop(true);
1177       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());})
1178     } else if (has_method() && method()->name() != 0 &&
1179                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1180       set_do_vector_loop(true);
1181     }
1182   }
1183   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1184   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());})
1185 
1186   set_age_code(has_method() && method()->profile_aging());
1187   set_rtm_state(NoRTM); // No RTM lock eliding by default
1188   _max_node_limit = _directive->MaxNodeLimitOption;
1189 
1190 #if INCLUDE_RTM_OPT
1191   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1192     int rtm_state = method()->method_data()->rtm_state();
1193     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1194       // Don't generate RTM lock eliding code.
1195       set_rtm_state(NoRTM);
1196     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1197       // Generate RTM lock eliding code without abort ratio calculation code.
1198       set_rtm_state(UseRTM);
1199     } else if (UseRTMDeopt) {
1200       // Generate RTM lock eliding code and include abort ratio calculation
1201       // code if UseRTMDeopt is on.
1202       set_rtm_state(ProfileRTM);
1203     }
1204   }
1205 #endif
1206   if (debug_info()->recording_non_safepoints()) {
1207     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1208                         (comp_arena(), 8, 0, NULL));
1209     set_default_node_notes(Node_Notes::make(this));
1210   }
1211 
1212   // // -- Initialize types before each compile --
1213   // // Update cached type information
1214   // if( _method && _method->constants() )
1215   //   Type::update_loaded_types(_method, _method->constants());
1216 
1217   // Init alias_type map.
1218   if (!_do_escape_analysis && aliaslevel == 3)
1219     aliaslevel = 2;  // No unique types without escape analysis
1220   _AliasLevel = aliaslevel;
1221   const int grow_ats = 16;
1222   _max_alias_types = grow_ats;
1223   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1224   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1225   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1226   {
1227     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1228   }
1229   // Initialize the first few types.
1230   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1231   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1232   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1233   _num_alias_types = AliasIdxRaw+1;
1234   // Zero out the alias type cache.
1235   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1236   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1237   probe_alias_cache(NULL)->_index = AliasIdxTop;
1238 
1239   _intrinsics = NULL;
1240   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1241   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1242   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1243   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1244   _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1245   _value_type_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
1246   register_library_intrinsics();
1247 }
1248 
1249 //---------------------------init_start----------------------------------------
1250 // Install the StartNode on this compile object.
1251 void Compile::init_start(StartNode* s) {
1252   if (failing())
1253     return; // already failing
1254   assert(s == start(), "");
1255 }
1256 
1257 /**
1258  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1259  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1260  * the ideal graph.
1261  */
1262 StartNode* Compile::start() const {
1263   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1264   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1265     Node* start = root()->fast_out(i);
1266     if (start->is_Start()) {
1267       return start->as_Start();
1268     }
1269   }
1270   fatal("Did not find Start node!");
1271   return NULL;
1272 }
1273 
1274 //-------------------------------immutable_memory-------------------------------------
1275 // Access immutable memory
1276 Node* Compile::immutable_memory() {
1277   if (_immutable_memory != NULL) {
1278     return _immutable_memory;
1279   }
1280   StartNode* s = start();
1281   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1282     Node *p = s->fast_out(i);
1283     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1284       _immutable_memory = p;
1285       return _immutable_memory;
1286     }
1287   }
1288   ShouldNotReachHere();
1289   return NULL;
1290 }
1291 
1292 //----------------------set_cached_top_node------------------------------------
1293 // Install the cached top node, and make sure Node::is_top works correctly.
1294 void Compile::set_cached_top_node(Node* tn) {
1295   if (tn != NULL)  verify_top(tn);
1296   Node* old_top = _top;
1297   _top = tn;
1298   // Calling Node::setup_is_top allows the nodes the chance to adjust
1299   // their _out arrays.
1300   if (_top != NULL)     _top->setup_is_top();
1301   if (old_top != NULL)  old_top->setup_is_top();
1302   assert(_top == NULL || top()->is_top(), "");
1303 }
1304 
1305 #ifdef ASSERT
1306 uint Compile::count_live_nodes_by_graph_walk() {
1307   Unique_Node_List useful(comp_arena());
1308   // Get useful node list by walking the graph.
1309   identify_useful_nodes(useful);
1310   return useful.size();
1311 }
1312 
1313 void Compile::print_missing_nodes() {
1314 
1315   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1316   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1317     return;
1318   }
1319 
1320   // This is an expensive function. It is executed only when the user
1321   // specifies VerifyIdealNodeCount option or otherwise knows the
1322   // additional work that needs to be done to identify reachable nodes
1323   // by walking the flow graph and find the missing ones using
1324   // _dead_node_list.
1325 
1326   Unique_Node_List useful(comp_arena());
1327   // Get useful node list by walking the graph.
1328   identify_useful_nodes(useful);
1329 
1330   uint l_nodes = C->live_nodes();
1331   uint l_nodes_by_walk = useful.size();
1332 
1333   if (l_nodes != l_nodes_by_walk) {
1334     if (_log != NULL) {
1335       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1336       _log->stamp();
1337       _log->end_head();
1338     }
1339     VectorSet& useful_member_set = useful.member_set();
1340     int last_idx = l_nodes_by_walk;
1341     for (int i = 0; i < last_idx; i++) {
1342       if (useful_member_set.test(i)) {
1343         if (_dead_node_list.test(i)) {
1344           if (_log != NULL) {
1345             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1346           }
1347           if (PrintIdealNodeCount) {
1348             // Print the log message to tty
1349               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1350               useful.at(i)->dump();
1351           }
1352         }
1353       }
1354       else if (! _dead_node_list.test(i)) {
1355         if (_log != NULL) {
1356           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1357         }
1358         if (PrintIdealNodeCount) {
1359           // Print the log message to tty
1360           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1361         }
1362       }
1363     }
1364     if (_log != NULL) {
1365       _log->tail("mismatched_nodes");
1366     }
1367   }
1368 }
1369 void Compile::record_modified_node(Node* n) {
1370   if (_modified_nodes != NULL && !_inlining_incrementally &&
1371       n->outcnt() != 0 && !n->is_Con()) {
1372     _modified_nodes->push(n);
1373   }
1374 }
1375 
1376 void Compile::remove_modified_node(Node* n) {
1377   if (_modified_nodes != NULL) {
1378     _modified_nodes->remove(n);
1379   }
1380 }
1381 #endif
1382 
1383 #ifndef PRODUCT
1384 void Compile::verify_top(Node* tn) const {
1385   if (tn != NULL) {
1386     assert(tn->is_Con(), "top node must be a constant");
1387     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1388     assert(tn->in(0) != NULL, "must have live top node");
1389   }
1390 }
1391 #endif
1392 
1393 
1394 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1395 
1396 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1397   guarantee(arr != NULL, "");
1398   int num_blocks = arr->length();
1399   if (grow_by < num_blocks)  grow_by = num_blocks;
1400   int num_notes = grow_by * _node_notes_block_size;
1401   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1402   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1403   while (num_notes > 0) {
1404     arr->append(notes);
1405     notes     += _node_notes_block_size;
1406     num_notes -= _node_notes_block_size;
1407   }
1408   assert(num_notes == 0, "exact multiple, please");
1409 }
1410 
1411 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1412   if (source == NULL || dest == NULL)  return false;
1413 
1414   if (dest->is_Con())
1415     return false;               // Do not push debug info onto constants.
1416 
1417 #ifdef ASSERT
1418   // Leave a bread crumb trail pointing to the original node:
1419   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1420     dest->set_debug_orig(source);
1421   }
1422 #endif
1423 
1424   if (node_note_array() == NULL)
1425     return false;               // Not collecting any notes now.
1426 
1427   // This is a copy onto a pre-existing node, which may already have notes.
1428   // If both nodes have notes, do not overwrite any pre-existing notes.
1429   Node_Notes* source_notes = node_notes_at(source->_idx);
1430   if (source_notes == NULL || source_notes->is_clear())  return false;
1431   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1432   if (dest_notes == NULL || dest_notes->is_clear()) {
1433     return set_node_notes_at(dest->_idx, source_notes);
1434   }
1435 
1436   Node_Notes merged_notes = (*source_notes);
1437   // The order of operations here ensures that dest notes will win...
1438   merged_notes.update_from(dest_notes);
1439   return set_node_notes_at(dest->_idx, &merged_notes);
1440 }
1441 
1442 
1443 //--------------------------allow_range_check_smearing-------------------------
1444 // Gating condition for coalescing similar range checks.
1445 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1446 // single covering check that is at least as strong as any of them.
1447 // If the optimization succeeds, the simplified (strengthened) range check
1448 // will always succeed.  If it fails, we will deopt, and then give up
1449 // on the optimization.
1450 bool Compile::allow_range_check_smearing() const {
1451   // If this method has already thrown a range-check,
1452   // assume it was because we already tried range smearing
1453   // and it failed.
1454   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1455   return !already_trapped;
1456 }
1457 
1458 
1459 //------------------------------flatten_alias_type-----------------------------
1460 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1461   int offset = tj->offset();
1462   TypePtr::PTR ptr = tj->ptr();
1463 
1464   // Known instance (scalarizable allocation) alias only with itself.
1465   bool is_known_inst = tj->isa_oopptr() != NULL &&
1466                        tj->is_oopptr()->is_known_instance();
1467 
1468   // Process weird unsafe references.
1469   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1470     bool default_value_load = EnableValhalla && tj->is_instptr()->klass() == ciEnv::current()->Class_klass();
1471     assert(InlineUnsafeOps || default_value_load, "indeterminate pointers come only from unsafe ops");
1472     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1473     tj = TypeOopPtr::BOTTOM;
1474     ptr = tj->ptr();
1475     offset = tj->offset();
1476   }
1477 
1478   // Array pointers need some flattening
1479   const TypeAryPtr *ta = tj->isa_aryptr();
1480   if (ta && ta->is_stable()) {
1481     // Erase stability property for alias analysis.
1482     tj = ta = ta->cast_to_stable(false);
1483   }
1484   if( ta && is_known_inst ) {
1485     if ( offset != Type::OffsetBot &&
1486          offset > arrayOopDesc::length_offset_in_bytes() ) {
1487       offset = Type::OffsetBot; // Flatten constant access into array body only
1488       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id());
1489     }
1490   } else if( ta && _AliasLevel >= 2 ) {
1491     // For arrays indexed by constant indices, we flatten the alias
1492     // space to include all of the array body.  Only the header, klass
1493     // and array length can be accessed un-aliased.
1494     // For flattened value type array, each field has its own slice so
1495     // we must include the field offset.
1496     if( offset != Type::OffsetBot ) {
1497       if( ta->const_oop() ) { // MethodData* or Method*
1498         offset = Type::OffsetBot;   // Flatten constant access into array body
1499         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1500       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1501         // range is OK as-is.
1502         tj = ta = TypeAryPtr::RANGE;
1503       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1504         tj = TypeInstPtr::KLASS; // all klass loads look alike
1505         ta = TypeAryPtr::RANGE; // generic ignored junk
1506         ptr = TypePtr::BotPTR;
1507       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1508         tj = TypeInstPtr::MARK;
1509         ta = TypeAryPtr::RANGE; // generic ignored junk
1510         ptr = TypePtr::BotPTR;
1511       } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1512         ta = tj->isa_aryptr();
1513       } else {                  // Random constant offset into array body
1514         offset = Type::OffsetBot;   // Flatten constant access into array body
1515         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1516       }
1517     }
1518     // Arrays of fixed size alias with arrays of unknown size.
1519     if (ta->size() != TypeInt::POS) {
1520       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1521       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset());
1522     }
1523     // Arrays of known objects become arrays of unknown objects.
1524     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1525       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1526       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1527     }
1528     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1529       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1530       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1531     }
1532     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1533     // cannot be distinguished by bytecode alone.
1534     if (ta->elem() == TypeInt::BOOL) {
1535       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1536       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1537       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1538     }
1539     // During the 2nd round of IterGVN, NotNull castings are removed.
1540     // Make sure the Bottom and NotNull variants alias the same.
1541     // Also, make sure exact and non-exact variants alias the same.
1542     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1543       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1544     }
1545   }
1546 
1547   // Oop pointers need some flattening
1548   const TypeInstPtr *to = tj->isa_instptr();
1549   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1550     ciInstanceKlass *k = to->klass()->as_instance_klass();
1551     if( ptr == TypePtr::Constant ) {
1552       if (to->klass() != ciEnv::current()->Class_klass() ||
1553           offset < k->size_helper() * wordSize) {
1554         // No constant oop pointers (such as Strings); they alias with
1555         // unknown strings.
1556         assert(!is_known_inst, "not scalarizable allocation");
1557         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1558       }
1559     } else if( is_known_inst ) {
1560       tj = to; // Keep NotNull and klass_is_exact for instance type
1561     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1562       // During the 2nd round of IterGVN, NotNull castings are removed.
1563       // Make sure the Bottom and NotNull variants alias the same.
1564       // Also, make sure exact and non-exact variants alias the same.
1565       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1566     }
1567     if (to->speculative() != NULL) {
1568       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id());
1569     }
1570     // Canonicalize the holder of this field
1571     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1572       // First handle header references such as a LoadKlassNode, even if the
1573       // object's klass is unloaded at compile time (4965979).
1574       if (!is_known_inst) { // Do it only for non-instance types
1575         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset));
1576       }
1577     } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1578       to = tj->is_instptr();
1579     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1580       // Static fields are in the space above the normal instance
1581       // fields in the java.lang.Class instance.
1582       if (to->klass() != ciEnv::current()->Class_klass()) {
1583         to = NULL;
1584         tj = TypeOopPtr::BOTTOM;
1585         offset = tj->offset();
1586       }
1587     } else {
1588       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1589       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1590         if( is_known_inst ) {
1591           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id());
1592         } else {
1593           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
1594         }
1595       }
1596     }
1597   }
1598 
1599   // Klass pointers to object array klasses need some flattening
1600   const TypeKlassPtr *tk = tj->isa_klassptr();
1601   if( tk ) {
1602     // If we are referencing a field within a Klass, we need
1603     // to assume the worst case of an Object.  Both exact and
1604     // inexact types must flatten to the same alias class so
1605     // use NotNull as the PTR.
1606     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1607 
1608       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1609                                    TypeKlassPtr::OBJECT->klass(),
1610                                    Type::Offset(offset));
1611     }
1612 
1613     ciKlass* klass = tk->klass();
1614     if (klass != NULL && klass->is_obj_array_klass()) {
1615       ciKlass* k = TypeAryPtr::OOPS->klass();
1616       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1617         k = TypeInstPtr::BOTTOM->klass();
1618       tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset));
1619     }
1620 
1621     // Check for precise loads from the primary supertype array and force them
1622     // to the supertype cache alias index.  Check for generic array loads from
1623     // the primary supertype array and also force them to the supertype cache
1624     // alias index.  Since the same load can reach both, we need to merge
1625     // these 2 disparate memories into the same alias class.  Since the
1626     // primary supertype array is read-only, there's no chance of confusion
1627     // where we bypass an array load and an array store.
1628     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1629     if (offset == Type::OffsetBot ||
1630         (offset >= primary_supers_offset &&
1631          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1632         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1633       offset = in_bytes(Klass::secondary_super_cache_offset());
1634       tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset));
1635     }
1636   }
1637 
1638   // Flatten all Raw pointers together.
1639   if (tj->base() == Type::RawPtr)
1640     tj = TypeRawPtr::BOTTOM;
1641 
1642   if (tj->base() == Type::AnyPtr)
1643     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1644 
1645   // Flatten all to bottom for now
1646   switch( _AliasLevel ) {
1647   case 0:
1648     tj = TypePtr::BOTTOM;
1649     break;
1650   case 1:                       // Flatten to: oop, static, field or array
1651     switch (tj->base()) {
1652     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1653     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1654     case Type::AryPtr:   // do not distinguish arrays at all
1655     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1656     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1657     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1658     default: ShouldNotReachHere();
1659     }
1660     break;
1661   case 2:                       // No collapsing at level 2; keep all splits
1662   case 3:                       // No collapsing at level 3; keep all splits
1663     break;
1664   default:
1665     Unimplemented();
1666   }
1667 
1668   offset = tj->offset();
1669   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1670 
1671   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1672           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1673           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1674           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1675           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1676           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1677           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1678           (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1679           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1680   assert( tj->ptr() != TypePtr::TopPTR &&
1681           tj->ptr() != TypePtr::AnyNull &&
1682           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1683 //    assert( tj->ptr() != TypePtr::Constant ||
1684 //            tj->base() == Type::RawPtr ||
1685 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1686 
1687   return tj;
1688 }
1689 
1690 void Compile::AliasType::Init(int i, const TypePtr* at) {
1691   _index = i;
1692   _adr_type = at;
1693   _field = NULL;
1694   _element = NULL;
1695   _is_rewritable = true; // default
1696   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1697   if (atoop != NULL && atoop->is_known_instance()) {
1698     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1699     _general_index = Compile::current()->get_alias_index(gt);
1700   } else {
1701     _general_index = 0;
1702   }
1703 }
1704 
1705 BasicType Compile::AliasType::basic_type() const {
1706   if (element() != NULL) {
1707     const Type* element = adr_type()->is_aryptr()->elem();
1708     return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1709   } if (field() != NULL) {
1710     return field()->layout_type();
1711   } else {
1712     return T_ILLEGAL; // unknown
1713   }
1714 }
1715 
1716 //---------------------------------print_on------------------------------------
1717 #ifndef PRODUCT
1718 void Compile::AliasType::print_on(outputStream* st) {
1719   if (index() < 10)
1720         st->print("@ <%d> ", index());
1721   else  st->print("@ <%d>",  index());
1722   st->print(is_rewritable() ? "   " : " RO");
1723   int offset = adr_type()->offset();
1724   if (offset == Type::OffsetBot)
1725         st->print(" +any");
1726   else  st->print(" +%-3d", offset);
1727   st->print(" in ");
1728   adr_type()->dump_on(st);
1729   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1730   if (field() != NULL && tjp) {
1731     if (tjp->klass()  != field()->holder() ||
1732         tjp->offset() != field()->offset_in_bytes()) {
1733       st->print(" != ");
1734       field()->print();
1735       st->print(" ***");
1736     }
1737   }
1738 }
1739 
1740 void print_alias_types() {
1741   Compile* C = Compile::current();
1742   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1743   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1744     C->alias_type(idx)->print_on(tty);
1745     tty->cr();
1746   }
1747 }
1748 #endif
1749 
1750 
1751 //----------------------------probe_alias_cache--------------------------------
1752 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1753   intptr_t key = (intptr_t) adr_type;
1754   key ^= key >> logAliasCacheSize;
1755   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1756 }
1757 
1758 
1759 //-----------------------------grow_alias_types--------------------------------
1760 void Compile::grow_alias_types() {
1761   const int old_ats  = _max_alias_types; // how many before?
1762   const int new_ats  = old_ats;          // how many more?
1763   const int grow_ats = old_ats+new_ats;  // how many now?
1764   _max_alias_types = grow_ats;
1765   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1766   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1767   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1768   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1769 }
1770 
1771 
1772 //--------------------------------find_alias_type------------------------------
1773 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1774   if (_AliasLevel == 0)
1775     return alias_type(AliasIdxBot);
1776 
1777   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1778   if (ace->_adr_type == adr_type) {
1779     return alias_type(ace->_index);
1780   }
1781 
1782   // Handle special cases.
1783   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1784   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1785 
1786   // Do it the slow way.
1787   const TypePtr* flat = flatten_alias_type(adr_type);
1788 
1789 #ifdef ASSERT
1790   {
1791     ResourceMark rm;
1792     assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1793            Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1794     assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1795            Type::str(adr_type));
1796     if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1797       const TypeOopPtr* foop = flat->is_oopptr();
1798       // Scalarizable allocations have exact klass always.
1799       bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1800       const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1801       assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1802              Type::str(foop), Type::str(xoop));
1803     }
1804   }
1805 #endif
1806 
1807   int idx = AliasIdxTop;
1808   for (int i = 0; i < num_alias_types(); i++) {
1809     if (alias_type(i)->adr_type() == flat) {
1810       idx = i;
1811       break;
1812     }
1813   }
1814 
1815   if (idx == AliasIdxTop) {
1816     if (no_create)  return NULL;
1817     // Grow the array if necessary.
1818     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1819     // Add a new alias type.
1820     idx = _num_alias_types++;
1821     _alias_types[idx]->Init(idx, flat);
1822     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1823     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1824     if (flat->isa_instptr()) {
1825       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1826           && flat->is_instptr()->klass() == env()->Class_klass())
1827         alias_type(idx)->set_rewritable(false);
1828     }
1829     ciField* field = NULL;
1830     if (flat->isa_aryptr()) {
1831 #ifdef ASSERT
1832       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1833       // (T_BYTE has the weakest alignment and size restrictions...)
1834       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1835 #endif
1836       const Type* elemtype = flat->is_aryptr()->elem();
1837       if (flat->offset() == TypePtr::OffsetBot) {
1838         alias_type(idx)->set_element(elemtype);
1839       }
1840       int field_offset = flat->is_aryptr()->field_offset().get();
1841       if (elemtype->isa_valuetype() && field_offset != Type::OffsetBot) {
1842         ciValueKlass* vk = elemtype->is_valuetype()->value_klass();
1843         field_offset += vk->first_field_offset();
1844         field = vk->get_field_by_offset(field_offset, false);
1845       }
1846     }
1847     if (flat->isa_klassptr()) {
1848       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1849         alias_type(idx)->set_rewritable(false);
1850       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1851         alias_type(idx)->set_rewritable(false);
1852       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1853         alias_type(idx)->set_rewritable(false);
1854       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1855         alias_type(idx)->set_rewritable(false);
1856     }
1857     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1858     // but the base pointer type is not distinctive enough to identify
1859     // references into JavaThread.)
1860 
1861     // Check for final fields.
1862     const TypeInstPtr* tinst = flat->isa_instptr();
1863     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1864       if (tinst->const_oop() != NULL &&
1865           tinst->klass() == ciEnv::current()->Class_klass() &&
1866           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1867         // static field
1868         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1869         field = k->get_field_by_offset(tinst->offset(), true);
1870       } else if (tinst->klass()->is_valuetype()) {
1871         // Value type field
1872         ciValueKlass* vk = tinst->value_klass();
1873         field = vk->get_field_by_offset(tinst->offset(), false);
1874       } else {
1875         ciInstanceKlass* k = tinst->klass()->as_instance_klass();
1876         field = k->get_field_by_offset(tinst->offset(), false);
1877       }
1878     }
1879     assert(field == NULL ||
1880            original_field == NULL ||
1881            (field->holder() == original_field->holder() &&
1882             field->offset() == original_field->offset() &&
1883             field->is_static() == original_field->is_static()), "wrong field?");
1884     // Set field() and is_rewritable() attributes.
1885     if (field != NULL)  alias_type(idx)->set_field(field);
1886   }
1887 
1888   // Fill the cache for next time.
1889   ace->_adr_type = adr_type;
1890   ace->_index    = idx;
1891   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1892 
1893   // Might as well try to fill the cache for the flattened version, too.
1894   AliasCacheEntry* face = probe_alias_cache(flat);
1895   if (face->_adr_type == NULL) {
1896     face->_adr_type = flat;
1897     face->_index    = idx;
1898     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1899   }
1900 
1901   return alias_type(idx);
1902 }
1903 
1904 
1905 Compile::AliasType* Compile::alias_type(ciField* field) {
1906   const TypeOopPtr* t;
1907   if (field->is_static())
1908     t = TypeInstPtr::make(field->holder()->java_mirror());
1909   else
1910     t = TypeOopPtr::make_from_klass_raw(field->holder());
1911   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1912   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1913   return atp;
1914 }
1915 
1916 
1917 //------------------------------have_alias_type--------------------------------
1918 bool Compile::have_alias_type(const TypePtr* adr_type) {
1919   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1920   if (ace->_adr_type == adr_type) {
1921     return true;
1922   }
1923 
1924   // Handle special cases.
1925   if (adr_type == NULL)             return true;
1926   if (adr_type == TypePtr::BOTTOM)  return true;
1927 
1928   return find_alias_type(adr_type, true, NULL) != NULL;
1929 }
1930 
1931 //-----------------------------must_alias--------------------------------------
1932 // True if all values of the given address type are in the given alias category.
1933 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1934   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1935   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1936   if (alias_idx == AliasIdxTop)         return false; // the empty category
1937   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1938 
1939   // the only remaining possible overlap is identity
1940   int adr_idx = get_alias_index(adr_type);
1941   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1942   assert(adr_idx == alias_idx ||
1943          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1944           && adr_type                       != TypeOopPtr::BOTTOM),
1945          "should not be testing for overlap with an unsafe pointer");
1946   return adr_idx == alias_idx;
1947 }
1948 
1949 //------------------------------can_alias--------------------------------------
1950 // True if any values of the given address type are in the given alias category.
1951 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1952   if (alias_idx == AliasIdxTop)         return false; // the empty category
1953   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1954   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1955   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1956 
1957   // the only remaining possible overlap is identity
1958   int adr_idx = get_alias_index(adr_type);
1959   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1960   return adr_idx == alias_idx;
1961 }
1962 
1963 
1964 
1965 //---------------------------pop_warm_call-------------------------------------
1966 WarmCallInfo* Compile::pop_warm_call() {
1967   WarmCallInfo* wci = _warm_calls;
1968   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1969   return wci;
1970 }
1971 
1972 //----------------------------Inline_Warm--------------------------------------
1973 int Compile::Inline_Warm() {
1974   // If there is room, try to inline some more warm call sites.
1975   // %%% Do a graph index compaction pass when we think we're out of space?
1976   if (!InlineWarmCalls)  return 0;
1977 
1978   int calls_made_hot = 0;
1979   int room_to_grow   = NodeCountInliningCutoff - unique();
1980   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1981   int amount_grown   = 0;
1982   WarmCallInfo* call;
1983   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1984     int est_size = (int)call->size();
1985     if (est_size > (room_to_grow - amount_grown)) {
1986       // This one won't fit anyway.  Get rid of it.
1987       call->make_cold();
1988       continue;
1989     }
1990     call->make_hot();
1991     calls_made_hot++;
1992     amount_grown   += est_size;
1993     amount_to_grow -= est_size;
1994   }
1995 
1996   if (calls_made_hot > 0)  set_major_progress();
1997   return calls_made_hot;
1998 }
1999 
2000 
2001 //----------------------------Finish_Warm--------------------------------------
2002 void Compile::Finish_Warm() {
2003   if (!InlineWarmCalls)  return;
2004   if (failing())  return;
2005   if (warm_calls() == NULL)  return;
2006 
2007   // Clean up loose ends, if we are out of space for inlining.
2008   WarmCallInfo* call;
2009   while ((call = pop_warm_call()) != NULL) {
2010     call->make_cold();
2011   }
2012 }
2013 
2014 //---------------------cleanup_loop_predicates-----------------------
2015 // Remove the opaque nodes that protect the predicates so that all unused
2016 // checks and uncommon_traps will be eliminated from the ideal graph
2017 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
2018   if (predicate_count()==0) return;
2019   for (int i = predicate_count(); i > 0; i--) {
2020     Node * n = predicate_opaque1_node(i-1);
2021     assert(n->Opcode() == Op_Opaque1, "must be");
2022     igvn.replace_node(n, n->in(1));
2023   }
2024   assert(predicate_count()==0, "should be clean!");
2025 }
2026 
2027 void Compile::add_range_check_cast(Node* n) {
2028   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2029   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
2030   _range_check_casts->append(n);
2031 }
2032 
2033 // Remove all range check dependent CastIINodes.
2034 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
2035   for (int i = range_check_cast_count(); i > 0; i--) {
2036     Node* cast = range_check_cast_node(i-1);
2037     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
2038     igvn.replace_node(cast, cast->in(1));
2039   }
2040   assert(range_check_cast_count() == 0, "should be empty");
2041 }
2042 
2043 void Compile::add_opaque4_node(Node* n) {
2044   assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
2045   assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
2046   _opaque4_nodes->append(n);
2047 }
2048 
2049 // Remove all Opaque4 nodes.
2050 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2051   for (int i = opaque4_count(); i > 0; i--) {
2052     Node* opaq = opaque4_node(i-1);
2053     assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2054     igvn.replace_node(opaq, opaq->in(2));
2055   }
2056   assert(opaque4_count() == 0, "should be empty");
2057 }
2058 
2059 void Compile::add_value_type(Node* n) {
2060   assert(n->is_ValueTypeBase(), "unexpected node");
2061   if (_value_type_nodes != NULL) {
2062     _value_type_nodes->push(n);
2063   }
2064 }
2065 
2066 void Compile::remove_value_type(Node* n) {
2067   assert(n->is_ValueTypeBase(), "unexpected node");
2068   if (_value_type_nodes != NULL) {
2069     _value_type_nodes->remove(n);
2070   }
2071 }
2072 
2073 // Does the return value keep otherwise useless value type allocations
2074 // alive?
2075 static bool return_val_keeps_allocations_alive(Node* ret_val) {
2076   ResourceMark rm;
2077   Unique_Node_List wq;
2078   wq.push(ret_val);
2079   bool some_allocations = false;
2080   for (uint i = 0; i < wq.size(); i++) {
2081     Node* n = wq.at(i);
2082     assert(!n->is_ValueTypeBase(), "chain of value type nodes");
2083     if (n->outcnt() > 1) {
2084       // Some other use for the allocation
2085       return false;
2086     } else if (n->is_Phi()) {
2087       for (uint j = 1; j < n->req(); j++) {
2088         wq.push(n->in(j));
2089       }
2090     } else if (n->is_CheckCastPP() &&
2091                n->in(1)->is_Proj() &&
2092                n->in(1)->in(0)->is_Allocate()) {
2093       some_allocations = true;
2094     }
2095   }
2096   return some_allocations;
2097 }
2098 
2099 void Compile::process_value_types(PhaseIterGVN &igvn) {
2100   // Make value types scalar in safepoints
2101   while (_value_type_nodes->size() != 0) {
2102     ValueTypeBaseNode* vt = _value_type_nodes->pop()->as_ValueTypeBase();
2103     vt->make_scalar_in_safepoints(&igvn);
2104     if (vt->is_ValueTypePtr()) {
2105       igvn.replace_node(vt, vt->get_oop());
2106     } else if (vt->outcnt() == 0) {
2107       igvn.remove_dead_node(vt);
2108     }
2109   }
2110   _value_type_nodes = NULL;
2111   if (tf()->returns_value_type_as_fields()) {
2112     Node* ret = NULL;
2113     for (uint i = 1; i < root()->req(); i++){
2114       Node* in = root()->in(i);
2115       if (in->Opcode() == Op_Return) {
2116         assert(ret == NULL, "only one return");
2117         ret = in;
2118       }
2119     }
2120     if (ret != NULL) {
2121       Node* ret_val = ret->in(TypeFunc::Parms);
2122       if (igvn.type(ret_val)->isa_oopptr() &&
2123           return_val_keeps_allocations_alive(ret_val)) {
2124         igvn.replace_input_of(ret, TypeFunc::Parms, ValueTypeNode::tagged_klass(igvn.type(ret_val)->value_klass(), igvn));
2125         assert(ret_val->outcnt() == 0, "should be dead now");
2126         igvn.remove_dead_node(ret_val);
2127       }
2128     }
2129   }
2130   igvn.optimize();
2131 }
2132 
2133 // StringOpts and late inlining of string methods
2134 void Compile::inline_string_calls(bool parse_time) {
2135   {
2136     // remove useless nodes to make the usage analysis simpler
2137     ResourceMark rm;
2138     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2139   }
2140 
2141   {
2142     ResourceMark rm;
2143     print_method(PHASE_BEFORE_STRINGOPTS, 3);
2144     PhaseStringOpts pso(initial_gvn(), for_igvn());
2145     print_method(PHASE_AFTER_STRINGOPTS, 3);
2146   }
2147 
2148   // now inline anything that we skipped the first time around
2149   if (!parse_time) {
2150     _late_inlines_pos = _late_inlines.length();
2151   }
2152 
2153   while (_string_late_inlines.length() > 0) {
2154     CallGenerator* cg = _string_late_inlines.pop();
2155     cg->do_late_inline();
2156     if (failing())  return;
2157   }
2158   _string_late_inlines.trunc_to(0);
2159 }
2160 
2161 // Late inlining of boxing methods
2162 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2163   if (_boxing_late_inlines.length() > 0) {
2164     assert(has_boxed_value(), "inconsistent");
2165 
2166     PhaseGVN* gvn = initial_gvn();
2167     set_inlining_incrementally(true);
2168 
2169     assert( igvn._worklist.size() == 0, "should be done with igvn" );
2170     for_igvn()->clear();
2171     gvn->replace_with(&igvn);
2172 
2173     _late_inlines_pos = _late_inlines.length();
2174 
2175     while (_boxing_late_inlines.length() > 0) {
2176       CallGenerator* cg = _boxing_late_inlines.pop();
2177       cg->do_late_inline();
2178       if (failing())  return;
2179     }
2180     _boxing_late_inlines.trunc_to(0);
2181 
2182     inline_incrementally_cleanup(igvn);
2183 
2184     set_inlining_incrementally(false);
2185   }
2186 }
2187 
2188 bool Compile::inline_incrementally_one() {
2189   assert(IncrementalInline, "incremental inlining should be on");
2190 
2191   TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2192   set_inlining_progress(false);
2193   set_do_cleanup(false);
2194   int i = 0;
2195   for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2196     CallGenerator* cg = _late_inlines.at(i);
2197     _late_inlines_pos = i+1;
2198     cg->do_late_inline();
2199     if (failing())  return false;
2200   }
2201   int j = 0;
2202   for (; i < _late_inlines.length(); i++, j++) {
2203     _late_inlines.at_put(j, _late_inlines.at(i));
2204   }
2205   _late_inlines.trunc_to(j);
2206   assert(inlining_progress() || _late_inlines.length() == 0, "");
2207 
2208   bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2209 
2210   set_inlining_progress(false);
2211   set_do_cleanup(false);
2212   return (_late_inlines.length() > 0) && !needs_cleanup;
2213 }
2214 
2215 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2216   {
2217     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2218     ResourceMark rm;
2219     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2220   }
2221   {
2222     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2223     igvn = PhaseIterGVN(initial_gvn());
2224     igvn.optimize();
2225   }
2226 }
2227 
2228 // Perform incremental inlining until bound on number of live nodes is reached
2229 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2230   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2231 
2232   set_inlining_incrementally(true);
2233   uint low_live_nodes = 0;
2234 
2235   while (_late_inlines.length() > 0) {
2236     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2237       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2238         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2239         // PhaseIdealLoop is expensive so we only try it once we are
2240         // out of live nodes and we only try it again if the previous
2241         // helped got the number of nodes down significantly
2242         PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2243         if (failing())  return;
2244         low_live_nodes = live_nodes();
2245         _major_progress = true;
2246       }
2247 
2248       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2249         break; // finish
2250       }
2251     }
2252 
2253     for_igvn()->clear();
2254     initial_gvn()->replace_with(&igvn);
2255 
2256     while (inline_incrementally_one()) {
2257       assert(!failing(), "inconsistent");
2258     }
2259 
2260     if (failing())  return;
2261 
2262     inline_incrementally_cleanup(igvn);
2263 
2264     if (failing())  return;
2265   }
2266   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2267 
2268   if (_string_late_inlines.length() > 0) {
2269     assert(has_stringbuilder(), "inconsistent");
2270     for_igvn()->clear();
2271     initial_gvn()->replace_with(&igvn);
2272 
2273     inline_string_calls(false);
2274 
2275     if (failing())  return;
2276 
2277     inline_incrementally_cleanup(igvn);
2278   }
2279 
2280   set_inlining_incrementally(false);
2281 }
2282 
2283 
2284 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2285   if(_loop_opts_cnt > 0) {
2286     debug_only( int cnt = 0; );
2287     while(major_progress() && (_loop_opts_cnt > 0)) {
2288       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2289       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2290       PhaseIdealLoop::optimize(igvn, mode);
2291       _loop_opts_cnt--;
2292       if (failing())  return false;
2293       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2294     }
2295   }
2296   return true;
2297 }
2298 
2299 // Remove edges from "root" to each SafePoint at a backward branch.
2300 // They were inserted during parsing (see add_safepoint()) to make
2301 // infinite loops without calls or exceptions visible to root, i.e.,
2302 // useful.
2303 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2304   Node *r = root();
2305   if (r != NULL) {
2306     for (uint i = r->req(); i < r->len(); ++i) {
2307       Node *n = r->in(i);
2308       if (n != NULL && n->is_SafePoint()) {
2309         r->rm_prec(i);
2310         if (n->outcnt() == 0) {
2311           igvn.remove_dead_node(n);
2312         }
2313         --i;
2314       }
2315     }
2316   }
2317 }
2318 
2319 //------------------------------Optimize---------------------------------------
2320 // Given a graph, optimize it.
2321 void Compile::Optimize() {
2322   TracePhase tp("optimizer", &timers[_t_optimizer]);
2323 
2324 #ifndef PRODUCT
2325   if (_directive->BreakAtCompileOption) {
2326     BREAKPOINT;
2327   }
2328 
2329 #endif
2330 
2331 #ifdef ASSERT
2332   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2333   bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2334 #endif
2335 
2336   ResourceMark rm;
2337 
2338   print_inlining_reinit();
2339 
2340   NOT_PRODUCT( verify_graph_edges(); )
2341 
2342   print_method(PHASE_AFTER_PARSING);
2343 
2344  {
2345   // Iterative Global Value Numbering, including ideal transforms
2346   // Initialize IterGVN with types and values from parse-time GVN
2347   PhaseIterGVN igvn(initial_gvn());
2348 #ifdef ASSERT
2349   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2350 #endif
2351   {
2352     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2353     igvn.optimize();
2354   }
2355 
2356   if (failing())  return;
2357 
2358   print_method(PHASE_ITER_GVN1, 2);
2359 
2360   inline_incrementally(igvn);
2361 
2362   print_method(PHASE_INCREMENTAL_INLINE, 2);
2363 
2364   if (failing())  return;
2365 
2366   if (eliminate_boxing()) {
2367     // Inline valueOf() methods now.
2368     inline_boxing_calls(igvn);
2369 
2370     if (AlwaysIncrementalInline) {
2371       inline_incrementally(igvn);
2372     }
2373 
2374     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2375 
2376     if (failing())  return;
2377   }
2378 
2379   // Now that all inlining is over, cut edge from root to loop
2380   // safepoints
2381   remove_root_to_sfpts_edges(igvn);
2382 
2383   // Remove the speculative part of types and clean up the graph from
2384   // the extra CastPP nodes whose only purpose is to carry them. Do
2385   // that early so that optimizations are not disrupted by the extra
2386   // CastPP nodes.
2387   remove_speculative_types(igvn);
2388 
2389   // No more new expensive nodes will be added to the list from here
2390   // so keep only the actual candidates for optimizations.
2391   cleanup_expensive_nodes(igvn);
2392 
2393   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2394     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2395     initial_gvn()->replace_with(&igvn);
2396     for_igvn()->clear();
2397     Unique_Node_List new_worklist(C->comp_arena());
2398     {
2399       ResourceMark rm;
2400       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2401     }
2402     set_for_igvn(&new_worklist);
2403     igvn = PhaseIterGVN(initial_gvn());
2404     igvn.optimize();
2405   }
2406 
2407   if (_value_type_nodes->size() > 0) {
2408     // Do this once all inlining is over to avoid getting inconsistent debug info
2409     process_value_types(igvn);
2410   }
2411 
2412   // Perform escape analysis
2413   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2414     if (has_loops()) {
2415       // Cleanup graph (remove dead nodes).
2416       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2417       PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2418       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2419       if (failing())  return;
2420     }
2421     ConnectionGraph::do_analysis(this, &igvn);
2422 
2423     if (failing())  return;
2424 
2425     // Optimize out fields loads from scalar replaceable allocations.
2426     igvn.optimize();
2427     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2428 
2429     if (failing())  return;
2430 
2431     if (congraph() != NULL && macro_count() > 0) {
2432       TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2433       PhaseMacroExpand mexp(igvn);
2434       mexp.eliminate_macro_nodes();
2435       igvn.set_delay_transform(false);
2436 
2437       igvn.optimize();
2438       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2439 
2440       if (failing())  return;
2441     }
2442   }
2443 
2444   // Loop transforms on the ideal graph.  Range Check Elimination,
2445   // peeling, unrolling, etc.
2446 
2447   // Set loop opts counter
2448   if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2449     {
2450       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2451       PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2452       _loop_opts_cnt--;
2453       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2454       if (failing())  return;
2455     }
2456     // Loop opts pass if partial peeling occurred in previous pass
2457     if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2458       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2459       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2460       _loop_opts_cnt--;
2461       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2462       if (failing())  return;
2463     }
2464     // Loop opts pass for loop-unrolling before CCP
2465     if(major_progress() && (_loop_opts_cnt > 0)) {
2466       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2467       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2468       _loop_opts_cnt--;
2469       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2470     }
2471     if (!failing()) {
2472       // Verify that last round of loop opts produced a valid graph
2473       TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2474       PhaseIdealLoop::verify(igvn);
2475     }
2476   }
2477   if (failing())  return;
2478 
2479   // Conditional Constant Propagation;
2480   PhaseCCP ccp( &igvn );
2481   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2482   {
2483     TracePhase tp("ccp", &timers[_t_ccp]);
2484     ccp.do_transform();
2485   }
2486   print_method(PHASE_CPP1, 2);
2487 
2488   assert( true, "Break here to ccp.dump_old2new_map()");
2489 
2490   // Iterative Global Value Numbering, including ideal transforms
2491   {
2492     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2493     igvn = ccp;
2494     igvn.optimize();
2495   }
2496 
2497   print_method(PHASE_ITER_GVN2, 2);
2498 
2499   if (failing())  return;
2500 
2501   // Loop transforms on the ideal graph.  Range Check Elimination,
2502   // peeling, unrolling, etc.
2503   if (!optimize_loops(igvn, LoopOptsDefault)) {
2504     return;
2505   }
2506 
2507 #if INCLUDE_ZGC
2508   if (UseZGC) {
2509     ZBarrierSetC2::find_dominating_barriers(igvn);
2510   }
2511 #endif
2512 
2513   if (failing())  return;
2514 
2515   // Ensure that major progress is now clear
2516   C->clear_major_progress();
2517 
2518   {
2519     // Verify that all previous optimizations produced a valid graph
2520     // at least to this point, even if no loop optimizations were done.
2521     TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2522     PhaseIdealLoop::verify(igvn);
2523   }
2524 
2525   if (range_check_cast_count() > 0) {
2526     // No more loop optimizations. Remove all range check dependent CastIINodes.
2527     C->remove_range_check_casts(igvn);
2528     igvn.optimize();
2529   }
2530 
2531 #ifdef ASSERT
2532   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2533   bs->verify_gc_barriers(this, BarrierSetC2::BeforeExpand);
2534 #endif
2535 
2536   {
2537     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2538     PhaseMacroExpand  mex(igvn);
2539     print_method(PHASE_BEFORE_MACRO_EXPANSION, 2);
2540     if (mex.expand_macro_nodes()) {
2541       assert(failing(), "must bail out w/ explicit message");
2542       return;
2543     }
2544   }
2545 
2546   {
2547     TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2548     print_method(PHASE_BEFORE_BARRIER_EXPAND, 2);
2549     BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2550     if (bs->expand_barriers(this, igvn)) {
2551       assert(failing(), "must bail out w/ explicit message");
2552       return;
2553     }
2554   }
2555 
2556   if (opaque4_count() > 0) {
2557     C->remove_opaque4_nodes(igvn);
2558     igvn.optimize();
2559   }
2560 
2561   DEBUG_ONLY( _modified_nodes = NULL; )
2562  } // (End scope of igvn; run destructor if necessary for asserts.)
2563 
2564  process_print_inlining();
2565  // A method with only infinite loops has no edges entering loops from root
2566  {
2567    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2568    if (final_graph_reshaping()) {
2569      assert(failing(), "must bail out w/ explicit message");
2570      return;
2571    }
2572  }
2573 
2574  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2575 }
2576 
2577 //------------------------------Code_Gen---------------------------------------
2578 // Given a graph, generate code for it
2579 void Compile::Code_Gen() {
2580   if (failing()) {
2581     return;
2582   }
2583 
2584   // Perform instruction selection.  You might think we could reclaim Matcher
2585   // memory PDQ, but actually the Matcher is used in generating spill code.
2586   // Internals of the Matcher (including some VectorSets) must remain live
2587   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2588   // set a bit in reclaimed memory.
2589 
2590   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2591   // nodes.  Mapping is only valid at the root of each matched subtree.
2592   NOT_PRODUCT( verify_graph_edges(); )
2593 
2594   Matcher matcher;
2595   _matcher = &matcher;
2596   {
2597     TracePhase tp("matcher", &timers[_t_matcher]);
2598     matcher.match();
2599   }
2600   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2601   // nodes.  Mapping is only valid at the root of each matched subtree.
2602   NOT_PRODUCT( verify_graph_edges(); )
2603 
2604   // If you have too many nodes, or if matching has failed, bail out
2605   check_node_count(0, "out of nodes matching instructions");
2606   if (failing()) {
2607     return;
2608   }
2609 
2610   print_method(PHASE_MATCHING, 2);
2611 
2612   // Build a proper-looking CFG
2613   PhaseCFG cfg(node_arena(), root(), matcher);
2614   _cfg = &cfg;
2615   {
2616     TracePhase tp("scheduler", &timers[_t_scheduler]);
2617     bool success = cfg.do_global_code_motion();
2618     if (!success) {
2619       return;
2620     }
2621 
2622     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2623     NOT_PRODUCT( verify_graph_edges(); )
2624     debug_only( cfg.verify(); )
2625   }
2626 
2627   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2628   _regalloc = &regalloc;
2629   {
2630     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2631     // Perform register allocation.  After Chaitin, use-def chains are
2632     // no longer accurate (at spill code) and so must be ignored.
2633     // Node->LRG->reg mappings are still accurate.
2634     _regalloc->Register_Allocate();
2635 
2636     // Bail out if the allocator builds too many nodes
2637     if (failing()) {
2638       return;
2639     }
2640   }
2641 
2642   // Prior to register allocation we kept empty basic blocks in case the
2643   // the allocator needed a place to spill.  After register allocation we
2644   // are not adding any new instructions.  If any basic block is empty, we
2645   // can now safely remove it.
2646   {
2647     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2648     cfg.remove_empty_blocks();
2649     if (do_freq_based_layout()) {
2650       PhaseBlockLayout layout(cfg);
2651     } else {
2652       cfg.set_loop_alignment();
2653     }
2654     cfg.fixup_flow();
2655   }
2656 
2657   // Apply peephole optimizations
2658   if( OptoPeephole ) {
2659     TracePhase tp("peephole", &timers[_t_peephole]);
2660     PhasePeephole peep( _regalloc, cfg);
2661     peep.do_transform();
2662   }
2663 
2664   // Do late expand if CPU requires this.
2665   if (Matcher::require_postalloc_expand) {
2666     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2667     cfg.postalloc_expand(_regalloc);
2668   }
2669 
2670   // Convert Nodes to instruction bits in a buffer
2671   {
2672     TraceTime tp("output", &timers[_t_output], CITime);
2673     Output();
2674   }
2675 
2676   print_method(PHASE_FINAL_CODE);
2677 
2678   // He's dead, Jim.
2679   _cfg     = (PhaseCFG*)((intptr_t)0xdeadbeef);
2680   _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2681 }
2682 
2683 
2684 //------------------------------dump_asm---------------------------------------
2685 // Dump formatted assembly
2686 #ifndef PRODUCT
2687 void Compile::dump_asm(int *pcs, uint pc_limit) {
2688   bool cut_short = false;
2689   tty->print_cr("#");
2690   tty->print("#  ");  _tf->dump();  tty->cr();
2691   tty->print_cr("#");
2692 
2693   // For all blocks
2694   int pc = 0x0;                 // Program counter
2695   char starts_bundle = ' ';
2696   _regalloc->dump_frame();
2697 
2698   Node *n = NULL;
2699   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2700     if (VMThread::should_terminate()) {
2701       cut_short = true;
2702       break;
2703     }
2704     Block* block = _cfg->get_block(i);
2705     if (block->is_connector() && !Verbose) {
2706       continue;
2707     }
2708     n = block->head();
2709     if (pcs && n->_idx < pc_limit) {
2710       tty->print("%3.3x   ", pcs[n->_idx]);
2711     } else {
2712       tty->print("      ");
2713     }
2714     block->dump_head(_cfg);
2715     if (block->is_connector()) {
2716       tty->print_cr("        # Empty connector block");
2717     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2718       tty->print_cr("        # Block is sole successor of call");
2719     }
2720 
2721     // For all instructions
2722     Node *delay = NULL;
2723     for (uint j = 0; j < block->number_of_nodes(); j++) {
2724       if (VMThread::should_terminate()) {
2725         cut_short = true;
2726         break;
2727       }
2728       n = block->get_node(j);
2729       if (valid_bundle_info(n)) {
2730         Bundle* bundle = node_bundling(n);
2731         if (bundle->used_in_unconditional_delay()) {
2732           delay = n;
2733           continue;
2734         }
2735         if (bundle->starts_bundle()) {
2736           starts_bundle = '+';
2737         }
2738       }
2739 
2740       if (WizardMode) {
2741         n->dump();
2742       }
2743 
2744       if( !n->is_Region() &&    // Dont print in the Assembly
2745           !n->is_Phi() &&       // a few noisely useless nodes
2746           !n->is_Proj() &&
2747           !n->is_MachTemp() &&
2748           !n->is_SafePointScalarObject() &&
2749           !n->is_Catch() &&     // Would be nice to print exception table targets
2750           !n->is_MergeMem() &&  // Not very interesting
2751           !n->is_top() &&       // Debug info table constants
2752           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2753           ) {
2754         if (pcs && n->_idx < pc_limit)
2755           tty->print("%3.3x", pcs[n->_idx]);
2756         else
2757           tty->print("   ");
2758         tty->print(" %c ", starts_bundle);
2759         starts_bundle = ' ';
2760         tty->print("\t");
2761         n->format(_regalloc, tty);
2762         tty->cr();
2763       }
2764 
2765       // If we have an instruction with a delay slot, and have seen a delay,
2766       // then back up and print it
2767       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2768         assert(delay != NULL, "no unconditional delay instruction");
2769         if (WizardMode) delay->dump();
2770 
2771         if (node_bundling(delay)->starts_bundle())
2772           starts_bundle = '+';
2773         if (pcs && n->_idx < pc_limit)
2774           tty->print("%3.3x", pcs[n->_idx]);
2775         else
2776           tty->print("   ");
2777         tty->print(" %c ", starts_bundle);
2778         starts_bundle = ' ';
2779         tty->print("\t");
2780         delay->format(_regalloc, tty);
2781         tty->cr();
2782         delay = NULL;
2783       }
2784 
2785       // Dump the exception table as well
2786       if( n->is_Catch() && (Verbose || WizardMode) ) {
2787         // Print the exception table for this offset
2788         _handler_table.print_subtable_for(pc);
2789       }
2790     }
2791 
2792     if (pcs && n->_idx < pc_limit)
2793       tty->print_cr("%3.3x", pcs[n->_idx]);
2794     else
2795       tty->cr();
2796 
2797     assert(cut_short || delay == NULL, "no unconditional delay branch");
2798 
2799   } // End of per-block dump
2800   tty->cr();
2801 
2802   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2803 }
2804 #endif
2805 
2806 //------------------------------Final_Reshape_Counts---------------------------
2807 // This class defines counters to help identify when a method
2808 // may/must be executed using hardware with only 24-bit precision.
2809 struct Final_Reshape_Counts : public StackObj {
2810   int  _call_count;             // count non-inlined 'common' calls
2811   int  _float_count;            // count float ops requiring 24-bit precision
2812   int  _double_count;           // count double ops requiring more precision
2813   int  _java_call_count;        // count non-inlined 'java' calls
2814   int  _inner_loop_count;       // count loops which need alignment
2815   VectorSet _visited;           // Visitation flags
2816   Node_List _tests;             // Set of IfNodes & PCTableNodes
2817 
2818   Final_Reshape_Counts() :
2819     _call_count(0), _float_count(0), _double_count(0),
2820     _java_call_count(0), _inner_loop_count(0),
2821     _visited( Thread::current()->resource_area() ) { }
2822 
2823   void inc_call_count  () { _call_count  ++; }
2824   void inc_float_count () { _float_count ++; }
2825   void inc_double_count() { _double_count++; }
2826   void inc_java_call_count() { _java_call_count++; }
2827   void inc_inner_loop_count() { _inner_loop_count++; }
2828 
2829   int  get_call_count  () const { return _call_count  ; }
2830   int  get_float_count () const { return _float_count ; }
2831   int  get_double_count() const { return _double_count; }
2832   int  get_java_call_count() const { return _java_call_count; }
2833   int  get_inner_loop_count() const { return _inner_loop_count; }
2834 };
2835 
2836 #ifdef ASSERT
2837 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2838   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2839   // Make sure the offset goes inside the instance layout.
2840   return k->contains_field_offset(tp->offset());
2841   // Note that OffsetBot and OffsetTop are very negative.
2842 }
2843 #endif
2844 
2845 // Eliminate trivially redundant StoreCMs and accumulate their
2846 // precedence edges.
2847 void Compile::eliminate_redundant_card_marks(Node* n) {
2848   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2849   if (n->in(MemNode::Address)->outcnt() > 1) {
2850     // There are multiple users of the same address so it might be
2851     // possible to eliminate some of the StoreCMs
2852     Node* mem = n->in(MemNode::Memory);
2853     Node* adr = n->in(MemNode::Address);
2854     Node* val = n->in(MemNode::ValueIn);
2855     Node* prev = n;
2856     bool done = false;
2857     // Walk the chain of StoreCMs eliminating ones that match.  As
2858     // long as it's a chain of single users then the optimization is
2859     // safe.  Eliminating partially redundant StoreCMs would require
2860     // cloning copies down the other paths.
2861     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2862       if (adr == mem->in(MemNode::Address) &&
2863           val == mem->in(MemNode::ValueIn)) {
2864         // redundant StoreCM
2865         if (mem->req() > MemNode::OopStore) {
2866           // Hasn't been processed by this code yet.
2867           n->add_prec(mem->in(MemNode::OopStore));
2868         } else {
2869           // Already converted to precedence edge
2870           for (uint i = mem->req(); i < mem->len(); i++) {
2871             // Accumulate any precedence edges
2872             if (mem->in(i) != NULL) {
2873               n->add_prec(mem->in(i));
2874             }
2875           }
2876           // Everything above this point has been processed.
2877           done = true;
2878         }
2879         // Eliminate the previous StoreCM
2880         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2881         assert(mem->outcnt() == 0, "should be dead");
2882         mem->disconnect_inputs(NULL, this);
2883       } else {
2884         prev = mem;
2885       }
2886       mem = prev->in(MemNode::Memory);
2887     }
2888   }
2889 }
2890 
2891 
2892 //------------------------------final_graph_reshaping_impl----------------------
2893 // Implement items 1-5 from final_graph_reshaping below.
2894 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2895 
2896   if ( n->outcnt() == 0 ) return; // dead node
2897   uint nop = n->Opcode();
2898 
2899   // Check for 2-input instruction with "last use" on right input.
2900   // Swap to left input.  Implements item (2).
2901   if( n->req() == 3 &&          // two-input instruction
2902       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2903       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2904       n->in(2)->outcnt() == 1 &&// right use IS a last use
2905       !n->in(2)->is_Con() ) {   // right use is not a constant
2906     // Check for commutative opcode
2907     switch( nop ) {
2908     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2909     case Op_MaxI:  case Op_MinI:
2910     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2911     case Op_AndL:  case Op_XorL:  case Op_OrL:
2912     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2913       // Move "last use" input to left by swapping inputs
2914       n->swap_edges(1, 2);
2915       break;
2916     }
2917     default:
2918       break;
2919     }
2920   }
2921 
2922 #ifdef ASSERT
2923   if( n->is_Mem() ) {
2924     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2925     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2926             // oop will be recorded in oop map if load crosses safepoint
2927             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2928                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2929             "raw memory operations should have control edge");
2930   }
2931   if (n->is_MemBar()) {
2932     MemBarNode* mb = n->as_MemBar();
2933     if (mb->trailing_store() || mb->trailing_load_store()) {
2934       assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2935       Node* mem = mb->in(MemBarNode::Precedent);
2936       assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2937              (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2938     } else if (mb->leading()) {
2939       assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2940     }
2941   }
2942 #endif
2943   // Count FPU ops and common calls, implements item (3)
2944   bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2945   if (!gc_handled) {
2946     final_graph_reshaping_main_switch(n, frc, nop);
2947   }
2948 
2949   // Collect CFG split points
2950   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2951     frc._tests.push(n);
2952   }
2953 }
2954 
2955 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2956   switch( nop ) {
2957   // Count all float operations that may use FPU
2958   case Op_AddF:
2959   case Op_SubF:
2960   case Op_MulF:
2961   case Op_DivF:
2962   case Op_NegF:
2963   case Op_ModF:
2964   case Op_ConvI2F:
2965   case Op_ConF:
2966   case Op_CmpF:
2967   case Op_CmpF3:
2968   // case Op_ConvL2F: // longs are split into 32-bit halves
2969     frc.inc_float_count();
2970     break;
2971 
2972   case Op_ConvF2D:
2973   case Op_ConvD2F:
2974     frc.inc_float_count();
2975     frc.inc_double_count();
2976     break;
2977 
2978   // Count all double operations that may use FPU
2979   case Op_AddD:
2980   case Op_SubD:
2981   case Op_MulD:
2982   case Op_DivD:
2983   case Op_NegD:
2984   case Op_ModD:
2985   case Op_ConvI2D:
2986   case Op_ConvD2I:
2987   // case Op_ConvL2D: // handled by leaf call
2988   // case Op_ConvD2L: // handled by leaf call
2989   case Op_ConD:
2990   case Op_CmpD:
2991   case Op_CmpD3:
2992     frc.inc_double_count();
2993     break;
2994   case Op_Opaque1:              // Remove Opaque Nodes before matching
2995   case Op_Opaque2:              // Remove Opaque Nodes before matching
2996   case Op_Opaque3:
2997     n->subsume_by(n->in(1), this);
2998     break;
2999   case Op_CallStaticJava:
3000   case Op_CallJava:
3001   case Op_CallDynamicJava:
3002     frc.inc_java_call_count(); // Count java call site;
3003   case Op_CallRuntime:
3004   case Op_CallLeaf:
3005   case Op_CallLeafNoFP: {
3006     assert (n->is_Call(), "");
3007     CallNode *call = n->as_Call();
3008     // Count call sites where the FP mode bit would have to be flipped.
3009     // Do not count uncommon runtime calls:
3010     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
3011     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
3012     if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
3013       frc.inc_call_count();   // Count the call site
3014     } else {                  // See if uncommon argument is shared
3015       Node *n = call->in(TypeFunc::Parms);
3016       int nop = n->Opcode();
3017       // Clone shared simple arguments to uncommon calls, item (1).
3018       if (n->outcnt() > 1 &&
3019           !n->is_Proj() &&
3020           nop != Op_CreateEx &&
3021           nop != Op_CheckCastPP &&
3022           nop != Op_DecodeN &&
3023           nop != Op_DecodeNKlass &&
3024           !n->is_Mem() &&
3025           !n->is_Phi()) {
3026         Node *x = n->clone();
3027         call->set_req(TypeFunc::Parms, x);
3028       }
3029     }
3030     break;
3031   }
3032 
3033   case Op_StoreD:
3034   case Op_LoadD:
3035   case Op_LoadD_unaligned:
3036     frc.inc_double_count();
3037     goto handle_mem;
3038   case Op_StoreF:
3039   case Op_LoadF:
3040     frc.inc_float_count();
3041     goto handle_mem;
3042 
3043   case Op_StoreCM:
3044     {
3045       // Convert OopStore dependence into precedence edge
3046       Node* prec = n->in(MemNode::OopStore);
3047       n->del_req(MemNode::OopStore);
3048       n->add_prec(prec);
3049       eliminate_redundant_card_marks(n);
3050     }
3051 
3052     // fall through
3053 
3054   case Op_StoreB:
3055   case Op_StoreC:
3056   case Op_StorePConditional:
3057   case Op_StoreI:
3058   case Op_StoreL:
3059   case Op_StoreIConditional:
3060   case Op_StoreLConditional:
3061   case Op_CompareAndSwapB:
3062   case Op_CompareAndSwapS:
3063   case Op_CompareAndSwapI:
3064   case Op_CompareAndSwapL:
3065   case Op_CompareAndSwapP:
3066   case Op_CompareAndSwapN:
3067   case Op_WeakCompareAndSwapB:
3068   case Op_WeakCompareAndSwapS:
3069   case Op_WeakCompareAndSwapI:
3070   case Op_WeakCompareAndSwapL:
3071   case Op_WeakCompareAndSwapP:
3072   case Op_WeakCompareAndSwapN:
3073   case Op_CompareAndExchangeB:
3074   case Op_CompareAndExchangeS:
3075   case Op_CompareAndExchangeI:
3076   case Op_CompareAndExchangeL:
3077   case Op_CompareAndExchangeP:
3078   case Op_CompareAndExchangeN:
3079   case Op_GetAndAddS:
3080   case Op_GetAndAddB:
3081   case Op_GetAndAddI:
3082   case Op_GetAndAddL:
3083   case Op_GetAndSetS:
3084   case Op_GetAndSetB:
3085   case Op_GetAndSetI:
3086   case Op_GetAndSetL:
3087   case Op_GetAndSetP:
3088   case Op_GetAndSetN:
3089   case Op_StoreP:
3090   case Op_StoreN:
3091   case Op_StoreNKlass:
3092   case Op_LoadB:
3093   case Op_LoadUB:
3094   case Op_LoadUS:
3095   case Op_LoadI:
3096   case Op_LoadKlass:
3097   case Op_LoadNKlass:
3098   case Op_LoadL:
3099   case Op_LoadL_unaligned:
3100   case Op_LoadPLocked:
3101   case Op_LoadP:
3102   case Op_LoadN:
3103   case Op_LoadRange:
3104   case Op_LoadS: {
3105   handle_mem:
3106 #ifdef ASSERT
3107     if( VerifyOptoOopOffsets ) {
3108       MemNode* mem  = n->as_Mem();
3109       // Check to see if address types have grounded out somehow.
3110       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
3111       assert( !tp || oop_offset_is_sane(tp), "" );
3112     }
3113 #endif
3114     break;
3115   }
3116 
3117   case Op_AddP: {               // Assert sane base pointers
3118     Node *addp = n->in(AddPNode::Address);
3119     assert( !addp->is_AddP() ||
3120             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3121             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3122             "Base pointers must match (addp %u)", addp->_idx );
3123 #ifdef _LP64
3124     if ((UseCompressedOops || UseCompressedClassPointers) &&
3125         addp->Opcode() == Op_ConP &&
3126         addp == n->in(AddPNode::Base) &&
3127         n->in(AddPNode::Offset)->is_Con()) {
3128       // If the transformation of ConP to ConN+DecodeN is beneficial depends
3129       // on the platform and on the compressed oops mode.
3130       // Use addressing with narrow klass to load with offset on x86.
3131       // Some platforms can use the constant pool to load ConP.
3132       // Do this transformation here since IGVN will convert ConN back to ConP.
3133       const Type* t = addp->bottom_type();
3134       bool is_oop   = t->isa_oopptr() != NULL;
3135       bool is_klass = t->isa_klassptr() != NULL;
3136 
3137       if ((is_oop   && Matcher::const_oop_prefer_decode()  ) ||
3138           (is_klass && Matcher::const_klass_prefer_decode())) {
3139         Node* nn = NULL;
3140 
3141         int op = is_oop ? Op_ConN : Op_ConNKlass;
3142 
3143         // Look for existing ConN node of the same exact type.
3144         Node* r  = root();
3145         uint cnt = r->outcnt();
3146         for (uint i = 0; i < cnt; i++) {
3147           Node* m = r->raw_out(i);
3148           if (m!= NULL && m->Opcode() == op &&
3149               m->bottom_type()->make_ptr() == t) {
3150             nn = m;
3151             break;
3152           }
3153         }
3154         if (nn != NULL) {
3155           // Decode a narrow oop to match address
3156           // [R12 + narrow_oop_reg<<3 + offset]
3157           if (is_oop) {
3158             nn = new DecodeNNode(nn, t);
3159           } else {
3160             nn = new DecodeNKlassNode(nn, t);
3161           }
3162           // Check for succeeding AddP which uses the same Base.
3163           // Otherwise we will run into the assertion above when visiting that guy.
3164           for (uint i = 0; i < n->outcnt(); ++i) {
3165             Node *out_i = n->raw_out(i);
3166             if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3167               out_i->set_req(AddPNode::Base, nn);
3168 #ifdef ASSERT
3169               for (uint j = 0; j < out_i->outcnt(); ++j) {
3170                 Node *out_j = out_i->raw_out(j);
3171                 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3172                        "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3173               }
3174 #endif
3175             }
3176           }
3177           n->set_req(AddPNode::Base, nn);
3178           n->set_req(AddPNode::Address, nn);
3179           if (addp->outcnt() == 0) {
3180             addp->disconnect_inputs(NULL, this);
3181           }
3182         }
3183       }
3184     }
3185 #endif
3186     // platform dependent reshaping of the address expression
3187     reshape_address(n->as_AddP());
3188     break;
3189   }
3190 
3191   case Op_CastPP: {
3192     // Remove CastPP nodes to gain more freedom during scheduling but
3193     // keep the dependency they encode as control or precedence edges
3194     // (if control is set already) on memory operations. Some CastPP
3195     // nodes don't have a control (don't carry a dependency): skip
3196     // those.
3197     if (n->in(0) != NULL) {
3198       ResourceMark rm;
3199       Unique_Node_List wq;
3200       wq.push(n);
3201       for (uint next = 0; next < wq.size(); ++next) {
3202         Node *m = wq.at(next);
3203         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3204           Node* use = m->fast_out(i);
3205           if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3206             use->ensure_control_or_add_prec(n->in(0));
3207           } else {
3208             switch(use->Opcode()) {
3209             case Op_AddP:
3210             case Op_DecodeN:
3211             case Op_DecodeNKlass:
3212             case Op_CheckCastPP:
3213             case Op_CastPP:
3214               wq.push(use);
3215               break;
3216             }
3217           }
3218         }
3219       }
3220     }
3221     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3222     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3223       Node* in1 = n->in(1);
3224       const Type* t = n->bottom_type();
3225       Node* new_in1 = in1->clone();
3226       new_in1->as_DecodeN()->set_type(t);
3227 
3228       if (!Matcher::narrow_oop_use_complex_address()) {
3229         //
3230         // x86, ARM and friends can handle 2 adds in addressing mode
3231         // and Matcher can fold a DecodeN node into address by using
3232         // a narrow oop directly and do implicit NULL check in address:
3233         //
3234         // [R12 + narrow_oop_reg<<3 + offset]
3235         // NullCheck narrow_oop_reg
3236         //
3237         // On other platforms (Sparc) we have to keep new DecodeN node and
3238         // use it to do implicit NULL check in address:
3239         //
3240         // decode_not_null narrow_oop_reg, base_reg
3241         // [base_reg + offset]
3242         // NullCheck base_reg
3243         //
3244         // Pin the new DecodeN node to non-null path on these platform (Sparc)
3245         // to keep the information to which NULL check the new DecodeN node
3246         // corresponds to use it as value in implicit_null_check().
3247         //
3248         new_in1->set_req(0, n->in(0));
3249       }
3250 
3251       n->subsume_by(new_in1, this);
3252       if (in1->outcnt() == 0) {
3253         in1->disconnect_inputs(NULL, this);
3254       }
3255     } else {
3256       n->subsume_by(n->in(1), this);
3257       if (n->outcnt() == 0) {
3258         n->disconnect_inputs(NULL, this);
3259       }
3260     }
3261     break;
3262   }
3263 #ifdef _LP64
3264   case Op_CmpP:
3265     // Do this transformation here to preserve CmpPNode::sub() and
3266     // other TypePtr related Ideal optimizations (for example, ptr nullness).
3267     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3268       Node* in1 = n->in(1);
3269       Node* in2 = n->in(2);
3270       if (!in1->is_DecodeNarrowPtr()) {
3271         in2 = in1;
3272         in1 = n->in(2);
3273       }
3274       assert(in1->is_DecodeNarrowPtr(), "sanity");
3275 
3276       Node* new_in2 = NULL;
3277       if (in2->is_DecodeNarrowPtr()) {
3278         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3279         new_in2 = in2->in(1);
3280       } else if (in2->Opcode() == Op_ConP) {
3281         const Type* t = in2->bottom_type();
3282         if (t == TypePtr::NULL_PTR) {
3283           assert(in1->is_DecodeN(), "compare klass to null?");
3284           // Don't convert CmpP null check into CmpN if compressed
3285           // oops implicit null check is not generated.
3286           // This will allow to generate normal oop implicit null check.
3287           if (Matcher::gen_narrow_oop_implicit_null_checks())
3288             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3289           //
3290           // This transformation together with CastPP transformation above
3291           // will generated code for implicit NULL checks for compressed oops.
3292           //
3293           // The original code after Optimize()
3294           //
3295           //    LoadN memory, narrow_oop_reg
3296           //    decode narrow_oop_reg, base_reg
3297           //    CmpP base_reg, NULL
3298           //    CastPP base_reg // NotNull
3299           //    Load [base_reg + offset], val_reg
3300           //
3301           // after these transformations will be
3302           //
3303           //    LoadN memory, narrow_oop_reg
3304           //    CmpN narrow_oop_reg, NULL
3305           //    decode_not_null narrow_oop_reg, base_reg
3306           //    Load [base_reg + offset], val_reg
3307           //
3308           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3309           // since narrow oops can be used in debug info now (see the code in
3310           // final_graph_reshaping_walk()).
3311           //
3312           // At the end the code will be matched to
3313           // on x86:
3314           //
3315           //    Load_narrow_oop memory, narrow_oop_reg
3316           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3317           //    NullCheck narrow_oop_reg
3318           //
3319           // and on sparc:
3320           //
3321           //    Load_narrow_oop memory, narrow_oop_reg
3322           //    decode_not_null narrow_oop_reg, base_reg
3323           //    Load [base_reg + offset], val_reg
3324           //    NullCheck base_reg
3325           //
3326         } else if (t->isa_oopptr()) {
3327           new_in2 = ConNode::make(t->make_narrowoop());
3328         } else if (t->isa_klassptr()) {
3329           new_in2 = ConNode::make(t->make_narrowklass());
3330         }
3331       }
3332       if (new_in2 != NULL) {
3333         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3334         n->subsume_by(cmpN, this);
3335         if (in1->outcnt() == 0) {
3336           in1->disconnect_inputs(NULL, this);
3337         }
3338         if (in2->outcnt() == 0) {
3339           in2->disconnect_inputs(NULL, this);
3340         }
3341       }
3342     }
3343     break;
3344 
3345   case Op_DecodeN:
3346   case Op_DecodeNKlass:
3347     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3348     // DecodeN could be pinned when it can't be fold into
3349     // an address expression, see the code for Op_CastPP above.
3350     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3351     break;
3352 
3353   case Op_EncodeP:
3354   case Op_EncodePKlass: {
3355     Node* in1 = n->in(1);
3356     if (in1->is_DecodeNarrowPtr()) {
3357       n->subsume_by(in1->in(1), this);
3358     } else if (in1->Opcode() == Op_ConP) {
3359       const Type* t = in1->bottom_type();
3360       if (t == TypePtr::NULL_PTR) {
3361         assert(t->isa_oopptr(), "null klass?");
3362         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3363       } else if (t->isa_oopptr()) {
3364         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3365       } else if (t->isa_klassptr()) {
3366         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3367       }
3368     }
3369     if (in1->outcnt() == 0) {
3370       in1->disconnect_inputs(NULL, this);
3371     }
3372     break;
3373   }
3374 
3375   case Op_Proj: {
3376     if (OptimizeStringConcat) {
3377       ProjNode* p = n->as_Proj();
3378       if (p->_is_io_use) {
3379         // Separate projections were used for the exception path which
3380         // are normally removed by a late inline.  If it wasn't inlined
3381         // then they will hang around and should just be replaced with
3382         // the original one.
3383         Node* proj = NULL;
3384         // Replace with just one
3385         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3386           Node *use = i.get();
3387           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3388             proj = use;
3389             break;
3390           }
3391         }
3392         assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3393         if (proj != NULL) {
3394           p->subsume_by(proj, this);
3395         }
3396       }
3397     }
3398     break;
3399   }
3400 
3401   case Op_Phi:
3402     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3403       // The EncodeP optimization may create Phi with the same edges
3404       // for all paths. It is not handled well by Register Allocator.
3405       Node* unique_in = n->in(1);
3406       assert(unique_in != NULL, "");
3407       uint cnt = n->req();
3408       for (uint i = 2; i < cnt; i++) {
3409         Node* m = n->in(i);
3410         assert(m != NULL, "");
3411         if (unique_in != m)
3412           unique_in = NULL;
3413       }
3414       if (unique_in != NULL) {
3415         n->subsume_by(unique_in, this);
3416       }
3417     }
3418     break;
3419 
3420 #endif
3421 
3422 #ifdef ASSERT
3423   case Op_CastII:
3424     // Verify that all range check dependent CastII nodes were removed.
3425     if (n->isa_CastII()->has_range_check()) {
3426       n->dump(3);
3427       assert(false, "Range check dependent CastII node was not removed");
3428     }
3429     break;
3430 #endif
3431 
3432   case Op_ModI:
3433     if (UseDivMod) {
3434       // Check if a%b and a/b both exist
3435       Node* d = n->find_similar(Op_DivI);
3436       if (d) {
3437         // Replace them with a fused divmod if supported
3438         if (Matcher::has_match_rule(Op_DivModI)) {
3439           DivModINode* divmod = DivModINode::make(n);
3440           d->subsume_by(divmod->div_proj(), this);
3441           n->subsume_by(divmod->mod_proj(), this);
3442         } else {
3443           // replace a%b with a-((a/b)*b)
3444           Node* mult = new MulINode(d, d->in(2));
3445           Node* sub  = new SubINode(d->in(1), mult);
3446           n->subsume_by(sub, this);
3447         }
3448       }
3449     }
3450     break;
3451 
3452   case Op_ModL:
3453     if (UseDivMod) {
3454       // Check if a%b and a/b both exist
3455       Node* d = n->find_similar(Op_DivL);
3456       if (d) {
3457         // Replace them with a fused divmod if supported
3458         if (Matcher::has_match_rule(Op_DivModL)) {
3459           DivModLNode* divmod = DivModLNode::make(n);
3460           d->subsume_by(divmod->div_proj(), this);
3461           n->subsume_by(divmod->mod_proj(), this);
3462         } else {
3463           // replace a%b with a-((a/b)*b)
3464           Node* mult = new MulLNode(d, d->in(2));
3465           Node* sub  = new SubLNode(d->in(1), mult);
3466           n->subsume_by(sub, this);
3467         }
3468       }
3469     }
3470     break;
3471 
3472   case Op_LoadVector:
3473   case Op_StoreVector:
3474     break;
3475 
3476   case Op_AddReductionVI:
3477   case Op_AddReductionVL:
3478   case Op_AddReductionVF:
3479   case Op_AddReductionVD:
3480   case Op_MulReductionVI:
3481   case Op_MulReductionVL:
3482   case Op_MulReductionVF:
3483   case Op_MulReductionVD:
3484   case Op_MinReductionV:
3485   case Op_MaxReductionV:
3486     break;
3487 
3488   case Op_PackB:
3489   case Op_PackS:
3490   case Op_PackI:
3491   case Op_PackF:
3492   case Op_PackL:
3493   case Op_PackD:
3494     if (n->req()-1 > 2) {
3495       // Replace many operand PackNodes with a binary tree for matching
3496       PackNode* p = (PackNode*) n;
3497       Node* btp = p->binary_tree_pack(1, n->req());
3498       n->subsume_by(btp, this);
3499     }
3500     break;
3501   case Op_Loop:
3502   case Op_CountedLoop:
3503   case Op_OuterStripMinedLoop:
3504     if (n->as_Loop()->is_inner_loop()) {
3505       frc.inc_inner_loop_count();
3506     }
3507     n->as_Loop()->verify_strip_mined(0);
3508     break;
3509   case Op_LShiftI:
3510   case Op_RShiftI:
3511   case Op_URShiftI:
3512   case Op_LShiftL:
3513   case Op_RShiftL:
3514   case Op_URShiftL:
3515     if (Matcher::need_masked_shift_count) {
3516       // The cpu's shift instructions don't restrict the count to the
3517       // lower 5/6 bits. We need to do the masking ourselves.
3518       Node* in2 = n->in(2);
3519       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3520       const TypeInt* t = in2->find_int_type();
3521       if (t != NULL && t->is_con()) {
3522         juint shift = t->get_con();
3523         if (shift > mask) { // Unsigned cmp
3524           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3525         }
3526       } else {
3527         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3528           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3529           n->set_req(2, shift);
3530         }
3531       }
3532       if (in2->outcnt() == 0) { // Remove dead node
3533         in2->disconnect_inputs(NULL, this);
3534       }
3535     }
3536     break;
3537   case Op_MemBarStoreStore:
3538   case Op_MemBarRelease:
3539     // Break the link with AllocateNode: it is no longer useful and
3540     // confuses register allocation.
3541     if (n->req() > MemBarNode::Precedent) {
3542       n->set_req(MemBarNode::Precedent, top());
3543     }
3544     break;
3545   case Op_MemBarAcquire: {
3546     if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3547       // At parse time, the trailing MemBarAcquire for a volatile load
3548       // is created with an edge to the load. After optimizations,
3549       // that input may be a chain of Phis. If those phis have no
3550       // other use, then the MemBarAcquire keeps them alive and
3551       // register allocation can be confused.
3552       ResourceMark rm;
3553       Unique_Node_List wq;
3554       wq.push(n->in(MemBarNode::Precedent));
3555       n->set_req(MemBarNode::Precedent, top());
3556       while (wq.size() > 0) {
3557         Node* m = wq.pop();
3558         if (m->outcnt() == 0) {
3559           for (uint j = 0; j < m->req(); j++) {
3560             Node* in = m->in(j);
3561             if (in != NULL) {
3562               wq.push(in);
3563             }
3564           }
3565           m->disconnect_inputs(NULL, this);
3566         }
3567       }
3568     }
3569     break;
3570   }
3571   case Op_RangeCheck: {
3572     RangeCheckNode* rc = n->as_RangeCheck();
3573     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3574     n->subsume_by(iff, this);
3575     frc._tests.push(iff);
3576     break;
3577   }
3578   case Op_ConvI2L: {
3579     if (!Matcher::convi2l_type_required) {
3580       // Code generation on some platforms doesn't need accurate
3581       // ConvI2L types. Widening the type can help remove redundant
3582       // address computations.
3583       n->as_Type()->set_type(TypeLong::INT);
3584       ResourceMark rm;
3585       Node_List wq;
3586       wq.push(n);
3587       for (uint next = 0; next < wq.size(); next++) {
3588         Node *m = wq.at(next);
3589 
3590         for(;;) {
3591           // Loop over all nodes with identical inputs edges as m
3592           Node* k = m->find_similar(m->Opcode());
3593           if (k == NULL) {
3594             break;
3595           }
3596           // Push their uses so we get a chance to remove node made
3597           // redundant
3598           for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3599             Node* u = k->fast_out(i);
3600             assert(!wq.contains(u), "shouldn't process one node several times");
3601             if (u->Opcode() == Op_LShiftL ||
3602                 u->Opcode() == Op_AddL ||
3603                 u->Opcode() == Op_SubL ||
3604                 u->Opcode() == Op_AddP) {
3605               wq.push(u);
3606             }
3607           }
3608           // Replace all nodes with identical edges as m with m
3609           k->subsume_by(m, this);
3610         }
3611       }
3612     }
3613     break;
3614   }
3615   case Op_CmpUL: {
3616     if (!Matcher::has_match_rule(Op_CmpUL)) {
3617       // No support for unsigned long comparisons
3618       ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3619       Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3620       Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3621       ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3622       Node* andl = new AndLNode(orl, remove_sign_mask);
3623       Node* cmp = new CmpLNode(andl, n->in(2));
3624       n->subsume_by(cmp, this);
3625     }
3626     break;
3627   }
3628 #ifdef ASSERT
3629   case Op_ValueTypePtr:
3630   case Op_ValueType: {
3631     n->dump(-1);
3632     assert(false, "value type node was not removed");
3633     break;
3634   }
3635 #endif
3636   default:
3637     assert(!n->is_Call(), "");
3638     assert(!n->is_Mem(), "");
3639     assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3640     break;
3641   }
3642 }
3643 
3644 //------------------------------final_graph_reshaping_walk---------------------
3645 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3646 // requires that the walk visits a node's inputs before visiting the node.
3647 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3648   ResourceArea *area = Thread::current()->resource_area();
3649   Unique_Node_List sfpt(area);
3650 
3651   frc._visited.set(root->_idx); // first, mark node as visited
3652   uint cnt = root->req();
3653   Node *n = root;
3654   uint  i = 0;
3655   while (true) {
3656     if (i < cnt) {
3657       // Place all non-visited non-null inputs onto stack
3658       Node* m = n->in(i);
3659       ++i;
3660       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3661         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3662           // compute worst case interpreter size in case of a deoptimization
3663           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3664 
3665           sfpt.push(m);
3666         }
3667         cnt = m->req();
3668         nstack.push(n, i); // put on stack parent and next input's index
3669         n = m;
3670         i = 0;
3671       }
3672     } else {
3673       // Now do post-visit work
3674       final_graph_reshaping_impl( n, frc );
3675       if (nstack.is_empty())
3676         break;             // finished
3677       n = nstack.node();   // Get node from stack
3678       cnt = n->req();
3679       i = nstack.index();
3680       nstack.pop();        // Shift to the next node on stack
3681     }
3682   }
3683 
3684   // Skip next transformation if compressed oops are not used.
3685   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3686       (!UseCompressedOops && !UseCompressedClassPointers))
3687     return;
3688 
3689   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3690   // It could be done for an uncommon traps or any safepoints/calls
3691   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3692   while (sfpt.size() > 0) {
3693     n = sfpt.pop();
3694     JVMState *jvms = n->as_SafePoint()->jvms();
3695     assert(jvms != NULL, "sanity");
3696     int start = jvms->debug_start();
3697     int end   = n->req();
3698     bool is_uncommon = (n->is_CallStaticJava() &&
3699                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3700     for (int j = start; j < end; j++) {
3701       Node* in = n->in(j);
3702       if (in->is_DecodeNarrowPtr()) {
3703         bool safe_to_skip = true;
3704         if (!is_uncommon ) {
3705           // Is it safe to skip?
3706           for (uint i = 0; i < in->outcnt(); i++) {
3707             Node* u = in->raw_out(i);
3708             if (!u->is_SafePoint() ||
3709                 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3710               safe_to_skip = false;
3711             }
3712           }
3713         }
3714         if (safe_to_skip) {
3715           n->set_req(j, in->in(1));
3716         }
3717         if (in->outcnt() == 0) {
3718           in->disconnect_inputs(NULL, this);
3719         }
3720       }
3721     }
3722   }
3723 }
3724 
3725 //------------------------------final_graph_reshaping--------------------------
3726 // Final Graph Reshaping.
3727 //
3728 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3729 //     and not commoned up and forced early.  Must come after regular
3730 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3731 //     inputs to Loop Phis; these will be split by the allocator anyways.
3732 //     Remove Opaque nodes.
3733 // (2) Move last-uses by commutative operations to the left input to encourage
3734 //     Intel update-in-place two-address operations and better register usage
3735 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3736 //     calls canonicalizing them back.
3737 // (3) Count the number of double-precision FP ops, single-precision FP ops
3738 //     and call sites.  On Intel, we can get correct rounding either by
3739 //     forcing singles to memory (requires extra stores and loads after each
3740 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3741 //     clearing the mode bit around call sites).  The mode bit is only used
3742 //     if the relative frequency of single FP ops to calls is low enough.
3743 //     This is a key transform for SPEC mpeg_audio.
3744 // (4) Detect infinite loops; blobs of code reachable from above but not
3745 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3746 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3747 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3748 //     Detection is by looking for IfNodes where only 1 projection is
3749 //     reachable from below or CatchNodes missing some targets.
3750 // (5) Assert for insane oop offsets in debug mode.
3751 
3752 bool Compile::final_graph_reshaping() {
3753   // an infinite loop may have been eliminated by the optimizer,
3754   // in which case the graph will be empty.
3755   if (root()->req() == 1) {
3756     record_method_not_compilable("trivial infinite loop");
3757     return true;
3758   }
3759 
3760   // Expensive nodes have their control input set to prevent the GVN
3761   // from freely commoning them. There's no GVN beyond this point so
3762   // no need to keep the control input. We want the expensive nodes to
3763   // be freely moved to the least frequent code path by gcm.
3764   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3765   for (int i = 0; i < expensive_count(); i++) {
3766     _expensive_nodes->at(i)->set_req(0, NULL);
3767   }
3768 
3769   Final_Reshape_Counts frc;
3770 
3771   // Visit everybody reachable!
3772   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3773   Node_Stack nstack(live_nodes() >> 1);
3774   final_graph_reshaping_walk(nstack, root(), frc);
3775 
3776   // Check for unreachable (from below) code (i.e., infinite loops).
3777   for( uint i = 0; i < frc._tests.size(); i++ ) {
3778     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3779     // Get number of CFG targets.
3780     // Note that PCTables include exception targets after calls.
3781     uint required_outcnt = n->required_outcnt();
3782     if (n->outcnt() != required_outcnt) {
3783       // Check for a few special cases.  Rethrow Nodes never take the
3784       // 'fall-thru' path, so expected kids is 1 less.
3785       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3786         if (n->in(0)->in(0)->is_Call()) {
3787           CallNode *call = n->in(0)->in(0)->as_Call();
3788           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3789             required_outcnt--;      // Rethrow always has 1 less kid
3790           } else if (call->req() > TypeFunc::Parms &&
3791                      call->is_CallDynamicJava()) {
3792             // Check for null receiver. In such case, the optimizer has
3793             // detected that the virtual call will always result in a null
3794             // pointer exception. The fall-through projection of this CatchNode
3795             // will not be populated.
3796             Node *arg0 = call->in(TypeFunc::Parms);
3797             if (arg0->is_Type() &&
3798                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3799               required_outcnt--;
3800             }
3801           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3802                      call->req() > TypeFunc::Parms+1 &&
3803                      call->is_CallStaticJava()) {
3804             // Check for negative array length. In such case, the optimizer has
3805             // detected that the allocation attempt will always result in an
3806             // exception. There is no fall-through projection of this CatchNode .
3807             Node *arg1 = call->in(TypeFunc::Parms+1);
3808             if (arg1->is_Type() &&
3809                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3810               required_outcnt--;
3811             }
3812           }
3813         }
3814       }
3815       // Recheck with a better notion of 'required_outcnt'
3816       if (n->outcnt() != required_outcnt) {
3817         record_method_not_compilable("malformed control flow");
3818         return true;            // Not all targets reachable!
3819       }
3820     }
3821     // Check that I actually visited all kids.  Unreached kids
3822     // must be infinite loops.
3823     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3824       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3825         record_method_not_compilable("infinite loop");
3826         return true;            // Found unvisited kid; must be unreach
3827       }
3828 
3829     // Here so verification code in final_graph_reshaping_walk()
3830     // always see an OuterStripMinedLoopEnd
3831     if (n->is_OuterStripMinedLoopEnd()) {
3832       IfNode* init_iff = n->as_If();
3833       Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3834       n->subsume_by(iff, this);
3835     }
3836   }
3837 
3838   // If original bytecodes contained a mixture of floats and doubles
3839   // check if the optimizer has made it homogenous, item (3).
3840   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3841       frc.get_float_count() > 32 &&
3842       frc.get_double_count() == 0 &&
3843       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3844     set_24_bit_selection_and_mode( false,  true );
3845   }
3846 
3847   set_java_calls(frc.get_java_call_count());
3848   set_inner_loops(frc.get_inner_loop_count());
3849 
3850   // No infinite loops, no reason to bail out.
3851   return false;
3852 }
3853 
3854 //-----------------------------too_many_traps----------------------------------
3855 // Report if there are too many traps at the current method and bci.
3856 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3857 bool Compile::too_many_traps(ciMethod* method,
3858                              int bci,
3859                              Deoptimization::DeoptReason reason) {
3860   ciMethodData* md = method->method_data();
3861   if (md->is_empty()) {
3862     // Assume the trap has not occurred, or that it occurred only
3863     // because of a transient condition during start-up in the interpreter.
3864     return false;
3865   }
3866   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3867   if (md->has_trap_at(bci, m, reason) != 0) {
3868     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3869     // Also, if there are multiple reasons, or if there is no per-BCI record,
3870     // assume the worst.
3871     if (log())
3872       log()->elem("observe trap='%s' count='%d'",
3873                   Deoptimization::trap_reason_name(reason),
3874                   md->trap_count(reason));
3875     return true;
3876   } else {
3877     // Ignore method/bci and see if there have been too many globally.
3878     return too_many_traps(reason, md);
3879   }
3880 }
3881 
3882 // Less-accurate variant which does not require a method and bci.
3883 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3884                              ciMethodData* logmd) {
3885   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3886     // Too many traps globally.
3887     // Note that we use cumulative trap_count, not just md->trap_count.
3888     if (log()) {
3889       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3890       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3891                   Deoptimization::trap_reason_name(reason),
3892                   mcount, trap_count(reason));
3893     }
3894     return true;
3895   } else {
3896     // The coast is clear.
3897     return false;
3898   }
3899 }
3900 
3901 //--------------------------too_many_recompiles--------------------------------
3902 // Report if there are too many recompiles at the current method and bci.
3903 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3904 // Is not eager to return true, since this will cause the compiler to use
3905 // Action_none for a trap point, to avoid too many recompilations.
3906 bool Compile::too_many_recompiles(ciMethod* method,
3907                                   int bci,
3908                                   Deoptimization::DeoptReason reason) {
3909   ciMethodData* md = method->method_data();
3910   if (md->is_empty()) {
3911     // Assume the trap has not occurred, or that it occurred only
3912     // because of a transient condition during start-up in the interpreter.
3913     return false;
3914   }
3915   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3916   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3917   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3918   Deoptimization::DeoptReason per_bc_reason
3919     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3920   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3921   if ((per_bc_reason == Deoptimization::Reason_none
3922        || md->has_trap_at(bci, m, reason) != 0)
3923       // The trap frequency measure we care about is the recompile count:
3924       && md->trap_recompiled_at(bci, m)
3925       && md->overflow_recompile_count() >= bc_cutoff) {
3926     // Do not emit a trap here if it has already caused recompilations.
3927     // Also, if there are multiple reasons, or if there is no per-BCI record,
3928     // assume the worst.
3929     if (log())
3930       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3931                   Deoptimization::trap_reason_name(reason),
3932                   md->trap_count(reason),
3933                   md->overflow_recompile_count());
3934     return true;
3935   } else if (trap_count(reason) != 0
3936              && decompile_count() >= m_cutoff) {
3937     // Too many recompiles globally, and we have seen this sort of trap.
3938     // Use cumulative decompile_count, not just md->decompile_count.
3939     if (log())
3940       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3941                   Deoptimization::trap_reason_name(reason),
3942                   md->trap_count(reason), trap_count(reason),
3943                   md->decompile_count(), decompile_count());
3944     return true;
3945   } else {
3946     // The coast is clear.
3947     return false;
3948   }
3949 }
3950 
3951 // Compute when not to trap. Used by matching trap based nodes and
3952 // NullCheck optimization.
3953 void Compile::set_allowed_deopt_reasons() {
3954   _allowed_reasons = 0;
3955   if (is_method_compilation()) {
3956     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3957       assert(rs < BitsPerInt, "recode bit map");
3958       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3959         _allowed_reasons |= nth_bit(rs);
3960       }
3961     }
3962   }
3963 }
3964 
3965 bool Compile::is_compiling_clinit_for(ciKlass* k) {
3966   ciMethod* root = method(); // the root method of compilation
3967   return root->is_static_initializer() && root->holder() == k; // access in the context of clinit
3968 }
3969 
3970 #ifndef PRODUCT
3971 //------------------------------verify_graph_edges---------------------------
3972 // Walk the Graph and verify that there is a one-to-one correspondence
3973 // between Use-Def edges and Def-Use edges in the graph.
3974 void Compile::verify_graph_edges(bool no_dead_code) {
3975   if (VerifyGraphEdges) {
3976     ResourceArea *area = Thread::current()->resource_area();
3977     Unique_Node_List visited(area);
3978     // Call recursive graph walk to check edges
3979     _root->verify_edges(visited);
3980     if (no_dead_code) {
3981       // Now make sure that no visited node is used by an unvisited node.
3982       bool dead_nodes = false;
3983       Unique_Node_List checked(area);
3984       while (visited.size() > 0) {
3985         Node* n = visited.pop();
3986         checked.push(n);
3987         for (uint i = 0; i < n->outcnt(); i++) {
3988           Node* use = n->raw_out(i);
3989           if (checked.member(use))  continue;  // already checked
3990           if (visited.member(use))  continue;  // already in the graph
3991           if (use->is_Con())        continue;  // a dead ConNode is OK
3992           // At this point, we have found a dead node which is DU-reachable.
3993           if (!dead_nodes) {
3994             tty->print_cr("*** Dead nodes reachable via DU edges:");
3995             dead_nodes = true;
3996           }
3997           use->dump(2);
3998           tty->print_cr("---");
3999           checked.push(use);  // No repeats; pretend it is now checked.
4000         }
4001       }
4002       assert(!dead_nodes, "using nodes must be reachable from root");
4003     }
4004   }
4005 }
4006 #endif
4007 
4008 // The Compile object keeps track of failure reasons separately from the ciEnv.
4009 // This is required because there is not quite a 1-1 relation between the
4010 // ciEnv and its compilation task and the Compile object.  Note that one
4011 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
4012 // to backtrack and retry without subsuming loads.  Other than this backtracking
4013 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
4014 // by the logic in C2Compiler.
4015 void Compile::record_failure(const char* reason) {
4016   if (log() != NULL) {
4017     log()->elem("failure reason='%s' phase='compile'", reason);
4018   }
4019   if (_failure_reason == NULL) {
4020     // Record the first failure reason.
4021     _failure_reason = reason;
4022   }
4023 
4024   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
4025     C->print_method(PHASE_FAILURE);
4026   }
4027   _root = NULL;  // flush the graph, too
4028 }
4029 
4030 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
4031   : TraceTime(name, accumulator, CITime, CITimeVerbose),
4032     _phase_name(name), _dolog(CITimeVerbose)
4033 {
4034   if (_dolog) {
4035     C = Compile::current();
4036     _log = C->log();
4037   } else {
4038     C = NULL;
4039     _log = NULL;
4040   }
4041   if (_log != NULL) {
4042     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4043     _log->stamp();
4044     _log->end_head();
4045   }
4046 }
4047 
4048 Compile::TracePhase::~TracePhase() {
4049 
4050   C = Compile::current();
4051   if (_dolog) {
4052     _log = C->log();
4053   } else {
4054     _log = NULL;
4055   }
4056 
4057 #ifdef ASSERT
4058   if (PrintIdealNodeCount) {
4059     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
4060                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
4061   }
4062 
4063   if (VerifyIdealNodeCount) {
4064     Compile::current()->print_missing_nodes();
4065   }
4066 #endif
4067 
4068   if (_log != NULL) {
4069     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
4070   }
4071 }
4072 
4073 //=============================================================================
4074 // Two Constant's are equal when the type and the value are equal.
4075 bool Compile::Constant::operator==(const Constant& other) {
4076   if (type()          != other.type()         )  return false;
4077   if (can_be_reused() != other.can_be_reused())  return false;
4078   // For floating point values we compare the bit pattern.
4079   switch (type()) {
4080   case T_INT:
4081   case T_FLOAT:   return (_v._value.i == other._v._value.i);
4082   case T_LONG:
4083   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
4084   case T_OBJECT:
4085   case T_ADDRESS: return (_v._value.l == other._v._value.l);
4086   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
4087   case T_METADATA: return (_v._metadata == other._v._metadata);
4088   default: ShouldNotReachHere(); return false;
4089   }
4090 }
4091 
4092 static int type_to_size_in_bytes(BasicType t) {
4093   switch (t) {
4094   case T_INT:     return sizeof(jint   );
4095   case T_LONG:    return sizeof(jlong  );
4096   case T_FLOAT:   return sizeof(jfloat );
4097   case T_DOUBLE:  return sizeof(jdouble);
4098   case T_METADATA: return sizeof(Metadata*);
4099     // We use T_VOID as marker for jump-table entries (labels) which
4100     // need an internal word relocation.
4101   case T_VOID:
4102   case T_ADDRESS:
4103   case T_OBJECT:  return sizeof(jobject);
4104   default:
4105     ShouldNotReachHere();
4106     return -1;
4107   }
4108 }
4109 
4110 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4111   // sort descending
4112   if (a->freq() > b->freq())  return -1;
4113   if (a->freq() < b->freq())  return  1;
4114   return 0;
4115 }
4116 
4117 void Compile::ConstantTable::calculate_offsets_and_size() {
4118   // First, sort the array by frequencies.
4119   _constants.sort(qsort_comparator);
4120 
4121 #ifdef ASSERT
4122   // Make sure all jump-table entries were sorted to the end of the
4123   // array (they have a negative frequency).
4124   bool found_void = false;
4125   for (int i = 0; i < _constants.length(); i++) {
4126     Constant con = _constants.at(i);
4127     if (con.type() == T_VOID)
4128       found_void = true;  // jump-tables
4129     else
4130       assert(!found_void, "wrong sorting");
4131   }
4132 #endif
4133 
4134   int offset = 0;
4135   for (int i = 0; i < _constants.length(); i++) {
4136     Constant* con = _constants.adr_at(i);
4137 
4138     // Align offset for type.
4139     int typesize = type_to_size_in_bytes(con->type());
4140     offset = align_up(offset, typesize);
4141     con->set_offset(offset);   // set constant's offset
4142 
4143     if (con->type() == T_VOID) {
4144       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4145       offset = offset + typesize * n->outcnt();  // expand jump-table
4146     } else {
4147       offset = offset + typesize;
4148     }
4149   }
4150 
4151   // Align size up to the next section start (which is insts; see
4152   // CodeBuffer::align_at_start).
4153   assert(_size == -1, "already set?");
4154   _size = align_up(offset, (int)CodeEntryAlignment);
4155 }
4156 
4157 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4158   MacroAssembler _masm(&cb);
4159   for (int i = 0; i < _constants.length(); i++) {
4160     Constant con = _constants.at(i);
4161     address constant_addr = NULL;
4162     switch (con.type()) {
4163     case T_INT:    constant_addr = _masm.int_constant(   con.get_jint()   ); break;
4164     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
4165     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4166     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4167     case T_OBJECT: {
4168       jobject obj = con.get_jobject();
4169       int oop_index = _masm.oop_recorder()->find_index(obj);
4170       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4171       break;
4172     }
4173     case T_ADDRESS: {
4174       address addr = (address) con.get_jobject();
4175       constant_addr = _masm.address_constant(addr);
4176       break;
4177     }
4178     // We use T_VOID as marker for jump-table entries (labels) which
4179     // need an internal word relocation.
4180     case T_VOID: {
4181       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4182       // Fill the jump-table with a dummy word.  The real value is
4183       // filled in later in fill_jump_table.
4184       address dummy = (address) n;
4185       constant_addr = _masm.address_constant(dummy);
4186       // Expand jump-table
4187       for (uint i = 1; i < n->outcnt(); i++) {
4188         address temp_addr = _masm.address_constant(dummy + i);
4189         assert(temp_addr, "consts section too small");
4190       }
4191       break;
4192     }
4193     case T_METADATA: {
4194       Metadata* obj = con.get_metadata();
4195       int metadata_index = _masm.oop_recorder()->find_index(obj);
4196       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4197       break;
4198     }
4199     default: ShouldNotReachHere();
4200     }
4201     assert(constant_addr, "consts section too small");
4202     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4203             "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4204   }
4205 }
4206 
4207 int Compile::ConstantTable::find_offset(Constant& con) const {
4208   int idx = _constants.find(con);
4209   guarantee(idx != -1, "constant must be in constant table");
4210   int offset = _constants.at(idx).offset();
4211   guarantee(offset != -1, "constant table not emitted yet?");
4212   return offset;
4213 }
4214 
4215 void Compile::ConstantTable::add(Constant& con) {
4216   if (con.can_be_reused()) {
4217     int idx = _constants.find(con);
4218     if (idx != -1 && _constants.at(idx).can_be_reused()) {
4219       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
4220       return;
4221     }
4222   }
4223   (void) _constants.append(con);
4224 }
4225 
4226 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4227   Block* b = Compile::current()->cfg()->get_block_for_node(n);
4228   Constant con(type, value, b->_freq);
4229   add(con);
4230   return con;
4231 }
4232 
4233 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4234   Constant con(metadata);
4235   add(con);
4236   return con;
4237 }
4238 
4239 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4240   jvalue value;
4241   BasicType type = oper->type()->basic_type();
4242   switch (type) {
4243   case T_LONG:    value.j = oper->constantL(); break;
4244   case T_FLOAT:   value.f = oper->constantF(); break;
4245   case T_DOUBLE:  value.d = oper->constantD(); break;
4246   case T_OBJECT:
4247   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4248   case T_METADATA: return add((Metadata*)oper->constant()); break;
4249   default: guarantee(false, "unhandled type: %s", type2name(type));
4250   }
4251   return add(n, type, value);
4252 }
4253 
4254 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4255   jvalue value;
4256   // We can use the node pointer here to identify the right jump-table
4257   // as this method is called from Compile::Fill_buffer right before
4258   // the MachNodes are emitted and the jump-table is filled (means the
4259   // MachNode pointers do not change anymore).
4260   value.l = (jobject) n;
4261   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
4262   add(con);
4263   return con;
4264 }
4265 
4266 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4267   // If called from Compile::scratch_emit_size do nothing.
4268   if (Compile::current()->in_scratch_emit_size())  return;
4269 
4270   assert(labels.is_nonempty(), "must be");
4271   assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4272 
4273   // Since MachConstantNode::constant_offset() also contains
4274   // table_base_offset() we need to subtract the table_base_offset()
4275   // to get the plain offset into the constant table.
4276   int offset = n->constant_offset() - table_base_offset();
4277 
4278   MacroAssembler _masm(&cb);
4279   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4280 
4281   for (uint i = 0; i < n->outcnt(); i++) {
4282     address* constant_addr = &jump_table_base[i];
4283     assert(*constant_addr == (((address) n) + i), "all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i));
4284     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4285     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4286   }
4287 }
4288 
4289 //----------------------------static_subtype_check-----------------------------
4290 // Shortcut important common cases when superklass is exact:
4291 // (0) superklass is java.lang.Object (can occur in reflective code)
4292 // (1) subklass is already limited to a subtype of superklass => always ok
4293 // (2) subklass does not overlap with superklass => always fail
4294 // (3) superklass has NO subtypes and we can check with a simple compare.
4295 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4296   if (StressReflectiveCode || superk == NULL || subk == NULL) {
4297     return SSC_full_test;       // Let caller generate the general case.
4298   }
4299 
4300   if (superk == env()->Object_klass()) {
4301     return SSC_always_true;     // (0) this test cannot fail
4302   }
4303 
4304   ciType* superelem = superk;
4305   if (superelem->is_array_klass())
4306     superelem = superelem->as_array_klass()->base_element_type();
4307 
4308   if (!subk->is_interface()) {  // cannot trust static interface types yet
4309     if (subk->is_subtype_of(superk)) {
4310       return SSC_always_true;   // (1) false path dead; no dynamic test needed
4311     }
4312     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4313         !superk->is_subtype_of(subk)) {
4314       return SSC_always_false;
4315     }
4316   }
4317 
4318   // If casting to an instance klass, it must have no subtypes
4319   if (superk->is_interface()) {
4320     // Cannot trust interfaces yet.
4321     // %%% S.B. superk->nof_implementors() == 1
4322   } else if (superelem->is_instance_klass()) {
4323     ciInstanceKlass* ik = superelem->as_instance_klass();
4324     if (!ik->has_subklass() && !ik->is_interface()) {
4325       if (!ik->is_final()) {
4326         // Add a dependency if there is a chance of a later subclass.
4327         dependencies()->assert_leaf_type(ik);
4328       }
4329       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
4330     }
4331   } else {
4332     // A primitive array type has no subtypes.
4333     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
4334   }
4335 
4336   return SSC_full_test;
4337 }
4338 
4339 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4340 #ifdef _LP64
4341   // The scaled index operand to AddP must be a clean 64-bit value.
4342   // Java allows a 32-bit int to be incremented to a negative
4343   // value, which appears in a 64-bit register as a large
4344   // positive number.  Using that large positive number as an
4345   // operand in pointer arithmetic has bad consequences.
4346   // On the other hand, 32-bit overflow is rare, and the possibility
4347   // can often be excluded, if we annotate the ConvI2L node with
4348   // a type assertion that its value is known to be a small positive
4349   // number.  (The prior range check has ensured this.)
4350   // This assertion is used by ConvI2LNode::Ideal.
4351   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4352   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4353   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4354   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4355 #endif
4356   return idx;
4357 }
4358 
4359 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4360 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4361   if (ctrl != NULL) {
4362     // Express control dependency by a CastII node with a narrow type.
4363     value = new CastIINode(value, itype, false, true /* range check dependency */);
4364     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4365     // node from floating above the range check during loop optimizations. Otherwise, the
4366     // ConvI2L node may be eliminated independently of the range check, causing the data path
4367     // to become TOP while the control path is still there (although it's unreachable).
4368     value->set_req(0, ctrl);
4369     // Save CastII node to remove it after loop optimizations.
4370     phase->C->add_range_check_cast(value);
4371     value = phase->transform(value);
4372   }
4373   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4374   return phase->transform(new ConvI2LNode(value, ltype));
4375 }
4376 
4377 // The message about the current inlining is accumulated in
4378 // _print_inlining_stream and transfered into the _print_inlining_list
4379 // once we know whether inlining succeeds or not. For regular
4380 // inlining, messages are appended to the buffer pointed by
4381 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4382 // a new buffer is added after _print_inlining_idx in the list. This
4383 // way we can update the inlining message for late inlining call site
4384 // when the inlining is attempted again.
4385 void Compile::print_inlining_init() {
4386   if (print_inlining() || print_intrinsics()) {
4387     _print_inlining_stream = new stringStream();
4388     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4389   }
4390 }
4391 
4392 void Compile::print_inlining_reinit() {
4393   if (print_inlining() || print_intrinsics()) {
4394     // Re allocate buffer when we change ResourceMark
4395     _print_inlining_stream = new stringStream();
4396   }
4397 }
4398 
4399 void Compile::print_inlining_reset() {
4400   _print_inlining_stream->reset();
4401 }
4402 
4403 void Compile::print_inlining_commit() {
4404   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4405   // Transfer the message from _print_inlining_stream to the current
4406   // _print_inlining_list buffer and clear _print_inlining_stream.
4407   _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4408   print_inlining_reset();
4409 }
4410 
4411 void Compile::print_inlining_push() {
4412   // Add new buffer to the _print_inlining_list at current position
4413   _print_inlining_idx++;
4414   _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4415 }
4416 
4417 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4418   return _print_inlining_list->at(_print_inlining_idx);
4419 }
4420 
4421 void Compile::print_inlining_update(CallGenerator* cg) {
4422   if (print_inlining() || print_intrinsics()) {
4423     if (!cg->is_late_inline()) {
4424       if (print_inlining_current().cg() != NULL) {
4425         print_inlining_push();
4426       }
4427       print_inlining_commit();
4428     } else {
4429       if (print_inlining_current().cg() != cg &&
4430           (print_inlining_current().cg() != NULL ||
4431            print_inlining_current().ss()->size() != 0)) {
4432         print_inlining_push();
4433       }
4434       print_inlining_commit();
4435       print_inlining_current().set_cg(cg);
4436     }
4437   }
4438 }
4439 
4440 void Compile::print_inlining_move_to(CallGenerator* cg) {
4441   // We resume inlining at a late inlining call site. Locate the
4442   // corresponding inlining buffer so that we can update it.
4443   if (print_inlining()) {
4444     for (int i = 0; i < _print_inlining_list->length(); i++) {
4445       if (_print_inlining_list->adr_at(i)->cg() == cg) {
4446         _print_inlining_idx = i;
4447         return;
4448       }
4449     }
4450     ShouldNotReachHere();
4451   }
4452 }
4453 
4454 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4455   if (print_inlining()) {
4456     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4457     assert(print_inlining_current().cg() == cg, "wrong entry");
4458     // replace message with new message
4459     _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4460     print_inlining_commit();
4461     print_inlining_current().set_cg(cg);
4462   }
4463 }
4464 
4465 void Compile::print_inlining_assert_ready() {
4466   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4467 }
4468 
4469 void Compile::process_print_inlining() {
4470   bool do_print_inlining = print_inlining() || print_intrinsics();
4471   if (do_print_inlining || log() != NULL) {
4472     // Print inlining message for candidates that we couldn't inline
4473     // for lack of space
4474     for (int i = 0; i < _late_inlines.length(); i++) {
4475       CallGenerator* cg = _late_inlines.at(i);
4476       if (!cg->is_mh_late_inline()) {
4477         const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4478         if (do_print_inlining) {
4479           cg->print_inlining_late(msg);
4480         }
4481         log_late_inline_failure(cg, msg);
4482       }
4483     }
4484   }
4485   if (do_print_inlining) {
4486     ResourceMark rm;
4487     stringStream ss;
4488     for (int i = 0; i < _print_inlining_list->length(); i++) {
4489       ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4490     }
4491     size_t end = ss.size();
4492     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4493     strncpy(_print_inlining_output, ss.base(), end+1);
4494     _print_inlining_output[end] = 0;
4495   }
4496 }
4497 
4498 void Compile::dump_print_inlining() {
4499   if (_print_inlining_output != NULL) {
4500     tty->print_raw(_print_inlining_output);
4501   }
4502 }
4503 
4504 void Compile::log_late_inline(CallGenerator* cg) {
4505   if (log() != NULL) {
4506     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4507                 cg->unique_id());
4508     JVMState* p = cg->call_node()->jvms();
4509     while (p != NULL) {
4510       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4511       p = p->caller();
4512     }
4513     log()->tail("late_inline");
4514   }
4515 }
4516 
4517 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4518   log_late_inline(cg);
4519   if (log() != NULL) {
4520     log()->inline_fail(msg);
4521   }
4522 }
4523 
4524 void Compile::log_inline_id(CallGenerator* cg) {
4525   if (log() != NULL) {
4526     // The LogCompilation tool needs a unique way to identify late
4527     // inline call sites. This id must be unique for this call site in
4528     // this compilation. Try to have it unique across compilations as
4529     // well because it can be convenient when grepping through the log
4530     // file.
4531     // Distinguish OSR compilations from others in case CICountOSR is
4532     // on.
4533     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4534     cg->set_unique_id(id);
4535     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4536   }
4537 }
4538 
4539 void Compile::log_inline_failure(const char* msg) {
4540   if (C->log() != NULL) {
4541     C->log()->inline_fail(msg);
4542   }
4543 }
4544 
4545 
4546 // Dump inlining replay data to the stream.
4547 // Don't change thread state and acquire any locks.
4548 void Compile::dump_inline_data(outputStream* out) {
4549   InlineTree* inl_tree = ilt();
4550   if (inl_tree != NULL) {
4551     out->print(" inline %d", inl_tree->count());
4552     inl_tree->dump_replay_data(out);
4553   }
4554 }
4555 
4556 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4557   if (n1->Opcode() < n2->Opcode())      return -1;
4558   else if (n1->Opcode() > n2->Opcode()) return 1;
4559 
4560   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4561   for (uint i = 1; i < n1->req(); i++) {
4562     if (n1->in(i) < n2->in(i))      return -1;
4563     else if (n1->in(i) > n2->in(i)) return 1;
4564   }
4565 
4566   return 0;
4567 }
4568 
4569 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4570   Node* n1 = *n1p;
4571   Node* n2 = *n2p;
4572 
4573   return cmp_expensive_nodes(n1, n2);
4574 }
4575 
4576 void Compile::sort_expensive_nodes() {
4577   if (!expensive_nodes_sorted()) {
4578     _expensive_nodes->sort(cmp_expensive_nodes);
4579   }
4580 }
4581 
4582 bool Compile::expensive_nodes_sorted() const {
4583   for (int i = 1; i < _expensive_nodes->length(); i++) {
4584     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4585       return false;
4586     }
4587   }
4588   return true;
4589 }
4590 
4591 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4592   if (_expensive_nodes->length() == 0) {
4593     return false;
4594   }
4595 
4596   assert(OptimizeExpensiveOps, "optimization off?");
4597 
4598   // Take this opportunity to remove dead nodes from the list
4599   int j = 0;
4600   for (int i = 0; i < _expensive_nodes->length(); i++) {
4601     Node* n = _expensive_nodes->at(i);
4602     if (!n->is_unreachable(igvn)) {
4603       assert(n->is_expensive(), "should be expensive");
4604       _expensive_nodes->at_put(j, n);
4605       j++;
4606     }
4607   }
4608   _expensive_nodes->trunc_to(j);
4609 
4610   // Then sort the list so that similar nodes are next to each other
4611   // and check for at least two nodes of identical kind with same data
4612   // inputs.
4613   sort_expensive_nodes();
4614 
4615   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4616     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4617       return true;
4618     }
4619   }
4620 
4621   return false;
4622 }
4623 
4624 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4625   if (_expensive_nodes->length() == 0) {
4626     return;
4627   }
4628 
4629   assert(OptimizeExpensiveOps, "optimization off?");
4630 
4631   // Sort to bring similar nodes next to each other and clear the
4632   // control input of nodes for which there's only a single copy.
4633   sort_expensive_nodes();
4634 
4635   int j = 0;
4636   int identical = 0;
4637   int i = 0;
4638   bool modified = false;
4639   for (; i < _expensive_nodes->length()-1; i++) {
4640     assert(j <= i, "can't write beyond current index");
4641     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4642       identical++;
4643       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4644       continue;
4645     }
4646     if (identical > 0) {
4647       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4648       identical = 0;
4649     } else {
4650       Node* n = _expensive_nodes->at(i);
4651       igvn.replace_input_of(n, 0, NULL);
4652       igvn.hash_insert(n);
4653       modified = true;
4654     }
4655   }
4656   if (identical > 0) {
4657     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4658   } else if (_expensive_nodes->length() >= 1) {
4659     Node* n = _expensive_nodes->at(i);
4660     igvn.replace_input_of(n, 0, NULL);
4661     igvn.hash_insert(n);
4662     modified = true;
4663   }
4664   _expensive_nodes->trunc_to(j);
4665   if (modified) {
4666     igvn.optimize();
4667   }
4668 }
4669 
4670 void Compile::add_expensive_node(Node * n) {
4671   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4672   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4673   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4674   if (OptimizeExpensiveOps) {
4675     _expensive_nodes->append(n);
4676   } else {
4677     // Clear control input and let IGVN optimize expensive nodes if
4678     // OptimizeExpensiveOps is off.
4679     n->set_req(0, NULL);
4680   }
4681 }
4682 
4683 /**
4684  * Remove the speculative part of types and clean up the graph
4685  */
4686 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4687   if (UseTypeSpeculation) {
4688     Unique_Node_List worklist;
4689     worklist.push(root());
4690     int modified = 0;
4691     // Go over all type nodes that carry a speculative type, drop the
4692     // speculative part of the type and enqueue the node for an igvn
4693     // which may optimize it out.
4694     for (uint next = 0; next < worklist.size(); ++next) {
4695       Node *n  = worklist.at(next);
4696       if (n->is_Type()) {
4697         TypeNode* tn = n->as_Type();
4698         const Type* t = tn->type();
4699         const Type* t_no_spec = t->remove_speculative();
4700         if (t_no_spec != t) {
4701           bool in_hash = igvn.hash_delete(n);
4702           assert(in_hash, "node should be in igvn hash table");
4703           tn->set_type(t_no_spec);
4704           igvn.hash_insert(n);
4705           igvn._worklist.push(n); // give it a chance to go away
4706           modified++;
4707         }
4708       }
4709       uint max = n->len();
4710       for( uint i = 0; i < max; ++i ) {
4711         Node *m = n->in(i);
4712         if (not_a_node(m))  continue;
4713         worklist.push(m);
4714       }
4715     }
4716     // Drop the speculative part of all types in the igvn's type table
4717     igvn.remove_speculative_types();
4718     if (modified > 0) {
4719       igvn.optimize();
4720     }
4721 #ifdef ASSERT
4722     // Verify that after the IGVN is over no speculative type has resurfaced
4723     worklist.clear();
4724     worklist.push(root());
4725     for (uint next = 0; next < worklist.size(); ++next) {
4726       Node *n  = worklist.at(next);
4727       const Type* t = igvn.type_or_null(n);
4728       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4729       if (n->is_Type()) {
4730         t = n->as_Type()->type();
4731         assert(t == t->remove_speculative(), "no more speculative types");
4732       }
4733       uint max = n->len();
4734       for( uint i = 0; i < max; ++i ) {
4735         Node *m = n->in(i);
4736         if (not_a_node(m))  continue;
4737         worklist.push(m);
4738       }
4739     }
4740     igvn.check_no_speculative_types();
4741 #endif
4742   }
4743 }
4744 
4745 Node* Compile::optimize_acmp(PhaseGVN* phase, Node* a, Node* b) {
4746   const TypeInstPtr* ta = phase->type(a)->isa_instptr();
4747   const TypeInstPtr* tb = phase->type(b)->isa_instptr();
4748   if (!EnableValhalla || ta == NULL || tb == NULL ||
4749       ta->is_zero_type() || tb->is_zero_type() ||
4750       !ta->can_be_value_type() || !tb->can_be_value_type()) {
4751     // Use old acmp if one operand is null or not a value type
4752     return new CmpPNode(a, b);
4753   } else if (ta->is_valuetypeptr() || tb->is_valuetypeptr()) {
4754     // We know that one operand is a value type. Therefore,
4755     // new acmp will only return true if both operands are NULL.
4756     // Check if both operands are null by or'ing the oops.
4757     a = phase->transform(new CastP2XNode(NULL, a));
4758     b = phase->transform(new CastP2XNode(NULL, b));
4759     a = phase->transform(new OrXNode(a, b));
4760     return new CmpXNode(a, phase->MakeConX(0));
4761   }
4762   // Use new acmp
4763   return NULL;
4764 }
4765 
4766 // Auxiliary method to support randomized stressing/fuzzing.
4767 //
4768 // This method can be called the arbitrary number of times, with current count
4769 // as the argument. The logic allows selecting a single candidate from the
4770 // running list of candidates as follows:
4771 //    int count = 0;
4772 //    Cand* selected = null;
4773 //    while(cand = cand->next()) {
4774 //      if (randomized_select(++count)) {
4775 //        selected = cand;
4776 //      }
4777 //    }
4778 //
4779 // Including count equalizes the chances any candidate is "selected".
4780 // This is useful when we don't have the complete list of candidates to choose
4781 // from uniformly. In this case, we need to adjust the randomicity of the
4782 // selection, or else we will end up biasing the selection towards the latter
4783 // candidates.
4784 //
4785 // Quick back-envelope calculation shows that for the list of n candidates
4786 // the equal probability for the candidate to persist as "best" can be
4787 // achieved by replacing it with "next" k-th candidate with the probability
4788 // of 1/k. It can be easily shown that by the end of the run, the
4789 // probability for any candidate is converged to 1/n, thus giving the
4790 // uniform distribution among all the candidates.
4791 //
4792 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4793 #define RANDOMIZED_DOMAIN_POW 29
4794 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4795 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4796 bool Compile::randomized_select(int count) {
4797   assert(count > 0, "only positive");
4798   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4799 }
4800 
4801 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4802 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4803 
4804 void NodeCloneInfo::dump() const {
4805   tty->print(" {%d:%d} ", idx(), gen());
4806 }
4807 
4808 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4809   uint64_t val = value(old->_idx);
4810   NodeCloneInfo cio(val);
4811   assert(val != 0, "old node should be in the map");
4812   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4813   insert(nnn->_idx, cin.get());
4814 #ifndef PRODUCT
4815   if (is_debug()) {
4816     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4817   }
4818 #endif
4819 }
4820 
4821 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4822   NodeCloneInfo cio(value(old->_idx));
4823   if (cio.get() == 0) {
4824     cio.set(old->_idx, 0);
4825     insert(old->_idx, cio.get());
4826 #ifndef PRODUCT
4827     if (is_debug()) {
4828       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4829     }
4830 #endif
4831   }
4832   clone(old, nnn, gen);
4833 }
4834 
4835 int CloneMap::max_gen() const {
4836   int g = 0;
4837   DictI di(_dict);
4838   for(; di.test(); ++di) {
4839     int t = gen(di._key);
4840     if (g < t) {
4841       g = t;
4842 #ifndef PRODUCT
4843       if (is_debug()) {
4844         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4845       }
4846 #endif
4847     }
4848   }
4849   return g;
4850 }
4851 
4852 void CloneMap::dump(node_idx_t key) const {
4853   uint64_t val = value(key);
4854   if (val != 0) {
4855     NodeCloneInfo ni(val);
4856     ni.dump();
4857   }
4858 }
--- EOF ---