1 /*
   2  * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.hpp"
  27 #include "asm/macroAssembler.inline.hpp"
  28 #include "ci/ciReplay.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/exceptionHandlerTable.hpp"
  31 #include "code/nmethod.hpp"
  32 #include "compiler/compileBroker.hpp"
  33 #include "compiler/compileLog.hpp"
  34 #include "compiler/disassembler.hpp"
  35 #include "compiler/oopMap.hpp"
  36 #include "gc/shared/barrierSet.hpp"
  37 #include "gc/shared/c2/barrierSetC2.hpp"
  38 #include "memory/resourceArea.hpp"
  39 #include "opto/addnode.hpp"
  40 #include "opto/block.hpp"
  41 #include "opto/c2compiler.hpp"
  42 #include "opto/callGenerator.hpp"
  43 #include "opto/callnode.hpp"
  44 #include "opto/castnode.hpp"
  45 #include "opto/cfgnode.hpp"
  46 #include "opto/chaitin.hpp"
  47 #include "opto/compile.hpp"
  48 #include "opto/connode.hpp"
  49 #include "opto/convertnode.hpp"
  50 #include "opto/divnode.hpp"
  51 #include "opto/escape.hpp"
  52 #include "opto/idealGraphPrinter.hpp"
  53 #include "opto/loopnode.hpp"
  54 #include "opto/machnode.hpp"
  55 #include "opto/macro.hpp"
  56 #include "opto/matcher.hpp"
  57 #include "opto/mathexactnode.hpp"
  58 #include "opto/memnode.hpp"
  59 #include "opto/mulnode.hpp"
  60 #include "opto/narrowptrnode.hpp"
  61 #include "opto/node.hpp"
  62 #include "opto/opcodes.hpp"
  63 #include "opto/output.hpp"
  64 #include "opto/parse.hpp"
  65 #include "opto/phaseX.hpp"
  66 #include "opto/rootnode.hpp"
  67 #include "opto/runtime.hpp"
  68 #include "opto/stringopts.hpp"
  69 #include "opto/type.hpp"
  70 #include "opto/vectornode.hpp"
  71 #include "runtime/arguments.hpp"
  72 #include "runtime/sharedRuntime.hpp"
  73 #include "runtime/signature.hpp"
  74 #include "runtime/stubRoutines.hpp"
  75 #include "runtime/timer.hpp"
  76 #include "utilities/align.hpp"
  77 #include "utilities/copy.hpp"
  78 #include "utilities/macros.hpp"
  79 #if INCLUDE_ZGC
  80 #include "gc/z/c2/zBarrierSetC2.hpp"
  81 #endif
  82 
  83 
  84 // -------------------- Compile::mach_constant_base_node -----------------------
  85 // Constant table base node singleton.
  86 MachConstantBaseNode* Compile::mach_constant_base_node() {
  87   if (_mach_constant_base_node == NULL) {
  88     _mach_constant_base_node = new MachConstantBaseNode();
  89     _mach_constant_base_node->add_req(C->root());
  90   }
  91   return _mach_constant_base_node;
  92 }
  93 
  94 
  95 /// Support for intrinsics.
  96 
  97 // Return the index at which m must be inserted (or already exists).
  98 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
  99 class IntrinsicDescPair {
 100  private:
 101   ciMethod* _m;
 102   bool _is_virtual;
 103  public:
 104   IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
 105   static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
 106     ciMethod* m= elt->method();
 107     ciMethod* key_m = key->_m;
 108     if (key_m < m)      return -1;
 109     else if (key_m > m) return 1;
 110     else {
 111       bool is_virtual = elt->is_virtual();
 112       bool key_virtual = key->_is_virtual;
 113       if (key_virtual < is_virtual)      return -1;
 114       else if (key_virtual > is_virtual) return 1;
 115       else                               return 0;
 116     }
 117   }
 118 };
 119 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
 120 #ifdef ASSERT
 121   for (int i = 1; i < _intrinsics->length(); i++) {
 122     CallGenerator* cg1 = _intrinsics->at(i-1);
 123     CallGenerator* cg2 = _intrinsics->at(i);
 124     assert(cg1->method() != cg2->method()
 125            ? cg1->method()     < cg2->method()
 126            : cg1->is_virtual() < cg2->is_virtual(),
 127            "compiler intrinsics list must stay sorted");
 128   }
 129 #endif
 130   IntrinsicDescPair pair(m, is_virtual);
 131   return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
 132 }
 133 
 134 void Compile::register_intrinsic(CallGenerator* cg) {
 135   if (_intrinsics == NULL) {
 136     _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 137   }
 138   int len = _intrinsics->length();
 139   bool found = false;
 140   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
 141   assert(!found, "registering twice");
 142   _intrinsics->insert_before(index, cg);
 143   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 144 }
 145 
 146 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 147   assert(m->is_loaded(), "don't try this on unloaded methods");
 148   if (_intrinsics != NULL) {
 149     bool found = false;
 150     int index = intrinsic_insertion_index(m, is_virtual, found);
 151      if (found) {
 152       return _intrinsics->at(index);
 153     }
 154   }
 155   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 156   if (m->intrinsic_id() != vmIntrinsics::_none &&
 157       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 158     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 159     if (cg != NULL) {
 160       // Save it for next time:
 161       register_intrinsic(cg);
 162       return cg;
 163     } else {
 164       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 165     }
 166   }
 167   return NULL;
 168 }
 169 
 170 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 171 // in library_call.cpp.
 172 
 173 
 174 #ifndef PRODUCT
 175 // statistics gathering...
 176 
 177 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 178 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 179 
 180 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 181   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 182   int oflags = _intrinsic_hist_flags[id];
 183   assert(flags != 0, "what happened?");
 184   if (is_virtual) {
 185     flags |= _intrinsic_virtual;
 186   }
 187   bool changed = (flags != oflags);
 188   if ((flags & _intrinsic_worked) != 0) {
 189     juint count = (_intrinsic_hist_count[id] += 1);
 190     if (count == 1) {
 191       changed = true;           // first time
 192     }
 193     // increment the overall count also:
 194     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 195   }
 196   if (changed) {
 197     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 198       // Something changed about the intrinsic's virtuality.
 199       if ((flags & _intrinsic_virtual) != 0) {
 200         // This is the first use of this intrinsic as a virtual call.
 201         if (oflags != 0) {
 202           // We already saw it as a non-virtual, so note both cases.
 203           flags |= _intrinsic_both;
 204         }
 205       } else if ((oflags & _intrinsic_both) == 0) {
 206         // This is the first use of this intrinsic as a non-virtual
 207         flags |= _intrinsic_both;
 208       }
 209     }
 210     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 211   }
 212   // update the overall flags also:
 213   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 214   return changed;
 215 }
 216 
 217 static char* format_flags(int flags, char* buf) {
 218   buf[0] = 0;
 219   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 220   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 221   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 222   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 223   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 224   if (buf[0] == 0)  strcat(buf, ",");
 225   assert(buf[0] == ',', "must be");
 226   return &buf[1];
 227 }
 228 
 229 void Compile::print_intrinsic_statistics() {
 230   char flagsbuf[100];
 231   ttyLocker ttyl;
 232   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 233   tty->print_cr("Compiler intrinsic usage:");
 234   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 235   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 236   #define PRINT_STAT_LINE(name, c, f) \
 237     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 238   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 239     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 240     int   flags = _intrinsic_hist_flags[id];
 241     juint count = _intrinsic_hist_count[id];
 242     if ((flags | count) != 0) {
 243       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 244     }
 245   }
 246   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 247   if (xtty != NULL)  xtty->tail("statistics");
 248 }
 249 
 250 void Compile::print_statistics() {
 251   { ttyLocker ttyl;
 252     if (xtty != NULL)  xtty->head("statistics type='opto'");
 253     Parse::print_statistics();
 254     PhaseCCP::print_statistics();
 255     PhaseRegAlloc::print_statistics();
 256     Scheduling::print_statistics();
 257     PhasePeephole::print_statistics();
 258     PhaseIdealLoop::print_statistics();
 259     if (xtty != NULL)  xtty->tail("statistics");
 260   }
 261   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 262     // put this under its own <statistics> element.
 263     print_intrinsic_statistics();
 264   }
 265 }
 266 #endif //PRODUCT
 267 
 268 // Support for bundling info
 269 Bundle* Compile::node_bundling(const Node *n) {
 270   assert(valid_bundle_info(n), "oob");
 271   return &_node_bundling_base[n->_idx];
 272 }
 273 
 274 bool Compile::valid_bundle_info(const Node *n) {
 275   return (_node_bundling_limit > n->_idx);
 276 }
 277 
 278 
 279 void Compile::gvn_replace_by(Node* n, Node* nn) {
 280   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 281     Node* use = n->last_out(i);
 282     bool is_in_table = initial_gvn()->hash_delete(use);
 283     uint uses_found = 0;
 284     for (uint j = 0; j < use->len(); j++) {
 285       if (use->in(j) == n) {
 286         if (j < use->req())
 287           use->set_req(j, nn);
 288         else
 289           use->set_prec(j, nn);
 290         uses_found++;
 291       }
 292     }
 293     if (is_in_table) {
 294       // reinsert into table
 295       initial_gvn()->hash_find_insert(use);
 296     }
 297     record_for_igvn(use);
 298     i -= uses_found;    // we deleted 1 or more copies of this edge
 299   }
 300 }
 301 
 302 
 303 static inline bool not_a_node(const Node* n) {
 304   if (n == NULL)                   return true;
 305   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 306   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 307   return false;
 308 }
 309 
 310 // Identify all nodes that are reachable from below, useful.
 311 // Use breadth-first pass that records state in a Unique_Node_List,
 312 // recursive traversal is slower.
 313 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 314   int estimated_worklist_size = live_nodes();
 315   useful.map( estimated_worklist_size, NULL );  // preallocate space
 316 
 317   // Initialize worklist
 318   if (root() != NULL)     { useful.push(root()); }
 319   // If 'top' is cached, declare it useful to preserve cached node
 320   if( cached_top_node() ) { useful.push(cached_top_node()); }
 321 
 322   // Push all useful nodes onto the list, breadthfirst
 323   for( uint next = 0; next < useful.size(); ++next ) {
 324     assert( next < unique(), "Unique useful nodes < total nodes");
 325     Node *n  = useful.at(next);
 326     uint max = n->len();
 327     for( uint i = 0; i < max; ++i ) {
 328       Node *m = n->in(i);
 329       if (not_a_node(m))  continue;
 330       useful.push(m);
 331     }
 332   }
 333 }
 334 
 335 // Update dead_node_list with any missing dead nodes using useful
 336 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 337 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 338   uint max_idx = unique();
 339   VectorSet& useful_node_set = useful.member_set();
 340 
 341   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 342     // If node with index node_idx is not in useful set,
 343     // mark it as dead in dead node list.
 344     if (! useful_node_set.test(node_idx) ) {
 345       record_dead_node(node_idx);
 346     }
 347   }
 348 }
 349 
 350 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 351   int shift = 0;
 352   for (int i = 0; i < inlines->length(); i++) {
 353     CallGenerator* cg = inlines->at(i);
 354     CallNode* call = cg->call_node();
 355     if (shift > 0) {
 356       inlines->at_put(i-shift, cg);
 357     }
 358     if (!useful.member(call)) {
 359       shift++;
 360     }
 361   }
 362   inlines->trunc_to(inlines->length()-shift);
 363 }
 364 
 365 // Disconnect all useless nodes by disconnecting those at the boundary.
 366 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 367   uint next = 0;
 368   while (next < useful.size()) {
 369     Node *n = useful.at(next++);
 370     if (n->is_SafePoint()) {
 371       // We're done with a parsing phase. Replaced nodes are not valid
 372       // beyond that point.
 373       n->as_SafePoint()->delete_replaced_nodes();
 374     }
 375     // Use raw traversal of out edges since this code removes out edges
 376     int max = n->outcnt();
 377     for (int j = 0; j < max; ++j) {
 378       Node* child = n->raw_out(j);
 379       if (! useful.member(child)) {
 380         assert(!child->is_top() || child != top(),
 381                "If top is cached in Compile object it is in useful list");
 382         // Only need to remove this out-edge to the useless node
 383         n->raw_del_out(j);
 384         --j;
 385         --max;
 386       }
 387     }
 388     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 389       record_for_igvn(n->unique_out());
 390     }
 391   }
 392   // Remove useless macro and predicate opaq nodes
 393   for (int i = C->macro_count()-1; i >= 0; i--) {
 394     Node* n = C->macro_node(i);
 395     if (!useful.member(n)) {
 396       remove_macro_node(n);
 397     }
 398   }
 399   // Remove useless CastII nodes with range check dependency
 400   for (int i = range_check_cast_count() - 1; i >= 0; i--) {
 401     Node* cast = range_check_cast_node(i);
 402     if (!useful.member(cast)) {
 403       remove_range_check_cast(cast);
 404     }
 405   }
 406   // Remove useless expensive nodes
 407   for (int i = C->expensive_count()-1; i >= 0; i--) {
 408     Node* n = C->expensive_node(i);
 409     if (!useful.member(n)) {
 410       remove_expensive_node(n);
 411     }
 412   }
 413   // Remove useless Opaque4 nodes
 414   for (int i = opaque4_count() - 1; i >= 0; i--) {
 415     Node* opaq = opaque4_node(i);
 416     if (!useful.member(opaq)) {
 417       remove_opaque4_node(opaq);
 418     }
 419   }
 420   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 421   bs->eliminate_useless_gc_barriers(useful, this);
 422   // clean up the late inline lists
 423   remove_useless_late_inlines(&_string_late_inlines, useful);
 424   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 425   remove_useless_late_inlines(&_late_inlines, useful);
 426   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 427 }
 428 
 429 //------------------------------frame_size_in_words-----------------------------
 430 // frame_slots in units of words
 431 int Compile::frame_size_in_words() const {
 432   // shift is 0 in LP32 and 1 in LP64
 433   const int shift = (LogBytesPerWord - LogBytesPerInt);
 434   int words = _frame_slots >> shift;
 435   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 436   return words;
 437 }
 438 
 439 // To bang the stack of this compiled method we use the stack size
 440 // that the interpreter would need in case of a deoptimization. This
 441 // removes the need to bang the stack in the deoptimization blob which
 442 // in turn simplifies stack overflow handling.
 443 int Compile::bang_size_in_bytes() const {
 444   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
 445 }
 446 
 447 // ============================================================================
 448 //------------------------------CompileWrapper---------------------------------
 449 class CompileWrapper : public StackObj {
 450   Compile *const _compile;
 451  public:
 452   CompileWrapper(Compile* compile);
 453 
 454   ~CompileWrapper();
 455 };
 456 
 457 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 458   // the Compile* pointer is stored in the current ciEnv:
 459   ciEnv* env = compile->env();
 460   assert(env == ciEnv::current(), "must already be a ciEnv active");
 461   assert(env->compiler_data() == NULL, "compile already active?");
 462   env->set_compiler_data(compile);
 463   assert(compile == Compile::current(), "sanity");
 464 
 465   compile->set_type_dict(NULL);
 466   compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
 467   compile->clone_map().set_clone_idx(0);
 468   compile->set_type_last_size(0);
 469   compile->set_last_tf(NULL, NULL);
 470   compile->set_indexSet_arena(NULL);
 471   compile->set_indexSet_free_block_list(NULL);
 472   compile->init_type_arena();
 473   Type::Initialize(compile);
 474   _compile->set_scratch_buffer_blob(NULL);
 475   _compile->begin_method();
 476   _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
 477 }
 478 CompileWrapper::~CompileWrapper() {
 479   _compile->end_method();
 480   if (_compile->scratch_buffer_blob() != NULL)
 481     BufferBlob::free(_compile->scratch_buffer_blob());
 482   _compile->env()->set_compiler_data(NULL);
 483 }
 484 
 485 
 486 //----------------------------print_compile_messages---------------------------
 487 void Compile::print_compile_messages() {
 488 #ifndef PRODUCT
 489   // Check if recompiling
 490   if (_subsume_loads == false && PrintOpto) {
 491     // Recompiling without allowing machine instructions to subsume loads
 492     tty->print_cr("*********************************************************");
 493     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 494     tty->print_cr("*********************************************************");
 495   }
 496   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 497     // Recompiling without escape analysis
 498     tty->print_cr("*********************************************************");
 499     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 500     tty->print_cr("*********************************************************");
 501   }
 502   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 503     // Recompiling without boxing elimination
 504     tty->print_cr("*********************************************************");
 505     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 506     tty->print_cr("*********************************************************");
 507   }
 508   if (C->directive()->BreakAtCompileOption) {
 509     // Open the debugger when compiling this method.
 510     tty->print("### Breaking when compiling: ");
 511     method()->print_short_name();
 512     tty->cr();
 513     BREAKPOINT;
 514   }
 515 
 516   if( PrintOpto ) {
 517     if (is_osr_compilation()) {
 518       tty->print("[OSR]%3d", _compile_id);
 519     } else {
 520       tty->print("%3d", _compile_id);
 521     }
 522   }
 523 #endif
 524 }
 525 
 526 
 527 //-----------------------init_scratch_buffer_blob------------------------------
 528 // Construct a temporary BufferBlob and cache it for this compile.
 529 void Compile::init_scratch_buffer_blob(int const_size) {
 530   // If there is already a scratch buffer blob allocated and the
 531   // constant section is big enough, use it.  Otherwise free the
 532   // current and allocate a new one.
 533   BufferBlob* blob = scratch_buffer_blob();
 534   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 535     // Use the current blob.
 536   } else {
 537     if (blob != NULL) {
 538       BufferBlob::free(blob);
 539     }
 540 
 541     ResourceMark rm;
 542     _scratch_const_size = const_size;
 543     int size = C2Compiler::initial_code_buffer_size(const_size);
 544     blob = BufferBlob::create("Compile::scratch_buffer", size);
 545     // Record the buffer blob for next time.
 546     set_scratch_buffer_blob(blob);
 547     // Have we run out of code space?
 548     if (scratch_buffer_blob() == NULL) {
 549       // Let CompilerBroker disable further compilations.
 550       record_failure("Not enough space for scratch buffer in CodeCache");
 551       return;
 552     }
 553   }
 554 
 555   // Initialize the relocation buffers
 556   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 557   set_scratch_locs_memory(locs_buf);
 558 }
 559 
 560 
 561 //-----------------------scratch_emit_size-------------------------------------
 562 // Helper function that computes size by emitting code
 563 uint Compile::scratch_emit_size(const Node* n) {
 564   // Start scratch_emit_size section.
 565   set_in_scratch_emit_size(true);
 566 
 567   // Emit into a trash buffer and count bytes emitted.
 568   // This is a pretty expensive way to compute a size,
 569   // but it works well enough if seldom used.
 570   // All common fixed-size instructions are given a size
 571   // method by the AD file.
 572   // Note that the scratch buffer blob and locs memory are
 573   // allocated at the beginning of the compile task, and
 574   // may be shared by several calls to scratch_emit_size.
 575   // The allocation of the scratch buffer blob is particularly
 576   // expensive, since it has to grab the code cache lock.
 577   BufferBlob* blob = this->scratch_buffer_blob();
 578   assert(blob != NULL, "Initialize BufferBlob at start");
 579   assert(blob->size() > MAX_inst_size, "sanity");
 580   relocInfo* locs_buf = scratch_locs_memory();
 581   address blob_begin = blob->content_begin();
 582   address blob_end   = (address)locs_buf;
 583   assert(blob->contains(blob_end), "sanity");
 584   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 585   buf.initialize_consts_size(_scratch_const_size);
 586   buf.initialize_stubs_size(MAX_stubs_size);
 587   assert(locs_buf != NULL, "sanity");
 588   int lsize = MAX_locs_size / 3;
 589   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 590   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 591   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 592   // Mark as scratch buffer.
 593   buf.consts()->set_scratch_emit();
 594   buf.insts()->set_scratch_emit();
 595   buf.stubs()->set_scratch_emit();
 596 
 597   // Do the emission.
 598 
 599   Label fakeL; // Fake label for branch instructions.
 600   Label*   saveL = NULL;
 601   uint save_bnum = 0;
 602   bool is_branch = n->is_MachBranch();
 603   if (is_branch) {
 604     MacroAssembler masm(&buf);
 605     masm.bind(fakeL);
 606     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 607     n->as_MachBranch()->label_set(&fakeL, 0);
 608   }
 609   n->emit(buf, this->regalloc());
 610 
 611   // Emitting into the scratch buffer should not fail
 612   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
 613 
 614   if (is_branch) // Restore label.
 615     n->as_MachBranch()->label_set(saveL, save_bnum);
 616 
 617   // End scratch_emit_size section.
 618   set_in_scratch_emit_size(false);
 619 
 620   return buf.insts_size();
 621 }
 622 
 623 
 624 // ============================================================================
 625 //------------------------------Compile standard-------------------------------
 626 debug_only( int Compile::_debug_idx = 100000; )
 627 
 628 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 629 // the continuation bci for on stack replacement.
 630 
 631 
 632 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 633                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
 634                 : Phase(Compiler),
 635                   _compile_id(ci_env->compile_id()),
 636                   _save_argument_registers(false),
 637                   _subsume_loads(subsume_loads),
 638                   _do_escape_analysis(do_escape_analysis),
 639                   _eliminate_boxing(eliminate_boxing),
 640                   _method(target),
 641                   _entry_bci(osr_bci),
 642                   _stub_function(NULL),
 643                   _stub_name(NULL),
 644                   _stub_entry_point(NULL),
 645                   _max_node_limit(MaxNodeLimit),
 646                   _orig_pc_slot(0),
 647                   _orig_pc_slot_offset_in_bytes(0),
 648                   _inlining_progress(false),
 649                   _inlining_incrementally(false),
 650                   _do_cleanup(false),
 651                   _has_reserved_stack_access(target->has_reserved_stack_access()),
 652 #ifndef PRODUCT
 653                   _trace_opto_output(directive->TraceOptoOutputOption),
 654 #endif
 655                   _has_method_handle_invokes(false),
 656                   _clinit_barrier_on_entry(false),
 657                   _comp_arena(mtCompiler),
 658                   _barrier_set_state(BarrierSet::barrier_set()->barrier_set_c2()->create_barrier_state(comp_arena())),
 659                   _env(ci_env),
 660                   _directive(directive),
 661                   _log(ci_env->log()),
 662                   _failure_reason(NULL),
 663                   _congraph(NULL),
 664 #ifndef PRODUCT
 665                   _printer(IdealGraphPrinter::printer()),
 666 #endif
 667                   _dead_node_list(comp_arena()),
 668                   _dead_node_count(0),
 669                   _node_arena(mtCompiler),
 670                   _old_arena(mtCompiler),
 671                   _mach_constant_base_node(NULL),
 672                   _Compile_types(mtCompiler),
 673                   _initial_gvn(NULL),
 674                   _for_igvn(NULL),
 675                   _warm_calls(NULL),
 676                   _late_inlines(comp_arena(), 2, 0, NULL),
 677                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 678                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 679                   _late_inlines_pos(0),
 680                   _number_of_mh_late_inlines(0),
 681                   _print_inlining_stream(NULL),
 682                   _print_inlining_list(NULL),
 683                   _print_inlining_idx(0),
 684                   _print_inlining_output(NULL),
 685                   _replay_inline_data(NULL),
 686                   _java_calls(0),
 687                   _inner_loops(0),
 688                   _interpreter_frame_size(0),
 689                   _node_bundling_limit(0),
 690                   _node_bundling_base(NULL),
 691                   _code_buffer("Compile::Fill_buffer"),
 692                   _scratch_const_size(-1),
 693                   _in_scratch_emit_size(false)
 694 #ifndef PRODUCT
 695                   , _in_dump_cnt(0)
 696 #endif
 697 {
 698   C = this;
 699 #ifndef PRODUCT
 700   if (_printer != NULL) {
 701     _printer->set_compile(this);
 702   }
 703 #endif
 704   CompileWrapper cw(this);
 705 
 706   if (CITimeVerbose) {
 707     tty->print(" ");
 708     target->holder()->name()->print();
 709     tty->print(".");
 710     target->print_short_name();
 711     tty->print("  ");
 712   }
 713   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 714   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 715 
 716 #if defined(SUPPORT_ASSEMBLY) || defined(SUPPORT_ABSTRACT_ASSEMBLY)
 717   bool print_opto_assembly = directive->PrintOptoAssemblyOption;
 718   // We can always print a disassembly, either abstract (hex dump) or
 719   // with the help of a suitable hsdis library. Thus, we should not
 720   // couple print_assembly and print_opto_assembly controls.
 721   // But: always print opto and regular assembly on compile command 'print'.
 722   bool print_assembly = directive->PrintAssemblyOption;
 723   set_print_assembly(print_opto_assembly || print_assembly);
 724 #else
 725   set_print_assembly(false); // must initialize.
 726 #endif
 727 
 728 #ifndef PRODUCT
 729   set_parsed_irreducible_loop(false);
 730 
 731   if (directive->ReplayInlineOption) {
 732     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 733   }
 734 #endif
 735   set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
 736   set_print_intrinsics(directive->PrintIntrinsicsOption);
 737   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 738 
 739   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 740     // Make sure the method being compiled gets its own MDO,
 741     // so we can at least track the decompile_count().
 742     // Need MDO to record RTM code generation state.
 743     method()->ensure_method_data();
 744   }
 745 
 746   Init(::AliasLevel);
 747 
 748 
 749   print_compile_messages();
 750 
 751   _ilt = InlineTree::build_inline_tree_root();
 752 
 753   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 754   assert(num_alias_types() >= AliasIdxRaw, "");
 755 
 756 #define MINIMUM_NODE_HASH  1023
 757   // Node list that Iterative GVN will start with
 758   Unique_Node_List for_igvn(comp_arena());
 759   set_for_igvn(&for_igvn);
 760 
 761   // GVN that will be run immediately on new nodes
 762   uint estimated_size = method()->code_size()*4+64;
 763   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 764   PhaseGVN gvn(node_arena(), estimated_size);
 765   set_initial_gvn(&gvn);
 766 
 767   print_inlining_init();
 768   { // Scope for timing the parser
 769     TracePhase tp("parse", &timers[_t_parser]);
 770 
 771     // Put top into the hash table ASAP.
 772     initial_gvn()->transform_no_reclaim(top());
 773 
 774     // Set up tf(), start(), and find a CallGenerator.
 775     CallGenerator* cg = NULL;
 776     if (is_osr_compilation()) {
 777       const TypeTuple *domain = StartOSRNode::osr_domain();
 778       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 779       init_tf(TypeFunc::make(domain, range));
 780       StartNode* s = new StartOSRNode(root(), domain);
 781       initial_gvn()->set_type_bottom(s);
 782       init_start(s);
 783       cg = CallGenerator::for_osr(method(), entry_bci());
 784     } else {
 785       // Normal case.
 786       init_tf(TypeFunc::make(method()));
 787       StartNode* s = new StartNode(root(), tf()->domain());
 788       initial_gvn()->set_type_bottom(s);
 789       init_start(s);
 790       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get) {
 791         // With java.lang.ref.reference.get() we must go through the
 792         // intrinsic - even when get() is the root
 793         // method of the compile - so that, if necessary, the value in
 794         // the referent field of the reference object gets recorded by
 795         // the pre-barrier code.
 796         cg = find_intrinsic(method(), false);
 797       }
 798       if (cg == NULL) {
 799         float past_uses = method()->interpreter_invocation_count();
 800         float expected_uses = past_uses;
 801         cg = CallGenerator::for_inline(method(), expected_uses);
 802       }
 803     }
 804     if (failing())  return;
 805     if (cg == NULL) {
 806       record_method_not_compilable("cannot parse method");
 807       return;
 808     }
 809     JVMState* jvms = build_start_state(start(), tf());
 810     if ((jvms = cg->generate(jvms)) == NULL) {
 811       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 812         record_method_not_compilable("method parse failed");
 813       }
 814       return;
 815     }
 816     GraphKit kit(jvms);
 817 
 818     if (!kit.stopped()) {
 819       // Accept return values, and transfer control we know not where.
 820       // This is done by a special, unique ReturnNode bound to root.
 821       return_values(kit.jvms());
 822     }
 823 
 824     if (kit.has_exceptions()) {
 825       // Any exceptions that escape from this call must be rethrown
 826       // to whatever caller is dynamically above us on the stack.
 827       // This is done by a special, unique RethrowNode bound to root.
 828       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 829     }
 830 
 831     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 832 
 833     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 834       inline_string_calls(true);
 835     }
 836 
 837     if (failing())  return;
 838 
 839     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 840 
 841     // Remove clutter produced by parsing.
 842     if (!failing()) {
 843       ResourceMark rm;
 844       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 845     }
 846   }
 847 
 848   // Note:  Large methods are capped off in do_one_bytecode().
 849   if (failing())  return;
 850 
 851   // After parsing, node notes are no longer automagic.
 852   // They must be propagated by register_new_node_with_optimizer(),
 853   // clone(), or the like.
 854   set_default_node_notes(NULL);
 855 
 856   for (;;) {
 857     int successes = Inline_Warm();
 858     if (failing())  return;
 859     if (successes == 0)  break;
 860   }
 861 
 862   // Drain the list.
 863   Finish_Warm();
 864 #ifndef PRODUCT
 865   if (_printer && _printer->should_print(1)) {
 866     _printer->print_inlining();
 867   }
 868 #endif
 869 
 870   if (failing())  return;
 871   NOT_PRODUCT( verify_graph_edges(); )
 872 
 873   // Now optimize
 874   Optimize();
 875   if (failing())  return;
 876   NOT_PRODUCT( verify_graph_edges(); )
 877 
 878 #ifndef PRODUCT
 879   if (PrintIdeal) {
 880     ttyLocker ttyl;  // keep the following output all in one block
 881     // This output goes directly to the tty, not the compiler log.
 882     // To enable tools to match it up with the compilation activity,
 883     // be sure to tag this tty output with the compile ID.
 884     if (xtty != NULL) {
 885       xtty->head("ideal compile_id='%d'%s", compile_id(),
 886                  is_osr_compilation()    ? " compile_kind='osr'" :
 887                  "");
 888     }
 889     root()->dump(9999);
 890     if (xtty != NULL) {
 891       xtty->tail("ideal");
 892     }
 893   }
 894 #endif
 895 
 896 #ifdef ASSERT
 897   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
 898   bs->verify_gc_barriers(this, BarrierSetC2::BeforeCodeGen);
 899 #endif
 900 
 901   // Dump compilation data to replay it.
 902   if (directive->DumpReplayOption) {
 903     env()->dump_replay_data(_compile_id);
 904   }
 905   if (directive->DumpInlineOption && (ilt() != NULL)) {
 906     env()->dump_inline_data(_compile_id);
 907   }
 908 
 909   // Now that we know the size of all the monitors we can add a fixed slot
 910   // for the original deopt pc.
 911 
 912   _orig_pc_slot =  fixed_slots();
 913   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 914   set_fixed_slots(next_slot);
 915 
 916   // Compute when to use implicit null checks. Used by matching trap based
 917   // nodes and NullCheck optimization.
 918   set_allowed_deopt_reasons();
 919 
 920   // Now generate code
 921   Code_Gen();
 922   if (failing())  return;
 923 
 924   // Check if we want to skip execution of all compiled code.
 925   {
 926 #ifndef PRODUCT
 927     if (OptoNoExecute) {
 928       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 929       return;
 930     }
 931 #endif
 932     TracePhase tp("install_code", &timers[_t_registerMethod]);
 933 
 934     if (is_osr_compilation()) {
 935       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 936       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 937     } else {
 938       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 939       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 940     }
 941 
 942     env()->register_method(_method, _entry_bci,
 943                            &_code_offsets,
 944                            _orig_pc_slot_offset_in_bytes,
 945                            code_buffer(),
 946                            frame_size_in_words(), _oop_map_set,
 947                            &_handler_table, &_inc_table,
 948                            compiler,
 949                            has_unsafe_access(),
 950                            SharedRuntime::is_wide_vector(max_vector_size()),
 951                            rtm_state()
 952                            );
 953 
 954     if (log() != NULL) // Print code cache state into compiler log
 955       log()->code_cache_state();
 956   }
 957 }
 958 
 959 //------------------------------Compile----------------------------------------
 960 // Compile a runtime stub
 961 Compile::Compile( ciEnv* ci_env,
 962                   TypeFunc_generator generator,
 963                   address stub_function,
 964                   const char *stub_name,
 965                   int is_fancy_jump,
 966                   bool pass_tls,
 967                   bool save_arg_registers,
 968                   bool return_pc,
 969                   DirectiveSet* directive)
 970   : Phase(Compiler),
 971     _compile_id(0),
 972     _save_argument_registers(save_arg_registers),
 973     _subsume_loads(true),
 974     _do_escape_analysis(false),
 975     _eliminate_boxing(false),
 976     _method(NULL),
 977     _entry_bci(InvocationEntryBci),
 978     _stub_function(stub_function),
 979     _stub_name(stub_name),
 980     _stub_entry_point(NULL),
 981     _max_node_limit(MaxNodeLimit),
 982     _orig_pc_slot(0),
 983     _orig_pc_slot_offset_in_bytes(0),
 984     _inlining_progress(false),
 985     _inlining_incrementally(false),
 986     _has_reserved_stack_access(false),
 987 #ifndef PRODUCT
 988     _trace_opto_output(directive->TraceOptoOutputOption),
 989 #endif
 990     _has_method_handle_invokes(false),
 991     _clinit_barrier_on_entry(false),
 992     _comp_arena(mtCompiler),
 993     _env(ci_env),
 994     _directive(directive),
 995     _log(ci_env->log()),
 996     _failure_reason(NULL),
 997     _congraph(NULL),
 998 #ifndef PRODUCT
 999     _printer(NULL),
1000 #endif
1001     _dead_node_list(comp_arena()),
1002     _dead_node_count(0),
1003     _node_arena(mtCompiler),
1004     _old_arena(mtCompiler),
1005     _mach_constant_base_node(NULL),
1006     _Compile_types(mtCompiler),
1007     _initial_gvn(NULL),
1008     _for_igvn(NULL),
1009     _warm_calls(NULL),
1010     _number_of_mh_late_inlines(0),
1011     _print_inlining_stream(NULL),
1012     _print_inlining_list(NULL),
1013     _print_inlining_idx(0),
1014     _print_inlining_output(NULL),
1015     _replay_inline_data(NULL),
1016     _java_calls(0),
1017     _inner_loops(0),
1018     _interpreter_frame_size(0),
1019     _node_bundling_limit(0),
1020     _node_bundling_base(NULL),
1021     _code_buffer("Compile::Fill_buffer"),
1022 #ifndef PRODUCT
1023     _in_dump_cnt(0),
1024 #endif
1025     _allowed_reasons(0) {
1026   C = this;
1027 
1028   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1029   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1030 
1031 #ifndef PRODUCT
1032   set_print_assembly(PrintFrameConverterAssembly);
1033   set_parsed_irreducible_loop(false);
1034 #else
1035   set_print_assembly(false); // Must initialize.
1036 #endif
1037   set_has_irreducible_loop(false); // no loops
1038 
1039   CompileWrapper cw(this);
1040   Init(/*AliasLevel=*/ 0);
1041   init_tf((*generator)());
1042 
1043   {
1044     // The following is a dummy for the sake of GraphKit::gen_stub
1045     Unique_Node_List for_igvn(comp_arena());
1046     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1047     PhaseGVN gvn(Thread::current()->resource_area(),255);
1048     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1049     gvn.transform_no_reclaim(top());
1050 
1051     GraphKit kit;
1052     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1053   }
1054 
1055   NOT_PRODUCT( verify_graph_edges(); )
1056   Code_Gen();
1057   if (failing())  return;
1058 
1059 
1060   // Entry point will be accessed using compile->stub_entry_point();
1061   if (code_buffer() == NULL) {
1062     Matcher::soft_match_failure();
1063   } else {
1064     if (PrintAssembly && (WizardMode || Verbose))
1065       tty->print_cr("### Stub::%s", stub_name);
1066 
1067     if (!failing()) {
1068       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1069 
1070       // Make the NMethod
1071       // For now we mark the frame as never safe for profile stackwalking
1072       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1073                                                       code_buffer(),
1074                                                       CodeOffsets::frame_never_safe,
1075                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1076                                                       frame_size_in_words(),
1077                                                       _oop_map_set,
1078                                                       save_arg_registers);
1079       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1080 
1081       _stub_entry_point = rs->entry_point();
1082     }
1083   }
1084 }
1085 
1086 //------------------------------Init-------------------------------------------
1087 // Prepare for a single compilation
1088 void Compile::Init(int aliaslevel) {
1089   _unique  = 0;
1090   _regalloc = NULL;
1091 
1092   _tf      = NULL;  // filled in later
1093   _top     = NULL;  // cached later
1094   _matcher = NULL;  // filled in later
1095   _cfg     = NULL;  // filled in later
1096 
1097   set_24_bit_selection_and_mode(Use24BitFP, false);
1098 
1099   _node_note_array = NULL;
1100   _default_node_notes = NULL;
1101   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1102 
1103   _immutable_memory = NULL; // filled in at first inquiry
1104 
1105   // Globally visible Nodes
1106   // First set TOP to NULL to give safe behavior during creation of RootNode
1107   set_cached_top_node(NULL);
1108   set_root(new RootNode());
1109   // Now that you have a Root to point to, create the real TOP
1110   set_cached_top_node( new ConNode(Type::TOP) );
1111   set_recent_alloc(NULL, NULL);
1112 
1113   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1114   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1115   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1116   env()->set_dependencies(new Dependencies(env()));
1117 
1118   _fixed_slots = 0;
1119   set_has_split_ifs(false);
1120   set_has_loops(has_method() && method()->has_loops()); // first approximation
1121   set_has_stringbuilder(false);
1122   set_has_boxed_value(false);
1123   _trap_can_recompile = false;  // no traps emitted yet
1124   _major_progress = true; // start out assuming good things will happen
1125   set_has_unsafe_access(false);
1126   set_max_vector_size(0);
1127   set_clear_upper_avx(false);  //false as default for clear upper bits of ymm registers
1128   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1129   set_decompile_count(0);
1130 
1131   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1132   _loop_opts_cnt = LoopOptsCount;
1133   set_do_inlining(Inline);
1134   set_max_inline_size(MaxInlineSize);
1135   set_freq_inline_size(FreqInlineSize);
1136   set_do_scheduling(OptoScheduling);
1137   set_do_count_invocations(false);
1138   set_do_method_data_update(false);
1139 
1140   set_do_vector_loop(false);
1141 
1142   if (AllowVectorizeOnDemand) {
1143     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1144       set_do_vector_loop(true);
1145       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());})
1146     } else if (has_method() && method()->name() != 0 &&
1147                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1148       set_do_vector_loop(true);
1149     }
1150   }
1151   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1152   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());})
1153 
1154   set_age_code(has_method() && method()->profile_aging());
1155   set_rtm_state(NoRTM); // No RTM lock eliding by default
1156   _max_node_limit = _directive->MaxNodeLimitOption;
1157 
1158 #if INCLUDE_RTM_OPT
1159   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1160     int rtm_state = method()->method_data()->rtm_state();
1161     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1162       // Don't generate RTM lock eliding code.
1163       set_rtm_state(NoRTM);
1164     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1165       // Generate RTM lock eliding code without abort ratio calculation code.
1166       set_rtm_state(UseRTM);
1167     } else if (UseRTMDeopt) {
1168       // Generate RTM lock eliding code and include abort ratio calculation
1169       // code if UseRTMDeopt is on.
1170       set_rtm_state(ProfileRTM);
1171     }
1172   }
1173 #endif
1174   if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1175     set_clinit_barrier_on_entry(true);
1176   }
1177   if (debug_info()->recording_non_safepoints()) {
1178     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1179                         (comp_arena(), 8, 0, NULL));
1180     set_default_node_notes(Node_Notes::make(this));
1181   }
1182 
1183   // // -- Initialize types before each compile --
1184   // // Update cached type information
1185   // if( _method && _method->constants() )
1186   //   Type::update_loaded_types(_method, _method->constants());
1187 
1188   // Init alias_type map.
1189   if (!_do_escape_analysis && aliaslevel == 3)
1190     aliaslevel = 2;  // No unique types without escape analysis
1191   _AliasLevel = aliaslevel;
1192   const int grow_ats = 16;
1193   _max_alias_types = grow_ats;
1194   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1195   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1196   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1197   {
1198     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1199   }
1200   // Initialize the first few types.
1201   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1202   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1203   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1204   _num_alias_types = AliasIdxRaw+1;
1205   // Zero out the alias type cache.
1206   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1207   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1208   probe_alias_cache(NULL)->_index = AliasIdxTop;
1209 
1210   _intrinsics = NULL;
1211   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1212   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1213   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1214   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1215   _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1216   register_library_intrinsics();
1217 }
1218 
1219 //---------------------------init_start----------------------------------------
1220 // Install the StartNode on this compile object.
1221 void Compile::init_start(StartNode* s) {
1222   if (failing())
1223     return; // already failing
1224   assert(s == start(), "");
1225 }
1226 
1227 /**
1228  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1229  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1230  * the ideal graph.
1231  */
1232 StartNode* Compile::start() const {
1233   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1234   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1235     Node* start = root()->fast_out(i);
1236     if (start->is_Start()) {
1237       return start->as_Start();
1238     }
1239   }
1240   fatal("Did not find Start node!");
1241   return NULL;
1242 }
1243 
1244 //-------------------------------immutable_memory-------------------------------------
1245 // Access immutable memory
1246 Node* Compile::immutable_memory() {
1247   if (_immutable_memory != NULL) {
1248     return _immutable_memory;
1249   }
1250   StartNode* s = start();
1251   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1252     Node *p = s->fast_out(i);
1253     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1254       _immutable_memory = p;
1255       return _immutable_memory;
1256     }
1257   }
1258   ShouldNotReachHere();
1259   return NULL;
1260 }
1261 
1262 //----------------------set_cached_top_node------------------------------------
1263 // Install the cached top node, and make sure Node::is_top works correctly.
1264 void Compile::set_cached_top_node(Node* tn) {
1265   if (tn != NULL)  verify_top(tn);
1266   Node* old_top = _top;
1267   _top = tn;
1268   // Calling Node::setup_is_top allows the nodes the chance to adjust
1269   // their _out arrays.
1270   if (_top != NULL)     _top->setup_is_top();
1271   if (old_top != NULL)  old_top->setup_is_top();
1272   assert(_top == NULL || top()->is_top(), "");
1273 }
1274 
1275 #ifdef ASSERT
1276 uint Compile::count_live_nodes_by_graph_walk() {
1277   Unique_Node_List useful(comp_arena());
1278   // Get useful node list by walking the graph.
1279   identify_useful_nodes(useful);
1280   return useful.size();
1281 }
1282 
1283 void Compile::print_missing_nodes() {
1284 
1285   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1286   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1287     return;
1288   }
1289 
1290   // This is an expensive function. It is executed only when the user
1291   // specifies VerifyIdealNodeCount option or otherwise knows the
1292   // additional work that needs to be done to identify reachable nodes
1293   // by walking the flow graph and find the missing ones using
1294   // _dead_node_list.
1295 
1296   Unique_Node_List useful(comp_arena());
1297   // Get useful node list by walking the graph.
1298   identify_useful_nodes(useful);
1299 
1300   uint l_nodes = C->live_nodes();
1301   uint l_nodes_by_walk = useful.size();
1302 
1303   if (l_nodes != l_nodes_by_walk) {
1304     if (_log != NULL) {
1305       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1306       _log->stamp();
1307       _log->end_head();
1308     }
1309     VectorSet& useful_member_set = useful.member_set();
1310     int last_idx = l_nodes_by_walk;
1311     for (int i = 0; i < last_idx; i++) {
1312       if (useful_member_set.test(i)) {
1313         if (_dead_node_list.test(i)) {
1314           if (_log != NULL) {
1315             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1316           }
1317           if (PrintIdealNodeCount) {
1318             // Print the log message to tty
1319               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1320               useful.at(i)->dump();
1321           }
1322         }
1323       }
1324       else if (! _dead_node_list.test(i)) {
1325         if (_log != NULL) {
1326           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1327         }
1328         if (PrintIdealNodeCount) {
1329           // Print the log message to tty
1330           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1331         }
1332       }
1333     }
1334     if (_log != NULL) {
1335       _log->tail("mismatched_nodes");
1336     }
1337   }
1338 }
1339 void Compile::record_modified_node(Node* n) {
1340   if (_modified_nodes != NULL && !_inlining_incrementally &&
1341       n->outcnt() != 0 && !n->is_Con()) {
1342     _modified_nodes->push(n);
1343   }
1344 }
1345 
1346 void Compile::remove_modified_node(Node* n) {
1347   if (_modified_nodes != NULL) {
1348     _modified_nodes->remove(n);
1349   }
1350 }
1351 #endif
1352 
1353 #ifndef PRODUCT
1354 void Compile::verify_top(Node* tn) const {
1355   if (tn != NULL) {
1356     assert(tn->is_Con(), "top node must be a constant");
1357     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1358     assert(tn->in(0) != NULL, "must have live top node");
1359   }
1360 }
1361 #endif
1362 
1363 
1364 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1365 
1366 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1367   guarantee(arr != NULL, "");
1368   int num_blocks = arr->length();
1369   if (grow_by < num_blocks)  grow_by = num_blocks;
1370   int num_notes = grow_by * _node_notes_block_size;
1371   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1372   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1373   while (num_notes > 0) {
1374     arr->append(notes);
1375     notes     += _node_notes_block_size;
1376     num_notes -= _node_notes_block_size;
1377   }
1378   assert(num_notes == 0, "exact multiple, please");
1379 }
1380 
1381 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1382   if (source == NULL || dest == NULL)  return false;
1383 
1384   if (dest->is_Con())
1385     return false;               // Do not push debug info onto constants.
1386 
1387 #ifdef ASSERT
1388   // Leave a bread crumb trail pointing to the original node:
1389   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1390     dest->set_debug_orig(source);
1391   }
1392 #endif
1393 
1394   if (node_note_array() == NULL)
1395     return false;               // Not collecting any notes now.
1396 
1397   // This is a copy onto a pre-existing node, which may already have notes.
1398   // If both nodes have notes, do not overwrite any pre-existing notes.
1399   Node_Notes* source_notes = node_notes_at(source->_idx);
1400   if (source_notes == NULL || source_notes->is_clear())  return false;
1401   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1402   if (dest_notes == NULL || dest_notes->is_clear()) {
1403     return set_node_notes_at(dest->_idx, source_notes);
1404   }
1405 
1406   Node_Notes merged_notes = (*source_notes);
1407   // The order of operations here ensures that dest notes will win...
1408   merged_notes.update_from(dest_notes);
1409   return set_node_notes_at(dest->_idx, &merged_notes);
1410 }
1411 
1412 
1413 //--------------------------allow_range_check_smearing-------------------------
1414 // Gating condition for coalescing similar range checks.
1415 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1416 // single covering check that is at least as strong as any of them.
1417 // If the optimization succeeds, the simplified (strengthened) range check
1418 // will always succeed.  If it fails, we will deopt, and then give up
1419 // on the optimization.
1420 bool Compile::allow_range_check_smearing() const {
1421   // If this method has already thrown a range-check,
1422   // assume it was because we already tried range smearing
1423   // and it failed.
1424   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1425   return !already_trapped;
1426 }
1427 
1428 
1429 //------------------------------flatten_alias_type-----------------------------
1430 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1431   int offset = tj->offset();
1432   TypePtr::PTR ptr = tj->ptr();
1433 
1434   // Known instance (scalarizable allocation) alias only with itself.
1435   bool is_known_inst = tj->isa_oopptr() != NULL &&
1436                        tj->is_oopptr()->is_known_instance();
1437 
1438   // Process weird unsafe references.
1439   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1440     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1441     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1442     tj = TypeOopPtr::BOTTOM;
1443     ptr = tj->ptr();
1444     offset = tj->offset();
1445   }
1446 
1447   // Array pointers need some flattening
1448   const TypeAryPtr *ta = tj->isa_aryptr();
1449   if (ta && ta->is_stable()) {
1450     // Erase stability property for alias analysis.
1451     tj = ta = ta->cast_to_stable(false);
1452   }
1453   if( ta && is_known_inst ) {
1454     if ( offset != Type::OffsetBot &&
1455          offset > arrayOopDesc::length_offset_in_bytes() ) {
1456       offset = Type::OffsetBot; // Flatten constant access into array body only
1457       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1458     }
1459   } else if( ta && _AliasLevel >= 2 ) {
1460     // For arrays indexed by constant indices, we flatten the alias
1461     // space to include all of the array body.  Only the header, klass
1462     // and array length can be accessed un-aliased.
1463     if( offset != Type::OffsetBot ) {
1464       if( ta->const_oop() ) { // MethodData* or Method*
1465         offset = Type::OffsetBot;   // Flatten constant access into array body
1466         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1467       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1468         // range is OK as-is.
1469         tj = ta = TypeAryPtr::RANGE;
1470       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1471         tj = TypeInstPtr::KLASS; // all klass loads look alike
1472         ta = TypeAryPtr::RANGE; // generic ignored junk
1473         ptr = TypePtr::BotPTR;
1474       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1475         tj = TypeInstPtr::MARK;
1476         ta = TypeAryPtr::RANGE; // generic ignored junk
1477         ptr = TypePtr::BotPTR;
1478       } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1479         ta = tj->isa_aryptr();
1480       } else {                  // Random constant offset into array body
1481         offset = Type::OffsetBot;   // Flatten constant access into array body
1482         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1483       }
1484     }
1485     // Arrays of fixed size alias with arrays of unknown size.
1486     if (ta->size() != TypeInt::POS) {
1487       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1488       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1489     }
1490     // Arrays of known objects become arrays of unknown objects.
1491     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1492       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1493       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1494     }
1495     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1496       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1497       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1498     }
1499     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1500     // cannot be distinguished by bytecode alone.
1501     if (ta->elem() == TypeInt::BOOL) {
1502       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1503       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1504       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1505     }
1506     // During the 2nd round of IterGVN, NotNull castings are removed.
1507     // Make sure the Bottom and NotNull variants alias the same.
1508     // Also, make sure exact and non-exact variants alias the same.
1509     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1510       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1511     }
1512   }
1513 
1514   // Oop pointers need some flattening
1515   const TypeInstPtr *to = tj->isa_instptr();
1516   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1517     ciInstanceKlass *k = to->klass()->as_instance_klass();
1518     if( ptr == TypePtr::Constant ) {
1519       if (to->klass() != ciEnv::current()->Class_klass() ||
1520           offset < k->size_helper() * wordSize) {
1521         // No constant oop pointers (such as Strings); they alias with
1522         // unknown strings.
1523         assert(!is_known_inst, "not scalarizable allocation");
1524         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1525       }
1526     } else if( is_known_inst ) {
1527       tj = to; // Keep NotNull and klass_is_exact for instance type
1528     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1529       // During the 2nd round of IterGVN, NotNull castings are removed.
1530       // Make sure the Bottom and NotNull variants alias the same.
1531       // Also, make sure exact and non-exact variants alias the same.
1532       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1533     }
1534     if (to->speculative() != NULL) {
1535       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1536     }
1537     // Canonicalize the holder of this field
1538     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1539       // First handle header references such as a LoadKlassNode, even if the
1540       // object's klass is unloaded at compile time (4965979).
1541       if (!is_known_inst) { // Do it only for non-instance types
1542         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1543       }
1544     } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1545       to = tj->is_instptr();
1546     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1547       // Static fields are in the space above the normal instance
1548       // fields in the java.lang.Class instance.
1549       if (to->klass() != ciEnv::current()->Class_klass()) {
1550         to = NULL;
1551         tj = TypeOopPtr::BOTTOM;
1552         offset = tj->offset();
1553       }
1554     } else {
1555       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1556       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1557         if( is_known_inst ) {
1558           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1559         } else {
1560           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1561         }
1562       }
1563     }
1564   }
1565 
1566   // Klass pointers to object array klasses need some flattening
1567   const TypeKlassPtr *tk = tj->isa_klassptr();
1568   if( tk ) {
1569     // If we are referencing a field within a Klass, we need
1570     // to assume the worst case of an Object.  Both exact and
1571     // inexact types must flatten to the same alias class so
1572     // use NotNull as the PTR.
1573     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1574 
1575       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1576                                    TypeKlassPtr::OBJECT->klass(),
1577                                    offset);
1578     }
1579 
1580     ciKlass* klass = tk->klass();
1581     if( klass->is_obj_array_klass() ) {
1582       ciKlass* k = TypeAryPtr::OOPS->klass();
1583       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1584         k = TypeInstPtr::BOTTOM->klass();
1585       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1586     }
1587 
1588     // Check for precise loads from the primary supertype array and force them
1589     // to the supertype cache alias index.  Check for generic array loads from
1590     // the primary supertype array and also force them to the supertype cache
1591     // alias index.  Since the same load can reach both, we need to merge
1592     // these 2 disparate memories into the same alias class.  Since the
1593     // primary supertype array is read-only, there's no chance of confusion
1594     // where we bypass an array load and an array store.
1595     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1596     if (offset == Type::OffsetBot ||
1597         (offset >= primary_supers_offset &&
1598          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1599         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1600       offset = in_bytes(Klass::secondary_super_cache_offset());
1601       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1602     }
1603   }
1604 
1605   // Flatten all Raw pointers together.
1606   if (tj->base() == Type::RawPtr)
1607     tj = TypeRawPtr::BOTTOM;
1608 
1609   if (tj->base() == Type::AnyPtr)
1610     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1611 
1612   // Flatten all to bottom for now
1613   switch( _AliasLevel ) {
1614   case 0:
1615     tj = TypePtr::BOTTOM;
1616     break;
1617   case 1:                       // Flatten to: oop, static, field or array
1618     switch (tj->base()) {
1619     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1620     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1621     case Type::AryPtr:   // do not distinguish arrays at all
1622     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1623     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1624     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1625     default: ShouldNotReachHere();
1626     }
1627     break;
1628   case 2:                       // No collapsing at level 2; keep all splits
1629   case 3:                       // No collapsing at level 3; keep all splits
1630     break;
1631   default:
1632     Unimplemented();
1633   }
1634 
1635   offset = tj->offset();
1636   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1637 
1638   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1639           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1640           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1641           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1642           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1643           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1644           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1645           (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1646           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1647   assert( tj->ptr() != TypePtr::TopPTR &&
1648           tj->ptr() != TypePtr::AnyNull &&
1649           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1650 //    assert( tj->ptr() != TypePtr::Constant ||
1651 //            tj->base() == Type::RawPtr ||
1652 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1653 
1654   return tj;
1655 }
1656 
1657 void Compile::AliasType::Init(int i, const TypePtr* at) {
1658   _index = i;
1659   _adr_type = at;
1660   _field = NULL;
1661   _element = NULL;
1662   _is_rewritable = true; // default
1663   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1664   if (atoop != NULL && atoop->is_known_instance()) {
1665     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1666     _general_index = Compile::current()->get_alias_index(gt);
1667   } else {
1668     _general_index = 0;
1669   }
1670 }
1671 
1672 BasicType Compile::AliasType::basic_type() const {
1673   if (element() != NULL) {
1674     const Type* element = adr_type()->is_aryptr()->elem();
1675     return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1676   } if (field() != NULL) {
1677     return field()->layout_type();
1678   } else {
1679     return T_ILLEGAL; // unknown
1680   }
1681 }
1682 
1683 //---------------------------------print_on------------------------------------
1684 #ifndef PRODUCT
1685 void Compile::AliasType::print_on(outputStream* st) {
1686   if (index() < 10)
1687         st->print("@ <%d> ", index());
1688   else  st->print("@ <%d>",  index());
1689   st->print(is_rewritable() ? "   " : " RO");
1690   int offset = adr_type()->offset();
1691   if (offset == Type::OffsetBot)
1692         st->print(" +any");
1693   else  st->print(" +%-3d", offset);
1694   st->print(" in ");
1695   adr_type()->dump_on(st);
1696   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1697   if (field() != NULL && tjp) {
1698     if (tjp->klass()  != field()->holder() ||
1699         tjp->offset() != field()->offset_in_bytes()) {
1700       st->print(" != ");
1701       field()->print();
1702       st->print(" ***");
1703     }
1704   }
1705 }
1706 
1707 void print_alias_types() {
1708   Compile* C = Compile::current();
1709   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1710   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1711     C->alias_type(idx)->print_on(tty);
1712     tty->cr();
1713   }
1714 }
1715 #endif
1716 
1717 
1718 //----------------------------probe_alias_cache--------------------------------
1719 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1720   intptr_t key = (intptr_t) adr_type;
1721   key ^= key >> logAliasCacheSize;
1722   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1723 }
1724 
1725 
1726 //-----------------------------grow_alias_types--------------------------------
1727 void Compile::grow_alias_types() {
1728   const int old_ats  = _max_alias_types; // how many before?
1729   const int new_ats  = old_ats;          // how many more?
1730   const int grow_ats = old_ats+new_ats;  // how many now?
1731   _max_alias_types = grow_ats;
1732   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1733   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1734   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1735   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1736 }
1737 
1738 
1739 //--------------------------------find_alias_type------------------------------
1740 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1741   if (_AliasLevel == 0)
1742     return alias_type(AliasIdxBot);
1743 
1744   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1745   if (ace->_adr_type == adr_type) {
1746     return alias_type(ace->_index);
1747   }
1748 
1749   // Handle special cases.
1750   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1751   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1752 
1753   // Do it the slow way.
1754   const TypePtr* flat = flatten_alias_type(adr_type);
1755 
1756 #ifdef ASSERT
1757   {
1758     ResourceMark rm;
1759     assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1760            Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1761     assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1762            Type::str(adr_type));
1763     if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1764       const TypeOopPtr* foop = flat->is_oopptr();
1765       // Scalarizable allocations have exact klass always.
1766       bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1767       const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1768       assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1769              Type::str(foop), Type::str(xoop));
1770     }
1771   }
1772 #endif
1773 
1774   int idx = AliasIdxTop;
1775   for (int i = 0; i < num_alias_types(); i++) {
1776     if (alias_type(i)->adr_type() == flat) {
1777       idx = i;
1778       break;
1779     }
1780   }
1781 
1782   if (idx == AliasIdxTop) {
1783     if (no_create)  return NULL;
1784     // Grow the array if necessary.
1785     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1786     // Add a new alias type.
1787     idx = _num_alias_types++;
1788     _alias_types[idx]->Init(idx, flat);
1789     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1790     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1791     if (flat->isa_instptr()) {
1792       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1793           && flat->is_instptr()->klass() == env()->Class_klass())
1794         alias_type(idx)->set_rewritable(false);
1795     }
1796     if (flat->isa_aryptr()) {
1797 #ifdef ASSERT
1798       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1799       // (T_BYTE has the weakest alignment and size restrictions...)
1800       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1801 #endif
1802       if (flat->offset() == TypePtr::OffsetBot) {
1803         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1804       }
1805     }
1806     if (flat->isa_klassptr()) {
1807       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1808         alias_type(idx)->set_rewritable(false);
1809       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1810         alias_type(idx)->set_rewritable(false);
1811       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1812         alias_type(idx)->set_rewritable(false);
1813       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1814         alias_type(idx)->set_rewritable(false);
1815     }
1816     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1817     // but the base pointer type is not distinctive enough to identify
1818     // references into JavaThread.)
1819 
1820     // Check for final fields.
1821     const TypeInstPtr* tinst = flat->isa_instptr();
1822     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1823       ciField* field;
1824       if (tinst->const_oop() != NULL &&
1825           tinst->klass() == ciEnv::current()->Class_klass() &&
1826           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1827         // static field
1828         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1829         field = k->get_field_by_offset(tinst->offset(), true);
1830       } else {
1831         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1832         field = k->get_field_by_offset(tinst->offset(), false);
1833       }
1834       assert(field == NULL ||
1835              original_field == NULL ||
1836              (field->holder() == original_field->holder() &&
1837               field->offset() == original_field->offset() &&
1838               field->is_static() == original_field->is_static()), "wrong field?");
1839       // Set field() and is_rewritable() attributes.
1840       if (field != NULL)  alias_type(idx)->set_field(field);
1841     }
1842   }
1843 
1844   // Fill the cache for next time.
1845   ace->_adr_type = adr_type;
1846   ace->_index    = idx;
1847   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1848 
1849   // Might as well try to fill the cache for the flattened version, too.
1850   AliasCacheEntry* face = probe_alias_cache(flat);
1851   if (face->_adr_type == NULL) {
1852     face->_adr_type = flat;
1853     face->_index    = idx;
1854     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1855   }
1856 
1857   return alias_type(idx);
1858 }
1859 
1860 
1861 Compile::AliasType* Compile::alias_type(ciField* field) {
1862   const TypeOopPtr* t;
1863   if (field->is_static())
1864     t = TypeInstPtr::make(field->holder()->java_mirror());
1865   else
1866     t = TypeOopPtr::make_from_klass_raw(field->holder());
1867   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1868   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1869   return atp;
1870 }
1871 
1872 
1873 //------------------------------have_alias_type--------------------------------
1874 bool Compile::have_alias_type(const TypePtr* adr_type) {
1875   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1876   if (ace->_adr_type == adr_type) {
1877     return true;
1878   }
1879 
1880   // Handle special cases.
1881   if (adr_type == NULL)             return true;
1882   if (adr_type == TypePtr::BOTTOM)  return true;
1883 
1884   return find_alias_type(adr_type, true, NULL) != NULL;
1885 }
1886 
1887 //-----------------------------must_alias--------------------------------------
1888 // True if all values of the given address type are in the given alias category.
1889 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1890   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1891   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1892   if (alias_idx == AliasIdxTop)         return false; // the empty category
1893   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1894 
1895   // the only remaining possible overlap is identity
1896   int adr_idx = get_alias_index(adr_type);
1897   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1898   assert(adr_idx == alias_idx ||
1899          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1900           && adr_type                       != TypeOopPtr::BOTTOM),
1901          "should not be testing for overlap with an unsafe pointer");
1902   return adr_idx == alias_idx;
1903 }
1904 
1905 //------------------------------can_alias--------------------------------------
1906 // True if any values of the given address type are in the given alias category.
1907 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1908   if (alias_idx == AliasIdxTop)         return false; // the empty category
1909   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1910   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1911   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1912 
1913   // the only remaining possible overlap is identity
1914   int adr_idx = get_alias_index(adr_type);
1915   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1916   return adr_idx == alias_idx;
1917 }
1918 
1919 
1920 
1921 //---------------------------pop_warm_call-------------------------------------
1922 WarmCallInfo* Compile::pop_warm_call() {
1923   WarmCallInfo* wci = _warm_calls;
1924   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1925   return wci;
1926 }
1927 
1928 //----------------------------Inline_Warm--------------------------------------
1929 int Compile::Inline_Warm() {
1930   // If there is room, try to inline some more warm call sites.
1931   // %%% Do a graph index compaction pass when we think we're out of space?
1932   if (!InlineWarmCalls)  return 0;
1933 
1934   int calls_made_hot = 0;
1935   int room_to_grow   = NodeCountInliningCutoff - unique();
1936   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1937   int amount_grown   = 0;
1938   WarmCallInfo* call;
1939   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1940     int est_size = (int)call->size();
1941     if (est_size > (room_to_grow - amount_grown)) {
1942       // This one won't fit anyway.  Get rid of it.
1943       call->make_cold();
1944       continue;
1945     }
1946     call->make_hot();
1947     calls_made_hot++;
1948     amount_grown   += est_size;
1949     amount_to_grow -= est_size;
1950   }
1951 
1952   if (calls_made_hot > 0)  set_major_progress();
1953   return calls_made_hot;
1954 }
1955 
1956 
1957 //----------------------------Finish_Warm--------------------------------------
1958 void Compile::Finish_Warm() {
1959   if (!InlineWarmCalls)  return;
1960   if (failing())  return;
1961   if (warm_calls() == NULL)  return;
1962 
1963   // Clean up loose ends, if we are out of space for inlining.
1964   WarmCallInfo* call;
1965   while ((call = pop_warm_call()) != NULL) {
1966     call->make_cold();
1967   }
1968 }
1969 
1970 //---------------------cleanup_loop_predicates-----------------------
1971 // Remove the opaque nodes that protect the predicates so that all unused
1972 // checks and uncommon_traps will be eliminated from the ideal graph
1973 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1974   if (predicate_count()==0) return;
1975   for (int i = predicate_count(); i > 0; i--) {
1976     Node * n = predicate_opaque1_node(i-1);
1977     assert(n->Opcode() == Op_Opaque1, "must be");
1978     igvn.replace_node(n, n->in(1));
1979   }
1980   assert(predicate_count()==0, "should be clean!");
1981 }
1982 
1983 void Compile::add_range_check_cast(Node* n) {
1984   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1985   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1986   _range_check_casts->append(n);
1987 }
1988 
1989 // Remove all range check dependent CastIINodes.
1990 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1991   for (int i = range_check_cast_count(); i > 0; i--) {
1992     Node* cast = range_check_cast_node(i-1);
1993     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1994     igvn.replace_node(cast, cast->in(1));
1995   }
1996   assert(range_check_cast_count() == 0, "should be empty");
1997 }
1998 
1999 void Compile::add_opaque4_node(Node* n) {
2000   assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
2001   assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
2002   _opaque4_nodes->append(n);
2003 }
2004 
2005 // Remove all Opaque4 nodes.
2006 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2007   for (int i = opaque4_count(); i > 0; i--) {
2008     Node* opaq = opaque4_node(i-1);
2009     assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2010     igvn.replace_node(opaq, opaq->in(2));
2011   }
2012   assert(opaque4_count() == 0, "should be empty");
2013 }
2014 
2015 // StringOpts and late inlining of string methods
2016 void Compile::inline_string_calls(bool parse_time) {
2017   {
2018     // remove useless nodes to make the usage analysis simpler
2019     ResourceMark rm;
2020     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2021   }
2022 
2023   {
2024     ResourceMark rm;
2025     print_method(PHASE_BEFORE_STRINGOPTS, 3);
2026     PhaseStringOpts pso(initial_gvn(), for_igvn());
2027     print_method(PHASE_AFTER_STRINGOPTS, 3);
2028   }
2029 
2030   // now inline anything that we skipped the first time around
2031   if (!parse_time) {
2032     _late_inlines_pos = _late_inlines.length();
2033   }
2034 
2035   while (_string_late_inlines.length() > 0) {
2036     CallGenerator* cg = _string_late_inlines.pop();
2037     cg->do_late_inline();
2038     if (failing())  return;
2039   }
2040   _string_late_inlines.trunc_to(0);
2041 }
2042 
2043 // Late inlining of boxing methods
2044 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2045   if (_boxing_late_inlines.length() > 0) {
2046     assert(has_boxed_value(), "inconsistent");
2047 
2048     PhaseGVN* gvn = initial_gvn();
2049     set_inlining_incrementally(true);
2050 
2051     assert( igvn._worklist.size() == 0, "should be done with igvn" );
2052     for_igvn()->clear();
2053     gvn->replace_with(&igvn);
2054 
2055     _late_inlines_pos = _late_inlines.length();
2056 
2057     while (_boxing_late_inlines.length() > 0) {
2058       CallGenerator* cg = _boxing_late_inlines.pop();
2059       cg->do_late_inline();
2060       if (failing())  return;
2061     }
2062     _boxing_late_inlines.trunc_to(0);
2063 
2064     inline_incrementally_cleanup(igvn);
2065 
2066     set_inlining_incrementally(false);
2067   }
2068 }
2069 
2070 bool Compile::inline_incrementally_one() {
2071   assert(IncrementalInline, "incremental inlining should be on");
2072 
2073   TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2074   set_inlining_progress(false);
2075   set_do_cleanup(false);
2076   int i = 0;
2077   for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2078     CallGenerator* cg = _late_inlines.at(i);
2079     _late_inlines_pos = i+1;
2080     cg->do_late_inline();
2081     if (failing())  return false;
2082   }
2083   int j = 0;
2084   for (; i < _late_inlines.length(); i++, j++) {
2085     _late_inlines.at_put(j, _late_inlines.at(i));
2086   }
2087   _late_inlines.trunc_to(j);
2088   assert(inlining_progress() || _late_inlines.length() == 0, "");
2089 
2090   bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2091 
2092   set_inlining_progress(false);
2093   set_do_cleanup(false);
2094   return (_late_inlines.length() > 0) && !needs_cleanup;
2095 }
2096 
2097 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2098   {
2099     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2100     ResourceMark rm;
2101     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2102   }
2103   {
2104     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2105     igvn = PhaseIterGVN(initial_gvn());
2106     igvn.optimize();
2107   }
2108 }
2109 
2110 // Perform incremental inlining until bound on number of live nodes is reached
2111 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2112   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2113 
2114   set_inlining_incrementally(true);
2115   uint low_live_nodes = 0;
2116 
2117   while (_late_inlines.length() > 0) {
2118     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2119       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2120         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2121         // PhaseIdealLoop is expensive so we only try it once we are
2122         // out of live nodes and we only try it again if the previous
2123         // helped got the number of nodes down significantly
2124         PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2125         if (failing())  return;
2126         low_live_nodes = live_nodes();
2127         _major_progress = true;
2128       }
2129 
2130       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2131         break; // finish
2132       }
2133     }
2134 
2135     for_igvn()->clear();
2136     initial_gvn()->replace_with(&igvn);
2137 
2138     while (inline_incrementally_one()) {
2139       assert(!failing(), "inconsistent");
2140     }
2141 
2142     if (failing())  return;
2143 
2144     inline_incrementally_cleanup(igvn);
2145 
2146     if (failing())  return;
2147   }
2148   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2149 
2150   if (_string_late_inlines.length() > 0) {
2151     assert(has_stringbuilder(), "inconsistent");
2152     for_igvn()->clear();
2153     initial_gvn()->replace_with(&igvn);
2154 
2155     inline_string_calls(false);
2156 
2157     if (failing())  return;
2158 
2159     inline_incrementally_cleanup(igvn);
2160   }
2161 
2162   set_inlining_incrementally(false);
2163 }
2164 
2165 
2166 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2167   if(_loop_opts_cnt > 0) {
2168     debug_only( int cnt = 0; );
2169     while(major_progress() && (_loop_opts_cnt > 0)) {
2170       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2171       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2172       PhaseIdealLoop::optimize(igvn, mode);
2173       _loop_opts_cnt--;
2174       if (failing())  return false;
2175       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2176     }
2177   }
2178   return true;
2179 }
2180 
2181 // Remove edges from "root" to each SafePoint at a backward branch.
2182 // They were inserted during parsing (see add_safepoint()) to make
2183 // infinite loops without calls or exceptions visible to root, i.e.,
2184 // useful.
2185 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2186   Node *r = root();
2187   if (r != NULL) {
2188     for (uint i = r->req(); i < r->len(); ++i) {
2189       Node *n = r->in(i);
2190       if (n != NULL && n->is_SafePoint()) {
2191         r->rm_prec(i);
2192         if (n->outcnt() == 0) {
2193           igvn.remove_dead_node(n);
2194         }
2195         --i;
2196       }
2197     }
2198   }
2199 }
2200 
2201 //------------------------------Optimize---------------------------------------
2202 // Given a graph, optimize it.
2203 void Compile::Optimize() {
2204   TracePhase tp("optimizer", &timers[_t_optimizer]);
2205 
2206 #ifndef PRODUCT
2207   if (_directive->BreakAtCompileOption) {
2208     BREAKPOINT;
2209   }
2210 
2211 #endif
2212 
2213   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2214 #ifdef ASSERT
2215   bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2216 #endif
2217 
2218   ResourceMark rm;
2219 
2220   print_inlining_reinit();
2221 
2222   NOT_PRODUCT( verify_graph_edges(); )
2223 
2224   print_method(PHASE_AFTER_PARSING);
2225 
2226  {
2227   // Iterative Global Value Numbering, including ideal transforms
2228   // Initialize IterGVN with types and values from parse-time GVN
2229   PhaseIterGVN igvn(initial_gvn());
2230 #ifdef ASSERT
2231   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2232 #endif
2233   {
2234     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2235     igvn.optimize();
2236   }
2237 
2238   if (failing())  return;
2239 
2240   print_method(PHASE_ITER_GVN1, 2);
2241 
2242   inline_incrementally(igvn);
2243 
2244   print_method(PHASE_INCREMENTAL_INLINE, 2);
2245 
2246   if (failing())  return;
2247 
2248   if (eliminate_boxing()) {
2249     // Inline valueOf() methods now.
2250     inline_boxing_calls(igvn);
2251 
2252     if (AlwaysIncrementalInline) {
2253       inline_incrementally(igvn);
2254     }
2255 
2256     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2257 
2258     if (failing())  return;
2259   }
2260 
2261   // Now that all inlining is over, cut edge from root to loop
2262   // safepoints
2263   remove_root_to_sfpts_edges(igvn);
2264 
2265   // Remove the speculative part of types and clean up the graph from
2266   // the extra CastPP nodes whose only purpose is to carry them. Do
2267   // that early so that optimizations are not disrupted by the extra
2268   // CastPP nodes.
2269   remove_speculative_types(igvn);
2270 
2271   // No more new expensive nodes will be added to the list from here
2272   // so keep only the actual candidates for optimizations.
2273   cleanup_expensive_nodes(igvn);
2274 
2275   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2276     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2277     initial_gvn()->replace_with(&igvn);
2278     for_igvn()->clear();
2279     Unique_Node_List new_worklist(C->comp_arena());
2280     {
2281       ResourceMark rm;
2282       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2283     }
2284     set_for_igvn(&new_worklist);
2285     igvn = PhaseIterGVN(initial_gvn());
2286     igvn.optimize();
2287   }
2288 
2289   // Perform escape analysis
2290   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2291     if (has_loops()) {
2292       // Cleanup graph (remove dead nodes).
2293       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2294       PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2295       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2296       if (failing())  return;
2297     }
2298     ConnectionGraph::do_analysis(this, &igvn);
2299 
2300     if (failing())  return;
2301 
2302     // Optimize out fields loads from scalar replaceable allocations.
2303     igvn.optimize();
2304     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2305 
2306     if (failing())  return;
2307 
2308     if (congraph() != NULL && macro_count() > 0) {
2309       TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2310       PhaseMacroExpand mexp(igvn);
2311       mexp.eliminate_macro_nodes();
2312       igvn.set_delay_transform(false);
2313 
2314       igvn.optimize();
2315       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2316 
2317       if (failing())  return;
2318     }
2319   }
2320 
2321   // Loop transforms on the ideal graph.  Range Check Elimination,
2322   // peeling, unrolling, etc.
2323 
2324   // Set loop opts counter
2325   if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2326     {
2327       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2328       PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2329       _loop_opts_cnt--;
2330       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2331       if (failing())  return;
2332     }
2333     // Loop opts pass if partial peeling occurred in previous pass
2334     if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2335       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2336       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2337       _loop_opts_cnt--;
2338       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2339       if (failing())  return;
2340     }
2341     // Loop opts pass for loop-unrolling before CCP
2342     if(major_progress() && (_loop_opts_cnt > 0)) {
2343       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2344       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2345       _loop_opts_cnt--;
2346       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2347     }
2348     if (!failing()) {
2349       // Verify that last round of loop opts produced a valid graph
2350       TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2351       PhaseIdealLoop::verify(igvn);
2352     }
2353   }
2354   if (failing())  return;
2355 
2356   // Conditional Constant Propagation;
2357   PhaseCCP ccp( &igvn );
2358   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2359   {
2360     TracePhase tp("ccp", &timers[_t_ccp]);
2361     ccp.do_transform();
2362   }
2363   print_method(PHASE_CPP1, 2);
2364 
2365   assert( true, "Break here to ccp.dump_old2new_map()");
2366 
2367   // Iterative Global Value Numbering, including ideal transforms
2368   {
2369     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2370     igvn = ccp;
2371     igvn.optimize();
2372   }
2373   print_method(PHASE_ITER_GVN2, 2);
2374 
2375   if (failing())  return;
2376 
2377   // Loop transforms on the ideal graph.  Range Check Elimination,
2378   // peeling, unrolling, etc.
2379   if (!optimize_loops(igvn, LoopOptsDefault)) {
2380     return;
2381   }
2382 
2383   if (failing())  return;
2384 
2385   // Ensure that major progress is now clear
2386   C->clear_major_progress();
2387 
2388   {
2389     // Verify that all previous optimizations produced a valid graph
2390     // at least to this point, even if no loop optimizations were done.
2391     TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2392     PhaseIdealLoop::verify(igvn);
2393   }
2394 
2395   if (range_check_cast_count() > 0) {
2396     // No more loop optimizations. Remove all range check dependent CastIINodes.
2397     C->remove_range_check_casts(igvn);
2398     igvn.optimize();
2399   }
2400 
2401 #ifdef ASSERT
2402   bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2403 #endif
2404 
2405   {
2406     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2407     PhaseMacroExpand  mex(igvn);
2408     if (mex.expand_macro_nodes()) {
2409       assert(failing(), "must bail out w/ explicit message");
2410       return;
2411     }
2412     print_method(PHASE_MACRO_EXPANSION, 2);
2413   }
2414 
2415 #ifdef ASSERT
2416   bs->verify_gc_barriers(this, BarrierSetC2::BeforeLateInsertion);
2417 #endif
2418 
2419   bs->barrier_insertion_phase(C, igvn);
2420   if (failing())  return;
2421 
2422   {
2423     TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2424     if (bs->expand_barriers(this, igvn)) {
2425       assert(failing(), "must bail out w/ explicit message");
2426       return;
2427     }
2428     print_method(PHASE_BARRIER_EXPANSION, 2);
2429   }
2430 
2431   if (opaque4_count() > 0) {
2432     C->remove_opaque4_nodes(igvn);
2433     igvn.optimize();
2434   }
2435 
2436   DEBUG_ONLY( _modified_nodes = NULL; )
2437  } // (End scope of igvn; run destructor if necessary for asserts.)
2438 
2439  process_print_inlining();
2440  // A method with only infinite loops has no edges entering loops from root
2441  {
2442    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2443    if (final_graph_reshaping()) {
2444      assert(failing(), "must bail out w/ explicit message");
2445      return;
2446    }
2447  }
2448 
2449  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2450 }
2451 
2452 
2453 //------------------------------Code_Gen---------------------------------------
2454 // Given a graph, generate code for it
2455 void Compile::Code_Gen() {
2456   if (failing()) {
2457     return;
2458   }
2459 
2460   // Perform instruction selection.  You might think we could reclaim Matcher
2461   // memory PDQ, but actually the Matcher is used in generating spill code.
2462   // Internals of the Matcher (including some VectorSets) must remain live
2463   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2464   // set a bit in reclaimed memory.
2465 
2466   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2467   // nodes.  Mapping is only valid at the root of each matched subtree.
2468   NOT_PRODUCT( verify_graph_edges(); )
2469 
2470   Matcher matcher;
2471   _matcher = &matcher;
2472   {
2473     TracePhase tp("matcher", &timers[_t_matcher]);
2474     matcher.match();
2475   }
2476   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2477   // nodes.  Mapping is only valid at the root of each matched subtree.
2478   NOT_PRODUCT( verify_graph_edges(); )
2479 
2480   // If you have too many nodes, or if matching has failed, bail out
2481   check_node_count(0, "out of nodes matching instructions");
2482   if (failing()) {
2483     return;
2484   }
2485 
2486   print_method(PHASE_MATCHING, 2);
2487 
2488   // Build a proper-looking CFG
2489   PhaseCFG cfg(node_arena(), root(), matcher);
2490   _cfg = &cfg;
2491   {
2492     TracePhase tp("scheduler", &timers[_t_scheduler]);
2493     bool success = cfg.do_global_code_motion();
2494     if (!success) {
2495       return;
2496     }
2497 
2498     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2499     NOT_PRODUCT( verify_graph_edges(); )
2500     debug_only( cfg.verify(); )
2501   }
2502 
2503   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2504   _regalloc = &regalloc;
2505   {
2506     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2507     // Perform register allocation.  After Chaitin, use-def chains are
2508     // no longer accurate (at spill code) and so must be ignored.
2509     // Node->LRG->reg mappings are still accurate.
2510     _regalloc->Register_Allocate();
2511 
2512     // Bail out if the allocator builds too many nodes
2513     if (failing()) {
2514       return;
2515     }
2516   }
2517 
2518   // Prior to register allocation we kept empty basic blocks in case the
2519   // the allocator needed a place to spill.  After register allocation we
2520   // are not adding any new instructions.  If any basic block is empty, we
2521   // can now safely remove it.
2522   {
2523     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2524     cfg.remove_empty_blocks();
2525     if (do_freq_based_layout()) {
2526       PhaseBlockLayout layout(cfg);
2527     } else {
2528       cfg.set_loop_alignment();
2529     }
2530     cfg.fixup_flow();
2531   }
2532 
2533   // Apply peephole optimizations
2534   if( OptoPeephole ) {
2535     TracePhase tp("peephole", &timers[_t_peephole]);
2536     PhasePeephole peep( _regalloc, cfg);
2537     peep.do_transform();
2538   }
2539 
2540   // Do late expand if CPU requires this.
2541   if (Matcher::require_postalloc_expand) {
2542     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2543     cfg.postalloc_expand(_regalloc);
2544   }
2545 
2546   // Convert Nodes to instruction bits in a buffer
2547   {
2548     TraceTime tp("output", &timers[_t_output], CITime);
2549     Output();
2550   }
2551 
2552   print_method(PHASE_FINAL_CODE);
2553 
2554   // He's dead, Jim.
2555   _cfg     = (PhaseCFG*)((intptr_t)0xdeadbeef);
2556   _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2557 }
2558 
2559 
2560 //------------------------------dump_asm---------------------------------------
2561 // Dump formatted assembly
2562 #if defined(SUPPORT_OPTO_ASSEMBLY)
2563 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
2564 
2565   int pc_digits = 3; // #chars required for pc
2566   int sb_chars  = 3; // #chars for "start bundle" indicator
2567   int tab_size  = 8;
2568   if (pcs != NULL) {
2569     int max_pc = 0;
2570     for (uint i = 0; i < pc_limit; i++) {
2571       max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
2572     }
2573     pc_digits  = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
2574   }
2575   int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
2576 
2577   bool cut_short = false;
2578   st->print_cr("#");
2579   st->print("#  ");  _tf->dump_on(st);  st->cr();
2580   st->print_cr("#");
2581 
2582   // For all blocks
2583   int pc = 0x0;                 // Program counter
2584   char starts_bundle = ' ';
2585   _regalloc->dump_frame();
2586 
2587   Node *n = NULL;
2588   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2589     if (VMThread::should_terminate()) {
2590       cut_short = true;
2591       break;
2592     }
2593     Block* block = _cfg->get_block(i);
2594     if (block->is_connector() && !Verbose) {
2595       continue;
2596     }
2597     n = block->head();
2598     if ((pcs != NULL) && (n->_idx < pc_limit)) {
2599       pc = pcs[n->_idx];
2600       st->print("%*.*x", pc_digits, pc_digits, pc);
2601     }
2602     st->fill_to(prefix_len);
2603     block->dump_head(_cfg, st);
2604     if (block->is_connector()) {
2605       st->fill_to(prefix_len);
2606       st->print_cr("# Empty connector block");
2607     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2608       st->fill_to(prefix_len);
2609       st->print_cr("# Block is sole successor of call");
2610     }
2611 
2612     // For all instructions
2613     Node *delay = NULL;
2614     for (uint j = 0; j < block->number_of_nodes(); j++) {
2615       if (VMThread::should_terminate()) {
2616         cut_short = true;
2617         break;
2618       }
2619       n = block->get_node(j);
2620       if (valid_bundle_info(n)) {
2621         Bundle* bundle = node_bundling(n);
2622         if (bundle->used_in_unconditional_delay()) {
2623           delay = n;
2624           continue;
2625         }
2626         if (bundle->starts_bundle()) {
2627           starts_bundle = '+';
2628         }
2629       }
2630 
2631       if (WizardMode) {
2632         n->dump();
2633       }
2634 
2635       if( !n->is_Region() &&    // Dont print in the Assembly
2636           !n->is_Phi() &&       // a few noisely useless nodes
2637           !n->is_Proj() &&
2638           !n->is_MachTemp() &&
2639           !n->is_SafePointScalarObject() &&
2640           !n->is_Catch() &&     // Would be nice to print exception table targets
2641           !n->is_MergeMem() &&  // Not very interesting
2642           !n->is_top() &&       // Debug info table constants
2643           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2644           ) {
2645         if ((pcs != NULL) && (n->_idx < pc_limit)) {
2646           pc = pcs[n->_idx];
2647           st->print("%*.*x", pc_digits, pc_digits, pc);
2648         } else {
2649           st->fill_to(pc_digits);
2650         }
2651         st->print(" %c ", starts_bundle);
2652         starts_bundle = ' ';
2653         st->fill_to(prefix_len);
2654         n->format(_regalloc, st);
2655         st->cr();
2656       }
2657 
2658       // If we have an instruction with a delay slot, and have seen a delay,
2659       // then back up and print it
2660       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2661         // Coverity finding - Explicit null dereferenced.
2662         guarantee(delay != NULL, "no unconditional delay instruction");
2663         if (WizardMode) delay->dump();
2664 
2665         if (node_bundling(delay)->starts_bundle())
2666           starts_bundle = '+';
2667         if ((pcs != NULL) && (n->_idx < pc_limit)) {
2668           pc = pcs[n->_idx];
2669           st->print("%*.*x", pc_digits, pc_digits, pc);
2670         } else {
2671           st->fill_to(pc_digits);
2672         }
2673         st->print(" %c ", starts_bundle);
2674         starts_bundle = ' ';
2675         st->fill_to(prefix_len);
2676         delay->format(_regalloc, st);
2677         st->cr();
2678         delay = NULL;
2679       }
2680 
2681       // Dump the exception table as well
2682       if( n->is_Catch() && (Verbose || WizardMode) ) {
2683         // Print the exception table for this offset
2684         _handler_table.print_subtable_for(pc);
2685       }
2686       st->bol(); // Make sure we start on a new line
2687     }
2688     st->cr(); // one empty line between blocks
2689     assert(cut_short || delay == NULL, "no unconditional delay branch");
2690   } // End of per-block dump
2691 
2692   if (cut_short)  st->print_cr("*** disassembly is cut short ***");
2693 }
2694 #endif
2695 
2696 //------------------------------Final_Reshape_Counts---------------------------
2697 // This class defines counters to help identify when a method
2698 // may/must be executed using hardware with only 24-bit precision.
2699 struct Final_Reshape_Counts : public StackObj {
2700   int  _call_count;             // count non-inlined 'common' calls
2701   int  _float_count;            // count float ops requiring 24-bit precision
2702   int  _double_count;           // count double ops requiring more precision
2703   int  _java_call_count;        // count non-inlined 'java' calls
2704   int  _inner_loop_count;       // count loops which need alignment
2705   VectorSet _visited;           // Visitation flags
2706   Node_List _tests;             // Set of IfNodes & PCTableNodes
2707 
2708   Final_Reshape_Counts() :
2709     _call_count(0), _float_count(0), _double_count(0),
2710     _java_call_count(0), _inner_loop_count(0),
2711     _visited( Thread::current()->resource_area() ) { }
2712 
2713   void inc_call_count  () { _call_count  ++; }
2714   void inc_float_count () { _float_count ++; }
2715   void inc_double_count() { _double_count++; }
2716   void inc_java_call_count() { _java_call_count++; }
2717   void inc_inner_loop_count() { _inner_loop_count++; }
2718 
2719   int  get_call_count  () const { return _call_count  ; }
2720   int  get_float_count () const { return _float_count ; }
2721   int  get_double_count() const { return _double_count; }
2722   int  get_java_call_count() const { return _java_call_count; }
2723   int  get_inner_loop_count() const { return _inner_loop_count; }
2724 };
2725 
2726 #ifdef ASSERT
2727 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2728   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2729   // Make sure the offset goes inside the instance layout.
2730   return k->contains_field_offset(tp->offset());
2731   // Note that OffsetBot and OffsetTop are very negative.
2732 }
2733 #endif
2734 
2735 // Eliminate trivially redundant StoreCMs and accumulate their
2736 // precedence edges.
2737 void Compile::eliminate_redundant_card_marks(Node* n) {
2738   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2739   if (n->in(MemNode::Address)->outcnt() > 1) {
2740     // There are multiple users of the same address so it might be
2741     // possible to eliminate some of the StoreCMs
2742     Node* mem = n->in(MemNode::Memory);
2743     Node* adr = n->in(MemNode::Address);
2744     Node* val = n->in(MemNode::ValueIn);
2745     Node* prev = n;
2746     bool done = false;
2747     // Walk the chain of StoreCMs eliminating ones that match.  As
2748     // long as it's a chain of single users then the optimization is
2749     // safe.  Eliminating partially redundant StoreCMs would require
2750     // cloning copies down the other paths.
2751     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2752       if (adr == mem->in(MemNode::Address) &&
2753           val == mem->in(MemNode::ValueIn)) {
2754         // redundant StoreCM
2755         if (mem->req() > MemNode::OopStore) {
2756           // Hasn't been processed by this code yet.
2757           n->add_prec(mem->in(MemNode::OopStore));
2758         } else {
2759           // Already converted to precedence edge
2760           for (uint i = mem->req(); i < mem->len(); i++) {
2761             // Accumulate any precedence edges
2762             if (mem->in(i) != NULL) {
2763               n->add_prec(mem->in(i));
2764             }
2765           }
2766           // Everything above this point has been processed.
2767           done = true;
2768         }
2769         // Eliminate the previous StoreCM
2770         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2771         assert(mem->outcnt() == 0, "should be dead");
2772         mem->disconnect_inputs(NULL, this);
2773       } else {
2774         prev = mem;
2775       }
2776       mem = prev->in(MemNode::Memory);
2777     }
2778   }
2779 }
2780 
2781 //------------------------------final_graph_reshaping_impl----------------------
2782 // Implement items 1-5 from final_graph_reshaping below.
2783 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2784 
2785   if ( n->outcnt() == 0 ) return; // dead node
2786   uint nop = n->Opcode();
2787 
2788   // Check for 2-input instruction with "last use" on right input.
2789   // Swap to left input.  Implements item (2).
2790   if( n->req() == 3 &&          // two-input instruction
2791       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2792       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2793       n->in(2)->outcnt() == 1 &&// right use IS a last use
2794       !n->in(2)->is_Con() ) {   // right use is not a constant
2795     // Check for commutative opcode
2796     switch( nop ) {
2797     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2798     case Op_MaxI:  case Op_MinI:
2799     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2800     case Op_AndL:  case Op_XorL:  case Op_OrL:
2801     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2802       // Move "last use" input to left by swapping inputs
2803       n->swap_edges(1, 2);
2804       break;
2805     }
2806     default:
2807       break;
2808     }
2809   }
2810 
2811 #ifdef ASSERT
2812   if( n->is_Mem() ) {
2813     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2814     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2815             // oop will be recorded in oop map if load crosses safepoint
2816             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2817                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2818             "raw memory operations should have control edge");
2819   }
2820   if (n->is_MemBar()) {
2821     MemBarNode* mb = n->as_MemBar();
2822     if (mb->trailing_store() || mb->trailing_load_store()) {
2823       assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2824       Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2825       assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2826              (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2827     } else if (mb->leading()) {
2828       assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2829     }
2830   }
2831 #endif
2832   // Count FPU ops and common calls, implements item (3)
2833   bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2834   if (!gc_handled) {
2835     final_graph_reshaping_main_switch(n, frc, nop);
2836   }
2837 
2838   // Collect CFG split points
2839   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2840     frc._tests.push(n);
2841   }
2842 }
2843 
2844 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2845   switch( nop ) {
2846   // Count all float operations that may use FPU
2847   case Op_AddF:
2848   case Op_SubF:
2849   case Op_MulF:
2850   case Op_DivF:
2851   case Op_NegF:
2852   case Op_ModF:
2853   case Op_ConvI2F:
2854   case Op_ConF:
2855   case Op_CmpF:
2856   case Op_CmpF3:
2857   // case Op_ConvL2F: // longs are split into 32-bit halves
2858     frc.inc_float_count();
2859     break;
2860 
2861   case Op_ConvF2D:
2862   case Op_ConvD2F:
2863     frc.inc_float_count();
2864     frc.inc_double_count();
2865     break;
2866 
2867   // Count all double operations that may use FPU
2868   case Op_AddD:
2869   case Op_SubD:
2870   case Op_MulD:
2871   case Op_DivD:
2872   case Op_NegD:
2873   case Op_ModD:
2874   case Op_ConvI2D:
2875   case Op_ConvD2I:
2876   // case Op_ConvL2D: // handled by leaf call
2877   // case Op_ConvD2L: // handled by leaf call
2878   case Op_ConD:
2879   case Op_CmpD:
2880   case Op_CmpD3:
2881     frc.inc_double_count();
2882     break;
2883   case Op_Opaque1:              // Remove Opaque Nodes before matching
2884   case Op_Opaque2:              // Remove Opaque Nodes before matching
2885   case Op_Opaque3:
2886     n->subsume_by(n->in(1), this);
2887     break;
2888   case Op_CallStaticJava:
2889   case Op_CallJava:
2890   case Op_CallDynamicJava:
2891     frc.inc_java_call_count(); // Count java call site;
2892   case Op_CallRuntime:
2893   case Op_CallLeaf:
2894   case Op_CallLeafNoFP: {
2895     assert (n->is_Call(), "");
2896     CallNode *call = n->as_Call();
2897     // Count call sites where the FP mode bit would have to be flipped.
2898     // Do not count uncommon runtime calls:
2899     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2900     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2901     if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2902       frc.inc_call_count();   // Count the call site
2903     } else {                  // See if uncommon argument is shared
2904       Node *n = call->in(TypeFunc::Parms);
2905       int nop = n->Opcode();
2906       // Clone shared simple arguments to uncommon calls, item (1).
2907       if (n->outcnt() > 1 &&
2908           !n->is_Proj() &&
2909           nop != Op_CreateEx &&
2910           nop != Op_CheckCastPP &&
2911           nop != Op_DecodeN &&
2912           nop != Op_DecodeNKlass &&
2913           !n->is_Mem() &&
2914           !n->is_Phi()) {
2915         Node *x = n->clone();
2916         call->set_req(TypeFunc::Parms, x);
2917       }
2918     }
2919     break;
2920   }
2921 
2922   case Op_StoreD:
2923   case Op_LoadD:
2924   case Op_LoadD_unaligned:
2925     frc.inc_double_count();
2926     goto handle_mem;
2927   case Op_StoreF:
2928   case Op_LoadF:
2929     frc.inc_float_count();
2930     goto handle_mem;
2931 
2932   case Op_StoreCM:
2933     {
2934       // Convert OopStore dependence into precedence edge
2935       Node* prec = n->in(MemNode::OopStore);
2936       n->del_req(MemNode::OopStore);
2937       n->add_prec(prec);
2938       eliminate_redundant_card_marks(n);
2939     }
2940 
2941     // fall through
2942 
2943   case Op_StoreB:
2944   case Op_StoreC:
2945   case Op_StorePConditional:
2946   case Op_StoreI:
2947   case Op_StoreL:
2948   case Op_StoreIConditional:
2949   case Op_StoreLConditional:
2950   case Op_CompareAndSwapB:
2951   case Op_CompareAndSwapS:
2952   case Op_CompareAndSwapI:
2953   case Op_CompareAndSwapL:
2954   case Op_CompareAndSwapP:
2955   case Op_CompareAndSwapN:
2956   case Op_WeakCompareAndSwapB:
2957   case Op_WeakCompareAndSwapS:
2958   case Op_WeakCompareAndSwapI:
2959   case Op_WeakCompareAndSwapL:
2960   case Op_WeakCompareAndSwapP:
2961   case Op_WeakCompareAndSwapN:
2962   case Op_CompareAndExchangeB:
2963   case Op_CompareAndExchangeS:
2964   case Op_CompareAndExchangeI:
2965   case Op_CompareAndExchangeL:
2966   case Op_CompareAndExchangeP:
2967   case Op_CompareAndExchangeN:
2968   case Op_GetAndAddS:
2969   case Op_GetAndAddB:
2970   case Op_GetAndAddI:
2971   case Op_GetAndAddL:
2972   case Op_GetAndSetS:
2973   case Op_GetAndSetB:
2974   case Op_GetAndSetI:
2975   case Op_GetAndSetL:
2976   case Op_GetAndSetP:
2977   case Op_GetAndSetN:
2978   case Op_StoreP:
2979   case Op_StoreN:
2980   case Op_StoreNKlass:
2981   case Op_LoadB:
2982   case Op_LoadUB:
2983   case Op_LoadUS:
2984   case Op_LoadI:
2985   case Op_LoadKlass:
2986   case Op_LoadNKlass:
2987   case Op_LoadL:
2988   case Op_LoadL_unaligned:
2989   case Op_LoadPLocked:
2990   case Op_LoadP:
2991   case Op_LoadN:
2992   case Op_LoadRange:
2993   case Op_LoadS: {
2994   handle_mem:
2995 #ifdef ASSERT
2996     if( VerifyOptoOopOffsets ) {
2997       MemNode* mem  = n->as_Mem();
2998       // Check to see if address types have grounded out somehow.
2999       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
3000       assert( !tp || oop_offset_is_sane(tp), "" );
3001     }
3002 #endif
3003     break;
3004   }
3005 
3006   case Op_AddP: {               // Assert sane base pointers
3007     Node *addp = n->in(AddPNode::Address);
3008     assert( !addp->is_AddP() ||
3009             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3010             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3011             "Base pointers must match (addp %u)", addp->_idx );
3012 #ifdef _LP64
3013     if ((UseCompressedOops || UseCompressedClassPointers) &&
3014         addp->Opcode() == Op_ConP &&
3015         addp == n->in(AddPNode::Base) &&
3016         n->in(AddPNode::Offset)->is_Con()) {
3017       // If the transformation of ConP to ConN+DecodeN is beneficial depends
3018       // on the platform and on the compressed oops mode.
3019       // Use addressing with narrow klass to load with offset on x86.
3020       // Some platforms can use the constant pool to load ConP.
3021       // Do this transformation here since IGVN will convert ConN back to ConP.
3022       const Type* t = addp->bottom_type();
3023       bool is_oop   = t->isa_oopptr() != NULL;
3024       bool is_klass = t->isa_klassptr() != NULL;
3025 
3026       if ((is_oop   && Matcher::const_oop_prefer_decode()  ) ||
3027           (is_klass && Matcher::const_klass_prefer_decode())) {
3028         Node* nn = NULL;
3029 
3030         int op = is_oop ? Op_ConN : Op_ConNKlass;
3031 
3032         // Look for existing ConN node of the same exact type.
3033         Node* r  = root();
3034         uint cnt = r->outcnt();
3035         for (uint i = 0; i < cnt; i++) {
3036           Node* m = r->raw_out(i);
3037           if (m!= NULL && m->Opcode() == op &&
3038               m->bottom_type()->make_ptr() == t) {
3039             nn = m;
3040             break;
3041           }
3042         }
3043         if (nn != NULL) {
3044           // Decode a narrow oop to match address
3045           // [R12 + narrow_oop_reg<<3 + offset]
3046           if (is_oop) {
3047             nn = new DecodeNNode(nn, t);
3048           } else {
3049             nn = new DecodeNKlassNode(nn, t);
3050           }
3051           // Check for succeeding AddP which uses the same Base.
3052           // Otherwise we will run into the assertion above when visiting that guy.
3053           for (uint i = 0; i < n->outcnt(); ++i) {
3054             Node *out_i = n->raw_out(i);
3055             if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3056               out_i->set_req(AddPNode::Base, nn);
3057 #ifdef ASSERT
3058               for (uint j = 0; j < out_i->outcnt(); ++j) {
3059                 Node *out_j = out_i->raw_out(j);
3060                 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3061                        "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3062               }
3063 #endif
3064             }
3065           }
3066           n->set_req(AddPNode::Base, nn);
3067           n->set_req(AddPNode::Address, nn);
3068           if (addp->outcnt() == 0) {
3069             addp->disconnect_inputs(NULL, this);
3070           }
3071         }
3072       }
3073     }
3074 #endif
3075     // platform dependent reshaping of the address expression
3076     reshape_address(n->as_AddP());
3077     break;
3078   }
3079 
3080   case Op_CastPP: {
3081     // Remove CastPP nodes to gain more freedom during scheduling but
3082     // keep the dependency they encode as control or precedence edges
3083     // (if control is set already) on memory operations. Some CastPP
3084     // nodes don't have a control (don't carry a dependency): skip
3085     // those.
3086     if (n->in(0) != NULL) {
3087       ResourceMark rm;
3088       Unique_Node_List wq;
3089       wq.push(n);
3090       for (uint next = 0; next < wq.size(); ++next) {
3091         Node *m = wq.at(next);
3092         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3093           Node* use = m->fast_out(i);
3094           if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3095             use->ensure_control_or_add_prec(n->in(0));
3096           } else {
3097             switch(use->Opcode()) {
3098             case Op_AddP:
3099             case Op_DecodeN:
3100             case Op_DecodeNKlass:
3101             case Op_CheckCastPP:
3102             case Op_CastPP:
3103               wq.push(use);
3104               break;
3105             }
3106           }
3107         }
3108       }
3109     }
3110     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3111     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3112       Node* in1 = n->in(1);
3113       const Type* t = n->bottom_type();
3114       Node* new_in1 = in1->clone();
3115       new_in1->as_DecodeN()->set_type(t);
3116 
3117       if (!Matcher::narrow_oop_use_complex_address()) {
3118         //
3119         // x86, ARM and friends can handle 2 adds in addressing mode
3120         // and Matcher can fold a DecodeN node into address by using
3121         // a narrow oop directly and do implicit NULL check in address:
3122         //
3123         // [R12 + narrow_oop_reg<<3 + offset]
3124         // NullCheck narrow_oop_reg
3125         //
3126         // On other platforms (Sparc) we have to keep new DecodeN node and
3127         // use it to do implicit NULL check in address:
3128         //
3129         // decode_not_null narrow_oop_reg, base_reg
3130         // [base_reg + offset]
3131         // NullCheck base_reg
3132         //
3133         // Pin the new DecodeN node to non-null path on these platform (Sparc)
3134         // to keep the information to which NULL check the new DecodeN node
3135         // corresponds to use it as value in implicit_null_check().
3136         //
3137         new_in1->set_req(0, n->in(0));
3138       }
3139 
3140       n->subsume_by(new_in1, this);
3141       if (in1->outcnt() == 0) {
3142         in1->disconnect_inputs(NULL, this);
3143       }
3144     } else {
3145       n->subsume_by(n->in(1), this);
3146       if (n->outcnt() == 0) {
3147         n->disconnect_inputs(NULL, this);
3148       }
3149     }
3150     break;
3151   }
3152 #ifdef _LP64
3153   case Op_CmpP:
3154     // Do this transformation here to preserve CmpPNode::sub() and
3155     // other TypePtr related Ideal optimizations (for example, ptr nullness).
3156     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3157       Node* in1 = n->in(1);
3158       Node* in2 = n->in(2);
3159       if (!in1->is_DecodeNarrowPtr()) {
3160         in2 = in1;
3161         in1 = n->in(2);
3162       }
3163       assert(in1->is_DecodeNarrowPtr(), "sanity");
3164 
3165       Node* new_in2 = NULL;
3166       if (in2->is_DecodeNarrowPtr()) {
3167         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3168         new_in2 = in2->in(1);
3169       } else if (in2->Opcode() == Op_ConP) {
3170         const Type* t = in2->bottom_type();
3171         if (t == TypePtr::NULL_PTR) {
3172           assert(in1->is_DecodeN(), "compare klass to null?");
3173           // Don't convert CmpP null check into CmpN if compressed
3174           // oops implicit null check is not generated.
3175           // This will allow to generate normal oop implicit null check.
3176           if (Matcher::gen_narrow_oop_implicit_null_checks())
3177             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3178           //
3179           // This transformation together with CastPP transformation above
3180           // will generated code for implicit NULL checks for compressed oops.
3181           //
3182           // The original code after Optimize()
3183           //
3184           //    LoadN memory, narrow_oop_reg
3185           //    decode narrow_oop_reg, base_reg
3186           //    CmpP base_reg, NULL
3187           //    CastPP base_reg // NotNull
3188           //    Load [base_reg + offset], val_reg
3189           //
3190           // after these transformations will be
3191           //
3192           //    LoadN memory, narrow_oop_reg
3193           //    CmpN narrow_oop_reg, NULL
3194           //    decode_not_null narrow_oop_reg, base_reg
3195           //    Load [base_reg + offset], val_reg
3196           //
3197           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3198           // since narrow oops can be used in debug info now (see the code in
3199           // final_graph_reshaping_walk()).
3200           //
3201           // At the end the code will be matched to
3202           // on x86:
3203           //
3204           //    Load_narrow_oop memory, narrow_oop_reg
3205           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3206           //    NullCheck narrow_oop_reg
3207           //
3208           // and on sparc:
3209           //
3210           //    Load_narrow_oop memory, narrow_oop_reg
3211           //    decode_not_null narrow_oop_reg, base_reg
3212           //    Load [base_reg + offset], val_reg
3213           //    NullCheck base_reg
3214           //
3215         } else if (t->isa_oopptr()) {
3216           new_in2 = ConNode::make(t->make_narrowoop());
3217         } else if (t->isa_klassptr()) {
3218           new_in2 = ConNode::make(t->make_narrowklass());
3219         }
3220       }
3221       if (new_in2 != NULL) {
3222         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3223         n->subsume_by(cmpN, this);
3224         if (in1->outcnt() == 0) {
3225           in1->disconnect_inputs(NULL, this);
3226         }
3227         if (in2->outcnt() == 0) {
3228           in2->disconnect_inputs(NULL, this);
3229         }
3230       }
3231     }
3232     break;
3233 
3234   case Op_DecodeN:
3235   case Op_DecodeNKlass:
3236     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3237     // DecodeN could be pinned when it can't be fold into
3238     // an address expression, see the code for Op_CastPP above.
3239     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3240     break;
3241 
3242   case Op_EncodeP:
3243   case Op_EncodePKlass: {
3244     Node* in1 = n->in(1);
3245     if (in1->is_DecodeNarrowPtr()) {
3246       n->subsume_by(in1->in(1), this);
3247     } else if (in1->Opcode() == Op_ConP) {
3248       const Type* t = in1->bottom_type();
3249       if (t == TypePtr::NULL_PTR) {
3250         assert(t->isa_oopptr(), "null klass?");
3251         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3252       } else if (t->isa_oopptr()) {
3253         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3254       } else if (t->isa_klassptr()) {
3255         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3256       }
3257     }
3258     if (in1->outcnt() == 0) {
3259       in1->disconnect_inputs(NULL, this);
3260     }
3261     break;
3262   }
3263 
3264   case Op_Proj: {
3265     if (OptimizeStringConcat) {
3266       ProjNode* p = n->as_Proj();
3267       if (p->_is_io_use) {
3268         // Separate projections were used for the exception path which
3269         // are normally removed by a late inline.  If it wasn't inlined
3270         // then they will hang around and should just be replaced with
3271         // the original one.
3272         Node* proj = NULL;
3273         // Replace with just one
3274         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3275           Node *use = i.get();
3276           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3277             proj = use;
3278             break;
3279           }
3280         }
3281         assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3282         if (proj != NULL) {
3283           p->subsume_by(proj, this);
3284         }
3285       }
3286     }
3287     break;
3288   }
3289 
3290   case Op_Phi:
3291     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3292       // The EncodeP optimization may create Phi with the same edges
3293       // for all paths. It is not handled well by Register Allocator.
3294       Node* unique_in = n->in(1);
3295       assert(unique_in != NULL, "");
3296       uint cnt = n->req();
3297       for (uint i = 2; i < cnt; i++) {
3298         Node* m = n->in(i);
3299         assert(m != NULL, "");
3300         if (unique_in != m)
3301           unique_in = NULL;
3302       }
3303       if (unique_in != NULL) {
3304         n->subsume_by(unique_in, this);
3305       }
3306     }
3307     break;
3308 
3309 #endif
3310 
3311 #ifdef ASSERT
3312   case Op_CastII:
3313     // Verify that all range check dependent CastII nodes were removed.
3314     if (n->isa_CastII()->has_range_check()) {
3315       n->dump(3);
3316       assert(false, "Range check dependent CastII node was not removed");
3317     }
3318     break;
3319 #endif
3320 
3321   case Op_ModI:
3322     if (UseDivMod) {
3323       // Check if a%b and a/b both exist
3324       Node* d = n->find_similar(Op_DivI);
3325       if (d) {
3326         // Replace them with a fused divmod if supported
3327         if (Matcher::has_match_rule(Op_DivModI)) {
3328           DivModINode* divmod = DivModINode::make(n);
3329           d->subsume_by(divmod->div_proj(), this);
3330           n->subsume_by(divmod->mod_proj(), this);
3331         } else {
3332           // replace a%b with a-((a/b)*b)
3333           Node* mult = new MulINode(d, d->in(2));
3334           Node* sub  = new SubINode(d->in(1), mult);
3335           n->subsume_by(sub, this);
3336         }
3337       }
3338     }
3339     break;
3340 
3341   case Op_ModL:
3342     if (UseDivMod) {
3343       // Check if a%b and a/b both exist
3344       Node* d = n->find_similar(Op_DivL);
3345       if (d) {
3346         // Replace them with a fused divmod if supported
3347         if (Matcher::has_match_rule(Op_DivModL)) {
3348           DivModLNode* divmod = DivModLNode::make(n);
3349           d->subsume_by(divmod->div_proj(), this);
3350           n->subsume_by(divmod->mod_proj(), this);
3351         } else {
3352           // replace a%b with a-((a/b)*b)
3353           Node* mult = new MulLNode(d, d->in(2));
3354           Node* sub  = new SubLNode(d->in(1), mult);
3355           n->subsume_by(sub, this);
3356         }
3357       }
3358     }
3359     break;
3360 
3361   case Op_LoadVector:
3362   case Op_StoreVector:
3363     break;
3364 
3365   case Op_AddReductionVI:
3366   case Op_AddReductionVL:
3367   case Op_AddReductionVF:
3368   case Op_AddReductionVD:
3369   case Op_MulReductionVI:
3370   case Op_MulReductionVL:
3371   case Op_MulReductionVF:
3372   case Op_MulReductionVD:
3373   case Op_MinReductionV:
3374   case Op_MaxReductionV:
3375     break;
3376 
3377   case Op_PackB:
3378   case Op_PackS:
3379   case Op_PackI:
3380   case Op_PackF:
3381   case Op_PackL:
3382   case Op_PackD:
3383     if (n->req()-1 > 2) {
3384       // Replace many operand PackNodes with a binary tree for matching
3385       PackNode* p = (PackNode*) n;
3386       Node* btp = p->binary_tree_pack(1, n->req());
3387       n->subsume_by(btp, this);
3388     }
3389     break;
3390   case Op_Loop:
3391   case Op_CountedLoop:
3392   case Op_OuterStripMinedLoop:
3393     if (n->as_Loop()->is_inner_loop()) {
3394       frc.inc_inner_loop_count();
3395     }
3396     n->as_Loop()->verify_strip_mined(0);
3397     break;
3398   case Op_LShiftI:
3399   case Op_RShiftI:
3400   case Op_URShiftI:
3401   case Op_LShiftL:
3402   case Op_RShiftL:
3403   case Op_URShiftL:
3404     if (Matcher::need_masked_shift_count) {
3405       // The cpu's shift instructions don't restrict the count to the
3406       // lower 5/6 bits. We need to do the masking ourselves.
3407       Node* in2 = n->in(2);
3408       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3409       const TypeInt* t = in2->find_int_type();
3410       if (t != NULL && t->is_con()) {
3411         juint shift = t->get_con();
3412         if (shift > mask) { // Unsigned cmp
3413           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3414         }
3415       } else {
3416         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3417           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3418           n->set_req(2, shift);
3419         }
3420       }
3421       if (in2->outcnt() == 0) { // Remove dead node
3422         in2->disconnect_inputs(NULL, this);
3423       }
3424     }
3425     break;
3426   case Op_MemBarStoreStore:
3427   case Op_MemBarRelease:
3428     // Break the link with AllocateNode: it is no longer useful and
3429     // confuses register allocation.
3430     if (n->req() > MemBarNode::Precedent) {
3431       n->set_req(MemBarNode::Precedent, top());
3432     }
3433     break;
3434   case Op_MemBarAcquire: {
3435     if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3436       // At parse time, the trailing MemBarAcquire for a volatile load
3437       // is created with an edge to the load. After optimizations,
3438       // that input may be a chain of Phis. If those phis have no
3439       // other use, then the MemBarAcquire keeps them alive and
3440       // register allocation can be confused.
3441       ResourceMark rm;
3442       Unique_Node_List wq;
3443       wq.push(n->in(MemBarNode::Precedent));
3444       n->set_req(MemBarNode::Precedent, top());
3445       while (wq.size() > 0) {
3446         Node* m = wq.pop();
3447         if (m->outcnt() == 0) {
3448           for (uint j = 0; j < m->req(); j++) {
3449             Node* in = m->in(j);
3450             if (in != NULL) {
3451               wq.push(in);
3452             }
3453           }
3454           m->disconnect_inputs(NULL, this);
3455         }
3456       }
3457     }
3458     break;
3459   }
3460   case Op_RangeCheck: {
3461     RangeCheckNode* rc = n->as_RangeCheck();
3462     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3463     n->subsume_by(iff, this);
3464     frc._tests.push(iff);
3465     break;
3466   }
3467   case Op_ConvI2L: {
3468     if (!Matcher::convi2l_type_required) {
3469       // Code generation on some platforms doesn't need accurate
3470       // ConvI2L types. Widening the type can help remove redundant
3471       // address computations.
3472       n->as_Type()->set_type(TypeLong::INT);
3473       ResourceMark rm;
3474       Node_List wq;
3475       wq.push(n);
3476       for (uint next = 0; next < wq.size(); next++) {
3477         Node *m = wq.at(next);
3478 
3479         for(;;) {
3480           // Loop over all nodes with identical inputs edges as m
3481           Node* k = m->find_similar(m->Opcode());
3482           if (k == NULL) {
3483             break;
3484           }
3485           // Push their uses so we get a chance to remove node made
3486           // redundant
3487           for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3488             Node* u = k->fast_out(i);
3489             assert(!wq.contains(u), "shouldn't process one node several times");
3490             if (u->Opcode() == Op_LShiftL ||
3491                 u->Opcode() == Op_AddL ||
3492                 u->Opcode() == Op_SubL ||
3493                 u->Opcode() == Op_AddP) {
3494               wq.push(u);
3495             }
3496           }
3497           // Replace all nodes with identical edges as m with m
3498           k->subsume_by(m, this);
3499         }
3500       }
3501     }
3502     break;
3503   }
3504   case Op_CmpUL: {
3505     if (!Matcher::has_match_rule(Op_CmpUL)) {
3506       // No support for unsigned long comparisons
3507       ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3508       Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3509       Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3510       ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3511       Node* andl = new AndLNode(orl, remove_sign_mask);
3512       Node* cmp = new CmpLNode(andl, n->in(2));
3513       n->subsume_by(cmp, this);
3514     }
3515     break;
3516   }
3517   default:
3518     assert(!n->is_Call(), "");
3519     assert(!n->is_Mem(), "");
3520     assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3521     break;
3522   }
3523 }
3524 
3525 //------------------------------final_graph_reshaping_walk---------------------
3526 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3527 // requires that the walk visits a node's inputs before visiting the node.
3528 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3529   ResourceArea *area = Thread::current()->resource_area();
3530   Unique_Node_List sfpt(area);
3531 
3532   frc._visited.set(root->_idx); // first, mark node as visited
3533   uint cnt = root->req();
3534   Node *n = root;
3535   uint  i = 0;
3536   while (true) {
3537     if (i < cnt) {
3538       // Place all non-visited non-null inputs onto stack
3539       Node* m = n->in(i);
3540       ++i;
3541       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3542         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3543           // compute worst case interpreter size in case of a deoptimization
3544           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3545 
3546           sfpt.push(m);
3547         }
3548         cnt = m->req();
3549         nstack.push(n, i); // put on stack parent and next input's index
3550         n = m;
3551         i = 0;
3552       }
3553     } else {
3554       // Now do post-visit work
3555       final_graph_reshaping_impl( n, frc );
3556       if (nstack.is_empty())
3557         break;             // finished
3558       n = nstack.node();   // Get node from stack
3559       cnt = n->req();
3560       i = nstack.index();
3561       nstack.pop();        // Shift to the next node on stack
3562     }
3563   }
3564 
3565   // Skip next transformation if compressed oops are not used.
3566   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3567       (!UseCompressedOops && !UseCompressedClassPointers))
3568     return;
3569 
3570   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3571   // It could be done for an uncommon traps or any safepoints/calls
3572   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3573   while (sfpt.size() > 0) {
3574     n = sfpt.pop();
3575     JVMState *jvms = n->as_SafePoint()->jvms();
3576     assert(jvms != NULL, "sanity");
3577     int start = jvms->debug_start();
3578     int end   = n->req();
3579     bool is_uncommon = (n->is_CallStaticJava() &&
3580                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3581     for (int j = start; j < end; j++) {
3582       Node* in = n->in(j);
3583       if (in->is_DecodeNarrowPtr()) {
3584         bool safe_to_skip = true;
3585         if (!is_uncommon ) {
3586           // Is it safe to skip?
3587           for (uint i = 0; i < in->outcnt(); i++) {
3588             Node* u = in->raw_out(i);
3589             if (!u->is_SafePoint() ||
3590                 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3591               safe_to_skip = false;
3592             }
3593           }
3594         }
3595         if (safe_to_skip) {
3596           n->set_req(j, in->in(1));
3597         }
3598         if (in->outcnt() == 0) {
3599           in->disconnect_inputs(NULL, this);
3600         }
3601       }
3602     }
3603   }
3604 }
3605 
3606 //------------------------------final_graph_reshaping--------------------------
3607 // Final Graph Reshaping.
3608 //
3609 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3610 //     and not commoned up and forced early.  Must come after regular
3611 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3612 //     inputs to Loop Phis; these will be split by the allocator anyways.
3613 //     Remove Opaque nodes.
3614 // (2) Move last-uses by commutative operations to the left input to encourage
3615 //     Intel update-in-place two-address operations and better register usage
3616 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3617 //     calls canonicalizing them back.
3618 // (3) Count the number of double-precision FP ops, single-precision FP ops
3619 //     and call sites.  On Intel, we can get correct rounding either by
3620 //     forcing singles to memory (requires extra stores and loads after each
3621 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3622 //     clearing the mode bit around call sites).  The mode bit is only used
3623 //     if the relative frequency of single FP ops to calls is low enough.
3624 //     This is a key transform for SPEC mpeg_audio.
3625 // (4) Detect infinite loops; blobs of code reachable from above but not
3626 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3627 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3628 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3629 //     Detection is by looking for IfNodes where only 1 projection is
3630 //     reachable from below or CatchNodes missing some targets.
3631 // (5) Assert for insane oop offsets in debug mode.
3632 
3633 bool Compile::final_graph_reshaping() {
3634   // an infinite loop may have been eliminated by the optimizer,
3635   // in which case the graph will be empty.
3636   if (root()->req() == 1) {
3637     record_method_not_compilable("trivial infinite loop");
3638     return true;
3639   }
3640 
3641   // Expensive nodes have their control input set to prevent the GVN
3642   // from freely commoning them. There's no GVN beyond this point so
3643   // no need to keep the control input. We want the expensive nodes to
3644   // be freely moved to the least frequent code path by gcm.
3645   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3646   for (int i = 0; i < expensive_count(); i++) {
3647     _expensive_nodes->at(i)->set_req(0, NULL);
3648   }
3649 
3650   Final_Reshape_Counts frc;
3651 
3652   // Visit everybody reachable!
3653   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3654   Node_Stack nstack(live_nodes() >> 1);
3655   final_graph_reshaping_walk(nstack, root(), frc);
3656 
3657   // Check for unreachable (from below) code (i.e., infinite loops).
3658   for( uint i = 0; i < frc._tests.size(); i++ ) {
3659     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3660     // Get number of CFG targets.
3661     // Note that PCTables include exception targets after calls.
3662     uint required_outcnt = n->required_outcnt();
3663     if (n->outcnt() != required_outcnt) {
3664       // Check for a few special cases.  Rethrow Nodes never take the
3665       // 'fall-thru' path, so expected kids is 1 less.
3666       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3667         if (n->in(0)->in(0)->is_Call()) {
3668           CallNode *call = n->in(0)->in(0)->as_Call();
3669           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3670             required_outcnt--;      // Rethrow always has 1 less kid
3671           } else if (call->req() > TypeFunc::Parms &&
3672                      call->is_CallDynamicJava()) {
3673             // Check for null receiver. In such case, the optimizer has
3674             // detected that the virtual call will always result in a null
3675             // pointer exception. The fall-through projection of this CatchNode
3676             // will not be populated.
3677             Node *arg0 = call->in(TypeFunc::Parms);
3678             if (arg0->is_Type() &&
3679                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3680               required_outcnt--;
3681             }
3682           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3683                      call->req() > TypeFunc::Parms+1 &&
3684                      call->is_CallStaticJava()) {
3685             // Check for negative array length. In such case, the optimizer has
3686             // detected that the allocation attempt will always result in an
3687             // exception. There is no fall-through projection of this CatchNode .
3688             Node *arg1 = call->in(TypeFunc::Parms+1);
3689             if (arg1->is_Type() &&
3690                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3691               required_outcnt--;
3692             }
3693           }
3694         }
3695       }
3696       // Recheck with a better notion of 'required_outcnt'
3697       if (n->outcnt() != required_outcnt) {
3698         record_method_not_compilable("malformed control flow");
3699         return true;            // Not all targets reachable!
3700       }
3701     }
3702     // Check that I actually visited all kids.  Unreached kids
3703     // must be infinite loops.
3704     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3705       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3706         record_method_not_compilable("infinite loop");
3707         return true;            // Found unvisited kid; must be unreach
3708       }
3709 
3710     // Here so verification code in final_graph_reshaping_walk()
3711     // always see an OuterStripMinedLoopEnd
3712     if (n->is_OuterStripMinedLoopEnd()) {
3713       IfNode* init_iff = n->as_If();
3714       Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3715       n->subsume_by(iff, this);
3716     }
3717   }
3718 
3719   // If original bytecodes contained a mixture of floats and doubles
3720   // check if the optimizer has made it homogenous, item (3).
3721   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3722       frc.get_float_count() > 32 &&
3723       frc.get_double_count() == 0 &&
3724       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3725     set_24_bit_selection_and_mode( false,  true );
3726   }
3727 
3728   set_java_calls(frc.get_java_call_count());
3729   set_inner_loops(frc.get_inner_loop_count());
3730 
3731   // No infinite loops, no reason to bail out.
3732   return false;
3733 }
3734 
3735 //-----------------------------too_many_traps----------------------------------
3736 // Report if there are too many traps at the current method and bci.
3737 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3738 bool Compile::too_many_traps(ciMethod* method,
3739                              int bci,
3740                              Deoptimization::DeoptReason reason) {
3741   ciMethodData* md = method->method_data();
3742   if (md->is_empty()) {
3743     // Assume the trap has not occurred, or that it occurred only
3744     // because of a transient condition during start-up in the interpreter.
3745     return false;
3746   }
3747   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3748   if (md->has_trap_at(bci, m, reason) != 0) {
3749     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3750     // Also, if there are multiple reasons, or if there is no per-BCI record,
3751     // assume the worst.
3752     if (log())
3753       log()->elem("observe trap='%s' count='%d'",
3754                   Deoptimization::trap_reason_name(reason),
3755                   md->trap_count(reason));
3756     return true;
3757   } else {
3758     // Ignore method/bci and see if there have been too many globally.
3759     return too_many_traps(reason, md);
3760   }
3761 }
3762 
3763 // Less-accurate variant which does not require a method and bci.
3764 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3765                              ciMethodData* logmd) {
3766   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3767     // Too many traps globally.
3768     // Note that we use cumulative trap_count, not just md->trap_count.
3769     if (log()) {
3770       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3771       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3772                   Deoptimization::trap_reason_name(reason),
3773                   mcount, trap_count(reason));
3774     }
3775     return true;
3776   } else {
3777     // The coast is clear.
3778     return false;
3779   }
3780 }
3781 
3782 //--------------------------too_many_recompiles--------------------------------
3783 // Report if there are too many recompiles at the current method and bci.
3784 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3785 // Is not eager to return true, since this will cause the compiler to use
3786 // Action_none for a trap point, to avoid too many recompilations.
3787 bool Compile::too_many_recompiles(ciMethod* method,
3788                                   int bci,
3789                                   Deoptimization::DeoptReason reason) {
3790   ciMethodData* md = method->method_data();
3791   if (md->is_empty()) {
3792     // Assume the trap has not occurred, or that it occurred only
3793     // because of a transient condition during start-up in the interpreter.
3794     return false;
3795   }
3796   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3797   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3798   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3799   Deoptimization::DeoptReason per_bc_reason
3800     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3801   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3802   if ((per_bc_reason == Deoptimization::Reason_none
3803        || md->has_trap_at(bci, m, reason) != 0)
3804       // The trap frequency measure we care about is the recompile count:
3805       && md->trap_recompiled_at(bci, m)
3806       && md->overflow_recompile_count() >= bc_cutoff) {
3807     // Do not emit a trap here if it has already caused recompilations.
3808     // Also, if there are multiple reasons, or if there is no per-BCI record,
3809     // assume the worst.
3810     if (log())
3811       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3812                   Deoptimization::trap_reason_name(reason),
3813                   md->trap_count(reason),
3814                   md->overflow_recompile_count());
3815     return true;
3816   } else if (trap_count(reason) != 0
3817              && decompile_count() >= m_cutoff) {
3818     // Too many recompiles globally, and we have seen this sort of trap.
3819     // Use cumulative decompile_count, not just md->decompile_count.
3820     if (log())
3821       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3822                   Deoptimization::trap_reason_name(reason),
3823                   md->trap_count(reason), trap_count(reason),
3824                   md->decompile_count(), decompile_count());
3825     return true;
3826   } else {
3827     // The coast is clear.
3828     return false;
3829   }
3830 }
3831 
3832 // Compute when not to trap. Used by matching trap based nodes and
3833 // NullCheck optimization.
3834 void Compile::set_allowed_deopt_reasons() {
3835   _allowed_reasons = 0;
3836   if (is_method_compilation()) {
3837     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3838       assert(rs < BitsPerInt, "recode bit map");
3839       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3840         _allowed_reasons |= nth_bit(rs);
3841       }
3842     }
3843   }
3844 }
3845 
3846 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
3847   return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
3848 }
3849 
3850 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
3851   return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
3852 }
3853 
3854 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
3855   if (holder->is_initialized()) {
3856     return false;
3857   }
3858   if (holder->is_being_initialized()) {
3859     if (accessing_method->holder() == holder) {
3860       // Access inside a class. The barrier can be elided when access happens in <clinit>,
3861       // <init>, or a static method. In all those cases, there was an initialization
3862       // barrier on the holder klass passed.
3863       if (accessing_method->is_static_initializer() ||
3864           accessing_method->is_object_initializer() ||
3865           accessing_method->is_static()) {
3866         return false;
3867       }
3868     } else if (accessing_method->holder()->is_subclass_of(holder)) {
3869       // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
3870       // In case of <init> or a static method, the barrier is on the subclass is not enough:
3871       // child class can become fully initialized while its parent class is still being initialized.
3872       if (accessing_method->is_static_initializer()) {
3873         return false;
3874       }
3875     }
3876     ciMethod* root = method(); // the root method of compilation
3877     if (root != accessing_method) {
3878       return needs_clinit_barrier(holder, root); // check access in the context of compilation root
3879     }
3880   }
3881   return true;
3882 }
3883 
3884 #ifndef PRODUCT
3885 //------------------------------verify_graph_edges---------------------------
3886 // Walk the Graph and verify that there is a one-to-one correspondence
3887 // between Use-Def edges and Def-Use edges in the graph.
3888 void Compile::verify_graph_edges(bool no_dead_code) {
3889   if (VerifyGraphEdges) {
3890     ResourceArea *area = Thread::current()->resource_area();
3891     Unique_Node_List visited(area);
3892     // Call recursive graph walk to check edges
3893     _root->verify_edges(visited);
3894     if (no_dead_code) {
3895       // Now make sure that no visited node is used by an unvisited node.
3896       bool dead_nodes = false;
3897       Unique_Node_List checked(area);
3898       while (visited.size() > 0) {
3899         Node* n = visited.pop();
3900         checked.push(n);
3901         for (uint i = 0; i < n->outcnt(); i++) {
3902           Node* use = n->raw_out(i);
3903           if (checked.member(use))  continue;  // already checked
3904           if (visited.member(use))  continue;  // already in the graph
3905           if (use->is_Con())        continue;  // a dead ConNode is OK
3906           // At this point, we have found a dead node which is DU-reachable.
3907           if (!dead_nodes) {
3908             tty->print_cr("*** Dead nodes reachable via DU edges:");
3909             dead_nodes = true;
3910           }
3911           use->dump(2);
3912           tty->print_cr("---");
3913           checked.push(use);  // No repeats; pretend it is now checked.
3914         }
3915       }
3916       assert(!dead_nodes, "using nodes must be reachable from root");
3917     }
3918   }
3919 }
3920 #endif
3921 
3922 // The Compile object keeps track of failure reasons separately from the ciEnv.
3923 // This is required because there is not quite a 1-1 relation between the
3924 // ciEnv and its compilation task and the Compile object.  Note that one
3925 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3926 // to backtrack and retry without subsuming loads.  Other than this backtracking
3927 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3928 // by the logic in C2Compiler.
3929 void Compile::record_failure(const char* reason) {
3930   if (log() != NULL) {
3931     log()->elem("failure reason='%s' phase='compile'", reason);
3932   }
3933   if (_failure_reason == NULL) {
3934     // Record the first failure reason.
3935     _failure_reason = reason;
3936   }
3937 
3938   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3939     C->print_method(PHASE_FAILURE);
3940   }
3941   _root = NULL;  // flush the graph, too
3942 }
3943 
3944 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3945   : TraceTime(name, accumulator, CITime, CITimeVerbose),
3946     _phase_name(name), _dolog(CITimeVerbose)
3947 {
3948   if (_dolog) {
3949     C = Compile::current();
3950     _log = C->log();
3951   } else {
3952     C = NULL;
3953     _log = NULL;
3954   }
3955   if (_log != NULL) {
3956     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3957     _log->stamp();
3958     _log->end_head();
3959   }
3960 }
3961 
3962 Compile::TracePhase::~TracePhase() {
3963 
3964   C = Compile::current();
3965   if (_dolog) {
3966     _log = C->log();
3967   } else {
3968     _log = NULL;
3969   }
3970 
3971 #ifdef ASSERT
3972   if (PrintIdealNodeCount) {
3973     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3974                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3975   }
3976 
3977   if (VerifyIdealNodeCount) {
3978     Compile::current()->print_missing_nodes();
3979   }
3980 #endif
3981 
3982   if (_log != NULL) {
3983     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3984   }
3985 }
3986 
3987 //=============================================================================
3988 // Two Constant's are equal when the type and the value are equal.
3989 bool Compile::Constant::operator==(const Constant& other) {
3990   if (type()          != other.type()         )  return false;
3991   if (can_be_reused() != other.can_be_reused())  return false;
3992   // For floating point values we compare the bit pattern.
3993   switch (type()) {
3994   case T_INT:
3995   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3996   case T_LONG:
3997   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
3998   case T_OBJECT:
3999   case T_ADDRESS: return (_v._value.l == other._v._value.l);
4000   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
4001   case T_METADATA: return (_v._metadata == other._v._metadata);
4002   default: ShouldNotReachHere(); return false;
4003   }
4004 }
4005 
4006 static int type_to_size_in_bytes(BasicType t) {
4007   switch (t) {
4008   case T_INT:     return sizeof(jint   );
4009   case T_LONG:    return sizeof(jlong  );
4010   case T_FLOAT:   return sizeof(jfloat );
4011   case T_DOUBLE:  return sizeof(jdouble);
4012   case T_METADATA: return sizeof(Metadata*);
4013     // We use T_VOID as marker for jump-table entries (labels) which
4014     // need an internal word relocation.
4015   case T_VOID:
4016   case T_ADDRESS:
4017   case T_OBJECT:  return sizeof(jobject);
4018   default:
4019     ShouldNotReachHere();
4020     return -1;
4021   }
4022 }
4023 
4024 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4025   // sort descending
4026   if (a->freq() > b->freq())  return -1;
4027   if (a->freq() < b->freq())  return  1;
4028   return 0;
4029 }
4030 
4031 void Compile::ConstantTable::calculate_offsets_and_size() {
4032   // First, sort the array by frequencies.
4033   _constants.sort(qsort_comparator);
4034 
4035 #ifdef ASSERT
4036   // Make sure all jump-table entries were sorted to the end of the
4037   // array (they have a negative frequency).
4038   bool found_void = false;
4039   for (int i = 0; i < _constants.length(); i++) {
4040     Constant con = _constants.at(i);
4041     if (con.type() == T_VOID)
4042       found_void = true;  // jump-tables
4043     else
4044       assert(!found_void, "wrong sorting");
4045   }
4046 #endif
4047 
4048   int offset = 0;
4049   for (int i = 0; i < _constants.length(); i++) {
4050     Constant* con = _constants.adr_at(i);
4051 
4052     // Align offset for type.
4053     int typesize = type_to_size_in_bytes(con->type());
4054     offset = align_up(offset, typesize);
4055     con->set_offset(offset);   // set constant's offset
4056 
4057     if (con->type() == T_VOID) {
4058       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4059       offset = offset + typesize * n->outcnt();  // expand jump-table
4060     } else {
4061       offset = offset + typesize;
4062     }
4063   }
4064 
4065   // Align size up to the next section start (which is insts; see
4066   // CodeBuffer::align_at_start).
4067   assert(_size == -1, "already set?");
4068   _size = align_up(offset, (int)CodeEntryAlignment);
4069 }
4070 
4071 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4072   MacroAssembler _masm(&cb);
4073   for (int i = 0; i < _constants.length(); i++) {
4074     Constant con = _constants.at(i);
4075     address constant_addr = NULL;
4076     switch (con.type()) {
4077     case T_INT:    constant_addr = _masm.int_constant(   con.get_jint()   ); break;
4078     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
4079     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4080     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4081     case T_OBJECT: {
4082       jobject obj = con.get_jobject();
4083       int oop_index = _masm.oop_recorder()->find_index(obj);
4084       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4085       break;
4086     }
4087     case T_ADDRESS: {
4088       address addr = (address) con.get_jobject();
4089       constant_addr = _masm.address_constant(addr);
4090       break;
4091     }
4092     // We use T_VOID as marker for jump-table entries (labels) which
4093     // need an internal word relocation.
4094     case T_VOID: {
4095       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4096       // Fill the jump-table with a dummy word.  The real value is
4097       // filled in later in fill_jump_table.
4098       address dummy = (address) n;
4099       constant_addr = _masm.address_constant(dummy);
4100       // Expand jump-table
4101       for (uint i = 1; i < n->outcnt(); i++) {
4102         address temp_addr = _masm.address_constant(dummy + i);
4103         assert(temp_addr, "consts section too small");
4104       }
4105       break;
4106     }
4107     case T_METADATA: {
4108       Metadata* obj = con.get_metadata();
4109       int metadata_index = _masm.oop_recorder()->find_index(obj);
4110       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4111       break;
4112     }
4113     default: ShouldNotReachHere();
4114     }
4115     assert(constant_addr, "consts section too small");
4116     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4117             "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4118   }
4119 }
4120 
4121 int Compile::ConstantTable::find_offset(Constant& con) const {
4122   int idx = _constants.find(con);
4123   guarantee(idx != -1, "constant must be in constant table");
4124   int offset = _constants.at(idx).offset();
4125   guarantee(offset != -1, "constant table not emitted yet?");
4126   return offset;
4127 }
4128 
4129 void Compile::ConstantTable::add(Constant& con) {
4130   if (con.can_be_reused()) {
4131     int idx = _constants.find(con);
4132     if (idx != -1 && _constants.at(idx).can_be_reused()) {
4133       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
4134       return;
4135     }
4136   }
4137   (void) _constants.append(con);
4138 }
4139 
4140 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4141   Block* b = Compile::current()->cfg()->get_block_for_node(n);
4142   Constant con(type, value, b->_freq);
4143   add(con);
4144   return con;
4145 }
4146 
4147 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4148   Constant con(metadata);
4149   add(con);
4150   return con;
4151 }
4152 
4153 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4154   jvalue value;
4155   BasicType type = oper->type()->basic_type();
4156   switch (type) {
4157   case T_LONG:    value.j = oper->constantL(); break;
4158   case T_FLOAT:   value.f = oper->constantF(); break;
4159   case T_DOUBLE:  value.d = oper->constantD(); break;
4160   case T_OBJECT:
4161   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4162   case T_METADATA: return add((Metadata*)oper->constant()); break;
4163   default: guarantee(false, "unhandled type: %s", type2name(type));
4164   }
4165   return add(n, type, value);
4166 }
4167 
4168 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4169   jvalue value;
4170   // We can use the node pointer here to identify the right jump-table
4171   // as this method is called from Compile::Fill_buffer right before
4172   // the MachNodes are emitted and the jump-table is filled (means the
4173   // MachNode pointers do not change anymore).
4174   value.l = (jobject) n;
4175   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
4176   add(con);
4177   return con;
4178 }
4179 
4180 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4181   // If called from Compile::scratch_emit_size do nothing.
4182   if (Compile::current()->in_scratch_emit_size())  return;
4183 
4184   assert(labels.is_nonempty(), "must be");
4185   assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4186 
4187   // Since MachConstantNode::constant_offset() also contains
4188   // table_base_offset() we need to subtract the table_base_offset()
4189   // to get the plain offset into the constant table.
4190   int offset = n->constant_offset() - table_base_offset();
4191 
4192   MacroAssembler _masm(&cb);
4193   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4194 
4195   for (uint i = 0; i < n->outcnt(); i++) {
4196     address* constant_addr = &jump_table_base[i];
4197     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));
4198     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4199     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4200   }
4201 }
4202 
4203 //----------------------------static_subtype_check-----------------------------
4204 // Shortcut important common cases when superklass is exact:
4205 // (0) superklass is java.lang.Object (can occur in reflective code)
4206 // (1) subklass is already limited to a subtype of superklass => always ok
4207 // (2) subklass does not overlap with superklass => always fail
4208 // (3) superklass has NO subtypes and we can check with a simple compare.
4209 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4210   if (StressReflectiveCode) {
4211     return SSC_full_test;       // Let caller generate the general case.
4212   }
4213 
4214   if (superk == env()->Object_klass()) {
4215     return SSC_always_true;     // (0) this test cannot fail
4216   }
4217 
4218   ciType* superelem = superk;
4219   if (superelem->is_array_klass())
4220     superelem = superelem->as_array_klass()->base_element_type();
4221 
4222   if (!subk->is_interface()) {  // cannot trust static interface types yet
4223     if (subk->is_subtype_of(superk)) {
4224       return SSC_always_true;   // (1) false path dead; no dynamic test needed
4225     }
4226     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4227         !superk->is_subtype_of(subk)) {
4228       return SSC_always_false;
4229     }
4230   }
4231 
4232   // If casting to an instance klass, it must have no subtypes
4233   if (superk->is_interface()) {
4234     // Cannot trust interfaces yet.
4235     // %%% S.B. superk->nof_implementors() == 1
4236   } else if (superelem->is_instance_klass()) {
4237     ciInstanceKlass* ik = superelem->as_instance_klass();
4238     if (!ik->has_subklass() && !ik->is_interface()) {
4239       if (!ik->is_final()) {
4240         // Add a dependency if there is a chance of a later subclass.
4241         dependencies()->assert_leaf_type(ik);
4242       }
4243       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
4244     }
4245   } else {
4246     // A primitive array type has no subtypes.
4247     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
4248   }
4249 
4250   return SSC_full_test;
4251 }
4252 
4253 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4254 #ifdef _LP64
4255   // The scaled index operand to AddP must be a clean 64-bit value.
4256   // Java allows a 32-bit int to be incremented to a negative
4257   // value, which appears in a 64-bit register as a large
4258   // positive number.  Using that large positive number as an
4259   // operand in pointer arithmetic has bad consequences.
4260   // On the other hand, 32-bit overflow is rare, and the possibility
4261   // can often be excluded, if we annotate the ConvI2L node with
4262   // a type assertion that its value is known to be a small positive
4263   // number.  (The prior range check has ensured this.)
4264   // This assertion is used by ConvI2LNode::Ideal.
4265   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4266   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4267   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4268   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4269 #endif
4270   return idx;
4271 }
4272 
4273 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4274 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4275   if (ctrl != NULL) {
4276     // Express control dependency by a CastII node with a narrow type.
4277     value = new CastIINode(value, itype, false, true /* range check dependency */);
4278     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4279     // node from floating above the range check during loop optimizations. Otherwise, the
4280     // ConvI2L node may be eliminated independently of the range check, causing the data path
4281     // to become TOP while the control path is still there (although it's unreachable).
4282     value->set_req(0, ctrl);
4283     // Save CastII node to remove it after loop optimizations.
4284     phase->C->add_range_check_cast(value);
4285     value = phase->transform(value);
4286   }
4287   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4288   return phase->transform(new ConvI2LNode(value, ltype));
4289 }
4290 
4291 // The message about the current inlining is accumulated in
4292 // _print_inlining_stream and transfered into the _print_inlining_list
4293 // once we know whether inlining succeeds or not. For regular
4294 // inlining, messages are appended to the buffer pointed by
4295 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4296 // a new buffer is added after _print_inlining_idx in the list. This
4297 // way we can update the inlining message for late inlining call site
4298 // when the inlining is attempted again.
4299 void Compile::print_inlining_init() {
4300   if (print_inlining() || print_intrinsics()) {
4301     _print_inlining_stream = new stringStream();
4302     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4303   }
4304 }
4305 
4306 void Compile::print_inlining_reinit() {
4307   if (print_inlining() || print_intrinsics()) {
4308     // Re allocate buffer when we change ResourceMark
4309     _print_inlining_stream = new stringStream();
4310   }
4311 }
4312 
4313 void Compile::print_inlining_reset() {
4314   _print_inlining_stream->reset();
4315 }
4316 
4317 void Compile::print_inlining_commit() {
4318   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4319   // Transfer the message from _print_inlining_stream to the current
4320   // _print_inlining_list buffer and clear _print_inlining_stream.
4321   _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4322   print_inlining_reset();
4323 }
4324 
4325 void Compile::print_inlining_push() {
4326   // Add new buffer to the _print_inlining_list at current position
4327   _print_inlining_idx++;
4328   _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4329 }
4330 
4331 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4332   return _print_inlining_list->at(_print_inlining_idx);
4333 }
4334 
4335 void Compile::print_inlining_update(CallGenerator* cg) {
4336   if (print_inlining() || print_intrinsics()) {
4337     if (!cg->is_late_inline()) {
4338       if (print_inlining_current().cg() != NULL) {
4339         print_inlining_push();
4340       }
4341       print_inlining_commit();
4342     } else {
4343       if (print_inlining_current().cg() != cg &&
4344           (print_inlining_current().cg() != NULL ||
4345            print_inlining_current().ss()->size() != 0)) {
4346         print_inlining_push();
4347       }
4348       print_inlining_commit();
4349       print_inlining_current().set_cg(cg);
4350     }
4351   }
4352 }
4353 
4354 void Compile::print_inlining_move_to(CallGenerator* cg) {
4355   // We resume inlining at a late inlining call site. Locate the
4356   // corresponding inlining buffer so that we can update it.
4357   if (print_inlining()) {
4358     for (int i = 0; i < _print_inlining_list->length(); i++) {
4359       if (_print_inlining_list->adr_at(i)->cg() == cg) {
4360         _print_inlining_idx = i;
4361         return;
4362       }
4363     }
4364     ShouldNotReachHere();
4365   }
4366 }
4367 
4368 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4369   if (print_inlining()) {
4370     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4371     assert(print_inlining_current().cg() == cg, "wrong entry");
4372     // replace message with new message
4373     _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4374     print_inlining_commit();
4375     print_inlining_current().set_cg(cg);
4376   }
4377 }
4378 
4379 void Compile::print_inlining_assert_ready() {
4380   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4381 }
4382 
4383 void Compile::process_print_inlining() {
4384   bool do_print_inlining = print_inlining() || print_intrinsics();
4385   if (do_print_inlining || log() != NULL) {
4386     // Print inlining message for candidates that we couldn't inline
4387     // for lack of space
4388     for (int i = 0; i < _late_inlines.length(); i++) {
4389       CallGenerator* cg = _late_inlines.at(i);
4390       if (!cg->is_mh_late_inline()) {
4391         const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4392         if (do_print_inlining) {
4393           cg->print_inlining_late(msg);
4394         }
4395         log_late_inline_failure(cg, msg);
4396       }
4397     }
4398   }
4399   if (do_print_inlining) {
4400     ResourceMark rm;
4401     stringStream ss;
4402     for (int i = 0; i < _print_inlining_list->length(); i++) {
4403       ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4404     }
4405     size_t end = ss.size();
4406     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4407     strncpy(_print_inlining_output, ss.base(), end+1);
4408     _print_inlining_output[end] = 0;
4409   }
4410 }
4411 
4412 void Compile::dump_print_inlining() {
4413   if (_print_inlining_output != NULL) {
4414     tty->print_raw(_print_inlining_output);
4415   }
4416 }
4417 
4418 void Compile::log_late_inline(CallGenerator* cg) {
4419   if (log() != NULL) {
4420     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4421                 cg->unique_id());
4422     JVMState* p = cg->call_node()->jvms();
4423     while (p != NULL) {
4424       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4425       p = p->caller();
4426     }
4427     log()->tail("late_inline");
4428   }
4429 }
4430 
4431 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4432   log_late_inline(cg);
4433   if (log() != NULL) {
4434     log()->inline_fail(msg);
4435   }
4436 }
4437 
4438 void Compile::log_inline_id(CallGenerator* cg) {
4439   if (log() != NULL) {
4440     // The LogCompilation tool needs a unique way to identify late
4441     // inline call sites. This id must be unique for this call site in
4442     // this compilation. Try to have it unique across compilations as
4443     // well because it can be convenient when grepping through the log
4444     // file.
4445     // Distinguish OSR compilations from others in case CICountOSR is
4446     // on.
4447     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4448     cg->set_unique_id(id);
4449     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4450   }
4451 }
4452 
4453 void Compile::log_inline_failure(const char* msg) {
4454   if (C->log() != NULL) {
4455     C->log()->inline_fail(msg);
4456   }
4457 }
4458 
4459 
4460 // Dump inlining replay data to the stream.
4461 // Don't change thread state and acquire any locks.
4462 void Compile::dump_inline_data(outputStream* out) {
4463   InlineTree* inl_tree = ilt();
4464   if (inl_tree != NULL) {
4465     out->print(" inline %d", inl_tree->count());
4466     inl_tree->dump_replay_data(out);
4467   }
4468 }
4469 
4470 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4471   if (n1->Opcode() < n2->Opcode())      return -1;
4472   else if (n1->Opcode() > n2->Opcode()) return 1;
4473 
4474   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4475   for (uint i = 1; i < n1->req(); i++) {
4476     if (n1->in(i) < n2->in(i))      return -1;
4477     else if (n1->in(i) > n2->in(i)) return 1;
4478   }
4479 
4480   return 0;
4481 }
4482 
4483 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4484   Node* n1 = *n1p;
4485   Node* n2 = *n2p;
4486 
4487   return cmp_expensive_nodes(n1, n2);
4488 }
4489 
4490 void Compile::sort_expensive_nodes() {
4491   if (!expensive_nodes_sorted()) {
4492     _expensive_nodes->sort(cmp_expensive_nodes);
4493   }
4494 }
4495 
4496 bool Compile::expensive_nodes_sorted() const {
4497   for (int i = 1; i < _expensive_nodes->length(); i++) {
4498     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4499       return false;
4500     }
4501   }
4502   return true;
4503 }
4504 
4505 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4506   if (_expensive_nodes->length() == 0) {
4507     return false;
4508   }
4509 
4510   assert(OptimizeExpensiveOps, "optimization off?");
4511 
4512   // Take this opportunity to remove dead nodes from the list
4513   int j = 0;
4514   for (int i = 0; i < _expensive_nodes->length(); i++) {
4515     Node* n = _expensive_nodes->at(i);
4516     if (!n->is_unreachable(igvn)) {
4517       assert(n->is_expensive(), "should be expensive");
4518       _expensive_nodes->at_put(j, n);
4519       j++;
4520     }
4521   }
4522   _expensive_nodes->trunc_to(j);
4523 
4524   // Then sort the list so that similar nodes are next to each other
4525   // and check for at least two nodes of identical kind with same data
4526   // inputs.
4527   sort_expensive_nodes();
4528 
4529   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4530     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4531       return true;
4532     }
4533   }
4534 
4535   return false;
4536 }
4537 
4538 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4539   if (_expensive_nodes->length() == 0) {
4540     return;
4541   }
4542 
4543   assert(OptimizeExpensiveOps, "optimization off?");
4544 
4545   // Sort to bring similar nodes next to each other and clear the
4546   // control input of nodes for which there's only a single copy.
4547   sort_expensive_nodes();
4548 
4549   int j = 0;
4550   int identical = 0;
4551   int i = 0;
4552   bool modified = false;
4553   for (; i < _expensive_nodes->length()-1; i++) {
4554     assert(j <= i, "can't write beyond current index");
4555     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4556       identical++;
4557       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4558       continue;
4559     }
4560     if (identical > 0) {
4561       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4562       identical = 0;
4563     } else {
4564       Node* n = _expensive_nodes->at(i);
4565       igvn.replace_input_of(n, 0, NULL);
4566       igvn.hash_insert(n);
4567       modified = true;
4568     }
4569   }
4570   if (identical > 0) {
4571     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4572   } else if (_expensive_nodes->length() >= 1) {
4573     Node* n = _expensive_nodes->at(i);
4574     igvn.replace_input_of(n, 0, NULL);
4575     igvn.hash_insert(n);
4576     modified = true;
4577   }
4578   _expensive_nodes->trunc_to(j);
4579   if (modified) {
4580     igvn.optimize();
4581   }
4582 }
4583 
4584 void Compile::add_expensive_node(Node * n) {
4585   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4586   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4587   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4588   if (OptimizeExpensiveOps) {
4589     _expensive_nodes->append(n);
4590   } else {
4591     // Clear control input and let IGVN optimize expensive nodes if
4592     // OptimizeExpensiveOps is off.
4593     n->set_req(0, NULL);
4594   }
4595 }
4596 
4597 /**
4598  * Remove the speculative part of types and clean up the graph
4599  */
4600 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4601   if (UseTypeSpeculation) {
4602     Unique_Node_List worklist;
4603     worklist.push(root());
4604     int modified = 0;
4605     // Go over all type nodes that carry a speculative type, drop the
4606     // speculative part of the type and enqueue the node for an igvn
4607     // which may optimize it out.
4608     for (uint next = 0; next < worklist.size(); ++next) {
4609       Node *n  = worklist.at(next);
4610       if (n->is_Type()) {
4611         TypeNode* tn = n->as_Type();
4612         const Type* t = tn->type();
4613         const Type* t_no_spec = t->remove_speculative();
4614         if (t_no_spec != t) {
4615           bool in_hash = igvn.hash_delete(n);
4616           assert(in_hash, "node should be in igvn hash table");
4617           tn->set_type(t_no_spec);
4618           igvn.hash_insert(n);
4619           igvn._worklist.push(n); // give it a chance to go away
4620           modified++;
4621         }
4622       }
4623       uint max = n->len();
4624       for( uint i = 0; i < max; ++i ) {
4625         Node *m = n->in(i);
4626         if (not_a_node(m))  continue;
4627         worklist.push(m);
4628       }
4629     }
4630     // Drop the speculative part of all types in the igvn's type table
4631     igvn.remove_speculative_types();
4632     if (modified > 0) {
4633       igvn.optimize();
4634     }
4635 #ifdef ASSERT
4636     // Verify that after the IGVN is over no speculative type has resurfaced
4637     worklist.clear();
4638     worklist.push(root());
4639     for (uint next = 0; next < worklist.size(); ++next) {
4640       Node *n  = worklist.at(next);
4641       const Type* t = igvn.type_or_null(n);
4642       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4643       if (n->is_Type()) {
4644         t = n->as_Type()->type();
4645         assert(t == t->remove_speculative(), "no more speculative types");
4646       }
4647       uint max = n->len();
4648       for( uint i = 0; i < max; ++i ) {
4649         Node *m = n->in(i);
4650         if (not_a_node(m))  continue;
4651         worklist.push(m);
4652       }
4653     }
4654     igvn.check_no_speculative_types();
4655 #endif
4656   }
4657 }
4658 
4659 // Auxiliary method to support randomized stressing/fuzzing.
4660 //
4661 // This method can be called the arbitrary number of times, with current count
4662 // as the argument. The logic allows selecting a single candidate from the
4663 // running list of candidates as follows:
4664 //    int count = 0;
4665 //    Cand* selected = null;
4666 //    while(cand = cand->next()) {
4667 //      if (randomized_select(++count)) {
4668 //        selected = cand;
4669 //      }
4670 //    }
4671 //
4672 // Including count equalizes the chances any candidate is "selected".
4673 // This is useful when we don't have the complete list of candidates to choose
4674 // from uniformly. In this case, we need to adjust the randomicity of the
4675 // selection, or else we will end up biasing the selection towards the latter
4676 // candidates.
4677 //
4678 // Quick back-envelope calculation shows that for the list of n candidates
4679 // the equal probability for the candidate to persist as "best" can be
4680 // achieved by replacing it with "next" k-th candidate with the probability
4681 // of 1/k. It can be easily shown that by the end of the run, the
4682 // probability for any candidate is converged to 1/n, thus giving the
4683 // uniform distribution among all the candidates.
4684 //
4685 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4686 #define RANDOMIZED_DOMAIN_POW 29
4687 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4688 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4689 bool Compile::randomized_select(int count) {
4690   assert(count > 0, "only positive");
4691   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4692 }
4693 
4694 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4695 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4696 
4697 void NodeCloneInfo::dump() const {
4698   tty->print(" {%d:%d} ", idx(), gen());
4699 }
4700 
4701 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4702   uint64_t val = value(old->_idx);
4703   NodeCloneInfo cio(val);
4704   assert(val != 0, "old node should be in the map");
4705   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4706   insert(nnn->_idx, cin.get());
4707 #ifndef PRODUCT
4708   if (is_debug()) {
4709     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4710   }
4711 #endif
4712 }
4713 
4714 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4715   NodeCloneInfo cio(value(old->_idx));
4716   if (cio.get() == 0) {
4717     cio.set(old->_idx, 0);
4718     insert(old->_idx, cio.get());
4719 #ifndef PRODUCT
4720     if (is_debug()) {
4721       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4722     }
4723 #endif
4724   }
4725   clone(old, nnn, gen);
4726 }
4727 
4728 int CloneMap::max_gen() const {
4729   int g = 0;
4730   DictI di(_dict);
4731   for(; di.test(); ++di) {
4732     int t = gen(di._key);
4733     if (g < t) {
4734       g = t;
4735 #ifndef PRODUCT
4736       if (is_debug()) {
4737         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4738       }
4739 #endif
4740     }
4741   }
4742   return g;
4743 }
4744 
4745 void CloneMap::dump(node_idx_t key) const {
4746   uint64_t val = value(key);
4747   if (val != 0) {
4748     NodeCloneInfo ni(val);
4749     ni.dump();
4750   }
4751 }