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                            has_monitors(),
 952                            0,
 953                            rtm_state()
 954                            );
 955 
 956     if (log() != NULL) // Print code cache state into compiler log
 957       log()->code_cache_state();
 958   }
 959 }
 960 
 961 //------------------------------Compile----------------------------------------
 962 // Compile a runtime stub
 963 Compile::Compile( ciEnv* ci_env,
 964                   TypeFunc_generator generator,
 965                   address stub_function,
 966                   const char *stub_name,
 967                   int is_fancy_jump,
 968                   bool pass_tls,
 969                   bool save_arg_registers,
 970                   bool return_pc,
 971                   DirectiveSet* directive)
 972   : Phase(Compiler),
 973     _compile_id(0),
 974     _save_argument_registers(save_arg_registers),
 975     _subsume_loads(true),
 976     _do_escape_analysis(false),
 977     _eliminate_boxing(false),
 978     _method(NULL),
 979     _entry_bci(InvocationEntryBci),
 980     _stub_function(stub_function),
 981     _stub_name(stub_name),
 982     _stub_entry_point(NULL),
 983     _max_node_limit(MaxNodeLimit),
 984     _orig_pc_slot(0),
 985     _orig_pc_slot_offset_in_bytes(0),
 986     _inlining_progress(false),
 987     _inlining_incrementally(false),
 988     _has_reserved_stack_access(false),
 989 #ifndef PRODUCT
 990     _trace_opto_output(directive->TraceOptoOutputOption),
 991 #endif
 992     _has_method_handle_invokes(false),
 993     _clinit_barrier_on_entry(false),
 994     _comp_arena(mtCompiler),
 995     _env(ci_env),
 996     _directive(directive),
 997     _log(ci_env->log()),
 998     _failure_reason(NULL),
 999     _congraph(NULL),
1000 #ifndef PRODUCT
1001     _printer(NULL),
1002 #endif
1003     _dead_node_list(comp_arena()),
1004     _dead_node_count(0),
1005     _node_arena(mtCompiler),
1006     _old_arena(mtCompiler),
1007     _mach_constant_base_node(NULL),
1008     _Compile_types(mtCompiler),
1009     _initial_gvn(NULL),
1010     _for_igvn(NULL),
1011     _warm_calls(NULL),
1012     _number_of_mh_late_inlines(0),
1013     _print_inlining_stream(NULL),
1014     _print_inlining_list(NULL),
1015     _print_inlining_idx(0),
1016     _print_inlining_output(NULL),
1017     _replay_inline_data(NULL),
1018     _java_calls(0),
1019     _inner_loops(0),
1020     _interpreter_frame_size(0),
1021     _node_bundling_limit(0),
1022     _node_bundling_base(NULL),
1023     _code_buffer("Compile::Fill_buffer"),
1024 #ifndef PRODUCT
1025     _in_dump_cnt(0),
1026 #endif
1027     _allowed_reasons(0) {
1028   C = this;
1029 
1030   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1031   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1032 
1033 #ifndef PRODUCT
1034   set_print_assembly(PrintFrameConverterAssembly);
1035   set_parsed_irreducible_loop(false);
1036 #else
1037   set_print_assembly(false); // Must initialize.
1038 #endif
1039   set_has_irreducible_loop(false); // no loops
1040 
1041   CompileWrapper cw(this);
1042   Init(/*AliasLevel=*/ 0);
1043   init_tf((*generator)());
1044 
1045   {
1046     // The following is a dummy for the sake of GraphKit::gen_stub
1047     Unique_Node_List for_igvn(comp_arena());
1048     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1049     PhaseGVN gvn(Thread::current()->resource_area(),255);
1050     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1051     gvn.transform_no_reclaim(top());
1052 
1053     GraphKit kit;
1054     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1055   }
1056 
1057   NOT_PRODUCT( verify_graph_edges(); )
1058   Code_Gen();
1059   if (failing())  return;
1060 
1061 
1062   // Entry point will be accessed using compile->stub_entry_point();
1063   if (code_buffer() == NULL) {
1064     Matcher::soft_match_failure();
1065   } else {
1066     if (PrintAssembly && (WizardMode || Verbose))
1067       tty->print_cr("### Stub::%s", stub_name);
1068 
1069     if (!failing()) {
1070       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1071 
1072       // Make the NMethod
1073       // For now we mark the frame as never safe for profile stackwalking
1074       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1075                                                       code_buffer(),
1076                                                       CodeOffsets::frame_never_safe,
1077                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1078                                                       frame_size_in_words(),
1079                                                       _oop_map_set,
1080                                                       save_arg_registers);
1081       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1082 
1083       _stub_entry_point = rs->entry_point();
1084     }
1085   }
1086 }
1087 
1088 //------------------------------Init-------------------------------------------
1089 // Prepare for a single compilation
1090 void Compile::Init(int aliaslevel) {
1091   _unique  = 0;
1092   _regalloc = NULL;
1093 
1094   _tf      = NULL;  // filled in later
1095   _top     = NULL;  // cached later
1096   _matcher = NULL;  // filled in later
1097   _cfg     = NULL;  // filled in later
1098 
1099   set_24_bit_selection_and_mode(Use24BitFP, false);
1100 
1101   _node_note_array = NULL;
1102   _default_node_notes = NULL;
1103   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1104 
1105   _immutable_memory = NULL; // filled in at first inquiry
1106 
1107   // Globally visible Nodes
1108   // First set TOP to NULL to give safe behavior during creation of RootNode
1109   set_cached_top_node(NULL);
1110   set_root(new RootNode());
1111   // Now that you have a Root to point to, create the real TOP
1112   set_cached_top_node( new ConNode(Type::TOP) );
1113   set_recent_alloc(NULL, NULL);
1114 
1115   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1116   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1117   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1118   env()->set_dependencies(new Dependencies(env()));
1119 
1120   _fixed_slots = 0;
1121   set_has_split_ifs(false);
1122   set_has_loops(has_method() && method()->has_loops()); // first approximation
1123   set_has_stringbuilder(false);
1124   set_has_boxed_value(false);
1125   _trap_can_recompile = false;  // no traps emitted yet
1126   _major_progress = true; // start out assuming good things will happen
1127   set_has_unsafe_access(false);
1128   set_max_vector_size(0);
1129   set_clear_upper_avx(false);  //false as default for clear upper bits of ymm registers
1130   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1131   set_decompile_count(0);
1132 
1133   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1134   _loop_opts_cnt = LoopOptsCount;
1135   set_do_inlining(Inline);
1136   set_max_inline_size(MaxInlineSize);
1137   set_freq_inline_size(FreqInlineSize);
1138   set_do_scheduling(OptoScheduling);
1139   set_do_count_invocations(false);
1140   set_do_method_data_update(false);
1141 
1142   set_do_vector_loop(false);
1143   set_has_monitors(false);
1144 
1145   if (AllowVectorizeOnDemand) {
1146     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1147       set_do_vector_loop(true);
1148       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());})
1149     } else if (has_method() && method()->name() != 0 &&
1150                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1151       set_do_vector_loop(true);
1152     }
1153   }
1154   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1155   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());})
1156 
1157   set_age_code(has_method() && method()->profile_aging());
1158   set_rtm_state(NoRTM); // No RTM lock eliding by default
1159   _max_node_limit = _directive->MaxNodeLimitOption;
1160 
1161 #if INCLUDE_RTM_OPT
1162   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1163     int rtm_state = method()->method_data()->rtm_state();
1164     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1165       // Don't generate RTM lock eliding code.
1166       set_rtm_state(NoRTM);
1167     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1168       // Generate RTM lock eliding code without abort ratio calculation code.
1169       set_rtm_state(UseRTM);
1170     } else if (UseRTMDeopt) {
1171       // Generate RTM lock eliding code and include abort ratio calculation
1172       // code if UseRTMDeopt is on.
1173       set_rtm_state(ProfileRTM);
1174     }
1175   }
1176 #endif
1177   if (VM_Version::supports_fast_class_init_checks() && has_method() && !is_osr_compilation() && method()->needs_clinit_barrier()) {
1178     set_clinit_barrier_on_entry(true);
1179   }
1180   if (debug_info()->recording_non_safepoints()) {
1181     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1182                         (comp_arena(), 8, 0, NULL));
1183     set_default_node_notes(Node_Notes::make(this));
1184   }
1185 
1186   // // -- Initialize types before each compile --
1187   // // Update cached type information
1188   // if( _method && _method->constants() )
1189   //   Type::update_loaded_types(_method, _method->constants());
1190 
1191   // Init alias_type map.
1192   if (!_do_escape_analysis && aliaslevel == 3)
1193     aliaslevel = 2;  // No unique types without escape analysis
1194   _AliasLevel = aliaslevel;
1195   const int grow_ats = 16;
1196   _max_alias_types = grow_ats;
1197   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1198   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1199   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1200   {
1201     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1202   }
1203   // Initialize the first few types.
1204   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1205   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1206   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1207   _num_alias_types = AliasIdxRaw+1;
1208   // Zero out the alias type cache.
1209   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1210   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1211   probe_alias_cache(NULL)->_index = AliasIdxTop;
1212 
1213   _intrinsics = NULL;
1214   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1215   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1216   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1217   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1218   _opaque4_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1219   register_library_intrinsics();
1220 }
1221 
1222 //---------------------------init_start----------------------------------------
1223 // Install the StartNode on this compile object.
1224 void Compile::init_start(StartNode* s) {
1225   if (failing())
1226     return; // already failing
1227   assert(s == start(), "");
1228 }
1229 
1230 /**
1231  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1232  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1233  * the ideal graph.
1234  */
1235 StartNode* Compile::start() const {
1236   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1237   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1238     Node* start = root()->fast_out(i);
1239     if (start->is_Start()) {
1240       return start->as_Start();
1241     }
1242   }
1243   fatal("Did not find Start node!");
1244   return NULL;
1245 }
1246 
1247 //-------------------------------immutable_memory-------------------------------------
1248 // Access immutable memory
1249 Node* Compile::immutable_memory() {
1250   if (_immutable_memory != NULL) {
1251     return _immutable_memory;
1252   }
1253   StartNode* s = start();
1254   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1255     Node *p = s->fast_out(i);
1256     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1257       _immutable_memory = p;
1258       return _immutable_memory;
1259     }
1260   }
1261   ShouldNotReachHere();
1262   return NULL;
1263 }
1264 
1265 //----------------------set_cached_top_node------------------------------------
1266 // Install the cached top node, and make sure Node::is_top works correctly.
1267 void Compile::set_cached_top_node(Node* tn) {
1268   if (tn != NULL)  verify_top(tn);
1269   Node* old_top = _top;
1270   _top = tn;
1271   // Calling Node::setup_is_top allows the nodes the chance to adjust
1272   // their _out arrays.
1273   if (_top != NULL)     _top->setup_is_top();
1274   if (old_top != NULL)  old_top->setup_is_top();
1275   assert(_top == NULL || top()->is_top(), "");
1276 }
1277 
1278 #ifdef ASSERT
1279 uint Compile::count_live_nodes_by_graph_walk() {
1280   Unique_Node_List useful(comp_arena());
1281   // Get useful node list by walking the graph.
1282   identify_useful_nodes(useful);
1283   return useful.size();
1284 }
1285 
1286 void Compile::print_missing_nodes() {
1287 
1288   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1289   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1290     return;
1291   }
1292 
1293   // This is an expensive function. It is executed only when the user
1294   // specifies VerifyIdealNodeCount option or otherwise knows the
1295   // additional work that needs to be done to identify reachable nodes
1296   // by walking the flow graph and find the missing ones using
1297   // _dead_node_list.
1298 
1299   Unique_Node_List useful(comp_arena());
1300   // Get useful node list by walking the graph.
1301   identify_useful_nodes(useful);
1302 
1303   uint l_nodes = C->live_nodes();
1304   uint l_nodes_by_walk = useful.size();
1305 
1306   if (l_nodes != l_nodes_by_walk) {
1307     if (_log != NULL) {
1308       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1309       _log->stamp();
1310       _log->end_head();
1311     }
1312     VectorSet& useful_member_set = useful.member_set();
1313     int last_idx = l_nodes_by_walk;
1314     for (int i = 0; i < last_idx; i++) {
1315       if (useful_member_set.test(i)) {
1316         if (_dead_node_list.test(i)) {
1317           if (_log != NULL) {
1318             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1319           }
1320           if (PrintIdealNodeCount) {
1321             // Print the log message to tty
1322               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1323               useful.at(i)->dump();
1324           }
1325         }
1326       }
1327       else if (! _dead_node_list.test(i)) {
1328         if (_log != NULL) {
1329           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1330         }
1331         if (PrintIdealNodeCount) {
1332           // Print the log message to tty
1333           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1334         }
1335       }
1336     }
1337     if (_log != NULL) {
1338       _log->tail("mismatched_nodes");
1339     }
1340   }
1341 }
1342 void Compile::record_modified_node(Node* n) {
1343   if (_modified_nodes != NULL && !_inlining_incrementally &&
1344       n->outcnt() != 0 && !n->is_Con()) {
1345     _modified_nodes->push(n);
1346   }
1347 }
1348 
1349 void Compile::remove_modified_node(Node* n) {
1350   if (_modified_nodes != NULL) {
1351     _modified_nodes->remove(n);
1352   }
1353 }
1354 #endif
1355 
1356 #ifndef PRODUCT
1357 void Compile::verify_top(Node* tn) const {
1358   if (tn != NULL) {
1359     assert(tn->is_Con(), "top node must be a constant");
1360     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1361     assert(tn->in(0) != NULL, "must have live top node");
1362   }
1363 }
1364 #endif
1365 
1366 
1367 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1368 
1369 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1370   guarantee(arr != NULL, "");
1371   int num_blocks = arr->length();
1372   if (grow_by < num_blocks)  grow_by = num_blocks;
1373   int num_notes = grow_by * _node_notes_block_size;
1374   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1375   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1376   while (num_notes > 0) {
1377     arr->append(notes);
1378     notes     += _node_notes_block_size;
1379     num_notes -= _node_notes_block_size;
1380   }
1381   assert(num_notes == 0, "exact multiple, please");
1382 }
1383 
1384 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1385   if (source == NULL || dest == NULL)  return false;
1386 
1387   if (dest->is_Con())
1388     return false;               // Do not push debug info onto constants.
1389 
1390 #ifdef ASSERT
1391   // Leave a bread crumb trail pointing to the original node:
1392   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1393     dest->set_debug_orig(source);
1394   }
1395 #endif
1396 
1397   if (node_note_array() == NULL)
1398     return false;               // Not collecting any notes now.
1399 
1400   // This is a copy onto a pre-existing node, which may already have notes.
1401   // If both nodes have notes, do not overwrite any pre-existing notes.
1402   Node_Notes* source_notes = node_notes_at(source->_idx);
1403   if (source_notes == NULL || source_notes->is_clear())  return false;
1404   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1405   if (dest_notes == NULL || dest_notes->is_clear()) {
1406     return set_node_notes_at(dest->_idx, source_notes);
1407   }
1408 
1409   Node_Notes merged_notes = (*source_notes);
1410   // The order of operations here ensures that dest notes will win...
1411   merged_notes.update_from(dest_notes);
1412   return set_node_notes_at(dest->_idx, &merged_notes);
1413 }
1414 
1415 
1416 //--------------------------allow_range_check_smearing-------------------------
1417 // Gating condition for coalescing similar range checks.
1418 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1419 // single covering check that is at least as strong as any of them.
1420 // If the optimization succeeds, the simplified (strengthened) range check
1421 // will always succeed.  If it fails, we will deopt, and then give up
1422 // on the optimization.
1423 bool Compile::allow_range_check_smearing() const {
1424   // If this method has already thrown a range-check,
1425   // assume it was because we already tried range smearing
1426   // and it failed.
1427   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1428   return !already_trapped;
1429 }
1430 
1431 
1432 //------------------------------flatten_alias_type-----------------------------
1433 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1434   int offset = tj->offset();
1435   TypePtr::PTR ptr = tj->ptr();
1436 
1437   // Known instance (scalarizable allocation) alias only with itself.
1438   bool is_known_inst = tj->isa_oopptr() != NULL &&
1439                        tj->is_oopptr()->is_known_instance();
1440 
1441   // Process weird unsafe references.
1442   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1443     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1444     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1445     tj = TypeOopPtr::BOTTOM;
1446     ptr = tj->ptr();
1447     offset = tj->offset();
1448   }
1449 
1450   // Array pointers need some flattening
1451   const TypeAryPtr *ta = tj->isa_aryptr();
1452   if (ta && ta->is_stable()) {
1453     // Erase stability property for alias analysis.
1454     tj = ta = ta->cast_to_stable(false);
1455   }
1456   if( ta && is_known_inst ) {
1457     if ( offset != Type::OffsetBot &&
1458          offset > arrayOopDesc::length_offset_in_bytes() ) {
1459       offset = Type::OffsetBot; // Flatten constant access into array body only
1460       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1461     }
1462   } else if( ta && _AliasLevel >= 2 ) {
1463     // For arrays indexed by constant indices, we flatten the alias
1464     // space to include all of the array body.  Only the header, klass
1465     // and array length can be accessed un-aliased.
1466     if( offset != Type::OffsetBot ) {
1467       if( ta->const_oop() ) { // MethodData* or Method*
1468         offset = Type::OffsetBot;   // Flatten constant access into array body
1469         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1470       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1471         // range is OK as-is.
1472         tj = ta = TypeAryPtr::RANGE;
1473       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1474         tj = TypeInstPtr::KLASS; // all klass loads look alike
1475         ta = TypeAryPtr::RANGE; // generic ignored junk
1476         ptr = TypePtr::BotPTR;
1477       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1478         tj = TypeInstPtr::MARK;
1479         ta = TypeAryPtr::RANGE; // generic ignored junk
1480         ptr = TypePtr::BotPTR;
1481       } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1482         ta = tj->isa_aryptr();
1483       } else {                  // Random constant offset into array body
1484         offset = Type::OffsetBot;   // Flatten constant access into array body
1485         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1486       }
1487     }
1488     // Arrays of fixed size alias with arrays of unknown size.
1489     if (ta->size() != TypeInt::POS) {
1490       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1491       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1492     }
1493     // Arrays of known objects become arrays of unknown objects.
1494     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1495       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1496       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1497     }
1498     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1499       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1500       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1501     }
1502     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1503     // cannot be distinguished by bytecode alone.
1504     if (ta->elem() == TypeInt::BOOL) {
1505       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1506       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1507       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1508     }
1509     // During the 2nd round of IterGVN, NotNull castings are removed.
1510     // Make sure the Bottom and NotNull variants alias the same.
1511     // Also, make sure exact and non-exact variants alias the same.
1512     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1513       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1514     }
1515   }
1516 
1517   // Oop pointers need some flattening
1518   const TypeInstPtr *to = tj->isa_instptr();
1519   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1520     ciInstanceKlass *k = to->klass()->as_instance_klass();
1521     if( ptr == TypePtr::Constant ) {
1522       if (to->klass() != ciEnv::current()->Class_klass() ||
1523           offset < k->size_helper() * wordSize) {
1524         // No constant oop pointers (such as Strings); they alias with
1525         // unknown strings.
1526         assert(!is_known_inst, "not scalarizable allocation");
1527         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1528       }
1529     } else if( is_known_inst ) {
1530       tj = to; // Keep NotNull and klass_is_exact for instance type
1531     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1532       // During the 2nd round of IterGVN, NotNull castings are removed.
1533       // Make sure the Bottom and NotNull variants alias the same.
1534       // Also, make sure exact and non-exact variants alias the same.
1535       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1536     }
1537     if (to->speculative() != NULL) {
1538       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1539     }
1540     // Canonicalize the holder of this field
1541     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1542       // First handle header references such as a LoadKlassNode, even if the
1543       // object's klass is unloaded at compile time (4965979).
1544       if (!is_known_inst) { // Do it only for non-instance types
1545         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1546       }
1547     } else if (BarrierSet::barrier_set()->barrier_set_c2()->flatten_gc_alias_type(tj)) {
1548       to = tj->is_instptr();
1549     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1550       // Static fields are in the space above the normal instance
1551       // fields in the java.lang.Class instance.
1552       if (to->klass() != ciEnv::current()->Class_klass()) {
1553         to = NULL;
1554         tj = TypeOopPtr::BOTTOM;
1555         offset = tj->offset();
1556       }
1557     } else {
1558       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1559       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1560         if( is_known_inst ) {
1561           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1562         } else {
1563           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1564         }
1565       }
1566     }
1567   }
1568 
1569   // Klass pointers to object array klasses need some flattening
1570   const TypeKlassPtr *tk = tj->isa_klassptr();
1571   if( tk ) {
1572     // If we are referencing a field within a Klass, we need
1573     // to assume the worst case of an Object.  Both exact and
1574     // inexact types must flatten to the same alias class so
1575     // use NotNull as the PTR.
1576     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1577 
1578       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1579                                    TypeKlassPtr::OBJECT->klass(),
1580                                    offset);
1581     }
1582 
1583     ciKlass* klass = tk->klass();
1584     if( klass->is_obj_array_klass() ) {
1585       ciKlass* k = TypeAryPtr::OOPS->klass();
1586       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1587         k = TypeInstPtr::BOTTOM->klass();
1588       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1589     }
1590 
1591     // Check for precise loads from the primary supertype array and force them
1592     // to the supertype cache alias index.  Check for generic array loads from
1593     // the primary supertype array and also force them to the supertype cache
1594     // alias index.  Since the same load can reach both, we need to merge
1595     // these 2 disparate memories into the same alias class.  Since the
1596     // primary supertype array is read-only, there's no chance of confusion
1597     // where we bypass an array load and an array store.
1598     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1599     if (offset == Type::OffsetBot ||
1600         (offset >= primary_supers_offset &&
1601          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1602         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1603       offset = in_bytes(Klass::secondary_super_cache_offset());
1604       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1605     }
1606   }
1607 
1608   // Flatten all Raw pointers together.
1609   if (tj->base() == Type::RawPtr)
1610     tj = TypeRawPtr::BOTTOM;
1611 
1612   if (tj->base() == Type::AnyPtr)
1613     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1614 
1615   // Flatten all to bottom for now
1616   switch( _AliasLevel ) {
1617   case 0:
1618     tj = TypePtr::BOTTOM;
1619     break;
1620   case 1:                       // Flatten to: oop, static, field or array
1621     switch (tj->base()) {
1622     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1623     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1624     case Type::AryPtr:   // do not distinguish arrays at all
1625     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1626     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1627     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1628     default: ShouldNotReachHere();
1629     }
1630     break;
1631   case 2:                       // No collapsing at level 2; keep all splits
1632   case 3:                       // No collapsing at level 3; keep all splits
1633     break;
1634   default:
1635     Unimplemented();
1636   }
1637 
1638   offset = tj->offset();
1639   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1640 
1641   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1642           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1643           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1644           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1645           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1646           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1647           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1648           (BarrierSet::barrier_set()->barrier_set_c2()->verify_gc_alias_type(tj, offset)),
1649           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1650   assert( tj->ptr() != TypePtr::TopPTR &&
1651           tj->ptr() != TypePtr::AnyNull &&
1652           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1653 //    assert( tj->ptr() != TypePtr::Constant ||
1654 //            tj->base() == Type::RawPtr ||
1655 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1656 
1657   return tj;
1658 }
1659 
1660 void Compile::AliasType::Init(int i, const TypePtr* at) {
1661   _index = i;
1662   _adr_type = at;
1663   _field = NULL;
1664   _element = NULL;
1665   _is_rewritable = true; // default
1666   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1667   if (atoop != NULL && atoop->is_known_instance()) {
1668     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1669     _general_index = Compile::current()->get_alias_index(gt);
1670   } else {
1671     _general_index = 0;
1672   }
1673 }
1674 
1675 BasicType Compile::AliasType::basic_type() const {
1676   if (element() != NULL) {
1677     const Type* element = adr_type()->is_aryptr()->elem();
1678     return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1679   } if (field() != NULL) {
1680     return field()->layout_type();
1681   } else {
1682     return T_ILLEGAL; // unknown
1683   }
1684 }
1685 
1686 //---------------------------------print_on------------------------------------
1687 #ifndef PRODUCT
1688 void Compile::AliasType::print_on(outputStream* st) {
1689   if (index() < 10)
1690         st->print("@ <%d> ", index());
1691   else  st->print("@ <%d>",  index());
1692   st->print(is_rewritable() ? "   " : " RO");
1693   int offset = adr_type()->offset();
1694   if (offset == Type::OffsetBot)
1695         st->print(" +any");
1696   else  st->print(" +%-3d", offset);
1697   st->print(" in ");
1698   adr_type()->dump_on(st);
1699   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1700   if (field() != NULL && tjp) {
1701     if (tjp->klass()  != field()->holder() ||
1702         tjp->offset() != field()->offset_in_bytes()) {
1703       st->print(" != ");
1704       field()->print();
1705       st->print(" ***");
1706     }
1707   }
1708 }
1709 
1710 void print_alias_types() {
1711   Compile* C = Compile::current();
1712   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1713   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1714     C->alias_type(idx)->print_on(tty);
1715     tty->cr();
1716   }
1717 }
1718 #endif
1719 
1720 
1721 //----------------------------probe_alias_cache--------------------------------
1722 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1723   intptr_t key = (intptr_t) adr_type;
1724   key ^= key >> logAliasCacheSize;
1725   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1726 }
1727 
1728 
1729 //-----------------------------grow_alias_types--------------------------------
1730 void Compile::grow_alias_types() {
1731   const int old_ats  = _max_alias_types; // how many before?
1732   const int new_ats  = old_ats;          // how many more?
1733   const int grow_ats = old_ats+new_ats;  // how many now?
1734   _max_alias_types = grow_ats;
1735   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1736   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1737   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1738   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1739 }
1740 
1741 
1742 //--------------------------------find_alias_type------------------------------
1743 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1744   if (_AliasLevel == 0)
1745     return alias_type(AliasIdxBot);
1746 
1747   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1748   if (ace->_adr_type == adr_type) {
1749     return alias_type(ace->_index);
1750   }
1751 
1752   // Handle special cases.
1753   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1754   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1755 
1756   // Do it the slow way.
1757   const TypePtr* flat = flatten_alias_type(adr_type);
1758 
1759 #ifdef ASSERT
1760   {
1761     ResourceMark rm;
1762     assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1763            Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1764     assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1765            Type::str(adr_type));
1766     if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1767       const TypeOopPtr* foop = flat->is_oopptr();
1768       // Scalarizable allocations have exact klass always.
1769       bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1770       const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1771       assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1772              Type::str(foop), Type::str(xoop));
1773     }
1774   }
1775 #endif
1776 
1777   int idx = AliasIdxTop;
1778   for (int i = 0; i < num_alias_types(); i++) {
1779     if (alias_type(i)->adr_type() == flat) {
1780       idx = i;
1781       break;
1782     }
1783   }
1784 
1785   if (idx == AliasIdxTop) {
1786     if (no_create)  return NULL;
1787     // Grow the array if necessary.
1788     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1789     // Add a new alias type.
1790     idx = _num_alias_types++;
1791     _alias_types[idx]->Init(idx, flat);
1792     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1793     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1794     if (flat->isa_instptr()) {
1795       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1796           && flat->is_instptr()->klass() == env()->Class_klass())
1797         alias_type(idx)->set_rewritable(false);
1798     }
1799     if (flat->isa_aryptr()) {
1800 #ifdef ASSERT
1801       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1802       // (T_BYTE has the weakest alignment and size restrictions...)
1803       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1804 #endif
1805       if (flat->offset() == TypePtr::OffsetBot) {
1806         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1807       }
1808     }
1809     if (flat->isa_klassptr()) {
1810       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1811         alias_type(idx)->set_rewritable(false);
1812       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1813         alias_type(idx)->set_rewritable(false);
1814       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1815         alias_type(idx)->set_rewritable(false);
1816       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1817         alias_type(idx)->set_rewritable(false);
1818     }
1819     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1820     // but the base pointer type is not distinctive enough to identify
1821     // references into JavaThread.)
1822 
1823     // Check for final fields.
1824     const TypeInstPtr* tinst = flat->isa_instptr();
1825     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1826       ciField* field;
1827       if (tinst->const_oop() != NULL &&
1828           tinst->klass() == ciEnv::current()->Class_klass() &&
1829           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1830         // static field
1831         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1832         field = k->get_field_by_offset(tinst->offset(), true);
1833       } else {
1834         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1835         field = k->get_field_by_offset(tinst->offset(), false);
1836       }
1837       assert(field == NULL ||
1838              original_field == NULL ||
1839              (field->holder() == original_field->holder() &&
1840               field->offset() == original_field->offset() &&
1841               field->is_static() == original_field->is_static()), "wrong field?");
1842       // Set field() and is_rewritable() attributes.
1843       if (field != NULL)  alias_type(idx)->set_field(field);
1844     }
1845   }
1846 
1847   // Fill the cache for next time.
1848   ace->_adr_type = adr_type;
1849   ace->_index    = idx;
1850   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1851 
1852   // Might as well try to fill the cache for the flattened version, too.
1853   AliasCacheEntry* face = probe_alias_cache(flat);
1854   if (face->_adr_type == NULL) {
1855     face->_adr_type = flat;
1856     face->_index    = idx;
1857     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1858   }
1859 
1860   return alias_type(idx);
1861 }
1862 
1863 
1864 Compile::AliasType* Compile::alias_type(ciField* field) {
1865   const TypeOopPtr* t;
1866   if (field->is_static())
1867     t = TypeInstPtr::make(field->holder()->java_mirror());
1868   else
1869     t = TypeOopPtr::make_from_klass_raw(field->holder());
1870   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1871   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1872   return atp;
1873 }
1874 
1875 
1876 //------------------------------have_alias_type--------------------------------
1877 bool Compile::have_alias_type(const TypePtr* adr_type) {
1878   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1879   if (ace->_adr_type == adr_type) {
1880     return true;
1881   }
1882 
1883   // Handle special cases.
1884   if (adr_type == NULL)             return true;
1885   if (adr_type == TypePtr::BOTTOM)  return true;
1886 
1887   return find_alias_type(adr_type, true, NULL) != NULL;
1888 }
1889 
1890 //-----------------------------must_alias--------------------------------------
1891 // True if all values of the given address type are in the given alias category.
1892 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1893   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1894   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1895   if (alias_idx == AliasIdxTop)         return false; // the empty category
1896   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1897 
1898   // the only remaining possible overlap is identity
1899   int adr_idx = get_alias_index(adr_type);
1900   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1901   assert(adr_idx == alias_idx ||
1902          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1903           && adr_type                       != TypeOopPtr::BOTTOM),
1904          "should not be testing for overlap with an unsafe pointer");
1905   return adr_idx == alias_idx;
1906 }
1907 
1908 //------------------------------can_alias--------------------------------------
1909 // True if any values of the given address type are in the given alias category.
1910 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1911   if (alias_idx == AliasIdxTop)         return false; // the empty category
1912   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1913   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1914   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1915 
1916   // the only remaining possible overlap is identity
1917   int adr_idx = get_alias_index(adr_type);
1918   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1919   return adr_idx == alias_idx;
1920 }
1921 
1922 
1923 
1924 //---------------------------pop_warm_call-------------------------------------
1925 WarmCallInfo* Compile::pop_warm_call() {
1926   WarmCallInfo* wci = _warm_calls;
1927   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1928   return wci;
1929 }
1930 
1931 //----------------------------Inline_Warm--------------------------------------
1932 int Compile::Inline_Warm() {
1933   // If there is room, try to inline some more warm call sites.
1934   // %%% Do a graph index compaction pass when we think we're out of space?
1935   if (!InlineWarmCalls)  return 0;
1936 
1937   int calls_made_hot = 0;
1938   int room_to_grow   = NodeCountInliningCutoff - unique();
1939   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1940   int amount_grown   = 0;
1941   WarmCallInfo* call;
1942   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1943     int est_size = (int)call->size();
1944     if (est_size > (room_to_grow - amount_grown)) {
1945       // This one won't fit anyway.  Get rid of it.
1946       call->make_cold();
1947       continue;
1948     }
1949     call->make_hot();
1950     calls_made_hot++;
1951     amount_grown   += est_size;
1952     amount_to_grow -= est_size;
1953   }
1954 
1955   if (calls_made_hot > 0)  set_major_progress();
1956   return calls_made_hot;
1957 }
1958 
1959 
1960 //----------------------------Finish_Warm--------------------------------------
1961 void Compile::Finish_Warm() {
1962   if (!InlineWarmCalls)  return;
1963   if (failing())  return;
1964   if (warm_calls() == NULL)  return;
1965 
1966   // Clean up loose ends, if we are out of space for inlining.
1967   WarmCallInfo* call;
1968   while ((call = pop_warm_call()) != NULL) {
1969     call->make_cold();
1970   }
1971 }
1972 
1973 //---------------------cleanup_loop_predicates-----------------------
1974 // Remove the opaque nodes that protect the predicates so that all unused
1975 // checks and uncommon_traps will be eliminated from the ideal graph
1976 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1977   if (predicate_count()==0) return;
1978   for (int i = predicate_count(); i > 0; i--) {
1979     Node * n = predicate_opaque1_node(i-1);
1980     assert(n->Opcode() == Op_Opaque1, "must be");
1981     igvn.replace_node(n, n->in(1));
1982   }
1983   assert(predicate_count()==0, "should be clean!");
1984 }
1985 
1986 void Compile::add_range_check_cast(Node* n) {
1987   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1988   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1989   _range_check_casts->append(n);
1990 }
1991 
1992 // Remove all range check dependent CastIINodes.
1993 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1994   for (int i = range_check_cast_count(); i > 0; i--) {
1995     Node* cast = range_check_cast_node(i-1);
1996     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1997     igvn.replace_node(cast, cast->in(1));
1998   }
1999   assert(range_check_cast_count() == 0, "should be empty");
2000 }
2001 
2002 void Compile::add_opaque4_node(Node* n) {
2003   assert(n->Opcode() == Op_Opaque4, "Opaque4 only");
2004   assert(!_opaque4_nodes->contains(n), "duplicate entry in Opaque4 list");
2005   _opaque4_nodes->append(n);
2006 }
2007 
2008 // Remove all Opaque4 nodes.
2009 void Compile::remove_opaque4_nodes(PhaseIterGVN &igvn) {
2010   for (int i = opaque4_count(); i > 0; i--) {
2011     Node* opaq = opaque4_node(i-1);
2012     assert(opaq->Opcode() == Op_Opaque4, "Opaque4 only");
2013     igvn.replace_node(opaq, opaq->in(2));
2014   }
2015   assert(opaque4_count() == 0, "should be empty");
2016 }
2017 
2018 // StringOpts and late inlining of string methods
2019 void Compile::inline_string_calls(bool parse_time) {
2020   {
2021     // remove useless nodes to make the usage analysis simpler
2022     ResourceMark rm;
2023     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2024   }
2025 
2026   {
2027     ResourceMark rm;
2028     print_method(PHASE_BEFORE_STRINGOPTS, 3);
2029     PhaseStringOpts pso(initial_gvn(), for_igvn());
2030     print_method(PHASE_AFTER_STRINGOPTS, 3);
2031   }
2032 
2033   // now inline anything that we skipped the first time around
2034   if (!parse_time) {
2035     _late_inlines_pos = _late_inlines.length();
2036   }
2037 
2038   while (_string_late_inlines.length() > 0) {
2039     CallGenerator* cg = _string_late_inlines.pop();
2040     cg->do_late_inline();
2041     if (failing())  return;
2042   }
2043   _string_late_inlines.trunc_to(0);
2044 }
2045 
2046 // Late inlining of boxing methods
2047 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
2048   if (_boxing_late_inlines.length() > 0) {
2049     assert(has_boxed_value(), "inconsistent");
2050 
2051     PhaseGVN* gvn = initial_gvn();
2052     set_inlining_incrementally(true);
2053 
2054     assert( igvn._worklist.size() == 0, "should be done with igvn" );
2055     for_igvn()->clear();
2056     gvn->replace_with(&igvn);
2057 
2058     _late_inlines_pos = _late_inlines.length();
2059 
2060     while (_boxing_late_inlines.length() > 0) {
2061       CallGenerator* cg = _boxing_late_inlines.pop();
2062       cg->do_late_inline();
2063       if (failing())  return;
2064     }
2065     _boxing_late_inlines.trunc_to(0);
2066 
2067     inline_incrementally_cleanup(igvn);
2068 
2069     set_inlining_incrementally(false);
2070   }
2071 }
2072 
2073 bool Compile::inline_incrementally_one() {
2074   assert(IncrementalInline, "incremental inlining should be on");
2075 
2076   TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2077   set_inlining_progress(false);
2078   set_do_cleanup(false);
2079   int i = 0;
2080   for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2081     CallGenerator* cg = _late_inlines.at(i);
2082     _late_inlines_pos = i+1;
2083     cg->do_late_inline();
2084     if (failing())  return false;
2085   }
2086   int j = 0;
2087   for (; i < _late_inlines.length(); i++, j++) {
2088     _late_inlines.at_put(j, _late_inlines.at(i));
2089   }
2090   _late_inlines.trunc_to(j);
2091   assert(inlining_progress() || _late_inlines.length() == 0, "");
2092 
2093   bool needs_cleanup = do_cleanup() || over_inlining_cutoff();
2094 
2095   set_inlining_progress(false);
2096   set_do_cleanup(false);
2097   return (_late_inlines.length() > 0) && !needs_cleanup;
2098 }
2099 
2100 void Compile::inline_incrementally_cleanup(PhaseIterGVN& igvn) {
2101   {
2102     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2103     ResourceMark rm;
2104     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2105   }
2106   {
2107     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2108     igvn = PhaseIterGVN(initial_gvn());
2109     igvn.optimize();
2110   }
2111 }
2112 
2113 // Perform incremental inlining until bound on number of live nodes is reached
2114 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2115   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2116 
2117   set_inlining_incrementally(true);
2118   uint low_live_nodes = 0;
2119 
2120   while (_late_inlines.length() > 0) {
2121     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2122       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2123         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2124         // PhaseIdealLoop is expensive so we only try it once we are
2125         // out of live nodes and we only try it again if the previous
2126         // helped got the number of nodes down significantly
2127         PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2128         if (failing())  return;
2129         low_live_nodes = live_nodes();
2130         _major_progress = true;
2131       }
2132 
2133       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2134         break; // finish
2135       }
2136     }
2137 
2138     for_igvn()->clear();
2139     initial_gvn()->replace_with(&igvn);
2140 
2141     while (inline_incrementally_one()) {
2142       assert(!failing(), "inconsistent");
2143     }
2144 
2145     if (failing())  return;
2146 
2147     inline_incrementally_cleanup(igvn);
2148 
2149     if (failing())  return;
2150   }
2151   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2152 
2153   if (_string_late_inlines.length() > 0) {
2154     assert(has_stringbuilder(), "inconsistent");
2155     for_igvn()->clear();
2156     initial_gvn()->replace_with(&igvn);
2157 
2158     inline_string_calls(false);
2159 
2160     if (failing())  return;
2161 
2162     inline_incrementally_cleanup(igvn);
2163   }
2164 
2165   set_inlining_incrementally(false);
2166 }
2167 
2168 
2169 bool Compile::optimize_loops(PhaseIterGVN& igvn, LoopOptsMode mode) {
2170   if(_loop_opts_cnt > 0) {
2171     debug_only( int cnt = 0; );
2172     while(major_progress() && (_loop_opts_cnt > 0)) {
2173       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2174       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2175       PhaseIdealLoop::optimize(igvn, mode);
2176       _loop_opts_cnt--;
2177       if (failing())  return false;
2178       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2179     }
2180   }
2181   return true;
2182 }
2183 
2184 // Remove edges from "root" to each SafePoint at a backward branch.
2185 // They were inserted during parsing (see add_safepoint()) to make
2186 // infinite loops without calls or exceptions visible to root, i.e.,
2187 // useful.
2188 void Compile::remove_root_to_sfpts_edges(PhaseIterGVN& igvn) {
2189   Node *r = root();
2190   if (r != NULL) {
2191     for (uint i = r->req(); i < r->len(); ++i) {
2192       Node *n = r->in(i);
2193       if (n != NULL && n->is_SafePoint()) {
2194         r->rm_prec(i);
2195         if (n->outcnt() == 0) {
2196           igvn.remove_dead_node(n);
2197         }
2198         --i;
2199       }
2200     }
2201   }
2202 }
2203 
2204 //------------------------------Optimize---------------------------------------
2205 // Given a graph, optimize it.
2206 void Compile::Optimize() {
2207   TracePhase tp("optimizer", &timers[_t_optimizer]);
2208 
2209 #ifndef PRODUCT
2210   if (_directive->BreakAtCompileOption) {
2211     BREAKPOINT;
2212   }
2213 
2214 #endif
2215 
2216   BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2217 #ifdef ASSERT
2218   bs->verify_gc_barriers(this, BarrierSetC2::BeforeOptimize);
2219 #endif
2220 
2221   ResourceMark rm;
2222 
2223   print_inlining_reinit();
2224 
2225   NOT_PRODUCT( verify_graph_edges(); )
2226 
2227   print_method(PHASE_AFTER_PARSING);
2228 
2229  {
2230   // Iterative Global Value Numbering, including ideal transforms
2231   // Initialize IterGVN with types and values from parse-time GVN
2232   PhaseIterGVN igvn(initial_gvn());
2233 #ifdef ASSERT
2234   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2235 #endif
2236   {
2237     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2238     igvn.optimize();
2239   }
2240 
2241   if (failing())  return;
2242 
2243   print_method(PHASE_ITER_GVN1, 2);
2244 
2245   inline_incrementally(igvn);
2246 
2247   print_method(PHASE_INCREMENTAL_INLINE, 2);
2248 
2249   if (failing())  return;
2250 
2251   if (eliminate_boxing()) {
2252     // Inline valueOf() methods now.
2253     inline_boxing_calls(igvn);
2254 
2255     if (AlwaysIncrementalInline) {
2256       inline_incrementally(igvn);
2257     }
2258 
2259     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2260 
2261     if (failing())  return;
2262   }
2263 
2264   // Now that all inlining is over, cut edge from root to loop
2265   // safepoints
2266   remove_root_to_sfpts_edges(igvn);
2267 
2268   // Remove the speculative part of types and clean up the graph from
2269   // the extra CastPP nodes whose only purpose is to carry them. Do
2270   // that early so that optimizations are not disrupted by the extra
2271   // CastPP nodes.
2272   remove_speculative_types(igvn);
2273 
2274   // No more new expensive nodes will be added to the list from here
2275   // so keep only the actual candidates for optimizations.
2276   cleanup_expensive_nodes(igvn);
2277 
2278   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2279     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2280     initial_gvn()->replace_with(&igvn);
2281     for_igvn()->clear();
2282     Unique_Node_List new_worklist(C->comp_arena());
2283     {
2284       ResourceMark rm;
2285       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2286     }
2287     set_for_igvn(&new_worklist);
2288     igvn = PhaseIterGVN(initial_gvn());
2289     igvn.optimize();
2290   }
2291 
2292   // Perform escape analysis
2293   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2294     if (has_loops()) {
2295       // Cleanup graph (remove dead nodes).
2296       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2297       PhaseIdealLoop::optimize(igvn, LoopOptsNone);
2298       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2299       if (failing())  return;
2300     }
2301     ConnectionGraph::do_analysis(this, &igvn);
2302 
2303     if (failing())  return;
2304 
2305     // Optimize out fields loads from scalar replaceable allocations.
2306     igvn.optimize();
2307     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2308 
2309     if (failing())  return;
2310 
2311     if (congraph() != NULL && macro_count() > 0) {
2312       TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2313       PhaseMacroExpand mexp(igvn);
2314       mexp.eliminate_macro_nodes();
2315       igvn.set_delay_transform(false);
2316 
2317       igvn.optimize();
2318       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2319 
2320       if (failing())  return;
2321     }
2322   }
2323 
2324   // Loop transforms on the ideal graph.  Range Check Elimination,
2325   // peeling, unrolling, etc.
2326 
2327   // Set loop opts counter
2328   if((_loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2329     {
2330       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2331       PhaseIdealLoop::optimize(igvn, LoopOptsDefault);
2332       _loop_opts_cnt--;
2333       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2334       if (failing())  return;
2335     }
2336     // Loop opts pass if partial peeling occurred in previous pass
2337     if(PartialPeelLoop && major_progress() && (_loop_opts_cnt > 0)) {
2338       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2339       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2340       _loop_opts_cnt--;
2341       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2342       if (failing())  return;
2343     }
2344     // Loop opts pass for loop-unrolling before CCP
2345     if(major_progress() && (_loop_opts_cnt > 0)) {
2346       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2347       PhaseIdealLoop::optimize(igvn, LoopOptsSkipSplitIf);
2348       _loop_opts_cnt--;
2349       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2350     }
2351     if (!failing()) {
2352       // Verify that last round of loop opts produced a valid graph
2353       TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2354       PhaseIdealLoop::verify(igvn);
2355     }
2356   }
2357   if (failing())  return;
2358 
2359   // Conditional Constant Propagation;
2360   PhaseCCP ccp( &igvn );
2361   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2362   {
2363     TracePhase tp("ccp", &timers[_t_ccp]);
2364     ccp.do_transform();
2365   }
2366   print_method(PHASE_CPP1, 2);
2367 
2368   assert( true, "Break here to ccp.dump_old2new_map()");
2369 
2370   // Iterative Global Value Numbering, including ideal transforms
2371   {
2372     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2373     igvn = ccp;
2374     igvn.optimize();
2375   }
2376   print_method(PHASE_ITER_GVN2, 2);
2377 
2378   if (failing())  return;
2379 
2380   // Loop transforms on the ideal graph.  Range Check Elimination,
2381   // peeling, unrolling, etc.
2382   if (!optimize_loops(igvn, LoopOptsDefault)) {
2383     return;
2384   }
2385 
2386   if (failing())  return;
2387 
2388   // Ensure that major progress is now clear
2389   C->clear_major_progress();
2390 
2391   {
2392     // Verify that all previous optimizations produced a valid graph
2393     // at least to this point, even if no loop optimizations were done.
2394     TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2395     PhaseIdealLoop::verify(igvn);
2396   }
2397 
2398   if (range_check_cast_count() > 0) {
2399     // No more loop optimizations. Remove all range check dependent CastIINodes.
2400     C->remove_range_check_casts(igvn);
2401     igvn.optimize();
2402   }
2403 
2404 #ifdef ASSERT
2405   bs->verify_gc_barriers(this, BarrierSetC2::BeforeMacroExpand);
2406 #endif
2407 
2408   {
2409     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2410     PhaseMacroExpand  mex(igvn);
2411     if (mex.expand_macro_nodes()) {
2412       assert(failing(), "must bail out w/ explicit message");
2413       return;
2414     }
2415     print_method(PHASE_MACRO_EXPANSION, 2);
2416   }
2417 
2418 #ifdef ASSERT
2419   bs->verify_gc_barriers(this, BarrierSetC2::BeforeLateInsertion);
2420 #endif
2421 
2422   bs->barrier_insertion_phase(C, igvn);
2423   if (failing())  return;
2424 
2425   {
2426     TracePhase tp("barrierExpand", &timers[_t_barrierExpand]);
2427     if (bs->expand_barriers(this, igvn)) {
2428       assert(failing(), "must bail out w/ explicit message");
2429       return;
2430     }
2431     print_method(PHASE_BARRIER_EXPANSION, 2);
2432   }
2433 
2434   if (opaque4_count() > 0) {
2435     C->remove_opaque4_nodes(igvn);
2436     igvn.optimize();
2437   }
2438 
2439   DEBUG_ONLY( _modified_nodes = NULL; )
2440  } // (End scope of igvn; run destructor if necessary for asserts.)
2441 
2442  process_print_inlining();
2443  // A method with only infinite loops has no edges entering loops from root
2444  {
2445    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2446    if (final_graph_reshaping()) {
2447      assert(failing(), "must bail out w/ explicit message");
2448      return;
2449    }
2450  }
2451 
2452  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2453 }
2454 
2455 
2456 //------------------------------Code_Gen---------------------------------------
2457 // Given a graph, generate code for it
2458 void Compile::Code_Gen() {
2459   if (failing()) {
2460     return;
2461   }
2462 
2463   // Perform instruction selection.  You might think we could reclaim Matcher
2464   // memory PDQ, but actually the Matcher is used in generating spill code.
2465   // Internals of the Matcher (including some VectorSets) must remain live
2466   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2467   // set a bit in reclaimed memory.
2468 
2469   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2470   // nodes.  Mapping is only valid at the root of each matched subtree.
2471   NOT_PRODUCT( verify_graph_edges(); )
2472 
2473   Matcher matcher;
2474   _matcher = &matcher;
2475   {
2476     TracePhase tp("matcher", &timers[_t_matcher]);
2477     matcher.match();
2478   }
2479   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2480   // nodes.  Mapping is only valid at the root of each matched subtree.
2481   NOT_PRODUCT( verify_graph_edges(); )
2482 
2483   // If you have too many nodes, or if matching has failed, bail out
2484   check_node_count(0, "out of nodes matching instructions");
2485   if (failing()) {
2486     return;
2487   }
2488 
2489   print_method(PHASE_MATCHING, 2);
2490 
2491   // Build a proper-looking CFG
2492   PhaseCFG cfg(node_arena(), root(), matcher);
2493   _cfg = &cfg;
2494   {
2495     TracePhase tp("scheduler", &timers[_t_scheduler]);
2496     bool success = cfg.do_global_code_motion();
2497     if (!success) {
2498       return;
2499     }
2500 
2501     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2502     NOT_PRODUCT( verify_graph_edges(); )
2503     debug_only( cfg.verify(); )
2504   }
2505 
2506   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2507   _regalloc = &regalloc;
2508   {
2509     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2510     // Perform register allocation.  After Chaitin, use-def chains are
2511     // no longer accurate (at spill code) and so must be ignored.
2512     // Node->LRG->reg mappings are still accurate.
2513     _regalloc->Register_Allocate();
2514 
2515     // Bail out if the allocator builds too many nodes
2516     if (failing()) {
2517       return;
2518     }
2519   }
2520 
2521   // Prior to register allocation we kept empty basic blocks in case the
2522   // the allocator needed a place to spill.  After register allocation we
2523   // are not adding any new instructions.  If any basic block is empty, we
2524   // can now safely remove it.
2525   {
2526     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2527     cfg.remove_empty_blocks();
2528     if (do_freq_based_layout()) {
2529       PhaseBlockLayout layout(cfg);
2530     } else {
2531       cfg.set_loop_alignment();
2532     }
2533     cfg.fixup_flow();
2534   }
2535 
2536   // Apply peephole optimizations
2537   if( OptoPeephole ) {
2538     TracePhase tp("peephole", &timers[_t_peephole]);
2539     PhasePeephole peep( _regalloc, cfg);
2540     peep.do_transform();
2541   }
2542 
2543   // Do late expand if CPU requires this.
2544   if (Matcher::require_postalloc_expand) {
2545     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2546     cfg.postalloc_expand(_regalloc);
2547   }
2548 
2549   // Convert Nodes to instruction bits in a buffer
2550   {
2551     TraceTime tp("output", &timers[_t_output], CITime);
2552     Output();
2553   }
2554 
2555   print_method(PHASE_FINAL_CODE);
2556 
2557   // He's dead, Jim.
2558   _cfg     = (PhaseCFG*)((intptr_t)0xdeadbeef);
2559   _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2560 }
2561 
2562 
2563 //------------------------------dump_asm---------------------------------------
2564 // Dump formatted assembly
2565 #if defined(SUPPORT_OPTO_ASSEMBLY)
2566 void Compile::dump_asm_on(outputStream* st, int* pcs, uint pc_limit) {
2567 
2568   int pc_digits = 3; // #chars required for pc
2569   int sb_chars  = 3; // #chars for "start bundle" indicator
2570   int tab_size  = 8;
2571   if (pcs != NULL) {
2572     int max_pc = 0;
2573     for (uint i = 0; i < pc_limit; i++) {
2574       max_pc = (max_pc < pcs[i]) ? pcs[i] : max_pc;
2575     }
2576     pc_digits  = ((max_pc < 4096) ? 3 : ((max_pc < 65536) ? 4 : ((max_pc < 65536*256) ? 6 : 8))); // #chars required for pc
2577   }
2578   int prefix_len = ((pc_digits + sb_chars + tab_size - 1)/tab_size)*tab_size;
2579 
2580   bool cut_short = false;
2581   st->print_cr("#");
2582   st->print("#  ");  _tf->dump_on(st);  st->cr();
2583   st->print_cr("#");
2584 
2585   // For all blocks
2586   int pc = 0x0;                 // Program counter
2587   char starts_bundle = ' ';
2588   _regalloc->dump_frame();
2589 
2590   Node *n = NULL;
2591   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2592     if (VMThread::should_terminate()) {
2593       cut_short = true;
2594       break;
2595     }
2596     Block* block = _cfg->get_block(i);
2597     if (block->is_connector() && !Verbose) {
2598       continue;
2599     }
2600     n = block->head();
2601     if ((pcs != NULL) && (n->_idx < pc_limit)) {
2602       pc = pcs[n->_idx];
2603       st->print("%*.*x", pc_digits, pc_digits, pc);
2604     }
2605     st->fill_to(prefix_len);
2606     block->dump_head(_cfg, st);
2607     if (block->is_connector()) {
2608       st->fill_to(prefix_len);
2609       st->print_cr("# Empty connector block");
2610     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2611       st->fill_to(prefix_len);
2612       st->print_cr("# Block is sole successor of call");
2613     }
2614 
2615     // For all instructions
2616     Node *delay = NULL;
2617     for (uint j = 0; j < block->number_of_nodes(); j++) {
2618       if (VMThread::should_terminate()) {
2619         cut_short = true;
2620         break;
2621       }
2622       n = block->get_node(j);
2623       if (valid_bundle_info(n)) {
2624         Bundle* bundle = node_bundling(n);
2625         if (bundle->used_in_unconditional_delay()) {
2626           delay = n;
2627           continue;
2628         }
2629         if (bundle->starts_bundle()) {
2630           starts_bundle = '+';
2631         }
2632       }
2633 
2634       if (WizardMode) {
2635         n->dump();
2636       }
2637 
2638       if( !n->is_Region() &&    // Dont print in the Assembly
2639           !n->is_Phi() &&       // a few noisely useless nodes
2640           !n->is_Proj() &&
2641           !n->is_MachTemp() &&
2642           !n->is_SafePointScalarObject() &&
2643           !n->is_Catch() &&     // Would be nice to print exception table targets
2644           !n->is_MergeMem() &&  // Not very interesting
2645           !n->is_top() &&       // Debug info table constants
2646           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2647           ) {
2648         if ((pcs != NULL) && (n->_idx < pc_limit)) {
2649           pc = pcs[n->_idx];
2650           st->print("%*.*x", pc_digits, pc_digits, pc);
2651         } else {
2652           st->fill_to(pc_digits);
2653         }
2654         st->print(" %c ", starts_bundle);
2655         starts_bundle = ' ';
2656         st->fill_to(prefix_len);
2657         n->format(_regalloc, st);
2658         st->cr();
2659       }
2660 
2661       // If we have an instruction with a delay slot, and have seen a delay,
2662       // then back up and print it
2663       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2664         // Coverity finding - Explicit null dereferenced.
2665         guarantee(delay != NULL, "no unconditional delay instruction");
2666         if (WizardMode) delay->dump();
2667 
2668         if (node_bundling(delay)->starts_bundle())
2669           starts_bundle = '+';
2670         if ((pcs != NULL) && (n->_idx < pc_limit)) {
2671           pc = pcs[n->_idx];
2672           st->print("%*.*x", pc_digits, pc_digits, pc);
2673         } else {
2674           st->fill_to(pc_digits);
2675         }
2676         st->print(" %c ", starts_bundle);
2677         starts_bundle = ' ';
2678         st->fill_to(prefix_len);
2679         delay->format(_regalloc, st);
2680         st->cr();
2681         delay = NULL;
2682       }
2683 
2684       // Dump the exception table as well
2685       if( n->is_Catch() && (Verbose || WizardMode) ) {
2686         // Print the exception table for this offset
2687         _handler_table.print_subtable_for(pc);
2688       }
2689       st->bol(); // Make sure we start on a new line
2690     }
2691     st->cr(); // one empty line between blocks
2692     assert(cut_short || delay == NULL, "no unconditional delay branch");
2693   } // End of per-block dump
2694 
2695   if (cut_short)  st->print_cr("*** disassembly is cut short ***");
2696 }
2697 #endif
2698 
2699 //------------------------------Final_Reshape_Counts---------------------------
2700 // This class defines counters to help identify when a method
2701 // may/must be executed using hardware with only 24-bit precision.
2702 struct Final_Reshape_Counts : public StackObj {
2703   int  _call_count;             // count non-inlined 'common' calls
2704   int  _float_count;            // count float ops requiring 24-bit precision
2705   int  _double_count;           // count double ops requiring more precision
2706   int  _java_call_count;        // count non-inlined 'java' calls
2707   int  _inner_loop_count;       // count loops which need alignment
2708   VectorSet _visited;           // Visitation flags
2709   Node_List _tests;             // Set of IfNodes & PCTableNodes
2710 
2711   Final_Reshape_Counts() :
2712     _call_count(0), _float_count(0), _double_count(0),
2713     _java_call_count(0), _inner_loop_count(0),
2714     _visited( Thread::current()->resource_area() ) { }
2715 
2716   void inc_call_count  () { _call_count  ++; }
2717   void inc_float_count () { _float_count ++; }
2718   void inc_double_count() { _double_count++; }
2719   void inc_java_call_count() { _java_call_count++; }
2720   void inc_inner_loop_count() { _inner_loop_count++; }
2721 
2722   int  get_call_count  () const { return _call_count  ; }
2723   int  get_float_count () const { return _float_count ; }
2724   int  get_double_count() const { return _double_count; }
2725   int  get_java_call_count() const { return _java_call_count; }
2726   int  get_inner_loop_count() const { return _inner_loop_count; }
2727 };
2728 
2729 #ifdef ASSERT
2730 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2731   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2732   // Make sure the offset goes inside the instance layout.
2733   return k->contains_field_offset(tp->offset());
2734   // Note that OffsetBot and OffsetTop are very negative.
2735 }
2736 #endif
2737 
2738 // Eliminate trivially redundant StoreCMs and accumulate their
2739 // precedence edges.
2740 void Compile::eliminate_redundant_card_marks(Node* n) {
2741   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2742   if (n->in(MemNode::Address)->outcnt() > 1) {
2743     // There are multiple users of the same address so it might be
2744     // possible to eliminate some of the StoreCMs
2745     Node* mem = n->in(MemNode::Memory);
2746     Node* adr = n->in(MemNode::Address);
2747     Node* val = n->in(MemNode::ValueIn);
2748     Node* prev = n;
2749     bool done = false;
2750     // Walk the chain of StoreCMs eliminating ones that match.  As
2751     // long as it's a chain of single users then the optimization is
2752     // safe.  Eliminating partially redundant StoreCMs would require
2753     // cloning copies down the other paths.
2754     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2755       if (adr == mem->in(MemNode::Address) &&
2756           val == mem->in(MemNode::ValueIn)) {
2757         // redundant StoreCM
2758         if (mem->req() > MemNode::OopStore) {
2759           // Hasn't been processed by this code yet.
2760           n->add_prec(mem->in(MemNode::OopStore));
2761         } else {
2762           // Already converted to precedence edge
2763           for (uint i = mem->req(); i < mem->len(); i++) {
2764             // Accumulate any precedence edges
2765             if (mem->in(i) != NULL) {
2766               n->add_prec(mem->in(i));
2767             }
2768           }
2769           // Everything above this point has been processed.
2770           done = true;
2771         }
2772         // Eliminate the previous StoreCM
2773         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2774         assert(mem->outcnt() == 0, "should be dead");
2775         mem->disconnect_inputs(NULL, this);
2776       } else {
2777         prev = mem;
2778       }
2779       mem = prev->in(MemNode::Memory);
2780     }
2781   }
2782 }
2783 
2784 //------------------------------final_graph_reshaping_impl----------------------
2785 // Implement items 1-5 from final_graph_reshaping below.
2786 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2787 
2788   if ( n->outcnt() == 0 ) return; // dead node
2789   uint nop = n->Opcode();
2790 
2791   // Check for 2-input instruction with "last use" on right input.
2792   // Swap to left input.  Implements item (2).
2793   if( n->req() == 3 &&          // two-input instruction
2794       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2795       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2796       n->in(2)->outcnt() == 1 &&// right use IS a last use
2797       !n->in(2)->is_Con() ) {   // right use is not a constant
2798     // Check for commutative opcode
2799     switch( nop ) {
2800     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2801     case Op_MaxI:  case Op_MinI:
2802     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2803     case Op_AndL:  case Op_XorL:  case Op_OrL:
2804     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2805       // Move "last use" input to left by swapping inputs
2806       n->swap_edges(1, 2);
2807       break;
2808     }
2809     default:
2810       break;
2811     }
2812   }
2813 
2814 #ifdef ASSERT
2815   if( n->is_Mem() ) {
2816     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2817     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2818             // oop will be recorded in oop map if load crosses safepoint
2819             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2820                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2821             "raw memory operations should have control edge");
2822   }
2823   if (n->is_MemBar()) {
2824     MemBarNode* mb = n->as_MemBar();
2825     if (mb->trailing_store() || mb->trailing_load_store()) {
2826       assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2827       Node* mem = BarrierSet::barrier_set()->barrier_set_c2()->step_over_gc_barrier(mb->in(MemBarNode::Precedent));
2828       assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2829              (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2830     } else if (mb->leading()) {
2831       assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2832     }
2833   }
2834 #endif
2835   // Count FPU ops and common calls, implements item (3)
2836   bool gc_handled = BarrierSet::barrier_set()->barrier_set_c2()->final_graph_reshaping(this, n, nop);
2837   if (!gc_handled) {
2838     final_graph_reshaping_main_switch(n, frc, nop);
2839   }
2840 
2841   // Collect CFG split points
2842   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
2843     frc._tests.push(n);
2844   }
2845 }
2846 
2847 void Compile::final_graph_reshaping_main_switch(Node* n, Final_Reshape_Counts& frc, uint nop) {
2848   switch( nop ) {
2849   // Count all float operations that may use FPU
2850   case Op_AddF:
2851   case Op_SubF:
2852   case Op_MulF:
2853   case Op_DivF:
2854   case Op_NegF:
2855   case Op_ModF:
2856   case Op_ConvI2F:
2857   case Op_ConF:
2858   case Op_CmpF:
2859   case Op_CmpF3:
2860   // case Op_ConvL2F: // longs are split into 32-bit halves
2861     frc.inc_float_count();
2862     break;
2863 
2864   case Op_ConvF2D:
2865   case Op_ConvD2F:
2866     frc.inc_float_count();
2867     frc.inc_double_count();
2868     break;
2869 
2870   // Count all double operations that may use FPU
2871   case Op_AddD:
2872   case Op_SubD:
2873   case Op_MulD:
2874   case Op_DivD:
2875   case Op_NegD:
2876   case Op_ModD:
2877   case Op_ConvI2D:
2878   case Op_ConvD2I:
2879   // case Op_ConvL2D: // handled by leaf call
2880   // case Op_ConvD2L: // handled by leaf call
2881   case Op_ConD:
2882   case Op_CmpD:
2883   case Op_CmpD3:
2884     frc.inc_double_count();
2885     break;
2886   case Op_Opaque1:              // Remove Opaque Nodes before matching
2887   case Op_Opaque2:              // Remove Opaque Nodes before matching
2888   case Op_Opaque3:
2889     n->subsume_by(n->in(1), this);
2890     break;
2891   case Op_CallStaticJava:
2892   case Op_CallJava:
2893   case Op_CallDynamicJava:
2894     frc.inc_java_call_count(); // Count java call site;
2895   case Op_CallRuntime:
2896   case Op_CallLeaf:
2897   case Op_CallLeafNoFP: {
2898     assert (n->is_Call(), "");
2899     CallNode *call = n->as_Call();
2900     // Count call sites where the FP mode bit would have to be flipped.
2901     // Do not count uncommon runtime calls:
2902     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2903     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2904     if (!call->is_CallStaticJava() || !call->as_CallStaticJava()->_name) {
2905       frc.inc_call_count();   // Count the call site
2906     } else {                  // See if uncommon argument is shared
2907       Node *n = call->in(TypeFunc::Parms);
2908       int nop = n->Opcode();
2909       // Clone shared simple arguments to uncommon calls, item (1).
2910       if (n->outcnt() > 1 &&
2911           !n->is_Proj() &&
2912           nop != Op_CreateEx &&
2913           nop != Op_CheckCastPP &&
2914           nop != Op_DecodeN &&
2915           nop != Op_DecodeNKlass &&
2916           !n->is_Mem() &&
2917           !n->is_Phi()) {
2918         Node *x = n->clone();
2919         call->set_req(TypeFunc::Parms, x);
2920       }
2921     }
2922     break;
2923   }
2924 
2925   case Op_StoreD:
2926   case Op_LoadD:
2927   case Op_LoadD_unaligned:
2928     frc.inc_double_count();
2929     goto handle_mem;
2930   case Op_StoreF:
2931   case Op_LoadF:
2932     frc.inc_float_count();
2933     goto handle_mem;
2934 
2935   case Op_StoreCM:
2936     {
2937       // Convert OopStore dependence into precedence edge
2938       Node* prec = n->in(MemNode::OopStore);
2939       n->del_req(MemNode::OopStore);
2940       n->add_prec(prec);
2941       eliminate_redundant_card_marks(n);
2942     }
2943 
2944     // fall through
2945 
2946   case Op_StoreB:
2947   case Op_StoreC:
2948   case Op_StorePConditional:
2949   case Op_StoreI:
2950   case Op_StoreL:
2951   case Op_StoreIConditional:
2952   case Op_StoreLConditional:
2953   case Op_CompareAndSwapB:
2954   case Op_CompareAndSwapS:
2955   case Op_CompareAndSwapI:
2956   case Op_CompareAndSwapL:
2957   case Op_CompareAndSwapP:
2958   case Op_CompareAndSwapN:
2959   case Op_WeakCompareAndSwapB:
2960   case Op_WeakCompareAndSwapS:
2961   case Op_WeakCompareAndSwapI:
2962   case Op_WeakCompareAndSwapL:
2963   case Op_WeakCompareAndSwapP:
2964   case Op_WeakCompareAndSwapN:
2965   case Op_CompareAndExchangeB:
2966   case Op_CompareAndExchangeS:
2967   case Op_CompareAndExchangeI:
2968   case Op_CompareAndExchangeL:
2969   case Op_CompareAndExchangeP:
2970   case Op_CompareAndExchangeN:
2971   case Op_GetAndAddS:
2972   case Op_GetAndAddB:
2973   case Op_GetAndAddI:
2974   case Op_GetAndAddL:
2975   case Op_GetAndSetS:
2976   case Op_GetAndSetB:
2977   case Op_GetAndSetI:
2978   case Op_GetAndSetL:
2979   case Op_GetAndSetP:
2980   case Op_GetAndSetN:
2981   case Op_StoreP:
2982   case Op_StoreN:
2983   case Op_StoreNKlass:
2984   case Op_LoadB:
2985   case Op_LoadUB:
2986   case Op_LoadUS:
2987   case Op_LoadI:
2988   case Op_LoadKlass:
2989   case Op_LoadNKlass:
2990   case Op_LoadL:
2991   case Op_LoadL_unaligned:
2992   case Op_LoadPLocked:
2993   case Op_LoadP:
2994   case Op_LoadN:
2995   case Op_LoadRange:
2996   case Op_LoadS: {
2997   handle_mem:
2998 #ifdef ASSERT
2999     if( VerifyOptoOopOffsets ) {
3000       MemNode* mem  = n->as_Mem();
3001       // Check to see if address types have grounded out somehow.
3002       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
3003       assert( !tp || oop_offset_is_sane(tp), "" );
3004     }
3005 #endif
3006     break;
3007   }
3008 
3009   case Op_AddP: {               // Assert sane base pointers
3010     Node *addp = n->in(AddPNode::Address);
3011     assert( !addp->is_AddP() ||
3012             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
3013             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
3014             "Base pointers must match (addp %u)", addp->_idx );
3015 #ifdef _LP64
3016     if ((UseCompressedOops || UseCompressedClassPointers) &&
3017         addp->Opcode() == Op_ConP &&
3018         addp == n->in(AddPNode::Base) &&
3019         n->in(AddPNode::Offset)->is_Con()) {
3020       // If the transformation of ConP to ConN+DecodeN is beneficial depends
3021       // on the platform and on the compressed oops mode.
3022       // Use addressing with narrow klass to load with offset on x86.
3023       // Some platforms can use the constant pool to load ConP.
3024       // Do this transformation here since IGVN will convert ConN back to ConP.
3025       const Type* t = addp->bottom_type();
3026       bool is_oop   = t->isa_oopptr() != NULL;
3027       bool is_klass = t->isa_klassptr() != NULL;
3028 
3029       if ((is_oop   && Matcher::const_oop_prefer_decode()  ) ||
3030           (is_klass && Matcher::const_klass_prefer_decode())) {
3031         Node* nn = NULL;
3032 
3033         int op = is_oop ? Op_ConN : Op_ConNKlass;
3034 
3035         // Look for existing ConN node of the same exact type.
3036         Node* r  = root();
3037         uint cnt = r->outcnt();
3038         for (uint i = 0; i < cnt; i++) {
3039           Node* m = r->raw_out(i);
3040           if (m!= NULL && m->Opcode() == op &&
3041               m->bottom_type()->make_ptr() == t) {
3042             nn = m;
3043             break;
3044           }
3045         }
3046         if (nn != NULL) {
3047           // Decode a narrow oop to match address
3048           // [R12 + narrow_oop_reg<<3 + offset]
3049           if (is_oop) {
3050             nn = new DecodeNNode(nn, t);
3051           } else {
3052             nn = new DecodeNKlassNode(nn, t);
3053           }
3054           // Check for succeeding AddP which uses the same Base.
3055           // Otherwise we will run into the assertion above when visiting that guy.
3056           for (uint i = 0; i < n->outcnt(); ++i) {
3057             Node *out_i = n->raw_out(i);
3058             if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
3059               out_i->set_req(AddPNode::Base, nn);
3060 #ifdef ASSERT
3061               for (uint j = 0; j < out_i->outcnt(); ++j) {
3062                 Node *out_j = out_i->raw_out(j);
3063                 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
3064                        "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
3065               }
3066 #endif
3067             }
3068           }
3069           n->set_req(AddPNode::Base, nn);
3070           n->set_req(AddPNode::Address, nn);
3071           if (addp->outcnt() == 0) {
3072             addp->disconnect_inputs(NULL, this);
3073           }
3074         }
3075       }
3076     }
3077 #endif
3078     // platform dependent reshaping of the address expression
3079     reshape_address(n->as_AddP());
3080     break;
3081   }
3082 
3083   case Op_CastPP: {
3084     // Remove CastPP nodes to gain more freedom during scheduling but
3085     // keep the dependency they encode as control or precedence edges
3086     // (if control is set already) on memory operations. Some CastPP
3087     // nodes don't have a control (don't carry a dependency): skip
3088     // those.
3089     if (n->in(0) != NULL) {
3090       ResourceMark rm;
3091       Unique_Node_List wq;
3092       wq.push(n);
3093       for (uint next = 0; next < wq.size(); ++next) {
3094         Node *m = wq.at(next);
3095         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3096           Node* use = m->fast_out(i);
3097           if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
3098             use->ensure_control_or_add_prec(n->in(0));
3099           } else {
3100             switch(use->Opcode()) {
3101             case Op_AddP:
3102             case Op_DecodeN:
3103             case Op_DecodeNKlass:
3104             case Op_CheckCastPP:
3105             case Op_CastPP:
3106               wq.push(use);
3107               break;
3108             }
3109           }
3110         }
3111       }
3112     }
3113     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
3114     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
3115       Node* in1 = n->in(1);
3116       const Type* t = n->bottom_type();
3117       Node* new_in1 = in1->clone();
3118       new_in1->as_DecodeN()->set_type(t);
3119 
3120       if (!Matcher::narrow_oop_use_complex_address()) {
3121         //
3122         // x86, ARM and friends can handle 2 adds in addressing mode
3123         // and Matcher can fold a DecodeN node into address by using
3124         // a narrow oop directly and do implicit NULL check in address:
3125         //
3126         // [R12 + narrow_oop_reg<<3 + offset]
3127         // NullCheck narrow_oop_reg
3128         //
3129         // On other platforms (Sparc) we have to keep new DecodeN node and
3130         // use it to do implicit NULL check in address:
3131         //
3132         // decode_not_null narrow_oop_reg, base_reg
3133         // [base_reg + offset]
3134         // NullCheck base_reg
3135         //
3136         // Pin the new DecodeN node to non-null path on these platform (Sparc)
3137         // to keep the information to which NULL check the new DecodeN node
3138         // corresponds to use it as value in implicit_null_check().
3139         //
3140         new_in1->set_req(0, n->in(0));
3141       }
3142 
3143       n->subsume_by(new_in1, this);
3144       if (in1->outcnt() == 0) {
3145         in1->disconnect_inputs(NULL, this);
3146       }
3147     } else {
3148       n->subsume_by(n->in(1), this);
3149       if (n->outcnt() == 0) {
3150         n->disconnect_inputs(NULL, this);
3151       }
3152     }
3153     break;
3154   }
3155 #ifdef _LP64
3156   case Op_CmpP:
3157     // Do this transformation here to preserve CmpPNode::sub() and
3158     // other TypePtr related Ideal optimizations (for example, ptr nullness).
3159     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3160       Node* in1 = n->in(1);
3161       Node* in2 = n->in(2);
3162       if (!in1->is_DecodeNarrowPtr()) {
3163         in2 = in1;
3164         in1 = n->in(2);
3165       }
3166       assert(in1->is_DecodeNarrowPtr(), "sanity");
3167 
3168       Node* new_in2 = NULL;
3169       if (in2->is_DecodeNarrowPtr()) {
3170         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3171         new_in2 = in2->in(1);
3172       } else if (in2->Opcode() == Op_ConP) {
3173         const Type* t = in2->bottom_type();
3174         if (t == TypePtr::NULL_PTR) {
3175           assert(in1->is_DecodeN(), "compare klass to null?");
3176           // Don't convert CmpP null check into CmpN if compressed
3177           // oops implicit null check is not generated.
3178           // This will allow to generate normal oop implicit null check.
3179           if (Matcher::gen_narrow_oop_implicit_null_checks())
3180             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3181           //
3182           // This transformation together with CastPP transformation above
3183           // will generated code for implicit NULL checks for compressed oops.
3184           //
3185           // The original code after Optimize()
3186           //
3187           //    LoadN memory, narrow_oop_reg
3188           //    decode narrow_oop_reg, base_reg
3189           //    CmpP base_reg, NULL
3190           //    CastPP base_reg // NotNull
3191           //    Load [base_reg + offset], val_reg
3192           //
3193           // after these transformations will be
3194           //
3195           //    LoadN memory, narrow_oop_reg
3196           //    CmpN narrow_oop_reg, NULL
3197           //    decode_not_null narrow_oop_reg, base_reg
3198           //    Load [base_reg + offset], val_reg
3199           //
3200           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3201           // since narrow oops can be used in debug info now (see the code in
3202           // final_graph_reshaping_walk()).
3203           //
3204           // At the end the code will be matched to
3205           // on x86:
3206           //
3207           //    Load_narrow_oop memory, narrow_oop_reg
3208           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3209           //    NullCheck narrow_oop_reg
3210           //
3211           // and on sparc:
3212           //
3213           //    Load_narrow_oop memory, narrow_oop_reg
3214           //    decode_not_null narrow_oop_reg, base_reg
3215           //    Load [base_reg + offset], val_reg
3216           //    NullCheck base_reg
3217           //
3218         } else if (t->isa_oopptr()) {
3219           new_in2 = ConNode::make(t->make_narrowoop());
3220         } else if (t->isa_klassptr()) {
3221           new_in2 = ConNode::make(t->make_narrowklass());
3222         }
3223       }
3224       if (new_in2 != NULL) {
3225         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3226         n->subsume_by(cmpN, this);
3227         if (in1->outcnt() == 0) {
3228           in1->disconnect_inputs(NULL, this);
3229         }
3230         if (in2->outcnt() == 0) {
3231           in2->disconnect_inputs(NULL, this);
3232         }
3233       }
3234     }
3235     break;
3236 
3237   case Op_DecodeN:
3238   case Op_DecodeNKlass:
3239     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3240     // DecodeN could be pinned when it can't be fold into
3241     // an address expression, see the code for Op_CastPP above.
3242     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3243     break;
3244 
3245   case Op_EncodeP:
3246   case Op_EncodePKlass: {
3247     Node* in1 = n->in(1);
3248     if (in1->is_DecodeNarrowPtr()) {
3249       n->subsume_by(in1->in(1), this);
3250     } else if (in1->Opcode() == Op_ConP) {
3251       const Type* t = in1->bottom_type();
3252       if (t == TypePtr::NULL_PTR) {
3253         assert(t->isa_oopptr(), "null klass?");
3254         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3255       } else if (t->isa_oopptr()) {
3256         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3257       } else if (t->isa_klassptr()) {
3258         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3259       }
3260     }
3261     if (in1->outcnt() == 0) {
3262       in1->disconnect_inputs(NULL, this);
3263     }
3264     break;
3265   }
3266 
3267   case Op_Proj: {
3268     if (OptimizeStringConcat) {
3269       ProjNode* p = n->as_Proj();
3270       if (p->_is_io_use) {
3271         // Separate projections were used for the exception path which
3272         // are normally removed by a late inline.  If it wasn't inlined
3273         // then they will hang around and should just be replaced with
3274         // the original one.
3275         Node* proj = NULL;
3276         // Replace with just one
3277         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3278           Node *use = i.get();
3279           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3280             proj = use;
3281             break;
3282           }
3283         }
3284         assert(proj != NULL || p->_con == TypeFunc::I_O, "io may be dropped at an infinite loop");
3285         if (proj != NULL) {
3286           p->subsume_by(proj, this);
3287         }
3288       }
3289     }
3290     break;
3291   }
3292 
3293   case Op_Phi:
3294     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3295       // The EncodeP optimization may create Phi with the same edges
3296       // for all paths. It is not handled well by Register Allocator.
3297       Node* unique_in = n->in(1);
3298       assert(unique_in != NULL, "");
3299       uint cnt = n->req();
3300       for (uint i = 2; i < cnt; i++) {
3301         Node* m = n->in(i);
3302         assert(m != NULL, "");
3303         if (unique_in != m)
3304           unique_in = NULL;
3305       }
3306       if (unique_in != NULL) {
3307         n->subsume_by(unique_in, this);
3308       }
3309     }
3310     break;
3311 
3312 #endif
3313 
3314 #ifdef ASSERT
3315   case Op_CastII:
3316     // Verify that all range check dependent CastII nodes were removed.
3317     if (n->isa_CastII()->has_range_check()) {
3318       n->dump(3);
3319       assert(false, "Range check dependent CastII node was not removed");
3320     }
3321     break;
3322 #endif
3323 
3324   case Op_ModI:
3325     if (UseDivMod) {
3326       // Check if a%b and a/b both exist
3327       Node* d = n->find_similar(Op_DivI);
3328       if (d) {
3329         // Replace them with a fused divmod if supported
3330         if (Matcher::has_match_rule(Op_DivModI)) {
3331           DivModINode* divmod = DivModINode::make(n);
3332           d->subsume_by(divmod->div_proj(), this);
3333           n->subsume_by(divmod->mod_proj(), this);
3334         } else {
3335           // replace a%b with a-((a/b)*b)
3336           Node* mult = new MulINode(d, d->in(2));
3337           Node* sub  = new SubINode(d->in(1), mult);
3338           n->subsume_by(sub, this);
3339         }
3340       }
3341     }
3342     break;
3343 
3344   case Op_ModL:
3345     if (UseDivMod) {
3346       // Check if a%b and a/b both exist
3347       Node* d = n->find_similar(Op_DivL);
3348       if (d) {
3349         // Replace them with a fused divmod if supported
3350         if (Matcher::has_match_rule(Op_DivModL)) {
3351           DivModLNode* divmod = DivModLNode::make(n);
3352           d->subsume_by(divmod->div_proj(), this);
3353           n->subsume_by(divmod->mod_proj(), this);
3354         } else {
3355           // replace a%b with a-((a/b)*b)
3356           Node* mult = new MulLNode(d, d->in(2));
3357           Node* sub  = new SubLNode(d->in(1), mult);
3358           n->subsume_by(sub, this);
3359         }
3360       }
3361     }
3362     break;
3363 
3364   case Op_LoadVector:
3365   case Op_StoreVector:
3366     break;
3367 
3368   case Op_AddReductionVI:
3369   case Op_AddReductionVL:
3370   case Op_AddReductionVF:
3371   case Op_AddReductionVD:
3372   case Op_MulReductionVI:
3373   case Op_MulReductionVL:
3374   case Op_MulReductionVF:
3375   case Op_MulReductionVD:
3376   case Op_MinReductionV:
3377   case Op_MaxReductionV:
3378     break;
3379 
3380   case Op_PackB:
3381   case Op_PackS:
3382   case Op_PackI:
3383   case Op_PackF:
3384   case Op_PackL:
3385   case Op_PackD:
3386     if (n->req()-1 > 2) {
3387       // Replace many operand PackNodes with a binary tree for matching
3388       PackNode* p = (PackNode*) n;
3389       Node* btp = p->binary_tree_pack(1, n->req());
3390       n->subsume_by(btp, this);
3391     }
3392     break;
3393   case Op_Loop:
3394   case Op_CountedLoop:
3395   case Op_OuterStripMinedLoop:
3396     if (n->as_Loop()->is_inner_loop()) {
3397       frc.inc_inner_loop_count();
3398     }
3399     n->as_Loop()->verify_strip_mined(0);
3400     break;
3401   case Op_LShiftI:
3402   case Op_RShiftI:
3403   case Op_URShiftI:
3404   case Op_LShiftL:
3405   case Op_RShiftL:
3406   case Op_URShiftL:
3407     if (Matcher::need_masked_shift_count) {
3408       // The cpu's shift instructions don't restrict the count to the
3409       // lower 5/6 bits. We need to do the masking ourselves.
3410       Node* in2 = n->in(2);
3411       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3412       const TypeInt* t = in2->find_int_type();
3413       if (t != NULL && t->is_con()) {
3414         juint shift = t->get_con();
3415         if (shift > mask) { // Unsigned cmp
3416           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3417         }
3418       } else {
3419         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3420           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3421           n->set_req(2, shift);
3422         }
3423       }
3424       if (in2->outcnt() == 0) { // Remove dead node
3425         in2->disconnect_inputs(NULL, this);
3426       }
3427     }
3428     break;
3429   case Op_MemBarStoreStore:
3430   case Op_MemBarRelease:
3431     // Break the link with AllocateNode: it is no longer useful and
3432     // confuses register allocation.
3433     if (n->req() > MemBarNode::Precedent) {
3434       n->set_req(MemBarNode::Precedent, top());
3435     }
3436     break;
3437   case Op_MemBarAcquire: {
3438     if (n->as_MemBar()->trailing_load() && n->req() > MemBarNode::Precedent) {
3439       // At parse time, the trailing MemBarAcquire for a volatile load
3440       // is created with an edge to the load. After optimizations,
3441       // that input may be a chain of Phis. If those phis have no
3442       // other use, then the MemBarAcquire keeps them alive and
3443       // register allocation can be confused.
3444       ResourceMark rm;
3445       Unique_Node_List wq;
3446       wq.push(n->in(MemBarNode::Precedent));
3447       n->set_req(MemBarNode::Precedent, top());
3448       while (wq.size() > 0) {
3449         Node* m = wq.pop();
3450         if (m->outcnt() == 0) {
3451           for (uint j = 0; j < m->req(); j++) {
3452             Node* in = m->in(j);
3453             if (in != NULL) {
3454               wq.push(in);
3455             }
3456           }
3457           m->disconnect_inputs(NULL, this);
3458         }
3459       }
3460     }
3461     break;
3462   }
3463   case Op_RangeCheck: {
3464     RangeCheckNode* rc = n->as_RangeCheck();
3465     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3466     n->subsume_by(iff, this);
3467     frc._tests.push(iff);
3468     break;
3469   }
3470   case Op_ConvI2L: {
3471     if (!Matcher::convi2l_type_required) {
3472       // Code generation on some platforms doesn't need accurate
3473       // ConvI2L types. Widening the type can help remove redundant
3474       // address computations.
3475       n->as_Type()->set_type(TypeLong::INT);
3476       ResourceMark rm;
3477       Node_List wq;
3478       wq.push(n);
3479       for (uint next = 0; next < wq.size(); next++) {
3480         Node *m = wq.at(next);
3481 
3482         for(;;) {
3483           // Loop over all nodes with identical inputs edges as m
3484           Node* k = m->find_similar(m->Opcode());
3485           if (k == NULL) {
3486             break;
3487           }
3488           // Push their uses so we get a chance to remove node made
3489           // redundant
3490           for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3491             Node* u = k->fast_out(i);
3492             assert(!wq.contains(u), "shouldn't process one node several times");
3493             if (u->Opcode() == Op_LShiftL ||
3494                 u->Opcode() == Op_AddL ||
3495                 u->Opcode() == Op_SubL ||
3496                 u->Opcode() == Op_AddP) {
3497               wq.push(u);
3498             }
3499           }
3500           // Replace all nodes with identical edges as m with m
3501           k->subsume_by(m, this);
3502         }
3503       }
3504     }
3505     break;
3506   }
3507   case Op_CmpUL: {
3508     if (!Matcher::has_match_rule(Op_CmpUL)) {
3509       // No support for unsigned long comparisons
3510       ConINode* sign_pos = new ConINode(TypeInt::make(BitsPerLong - 1));
3511       Node* sign_bit_mask = new RShiftLNode(n->in(1), sign_pos);
3512       Node* orl = new OrLNode(n->in(1), sign_bit_mask);
3513       ConLNode* remove_sign_mask = new ConLNode(TypeLong::make(max_jlong));
3514       Node* andl = new AndLNode(orl, remove_sign_mask);
3515       Node* cmp = new CmpLNode(andl, n->in(2));
3516       n->subsume_by(cmp, this);
3517     }
3518     break;
3519   }
3520   default:
3521     assert(!n->is_Call(), "");
3522     assert(!n->is_Mem(), "");
3523     assert(nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3524     break;
3525   }
3526 }
3527 
3528 //------------------------------final_graph_reshaping_walk---------------------
3529 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3530 // requires that the walk visits a node's inputs before visiting the node.
3531 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3532   ResourceArea *area = Thread::current()->resource_area();
3533   Unique_Node_List sfpt(area);
3534 
3535   frc._visited.set(root->_idx); // first, mark node as visited
3536   uint cnt = root->req();
3537   Node *n = root;
3538   uint  i = 0;
3539   while (true) {
3540     if (i < cnt) {
3541       // Place all non-visited non-null inputs onto stack
3542       Node* m = n->in(i);
3543       ++i;
3544       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3545         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3546           // compute worst case interpreter size in case of a deoptimization
3547           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3548 
3549           sfpt.push(m);
3550         }
3551         cnt = m->req();
3552         nstack.push(n, i); // put on stack parent and next input's index
3553         n = m;
3554         i = 0;
3555       }
3556     } else {
3557       // Now do post-visit work
3558       final_graph_reshaping_impl( n, frc );
3559       if (nstack.is_empty())
3560         break;             // finished
3561       n = nstack.node();   // Get node from stack
3562       cnt = n->req();
3563       i = nstack.index();
3564       nstack.pop();        // Shift to the next node on stack
3565     }
3566   }
3567 
3568   // Skip next transformation if compressed oops are not used.
3569   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3570       (!UseCompressedOops && !UseCompressedClassPointers))
3571     return;
3572 
3573   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3574   // It could be done for an uncommon traps or any safepoints/calls
3575   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3576   while (sfpt.size() > 0) {
3577     n = sfpt.pop();
3578     JVMState *jvms = n->as_SafePoint()->jvms();
3579     assert(jvms != NULL, "sanity");
3580     int start = jvms->debug_start();
3581     int end   = n->req();
3582     bool is_uncommon = (n->is_CallStaticJava() &&
3583                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3584     for (int j = start; j < end; j++) {
3585       Node* in = n->in(j);
3586       if (in->is_DecodeNarrowPtr()) {
3587         bool safe_to_skip = true;
3588         if (!is_uncommon ) {
3589           // Is it safe to skip?
3590           for (uint i = 0; i < in->outcnt(); i++) {
3591             Node* u = in->raw_out(i);
3592             if (!u->is_SafePoint() ||
3593                 (u->is_Call() && u->as_Call()->has_non_debug_use(n))) {
3594               safe_to_skip = false;
3595             }
3596           }
3597         }
3598         if (safe_to_skip) {
3599           n->set_req(j, in->in(1));
3600         }
3601         if (in->outcnt() == 0) {
3602           in->disconnect_inputs(NULL, this);
3603         }
3604       }
3605     }
3606   }
3607 }
3608 
3609 //------------------------------final_graph_reshaping--------------------------
3610 // Final Graph Reshaping.
3611 //
3612 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3613 //     and not commoned up and forced early.  Must come after regular
3614 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3615 //     inputs to Loop Phis; these will be split by the allocator anyways.
3616 //     Remove Opaque nodes.
3617 // (2) Move last-uses by commutative operations to the left input to encourage
3618 //     Intel update-in-place two-address operations and better register usage
3619 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3620 //     calls canonicalizing them back.
3621 // (3) Count the number of double-precision FP ops, single-precision FP ops
3622 //     and call sites.  On Intel, we can get correct rounding either by
3623 //     forcing singles to memory (requires extra stores and loads after each
3624 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3625 //     clearing the mode bit around call sites).  The mode bit is only used
3626 //     if the relative frequency of single FP ops to calls is low enough.
3627 //     This is a key transform for SPEC mpeg_audio.
3628 // (4) Detect infinite loops; blobs of code reachable from above but not
3629 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3630 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3631 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3632 //     Detection is by looking for IfNodes where only 1 projection is
3633 //     reachable from below or CatchNodes missing some targets.
3634 // (5) Assert for insane oop offsets in debug mode.
3635 
3636 bool Compile::final_graph_reshaping() {
3637   // an infinite loop may have been eliminated by the optimizer,
3638   // in which case the graph will be empty.
3639   if (root()->req() == 1) {
3640     record_method_not_compilable("trivial infinite loop");
3641     return true;
3642   }
3643 
3644   // Expensive nodes have their control input set to prevent the GVN
3645   // from freely commoning them. There's no GVN beyond this point so
3646   // no need to keep the control input. We want the expensive nodes to
3647   // be freely moved to the least frequent code path by gcm.
3648   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3649   for (int i = 0; i < expensive_count(); i++) {
3650     _expensive_nodes->at(i)->set_req(0, NULL);
3651   }
3652 
3653   Final_Reshape_Counts frc;
3654 
3655   // Visit everybody reachable!
3656   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3657   Node_Stack nstack(live_nodes() >> 1);
3658   final_graph_reshaping_walk(nstack, root(), frc);
3659 
3660   // Check for unreachable (from below) code (i.e., infinite loops).
3661   for( uint i = 0; i < frc._tests.size(); i++ ) {
3662     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3663     // Get number of CFG targets.
3664     // Note that PCTables include exception targets after calls.
3665     uint required_outcnt = n->required_outcnt();
3666     if (n->outcnt() != required_outcnt) {
3667       // Check for a few special cases.  Rethrow Nodes never take the
3668       // 'fall-thru' path, so expected kids is 1 less.
3669       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3670         if (n->in(0)->in(0)->is_Call()) {
3671           CallNode *call = n->in(0)->in(0)->as_Call();
3672           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3673             required_outcnt--;      // Rethrow always has 1 less kid
3674           } else if (call->req() > TypeFunc::Parms &&
3675                      call->is_CallDynamicJava()) {
3676             // Check for null receiver. In such case, the optimizer has
3677             // detected that the virtual call will always result in a null
3678             // pointer exception. The fall-through projection of this CatchNode
3679             // will not be populated.
3680             Node *arg0 = call->in(TypeFunc::Parms);
3681             if (arg0->is_Type() &&
3682                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3683               required_outcnt--;
3684             }
3685           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3686                      call->req() > TypeFunc::Parms+1 &&
3687                      call->is_CallStaticJava()) {
3688             // Check for negative array length. In such case, the optimizer has
3689             // detected that the allocation attempt will always result in an
3690             // exception. There is no fall-through projection of this CatchNode .
3691             Node *arg1 = call->in(TypeFunc::Parms+1);
3692             if (arg1->is_Type() &&
3693                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3694               required_outcnt--;
3695             }
3696           }
3697         }
3698       }
3699       // Recheck with a better notion of 'required_outcnt'
3700       if (n->outcnt() != required_outcnt) {
3701         record_method_not_compilable("malformed control flow");
3702         return true;            // Not all targets reachable!
3703       }
3704     }
3705     // Check that I actually visited all kids.  Unreached kids
3706     // must be infinite loops.
3707     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3708       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3709         record_method_not_compilable("infinite loop");
3710         return true;            // Found unvisited kid; must be unreach
3711       }
3712 
3713     // Here so verification code in final_graph_reshaping_walk()
3714     // always see an OuterStripMinedLoopEnd
3715     if (n->is_OuterStripMinedLoopEnd()) {
3716       IfNode* init_iff = n->as_If();
3717       Node* iff = new IfNode(init_iff->in(0), init_iff->in(1), init_iff->_prob, init_iff->_fcnt);
3718       n->subsume_by(iff, this);
3719     }
3720   }
3721 
3722   // If original bytecodes contained a mixture of floats and doubles
3723   // check if the optimizer has made it homogenous, item (3).
3724   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3725       frc.get_float_count() > 32 &&
3726       frc.get_double_count() == 0 &&
3727       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3728     set_24_bit_selection_and_mode( false,  true );
3729   }
3730 
3731   set_java_calls(frc.get_java_call_count());
3732   set_inner_loops(frc.get_inner_loop_count());
3733 
3734   // No infinite loops, no reason to bail out.
3735   return false;
3736 }
3737 
3738 //-----------------------------too_many_traps----------------------------------
3739 // Report if there are too many traps at the current method and bci.
3740 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3741 bool Compile::too_many_traps(ciMethod* method,
3742                              int bci,
3743                              Deoptimization::DeoptReason reason) {
3744   ciMethodData* md = method->method_data();
3745   if (md->is_empty()) {
3746     // Assume the trap has not occurred, or that it occurred only
3747     // because of a transient condition during start-up in the interpreter.
3748     return false;
3749   }
3750   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3751   if (md->has_trap_at(bci, m, reason) != 0) {
3752     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3753     // Also, if there are multiple reasons, or if there is no per-BCI record,
3754     // assume the worst.
3755     if (log())
3756       log()->elem("observe trap='%s' count='%d'",
3757                   Deoptimization::trap_reason_name(reason),
3758                   md->trap_count(reason));
3759     return true;
3760   } else {
3761     // Ignore method/bci and see if there have been too many globally.
3762     return too_many_traps(reason, md);
3763   }
3764 }
3765 
3766 // Less-accurate variant which does not require a method and bci.
3767 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3768                              ciMethodData* logmd) {
3769   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3770     // Too many traps globally.
3771     // Note that we use cumulative trap_count, not just md->trap_count.
3772     if (log()) {
3773       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3774       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3775                   Deoptimization::trap_reason_name(reason),
3776                   mcount, trap_count(reason));
3777     }
3778     return true;
3779   } else {
3780     // The coast is clear.
3781     return false;
3782   }
3783 }
3784 
3785 //--------------------------too_many_recompiles--------------------------------
3786 // Report if there are too many recompiles at the current method and bci.
3787 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3788 // Is not eager to return true, since this will cause the compiler to use
3789 // Action_none for a trap point, to avoid too many recompilations.
3790 bool Compile::too_many_recompiles(ciMethod* method,
3791                                   int bci,
3792                                   Deoptimization::DeoptReason reason) {
3793   ciMethodData* md = method->method_data();
3794   if (md->is_empty()) {
3795     // Assume the trap has not occurred, or that it occurred only
3796     // because of a transient condition during start-up in the interpreter.
3797     return false;
3798   }
3799   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3800   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3801   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3802   Deoptimization::DeoptReason per_bc_reason
3803     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3804   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3805   if ((per_bc_reason == Deoptimization::Reason_none
3806        || md->has_trap_at(bci, m, reason) != 0)
3807       // The trap frequency measure we care about is the recompile count:
3808       && md->trap_recompiled_at(bci, m)
3809       && md->overflow_recompile_count() >= bc_cutoff) {
3810     // Do not emit a trap here if it has already caused recompilations.
3811     // Also, if there are multiple reasons, or if there is no per-BCI record,
3812     // assume the worst.
3813     if (log())
3814       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3815                   Deoptimization::trap_reason_name(reason),
3816                   md->trap_count(reason),
3817                   md->overflow_recompile_count());
3818     return true;
3819   } else if (trap_count(reason) != 0
3820              && decompile_count() >= m_cutoff) {
3821     // Too many recompiles globally, and we have seen this sort of trap.
3822     // Use cumulative decompile_count, not just md->decompile_count.
3823     if (log())
3824       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3825                   Deoptimization::trap_reason_name(reason),
3826                   md->trap_count(reason), trap_count(reason),
3827                   md->decompile_count(), decompile_count());
3828     return true;
3829   } else {
3830     // The coast is clear.
3831     return false;
3832   }
3833 }
3834 
3835 // Compute when not to trap. Used by matching trap based nodes and
3836 // NullCheck optimization.
3837 void Compile::set_allowed_deopt_reasons() {
3838   _allowed_reasons = 0;
3839   if (is_method_compilation()) {
3840     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3841       assert(rs < BitsPerInt, "recode bit map");
3842       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3843         _allowed_reasons |= nth_bit(rs);
3844       }
3845     }
3846   }
3847 }
3848 
3849 bool Compile::needs_clinit_barrier(ciMethod* method, ciMethod* accessing_method) {
3850   return method->is_static() && needs_clinit_barrier(method->holder(), accessing_method);
3851 }
3852 
3853 bool Compile::needs_clinit_barrier(ciField* field, ciMethod* accessing_method) {
3854   return field->is_static() && needs_clinit_barrier(field->holder(), accessing_method);
3855 }
3856 
3857 bool Compile::needs_clinit_barrier(ciInstanceKlass* holder, ciMethod* accessing_method) {
3858   if (holder->is_initialized()) {
3859     return false;
3860   }
3861   if (holder->is_being_initialized()) {
3862     if (accessing_method->holder() == holder) {
3863       // Access inside a class. The barrier can be elided when access happens in <clinit>,
3864       // <init>, or a static method. In all those cases, there was an initialization
3865       // barrier on the holder klass passed.
3866       if (accessing_method->is_static_initializer() ||
3867           accessing_method->is_object_initializer() ||
3868           accessing_method->is_static()) {
3869         return false;
3870       }
3871     } else if (accessing_method->holder()->is_subclass_of(holder)) {
3872       // Access from a subclass. The barrier can be elided only when access happens in <clinit>.
3873       // In case of <init> or a static method, the barrier is on the subclass is not enough:
3874       // child class can become fully initialized while its parent class is still being initialized.
3875       if (accessing_method->is_static_initializer()) {
3876         return false;
3877       }
3878     }
3879     ciMethod* root = method(); // the root method of compilation
3880     if (root != accessing_method) {
3881       return needs_clinit_barrier(holder, root); // check access in the context of compilation root
3882     }
3883   }
3884   return true;
3885 }
3886 
3887 #ifndef PRODUCT
3888 //------------------------------verify_graph_edges---------------------------
3889 // Walk the Graph and verify that there is a one-to-one correspondence
3890 // between Use-Def edges and Def-Use edges in the graph.
3891 void Compile::verify_graph_edges(bool no_dead_code) {
3892   if (VerifyGraphEdges) {
3893     ResourceArea *area = Thread::current()->resource_area();
3894     Unique_Node_List visited(area);
3895     // Call recursive graph walk to check edges
3896     _root->verify_edges(visited);
3897     if (no_dead_code) {
3898       // Now make sure that no visited node is used by an unvisited node.
3899       bool dead_nodes = false;
3900       Unique_Node_List checked(area);
3901       while (visited.size() > 0) {
3902         Node* n = visited.pop();
3903         checked.push(n);
3904         for (uint i = 0; i < n->outcnt(); i++) {
3905           Node* use = n->raw_out(i);
3906           if (checked.member(use))  continue;  // already checked
3907           if (visited.member(use))  continue;  // already in the graph
3908           if (use->is_Con())        continue;  // a dead ConNode is OK
3909           // At this point, we have found a dead node which is DU-reachable.
3910           if (!dead_nodes) {
3911             tty->print_cr("*** Dead nodes reachable via DU edges:");
3912             dead_nodes = true;
3913           }
3914           use->dump(2);
3915           tty->print_cr("---");
3916           checked.push(use);  // No repeats; pretend it is now checked.
3917         }
3918       }
3919       assert(!dead_nodes, "using nodes must be reachable from root");
3920     }
3921   }
3922 }
3923 #endif
3924 
3925 // The Compile object keeps track of failure reasons separately from the ciEnv.
3926 // This is required because there is not quite a 1-1 relation between the
3927 // ciEnv and its compilation task and the Compile object.  Note that one
3928 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3929 // to backtrack and retry without subsuming loads.  Other than this backtracking
3930 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3931 // by the logic in C2Compiler.
3932 void Compile::record_failure(const char* reason) {
3933   if (log() != NULL) {
3934     log()->elem("failure reason='%s' phase='compile'", reason);
3935   }
3936   if (_failure_reason == NULL) {
3937     // Record the first failure reason.
3938     _failure_reason = reason;
3939   }
3940 
3941   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3942     C->print_method(PHASE_FAILURE);
3943   }
3944   _root = NULL;  // flush the graph, too
3945 }
3946 
3947 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3948   : TraceTime(name, accumulator, CITime, CITimeVerbose),
3949     _phase_name(name), _dolog(CITimeVerbose)
3950 {
3951   if (_dolog) {
3952     C = Compile::current();
3953     _log = C->log();
3954   } else {
3955     C = NULL;
3956     _log = NULL;
3957   }
3958   if (_log != NULL) {
3959     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3960     _log->stamp();
3961     _log->end_head();
3962   }
3963 }
3964 
3965 Compile::TracePhase::~TracePhase() {
3966 
3967   C = Compile::current();
3968   if (_dolog) {
3969     _log = C->log();
3970   } else {
3971     _log = NULL;
3972   }
3973 
3974 #ifdef ASSERT
3975   if (PrintIdealNodeCount) {
3976     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3977                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3978   }
3979 
3980   if (VerifyIdealNodeCount) {
3981     Compile::current()->print_missing_nodes();
3982   }
3983 #endif
3984 
3985   if (_log != NULL) {
3986     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3987   }
3988 }
3989 
3990 //=============================================================================
3991 // Two Constant's are equal when the type and the value are equal.
3992 bool Compile::Constant::operator==(const Constant& other) {
3993   if (type()          != other.type()         )  return false;
3994   if (can_be_reused() != other.can_be_reused())  return false;
3995   // For floating point values we compare the bit pattern.
3996   switch (type()) {
3997   case T_INT:
3998   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3999   case T_LONG:
4000   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
4001   case T_OBJECT:
4002   case T_ADDRESS: return (_v._value.l == other._v._value.l);
4003   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
4004   case T_METADATA: return (_v._metadata == other._v._metadata);
4005   default: ShouldNotReachHere(); return false;
4006   }
4007 }
4008 
4009 static int type_to_size_in_bytes(BasicType t) {
4010   switch (t) {
4011   case T_INT:     return sizeof(jint   );
4012   case T_LONG:    return sizeof(jlong  );
4013   case T_FLOAT:   return sizeof(jfloat );
4014   case T_DOUBLE:  return sizeof(jdouble);
4015   case T_METADATA: return sizeof(Metadata*);
4016     // We use T_VOID as marker for jump-table entries (labels) which
4017     // need an internal word relocation.
4018   case T_VOID:
4019   case T_ADDRESS:
4020   case T_OBJECT:  return sizeof(jobject);
4021   default:
4022     ShouldNotReachHere();
4023     return -1;
4024   }
4025 }
4026 
4027 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
4028   // sort descending
4029   if (a->freq() > b->freq())  return -1;
4030   if (a->freq() < b->freq())  return  1;
4031   return 0;
4032 }
4033 
4034 void Compile::ConstantTable::calculate_offsets_and_size() {
4035   // First, sort the array by frequencies.
4036   _constants.sort(qsort_comparator);
4037 
4038 #ifdef ASSERT
4039   // Make sure all jump-table entries were sorted to the end of the
4040   // array (they have a negative frequency).
4041   bool found_void = false;
4042   for (int i = 0; i < _constants.length(); i++) {
4043     Constant con = _constants.at(i);
4044     if (con.type() == T_VOID)
4045       found_void = true;  // jump-tables
4046     else
4047       assert(!found_void, "wrong sorting");
4048   }
4049 #endif
4050 
4051   int offset = 0;
4052   for (int i = 0; i < _constants.length(); i++) {
4053     Constant* con = _constants.adr_at(i);
4054 
4055     // Align offset for type.
4056     int typesize = type_to_size_in_bytes(con->type());
4057     offset = align_up(offset, typesize);
4058     con->set_offset(offset);   // set constant's offset
4059 
4060     if (con->type() == T_VOID) {
4061       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
4062       offset = offset + typesize * n->outcnt();  // expand jump-table
4063     } else {
4064       offset = offset + typesize;
4065     }
4066   }
4067 
4068   // Align size up to the next section start (which is insts; see
4069   // CodeBuffer::align_at_start).
4070   assert(_size == -1, "already set?");
4071   _size = align_up(offset, (int)CodeEntryAlignment);
4072 }
4073 
4074 void Compile::ConstantTable::emit(CodeBuffer& cb) {
4075   MacroAssembler _masm(&cb);
4076   for (int i = 0; i < _constants.length(); i++) {
4077     Constant con = _constants.at(i);
4078     address constant_addr = NULL;
4079     switch (con.type()) {
4080     case T_INT:    constant_addr = _masm.int_constant(   con.get_jint()   ); break;
4081     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
4082     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
4083     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
4084     case T_OBJECT: {
4085       jobject obj = con.get_jobject();
4086       int oop_index = _masm.oop_recorder()->find_index(obj);
4087       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
4088       break;
4089     }
4090     case T_ADDRESS: {
4091       address addr = (address) con.get_jobject();
4092       constant_addr = _masm.address_constant(addr);
4093       break;
4094     }
4095     // We use T_VOID as marker for jump-table entries (labels) which
4096     // need an internal word relocation.
4097     case T_VOID: {
4098       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
4099       // Fill the jump-table with a dummy word.  The real value is
4100       // filled in later in fill_jump_table.
4101       address dummy = (address) n;
4102       constant_addr = _masm.address_constant(dummy);
4103       // Expand jump-table
4104       for (uint i = 1; i < n->outcnt(); i++) {
4105         address temp_addr = _masm.address_constant(dummy + i);
4106         assert(temp_addr, "consts section too small");
4107       }
4108       break;
4109     }
4110     case T_METADATA: {
4111       Metadata* obj = con.get_metadata();
4112       int metadata_index = _masm.oop_recorder()->find_index(obj);
4113       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
4114       break;
4115     }
4116     default: ShouldNotReachHere();
4117     }
4118     assert(constant_addr, "consts section too small");
4119     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
4120             "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
4121   }
4122 }
4123 
4124 int Compile::ConstantTable::find_offset(Constant& con) const {
4125   int idx = _constants.find(con);
4126   guarantee(idx != -1, "constant must be in constant table");
4127   int offset = _constants.at(idx).offset();
4128   guarantee(offset != -1, "constant table not emitted yet?");
4129   return offset;
4130 }
4131 
4132 void Compile::ConstantTable::add(Constant& con) {
4133   if (con.can_be_reused()) {
4134     int idx = _constants.find(con);
4135     if (idx != -1 && _constants.at(idx).can_be_reused()) {
4136       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
4137       return;
4138     }
4139   }
4140   (void) _constants.append(con);
4141 }
4142 
4143 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
4144   Block* b = Compile::current()->cfg()->get_block_for_node(n);
4145   Constant con(type, value, b->_freq);
4146   add(con);
4147   return con;
4148 }
4149 
4150 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
4151   Constant con(metadata);
4152   add(con);
4153   return con;
4154 }
4155 
4156 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
4157   jvalue value;
4158   BasicType type = oper->type()->basic_type();
4159   switch (type) {
4160   case T_LONG:    value.j = oper->constantL(); break;
4161   case T_FLOAT:   value.f = oper->constantF(); break;
4162   case T_DOUBLE:  value.d = oper->constantD(); break;
4163   case T_OBJECT:
4164   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4165   case T_METADATA: return add((Metadata*)oper->constant()); break;
4166   default: guarantee(false, "unhandled type: %s", type2name(type));
4167   }
4168   return add(n, type, value);
4169 }
4170 
4171 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4172   jvalue value;
4173   // We can use the node pointer here to identify the right jump-table
4174   // as this method is called from Compile::Fill_buffer right before
4175   // the MachNodes are emitted and the jump-table is filled (means the
4176   // MachNode pointers do not change anymore).
4177   value.l = (jobject) n;
4178   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
4179   add(con);
4180   return con;
4181 }
4182 
4183 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4184   // If called from Compile::scratch_emit_size do nothing.
4185   if (Compile::current()->in_scratch_emit_size())  return;
4186 
4187   assert(labels.is_nonempty(), "must be");
4188   assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4189 
4190   // Since MachConstantNode::constant_offset() also contains
4191   // table_base_offset() we need to subtract the table_base_offset()
4192   // to get the plain offset into the constant table.
4193   int offset = n->constant_offset() - table_base_offset();
4194 
4195   MacroAssembler _masm(&cb);
4196   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4197 
4198   for (uint i = 0; i < n->outcnt(); i++) {
4199     address* constant_addr = &jump_table_base[i];
4200     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));
4201     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4202     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4203   }
4204 }
4205 
4206 //----------------------------static_subtype_check-----------------------------
4207 // Shortcut important common cases when superklass is exact:
4208 // (0) superklass is java.lang.Object (can occur in reflective code)
4209 // (1) subklass is already limited to a subtype of superklass => always ok
4210 // (2) subklass does not overlap with superklass => always fail
4211 // (3) superklass has NO subtypes and we can check with a simple compare.
4212 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4213   if (StressReflectiveCode) {
4214     return SSC_full_test;       // Let caller generate the general case.
4215   }
4216 
4217   if (superk == env()->Object_klass()) {
4218     return SSC_always_true;     // (0) this test cannot fail
4219   }
4220 
4221   ciType* superelem = superk;
4222   if (superelem->is_array_klass())
4223     superelem = superelem->as_array_klass()->base_element_type();
4224 
4225   if (!subk->is_interface()) {  // cannot trust static interface types yet
4226     if (subk->is_subtype_of(superk)) {
4227       return SSC_always_true;   // (1) false path dead; no dynamic test needed
4228     }
4229     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4230         !superk->is_subtype_of(subk)) {
4231       return SSC_always_false;
4232     }
4233   }
4234 
4235   // If casting to an instance klass, it must have no subtypes
4236   if (superk->is_interface()) {
4237     // Cannot trust interfaces yet.
4238     // %%% S.B. superk->nof_implementors() == 1
4239   } else if (superelem->is_instance_klass()) {
4240     ciInstanceKlass* ik = superelem->as_instance_klass();
4241     if (!ik->has_subklass() && !ik->is_interface()) {
4242       if (!ik->is_final()) {
4243         // Add a dependency if there is a chance of a later subclass.
4244         dependencies()->assert_leaf_type(ik);
4245       }
4246       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
4247     }
4248   } else {
4249     // A primitive array type has no subtypes.
4250     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
4251   }
4252 
4253   return SSC_full_test;
4254 }
4255 
4256 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4257 #ifdef _LP64
4258   // The scaled index operand to AddP must be a clean 64-bit value.
4259   // Java allows a 32-bit int to be incremented to a negative
4260   // value, which appears in a 64-bit register as a large
4261   // positive number.  Using that large positive number as an
4262   // operand in pointer arithmetic has bad consequences.
4263   // On the other hand, 32-bit overflow is rare, and the possibility
4264   // can often be excluded, if we annotate the ConvI2L node with
4265   // a type assertion that its value is known to be a small positive
4266   // number.  (The prior range check has ensured this.)
4267   // This assertion is used by ConvI2LNode::Ideal.
4268   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4269   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4270   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4271   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4272 #endif
4273   return idx;
4274 }
4275 
4276 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4277 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4278   if (ctrl != NULL) {
4279     // Express control dependency by a CastII node with a narrow type.
4280     value = new CastIINode(value, itype, false, true /* range check dependency */);
4281     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4282     // node from floating above the range check during loop optimizations. Otherwise, the
4283     // ConvI2L node may be eliminated independently of the range check, causing the data path
4284     // to become TOP while the control path is still there (although it's unreachable).
4285     value->set_req(0, ctrl);
4286     // Save CastII node to remove it after loop optimizations.
4287     phase->C->add_range_check_cast(value);
4288     value = phase->transform(value);
4289   }
4290   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4291   return phase->transform(new ConvI2LNode(value, ltype));
4292 }
4293 
4294 // The message about the current inlining is accumulated in
4295 // _print_inlining_stream and transfered into the _print_inlining_list
4296 // once we know whether inlining succeeds or not. For regular
4297 // inlining, messages are appended to the buffer pointed by
4298 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4299 // a new buffer is added after _print_inlining_idx in the list. This
4300 // way we can update the inlining message for late inlining call site
4301 // when the inlining is attempted again.
4302 void Compile::print_inlining_init() {
4303   if (print_inlining() || print_intrinsics()) {
4304     _print_inlining_stream = new stringStream();
4305     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4306   }
4307 }
4308 
4309 void Compile::print_inlining_reinit() {
4310   if (print_inlining() || print_intrinsics()) {
4311     // Re allocate buffer when we change ResourceMark
4312     _print_inlining_stream = new stringStream();
4313   }
4314 }
4315 
4316 void Compile::print_inlining_reset() {
4317   _print_inlining_stream->reset();
4318 }
4319 
4320 void Compile::print_inlining_commit() {
4321   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4322   // Transfer the message from _print_inlining_stream to the current
4323   // _print_inlining_list buffer and clear _print_inlining_stream.
4324   _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4325   print_inlining_reset();
4326 }
4327 
4328 void Compile::print_inlining_push() {
4329   // Add new buffer to the _print_inlining_list at current position
4330   _print_inlining_idx++;
4331   _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4332 }
4333 
4334 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4335   return _print_inlining_list->at(_print_inlining_idx);
4336 }
4337 
4338 void Compile::print_inlining_update(CallGenerator* cg) {
4339   if (print_inlining() || print_intrinsics()) {
4340     if (!cg->is_late_inline()) {
4341       if (print_inlining_current().cg() != NULL) {
4342         print_inlining_push();
4343       }
4344       print_inlining_commit();
4345     } else {
4346       if (print_inlining_current().cg() != cg &&
4347           (print_inlining_current().cg() != NULL ||
4348            print_inlining_current().ss()->size() != 0)) {
4349         print_inlining_push();
4350       }
4351       print_inlining_commit();
4352       print_inlining_current().set_cg(cg);
4353     }
4354   }
4355 }
4356 
4357 void Compile::print_inlining_move_to(CallGenerator* cg) {
4358   // We resume inlining at a late inlining call site. Locate the
4359   // corresponding inlining buffer so that we can update it.
4360   if (print_inlining()) {
4361     for (int i = 0; i < _print_inlining_list->length(); i++) {
4362       if (_print_inlining_list->adr_at(i)->cg() == cg) {
4363         _print_inlining_idx = i;
4364         return;
4365       }
4366     }
4367     ShouldNotReachHere();
4368   }
4369 }
4370 
4371 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4372   if (print_inlining()) {
4373     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4374     assert(print_inlining_current().cg() == cg, "wrong entry");
4375     // replace message with new message
4376     _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4377     print_inlining_commit();
4378     print_inlining_current().set_cg(cg);
4379   }
4380 }
4381 
4382 void Compile::print_inlining_assert_ready() {
4383   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4384 }
4385 
4386 void Compile::process_print_inlining() {
4387   bool do_print_inlining = print_inlining() || print_intrinsics();
4388   if (do_print_inlining || log() != NULL) {
4389     // Print inlining message for candidates that we couldn't inline
4390     // for lack of space
4391     for (int i = 0; i < _late_inlines.length(); i++) {
4392       CallGenerator* cg = _late_inlines.at(i);
4393       if (!cg->is_mh_late_inline()) {
4394         const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4395         if (do_print_inlining) {
4396           cg->print_inlining_late(msg);
4397         }
4398         log_late_inline_failure(cg, msg);
4399       }
4400     }
4401   }
4402   if (do_print_inlining) {
4403     ResourceMark rm;
4404     stringStream ss;
4405     for (int i = 0; i < _print_inlining_list->length(); i++) {
4406       ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4407     }
4408     size_t end = ss.size();
4409     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4410     strncpy(_print_inlining_output, ss.base(), end+1);
4411     _print_inlining_output[end] = 0;
4412   }
4413 }
4414 
4415 void Compile::dump_print_inlining() {
4416   if (_print_inlining_output != NULL) {
4417     tty->print_raw(_print_inlining_output);
4418   }
4419 }
4420 
4421 void Compile::log_late_inline(CallGenerator* cg) {
4422   if (log() != NULL) {
4423     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4424                 cg->unique_id());
4425     JVMState* p = cg->call_node()->jvms();
4426     while (p != NULL) {
4427       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4428       p = p->caller();
4429     }
4430     log()->tail("late_inline");
4431   }
4432 }
4433 
4434 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4435   log_late_inline(cg);
4436   if (log() != NULL) {
4437     log()->inline_fail(msg);
4438   }
4439 }
4440 
4441 void Compile::log_inline_id(CallGenerator* cg) {
4442   if (log() != NULL) {
4443     // The LogCompilation tool needs a unique way to identify late
4444     // inline call sites. This id must be unique for this call site in
4445     // this compilation. Try to have it unique across compilations as
4446     // well because it can be convenient when grepping through the log
4447     // file.
4448     // Distinguish OSR compilations from others in case CICountOSR is
4449     // on.
4450     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4451     cg->set_unique_id(id);
4452     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4453   }
4454 }
4455 
4456 void Compile::log_inline_failure(const char* msg) {
4457   if (C->log() != NULL) {
4458     C->log()->inline_fail(msg);
4459   }
4460 }
4461 
4462 
4463 // Dump inlining replay data to the stream.
4464 // Don't change thread state and acquire any locks.
4465 void Compile::dump_inline_data(outputStream* out) {
4466   InlineTree* inl_tree = ilt();
4467   if (inl_tree != NULL) {
4468     out->print(" inline %d", inl_tree->count());
4469     inl_tree->dump_replay_data(out);
4470   }
4471 }
4472 
4473 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4474   if (n1->Opcode() < n2->Opcode())      return -1;
4475   else if (n1->Opcode() > n2->Opcode()) return 1;
4476 
4477   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4478   for (uint i = 1; i < n1->req(); i++) {
4479     if (n1->in(i) < n2->in(i))      return -1;
4480     else if (n1->in(i) > n2->in(i)) return 1;
4481   }
4482 
4483   return 0;
4484 }
4485 
4486 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4487   Node* n1 = *n1p;
4488   Node* n2 = *n2p;
4489 
4490   return cmp_expensive_nodes(n1, n2);
4491 }
4492 
4493 void Compile::sort_expensive_nodes() {
4494   if (!expensive_nodes_sorted()) {
4495     _expensive_nodes->sort(cmp_expensive_nodes);
4496   }
4497 }
4498 
4499 bool Compile::expensive_nodes_sorted() const {
4500   for (int i = 1; i < _expensive_nodes->length(); i++) {
4501     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4502       return false;
4503     }
4504   }
4505   return true;
4506 }
4507 
4508 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4509   if (_expensive_nodes->length() == 0) {
4510     return false;
4511   }
4512 
4513   assert(OptimizeExpensiveOps, "optimization off?");
4514 
4515   // Take this opportunity to remove dead nodes from the list
4516   int j = 0;
4517   for (int i = 0; i < _expensive_nodes->length(); i++) {
4518     Node* n = _expensive_nodes->at(i);
4519     if (!n->is_unreachable(igvn)) {
4520       assert(n->is_expensive(), "should be expensive");
4521       _expensive_nodes->at_put(j, n);
4522       j++;
4523     }
4524   }
4525   _expensive_nodes->trunc_to(j);
4526 
4527   // Then sort the list so that similar nodes are next to each other
4528   // and check for at least two nodes of identical kind with same data
4529   // inputs.
4530   sort_expensive_nodes();
4531 
4532   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4533     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4534       return true;
4535     }
4536   }
4537 
4538   return false;
4539 }
4540 
4541 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4542   if (_expensive_nodes->length() == 0) {
4543     return;
4544   }
4545 
4546   assert(OptimizeExpensiveOps, "optimization off?");
4547 
4548   // Sort to bring similar nodes next to each other and clear the
4549   // control input of nodes for which there's only a single copy.
4550   sort_expensive_nodes();
4551 
4552   int j = 0;
4553   int identical = 0;
4554   int i = 0;
4555   bool modified = false;
4556   for (; i < _expensive_nodes->length()-1; i++) {
4557     assert(j <= i, "can't write beyond current index");
4558     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4559       identical++;
4560       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4561       continue;
4562     }
4563     if (identical > 0) {
4564       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4565       identical = 0;
4566     } else {
4567       Node* n = _expensive_nodes->at(i);
4568       igvn.replace_input_of(n, 0, NULL);
4569       igvn.hash_insert(n);
4570       modified = true;
4571     }
4572   }
4573   if (identical > 0) {
4574     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4575   } else if (_expensive_nodes->length() >= 1) {
4576     Node* n = _expensive_nodes->at(i);
4577     igvn.replace_input_of(n, 0, NULL);
4578     igvn.hash_insert(n);
4579     modified = true;
4580   }
4581   _expensive_nodes->trunc_to(j);
4582   if (modified) {
4583     igvn.optimize();
4584   }
4585 }
4586 
4587 void Compile::add_expensive_node(Node * n) {
4588   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4589   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4590   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4591   if (OptimizeExpensiveOps) {
4592     _expensive_nodes->append(n);
4593   } else {
4594     // Clear control input and let IGVN optimize expensive nodes if
4595     // OptimizeExpensiveOps is off.
4596     n->set_req(0, NULL);
4597   }
4598 }
4599 
4600 /**
4601  * Remove the speculative part of types and clean up the graph
4602  */
4603 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4604   if (UseTypeSpeculation) {
4605     Unique_Node_List worklist;
4606     worklist.push(root());
4607     int modified = 0;
4608     // Go over all type nodes that carry a speculative type, drop the
4609     // speculative part of the type and enqueue the node for an igvn
4610     // which may optimize it out.
4611     for (uint next = 0; next < worklist.size(); ++next) {
4612       Node *n  = worklist.at(next);
4613       if (n->is_Type()) {
4614         TypeNode* tn = n->as_Type();
4615         const Type* t = tn->type();
4616         const Type* t_no_spec = t->remove_speculative();
4617         if (t_no_spec != t) {
4618           bool in_hash = igvn.hash_delete(n);
4619           assert(in_hash, "node should be in igvn hash table");
4620           tn->set_type(t_no_spec);
4621           igvn.hash_insert(n);
4622           igvn._worklist.push(n); // give it a chance to go away
4623           modified++;
4624         }
4625       }
4626       uint max = n->len();
4627       for( uint i = 0; i < max; ++i ) {
4628         Node *m = n->in(i);
4629         if (not_a_node(m))  continue;
4630         worklist.push(m);
4631       }
4632     }
4633     // Drop the speculative part of all types in the igvn's type table
4634     igvn.remove_speculative_types();
4635     if (modified > 0) {
4636       igvn.optimize();
4637     }
4638 #ifdef ASSERT
4639     // Verify that after the IGVN is over no speculative type has resurfaced
4640     worklist.clear();
4641     worklist.push(root());
4642     for (uint next = 0; next < worklist.size(); ++next) {
4643       Node *n  = worklist.at(next);
4644       const Type* t = igvn.type_or_null(n);
4645       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4646       if (n->is_Type()) {
4647         t = n->as_Type()->type();
4648         assert(t == t->remove_speculative(), "no more speculative types");
4649       }
4650       uint max = n->len();
4651       for( uint i = 0; i < max; ++i ) {
4652         Node *m = n->in(i);
4653         if (not_a_node(m))  continue;
4654         worklist.push(m);
4655       }
4656     }
4657     igvn.check_no_speculative_types();
4658 #endif
4659   }
4660 }
4661 
4662 // Auxiliary method to support randomized stressing/fuzzing.
4663 //
4664 // This method can be called the arbitrary number of times, with current count
4665 // as the argument. The logic allows selecting a single candidate from the
4666 // running list of candidates as follows:
4667 //    int count = 0;
4668 //    Cand* selected = null;
4669 //    while(cand = cand->next()) {
4670 //      if (randomized_select(++count)) {
4671 //        selected = cand;
4672 //      }
4673 //    }
4674 //
4675 // Including count equalizes the chances any candidate is "selected".
4676 // This is useful when we don't have the complete list of candidates to choose
4677 // from uniformly. In this case, we need to adjust the randomicity of the
4678 // selection, or else we will end up biasing the selection towards the latter
4679 // candidates.
4680 //
4681 // Quick back-envelope calculation shows that for the list of n candidates
4682 // the equal probability for the candidate to persist as "best" can be
4683 // achieved by replacing it with "next" k-th candidate with the probability
4684 // of 1/k. It can be easily shown that by the end of the run, the
4685 // probability for any candidate is converged to 1/n, thus giving the
4686 // uniform distribution among all the candidates.
4687 //
4688 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4689 #define RANDOMIZED_DOMAIN_POW 29
4690 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4691 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4692 bool Compile::randomized_select(int count) {
4693   assert(count > 0, "only positive");
4694   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4695 }
4696 
4697 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4698 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4699 
4700 void NodeCloneInfo::dump() const {
4701   tty->print(" {%d:%d} ", idx(), gen());
4702 }
4703 
4704 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4705   uint64_t val = value(old->_idx);
4706   NodeCloneInfo cio(val);
4707   assert(val != 0, "old node should be in the map");
4708   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4709   insert(nnn->_idx, cin.get());
4710 #ifndef PRODUCT
4711   if (is_debug()) {
4712     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4713   }
4714 #endif
4715 }
4716 
4717 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4718   NodeCloneInfo cio(value(old->_idx));
4719   if (cio.get() == 0) {
4720     cio.set(old->_idx, 0);
4721     insert(old->_idx, cio.get());
4722 #ifndef PRODUCT
4723     if (is_debug()) {
4724       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4725     }
4726 #endif
4727   }
4728   clone(old, nnn, gen);
4729 }
4730 
4731 int CloneMap::max_gen() const {
4732   int g = 0;
4733   DictI di(_dict);
4734   for(; di.test(); ++di) {
4735     int t = gen(di._key);
4736     if (g < t) {
4737       g = t;
4738 #ifndef PRODUCT
4739       if (is_debug()) {
4740         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4741       }
4742 #endif
4743     }
4744   }
4745   return g;
4746 }
4747 
4748 void CloneMap::dump(node_idx_t key) const {
4749   uint64_t val = value(key);
4750   if (val != 0) {
4751     NodeCloneInfo ni(val);
4752     ni.dump();
4753   }
4754 }