1 /*
   2  * Copyright (c) 1997, 2015, 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 "opto/addnode.hpp"
  37 #include "opto/block.hpp"
  38 #include "opto/c2compiler.hpp"
  39 #include "opto/callGenerator.hpp"
  40 #include "opto/callnode.hpp"
  41 #include "opto/castnode.hpp"
  42 #include "opto/cfgnode.hpp"
  43 #include "opto/chaitin.hpp"
  44 #include "opto/compile.hpp"
  45 #include "opto/connode.hpp"
  46 #include "opto/convertnode.hpp"
  47 #include "opto/divnode.hpp"
  48 #include "opto/escape.hpp"
  49 #include "opto/idealGraphPrinter.hpp"
  50 #include "opto/loopnode.hpp"
  51 #include "opto/machnode.hpp"
  52 #include "opto/macro.hpp"
  53 #include "opto/matcher.hpp"
  54 #include "opto/mathexactnode.hpp"
  55 #include "opto/memnode.hpp"
  56 #include "opto/mulnode.hpp"
  57 #include "opto/narrowptrnode.hpp"
  58 #include "opto/node.hpp"
  59 #include "opto/opcodes.hpp"
  60 #include "opto/output.hpp"
  61 #include "opto/parse.hpp"
  62 #include "opto/phaseX.hpp"
  63 #include "opto/rootnode.hpp"
  64 #include "opto/runtime.hpp"
  65 #include "opto/stringopts.hpp"
  66 #include "opto/type.hpp"
  67 #include "opto/valuetypenode.hpp"
  68 #include "opto/vectornode.hpp"
  69 #include "runtime/arguments.hpp"
  70 #include "runtime/sharedRuntime.hpp"
  71 #include "runtime/signature.hpp"
  72 #include "runtime/stubRoutines.hpp"
  73 #include "runtime/timer.hpp"
  74 #include "utilities/copy.hpp"
  75 
  76 
  77 // -------------------- Compile::mach_constant_base_node -----------------------
  78 // Constant table base node singleton.
  79 MachConstantBaseNode* Compile::mach_constant_base_node() {
  80   if (_mach_constant_base_node == NULL) {
  81     _mach_constant_base_node = new MachConstantBaseNode();
  82     _mach_constant_base_node->add_req(C->root());
  83   }
  84   return _mach_constant_base_node;
  85 }
  86 
  87 
  88 /// Support for intrinsics.
  89 
  90 // Return the index at which m must be inserted (or already exists).
  91 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
  92 class IntrinsicDescPair {
  93  private:
  94   ciMethod* _m;
  95   bool _is_virtual;
  96  public:
  97   IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
  98   static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
  99     ciMethod* m= elt->method();
 100     ciMethod* key_m = key->_m;
 101     if (key_m < m)      return -1;
 102     else if (key_m > m) return 1;
 103     else {
 104       bool is_virtual = elt->is_virtual();
 105       bool key_virtual = key->_is_virtual;
 106       if (key_virtual < is_virtual)      return -1;
 107       else if (key_virtual > is_virtual) return 1;
 108       else                               return 0;
 109     }
 110   }
 111 };
 112 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
 113 #ifdef ASSERT
 114   for (int i = 1; i < _intrinsics->length(); i++) {
 115     CallGenerator* cg1 = _intrinsics->at(i-1);
 116     CallGenerator* cg2 = _intrinsics->at(i);
 117     assert(cg1->method() != cg2->method()
 118            ? cg1->method()     < cg2->method()
 119            : cg1->is_virtual() < cg2->is_virtual(),
 120            "compiler intrinsics list must stay sorted");
 121   }
 122 #endif
 123   IntrinsicDescPair pair(m, is_virtual);
 124   return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
 125 }
 126 
 127 void Compile::register_intrinsic(CallGenerator* cg) {
 128   if (_intrinsics == NULL) {
 129     _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 130   }
 131   int len = _intrinsics->length();
 132   bool found = false;
 133   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
 134   assert(!found, "registering twice");
 135   _intrinsics->insert_before(index, cg);
 136   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 137 }
 138 
 139 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 140   assert(m->is_loaded(), "don't try this on unloaded methods");
 141   if (_intrinsics != NULL) {
 142     bool found = false;
 143     int index = intrinsic_insertion_index(m, is_virtual, found);
 144      if (found) {
 145       return _intrinsics->at(index);
 146     }
 147   }
 148   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 149   if (m->intrinsic_id() != vmIntrinsics::_none &&
 150       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 151     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 152     if (cg != NULL) {
 153       // Save it for next time:
 154       register_intrinsic(cg);
 155       return cg;
 156     } else {
 157       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 158     }
 159   }
 160   return NULL;
 161 }
 162 
 163 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 164 // in library_call.cpp.
 165 
 166 
 167 #ifndef PRODUCT
 168 // statistics gathering...
 169 
 170 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 171 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 172 
 173 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 174   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 175   int oflags = _intrinsic_hist_flags[id];
 176   assert(flags != 0, "what happened?");
 177   if (is_virtual) {
 178     flags |= _intrinsic_virtual;
 179   }
 180   bool changed = (flags != oflags);
 181   if ((flags & _intrinsic_worked) != 0) {
 182     juint count = (_intrinsic_hist_count[id] += 1);
 183     if (count == 1) {
 184       changed = true;           // first time
 185     }
 186     // increment the overall count also:
 187     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 188   }
 189   if (changed) {
 190     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 191       // Something changed about the intrinsic's virtuality.
 192       if ((flags & _intrinsic_virtual) != 0) {
 193         // This is the first use of this intrinsic as a virtual call.
 194         if (oflags != 0) {
 195           // We already saw it as a non-virtual, so note both cases.
 196           flags |= _intrinsic_both;
 197         }
 198       } else if ((oflags & _intrinsic_both) == 0) {
 199         // This is the first use of this intrinsic as a non-virtual
 200         flags |= _intrinsic_both;
 201       }
 202     }
 203     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 204   }
 205   // update the overall flags also:
 206   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 207   return changed;
 208 }
 209 
 210 static char* format_flags(int flags, char* buf) {
 211   buf[0] = 0;
 212   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 213   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 214   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 215   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 216   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 217   if (buf[0] == 0)  strcat(buf, ",");
 218   assert(buf[0] == ',', "must be");
 219   return &buf[1];
 220 }
 221 
 222 void Compile::print_intrinsic_statistics() {
 223   char flagsbuf[100];
 224   ttyLocker ttyl;
 225   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 226   tty->print_cr("Compiler intrinsic usage:");
 227   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 228   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 229   #define PRINT_STAT_LINE(name, c, f) \
 230     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 231   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 232     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 233     int   flags = _intrinsic_hist_flags[id];
 234     juint count = _intrinsic_hist_count[id];
 235     if ((flags | count) != 0) {
 236       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 237     }
 238   }
 239   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 240   if (xtty != NULL)  xtty->tail("statistics");
 241 }
 242 
 243 void Compile::print_statistics() {
 244   { ttyLocker ttyl;
 245     if (xtty != NULL)  xtty->head("statistics type='opto'");
 246     Parse::print_statistics();
 247     PhaseCCP::print_statistics();
 248     PhaseRegAlloc::print_statistics();
 249     Scheduling::print_statistics();
 250     PhasePeephole::print_statistics();
 251     PhaseIdealLoop::print_statistics();
 252     if (xtty != NULL)  xtty->tail("statistics");
 253   }
 254   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 255     // put this under its own <statistics> element.
 256     print_intrinsic_statistics();
 257   }
 258 }
 259 #endif //PRODUCT
 260 
 261 // Support for bundling info
 262 Bundle* Compile::node_bundling(const Node *n) {
 263   assert(valid_bundle_info(n), "oob");
 264   return &_node_bundling_base[n->_idx];
 265 }
 266 
 267 bool Compile::valid_bundle_info(const Node *n) {
 268   return (_node_bundling_limit > n->_idx);
 269 }
 270 
 271 
 272 void Compile::gvn_replace_by(Node* n, Node* nn) {
 273   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 274     Node* use = n->last_out(i);
 275     bool is_in_table = initial_gvn()->hash_delete(use);
 276     uint uses_found = 0;
 277     for (uint j = 0; j < use->len(); j++) {
 278       if (use->in(j) == n) {
 279         if (j < use->req())
 280           use->set_req(j, nn);
 281         else
 282           use->set_prec(j, nn);
 283         uses_found++;
 284       }
 285     }
 286     if (is_in_table) {
 287       // reinsert into table
 288       initial_gvn()->hash_find_insert(use);
 289     }
 290     record_for_igvn(use);
 291     i -= uses_found;    // we deleted 1 or more copies of this edge
 292   }
 293 }
 294 
 295 
 296 static inline bool not_a_node(const Node* n) {
 297   if (n == NULL)                   return true;
 298   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 299   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 300   return false;
 301 }
 302 
 303 // Identify all nodes that are reachable from below, useful.
 304 // Use breadth-first pass that records state in a Unique_Node_List,
 305 // recursive traversal is slower.
 306 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 307   int estimated_worklist_size = live_nodes();
 308   useful.map( estimated_worklist_size, NULL );  // preallocate space
 309 
 310   // Initialize worklist
 311   if (root() != NULL)     { useful.push(root()); }
 312   // If 'top' is cached, declare it useful to preserve cached node
 313   if( cached_top_node() ) { useful.push(cached_top_node()); }
 314 
 315   // Push all useful nodes onto the list, breadthfirst
 316   for( uint next = 0; next < useful.size(); ++next ) {
 317     assert( next < unique(), "Unique useful nodes < total nodes");
 318     Node *n  = useful.at(next);
 319     uint max = n->len();
 320     for( uint i = 0; i < max; ++i ) {
 321       Node *m = n->in(i);
 322       if (not_a_node(m))  continue;
 323       useful.push(m);
 324     }
 325   }
 326 }
 327 
 328 // Update dead_node_list with any missing dead nodes using useful
 329 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 330 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 331   uint max_idx = unique();
 332   VectorSet& useful_node_set = useful.member_set();
 333 
 334   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 335     // If node with index node_idx is not in useful set,
 336     // mark it as dead in dead node list.
 337     if (! useful_node_set.test(node_idx) ) {
 338       record_dead_node(node_idx);
 339     }
 340   }
 341 }
 342 
 343 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 344   int shift = 0;
 345   for (int i = 0; i < inlines->length(); i++) {
 346     CallGenerator* cg = inlines->at(i);
 347     CallNode* call = cg->call_node();
 348     if (shift > 0) {
 349       inlines->at_put(i-shift, cg);
 350     }
 351     if (!useful.member(call)) {
 352       shift++;
 353     }
 354   }
 355   inlines->trunc_to(inlines->length()-shift);
 356 }
 357 
 358 // Disconnect all useless nodes by disconnecting those at the boundary.
 359 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 360   uint next = 0;
 361   while (next < useful.size()) {
 362     Node *n = useful.at(next++);
 363     if (n->is_SafePoint()) {
 364       // We're done with a parsing phase. Replaced nodes are not valid
 365       // beyond that point.
 366       n->as_SafePoint()->delete_replaced_nodes();
 367     }
 368     // Use raw traversal of out edges since this code removes out edges
 369     int max = n->outcnt();
 370     for (int j = 0; j < max; ++j) {
 371       Node* child = n->raw_out(j);
 372       if (! useful.member(child)) {
 373         assert(!child->is_top() || child != top(),
 374                "If top is cached in Compile object it is in useful list");
 375         // Only need to remove this out-edge to the useless node
 376         n->raw_del_out(j);
 377         --j;
 378         --max;
 379       }
 380     }
 381     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 382       record_for_igvn(n->unique_out());
 383     }
 384   }
 385   // Remove useless macro and predicate opaq nodes
 386   for (int i = C->macro_count()-1; i >= 0; i--) {
 387     Node* n = C->macro_node(i);
 388     if (!useful.member(n)) {
 389       remove_macro_node(n);
 390     }
 391   }
 392   // Remove useless CastII nodes with range check dependency
 393   for (int i = range_check_cast_count() - 1; i >= 0; i--) {
 394     Node* cast = range_check_cast_node(i);
 395     if (!useful.member(cast)) {
 396       remove_range_check_cast(cast);
 397     }
 398   }
 399   // Remove useless expensive node
 400   for (int i = C->expensive_count()-1; i >= 0; i--) {
 401     Node* n = C->expensive_node(i);
 402     if (!useful.member(n)) {
 403       remove_expensive_node(n);
 404     }
 405   }
 406   // clean up the late inline lists
 407   remove_useless_late_inlines(&_string_late_inlines, useful);
 408   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 409   remove_useless_late_inlines(&_late_inlines, useful);
 410   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 411 }
 412 
 413 //------------------------------frame_size_in_words-----------------------------
 414 // frame_slots in units of words
 415 int Compile::frame_size_in_words() const {
 416   // shift is 0 in LP32 and 1 in LP64
 417   const int shift = (LogBytesPerWord - LogBytesPerInt);
 418   int words = _frame_slots >> shift;
 419   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 420   return words;
 421 }
 422 
 423 // To bang the stack of this compiled method we use the stack size
 424 // that the interpreter would need in case of a deoptimization. This
 425 // removes the need to bang the stack in the deoptimization blob which
 426 // in turn simplifies stack overflow handling.
 427 int Compile::bang_size_in_bytes() const {
 428   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
 429 }
 430 
 431 // ============================================================================
 432 //------------------------------CompileWrapper---------------------------------
 433 class CompileWrapper : public StackObj {
 434   Compile *const _compile;
 435  public:
 436   CompileWrapper(Compile* compile);
 437 
 438   ~CompileWrapper();
 439 };
 440 
 441 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 442   // the Compile* pointer is stored in the current ciEnv:
 443   ciEnv* env = compile->env();
 444   assert(env == ciEnv::current(), "must already be a ciEnv active");
 445   assert(env->compiler_data() == NULL, "compile already active?");
 446   env->set_compiler_data(compile);
 447   assert(compile == Compile::current(), "sanity");
 448 
 449   compile->set_type_dict(NULL);
 450   compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
 451   compile->clone_map().set_clone_idx(0);
 452   compile->set_type_hwm(NULL);
 453   compile->set_type_last_size(0);
 454   compile->set_last_tf(NULL, NULL);
 455   compile->set_indexSet_arena(NULL);
 456   compile->set_indexSet_free_block_list(NULL);
 457   compile->init_type_arena();
 458   Type::Initialize(compile);
 459   _compile->set_scratch_buffer_blob(NULL);
 460   _compile->begin_method();
 461   _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
 462 }
 463 CompileWrapper::~CompileWrapper() {
 464   _compile->end_method();
 465   if (_compile->scratch_buffer_blob() != NULL)
 466     BufferBlob::free(_compile->scratch_buffer_blob());
 467   _compile->env()->set_compiler_data(NULL);
 468 }
 469 
 470 
 471 //----------------------------print_compile_messages---------------------------
 472 void Compile::print_compile_messages() {
 473 #ifndef PRODUCT
 474   // Check if recompiling
 475   if (_subsume_loads == false && PrintOpto) {
 476     // Recompiling without allowing machine instructions to subsume loads
 477     tty->print_cr("*********************************************************");
 478     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 479     tty->print_cr("*********************************************************");
 480   }
 481   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 482     // Recompiling without escape analysis
 483     tty->print_cr("*********************************************************");
 484     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 485     tty->print_cr("*********************************************************");
 486   }
 487   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 488     // Recompiling without boxing elimination
 489     tty->print_cr("*********************************************************");
 490     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 491     tty->print_cr("*********************************************************");
 492   }
 493   if (C->directive()->BreakAtCompileOption) {
 494     // Open the debugger when compiling this method.
 495     tty->print("### Breaking when compiling: ");
 496     method()->print_short_name();
 497     tty->cr();
 498     BREAKPOINT;
 499   }
 500 
 501   if( PrintOpto ) {
 502     if (is_osr_compilation()) {
 503       tty->print("[OSR]%3d", _compile_id);
 504     } else {
 505       tty->print("%3d", _compile_id);
 506     }
 507   }
 508 #endif
 509 }
 510 
 511 
 512 //-----------------------init_scratch_buffer_blob------------------------------
 513 // Construct a temporary BufferBlob and cache it for this compile.
 514 void Compile::init_scratch_buffer_blob(int const_size) {
 515   // If there is already a scratch buffer blob allocated and the
 516   // constant section is big enough, use it.  Otherwise free the
 517   // current and allocate a new one.
 518   BufferBlob* blob = scratch_buffer_blob();
 519   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 520     // Use the current blob.
 521   } else {
 522     if (blob != NULL) {
 523       BufferBlob::free(blob);
 524     }
 525 
 526     ResourceMark rm;
 527     _scratch_const_size = const_size;
 528     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
 529     blob = BufferBlob::create("Compile::scratch_buffer", size);
 530     // Record the buffer blob for next time.
 531     set_scratch_buffer_blob(blob);
 532     // Have we run out of code space?
 533     if (scratch_buffer_blob() == NULL) {
 534       // Let CompilerBroker disable further compilations.
 535       record_failure("Not enough space for scratch buffer in CodeCache");
 536       return;
 537     }
 538   }
 539 
 540   // Initialize the relocation buffers
 541   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 542   set_scratch_locs_memory(locs_buf);
 543 }
 544 
 545 
 546 //-----------------------scratch_emit_size-------------------------------------
 547 // Helper function that computes size by emitting code
 548 uint Compile::scratch_emit_size(const Node* n) {
 549   // Start scratch_emit_size section.
 550   set_in_scratch_emit_size(true);
 551 
 552   // Emit into a trash buffer and count bytes emitted.
 553   // This is a pretty expensive way to compute a size,
 554   // but it works well enough if seldom used.
 555   // All common fixed-size instructions are given a size
 556   // method by the AD file.
 557   // Note that the scratch buffer blob and locs memory are
 558   // allocated at the beginning of the compile task, and
 559   // may be shared by several calls to scratch_emit_size.
 560   // The allocation of the scratch buffer blob is particularly
 561   // expensive, since it has to grab the code cache lock.
 562   BufferBlob* blob = this->scratch_buffer_blob();
 563   assert(blob != NULL, "Initialize BufferBlob at start");
 564   assert(blob->size() > MAX_inst_size, "sanity");
 565   relocInfo* locs_buf = scratch_locs_memory();
 566   address blob_begin = blob->content_begin();
 567   address blob_end   = (address)locs_buf;
 568   assert(blob->content_contains(blob_end), "sanity");
 569   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 570   buf.initialize_consts_size(_scratch_const_size);
 571   buf.initialize_stubs_size(MAX_stubs_size);
 572   assert(locs_buf != NULL, "sanity");
 573   int lsize = MAX_locs_size / 3;
 574   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 575   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 576   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 577 
 578   // Do the emission.
 579 
 580   Label fakeL; // Fake label for branch instructions.
 581   Label*   saveL = NULL;
 582   uint save_bnum = 0;
 583   bool is_branch = n->is_MachBranch();
 584   if (is_branch) {
 585     MacroAssembler masm(&buf);
 586     masm.bind(fakeL);
 587     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 588     n->as_MachBranch()->label_set(&fakeL, 0);
 589   }
 590   n->emit(buf, this->regalloc());
 591 
 592   // Emitting into the scratch buffer should not fail
 593   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
 594 
 595   if (is_branch) // Restore label.
 596     n->as_MachBranch()->label_set(saveL, save_bnum);
 597 
 598   // End scratch_emit_size section.
 599   set_in_scratch_emit_size(false);
 600 
 601   return buf.insts_size();
 602 }
 603 
 604 
 605 // ============================================================================
 606 //------------------------------Compile standard-------------------------------
 607 debug_only( int Compile::_debug_idx = 100000; )
 608 
 609 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 610 // the continuation bci for on stack replacement.
 611 
 612 
 613 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 614                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
 615                 : Phase(Compiler),
 616                   _env(ci_env),
 617                   _directive(directive),
 618                   _log(ci_env->log()),
 619                   _compile_id(ci_env->compile_id()),
 620                   _save_argument_registers(false),
 621                   _stub_name(NULL),
 622                   _stub_function(NULL),
 623                   _stub_entry_point(NULL),
 624                   _method(target),
 625                   _entry_bci(osr_bci),
 626                   _initial_gvn(NULL),
 627                   _for_igvn(NULL),
 628                   _warm_calls(NULL),
 629                   _subsume_loads(subsume_loads),
 630                   _do_escape_analysis(do_escape_analysis),
 631                   _eliminate_boxing(eliminate_boxing),
 632                   _failure_reason(NULL),
 633                   _code_buffer("Compile::Fill_buffer"),
 634                   _orig_pc_slot(0),
 635                   _orig_pc_slot_offset_in_bytes(0),
 636                   _has_method_handle_invokes(false),
 637                   _mach_constant_base_node(NULL),
 638                   _node_bundling_limit(0),
 639                   _node_bundling_base(NULL),
 640                   _java_calls(0),
 641                   _inner_loops(0),
 642                   _scratch_const_size(-1),
 643                   _in_scratch_emit_size(false),
 644                   _dead_node_list(comp_arena()),
 645                   _dead_node_count(0),
 646 #ifndef PRODUCT
 647                   _trace_opto_output(directive->TraceOptoOutputOption),
 648                   _in_dump_cnt(0),
 649                   _printer(IdealGraphPrinter::printer()),
 650 #endif
 651                   _congraph(NULL),
 652                   _comp_arena(mtCompiler),
 653                   _node_arena(mtCompiler),
 654                   _old_arena(mtCompiler),
 655                   _Compile_types(mtCompiler),
 656                   _replay_inline_data(NULL),
 657                   _late_inlines(comp_arena(), 2, 0, NULL),
 658                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 659                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 660                   _late_inlines_pos(0),
 661                   _number_of_mh_late_inlines(0),
 662                   _inlining_progress(false),
 663                   _inlining_incrementally(false),
 664                   _print_inlining_list(NULL),
 665                   _print_inlining_stream(NULL),
 666                   _print_inlining_idx(0),
 667                   _print_inlining_output(NULL),
 668                   _interpreter_frame_size(0),
 669                   _max_node_limit(MaxNodeLimit),
 670                   _has_reserved_stack_access(target->has_reserved_stack_access()) {
 671   C = this;
 672 #ifndef PRODUCT
 673   if (_printer != NULL) {
 674     _printer->set_compile(this);
 675   }
 676 #endif
 677   CompileWrapper cw(this);
 678 
 679   if (CITimeVerbose) {
 680     tty->print(" ");
 681     target->holder()->name()->print();
 682     tty->print(".");
 683     target->print_short_name();
 684     tty->print("  ");
 685   }
 686   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 687   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 688 
 689 #ifndef PRODUCT
 690   bool print_opto_assembly = directive->PrintOptoAssemblyOption;
 691   if (!print_opto_assembly) {
 692     bool print_assembly = directive->PrintAssemblyOption;
 693     if (print_assembly && !Disassembler::can_decode()) {
 694       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 695       print_opto_assembly = true;
 696     }
 697   }
 698   set_print_assembly(print_opto_assembly);
 699   set_parsed_irreducible_loop(false);
 700 
 701   if (directive->ReplayInlineOption) {
 702     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 703   }
 704 #endif
 705   set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
 706   set_print_intrinsics(directive->PrintIntrinsicsOption);
 707   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 708 
 709   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 710     // Make sure the method being compiled gets its own MDO,
 711     // so we can at least track the decompile_count().
 712     // Need MDO to record RTM code generation state.
 713     method()->ensure_method_data();
 714   }
 715 
 716   Init(::AliasLevel);
 717 
 718 
 719   print_compile_messages();
 720 
 721   _ilt = InlineTree::build_inline_tree_root();
 722 
 723   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 724   assert(num_alias_types() >= AliasIdxRaw, "");
 725 
 726 #define MINIMUM_NODE_HASH  1023
 727   // Node list that Iterative GVN will start with
 728   Unique_Node_List for_igvn(comp_arena());
 729   set_for_igvn(&for_igvn);
 730 
 731   // GVN that will be run immediately on new nodes
 732   uint estimated_size = method()->code_size()*4+64;
 733   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 734   PhaseGVN gvn(node_arena(), estimated_size);
 735   set_initial_gvn(&gvn);
 736 
 737   print_inlining_init();
 738   { // Scope for timing the parser
 739     TracePhase tp("parse", &timers[_t_parser]);
 740 
 741     // Put top into the hash table ASAP.
 742     initial_gvn()->transform_no_reclaim(top());
 743 
 744     // Set up tf(), start(), and find a CallGenerator.
 745     CallGenerator* cg = NULL;
 746     if (is_osr_compilation()) {
 747       const TypeTuple *domain = StartOSRNode::osr_domain();
 748       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 749       init_tf(TypeFunc::make(domain, range));
 750       StartNode* s = new StartOSRNode(root(), domain);
 751       initial_gvn()->set_type_bottom(s);
 752       init_start(s);
 753       cg = CallGenerator::for_osr(method(), entry_bci());
 754     } else {
 755       // Normal case.
 756       init_tf(TypeFunc::make(method()));
 757       StartNode* s = new StartNode(root(), tf()->domain_cc());
 758       initial_gvn()->set_type_bottom(s);
 759       init_start(s);
 760       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
 761         // With java.lang.ref.reference.get() we must go through the
 762         // intrinsic when G1 is enabled - even when get() is the root
 763         // method of the compile - so that, if necessary, the value in
 764         // the referent field of the reference object gets recorded by
 765         // the pre-barrier code.
 766         // Specifically, if G1 is enabled, the value in the referent
 767         // field is recorded by the G1 SATB pre barrier. This will
 768         // result in the referent being marked live and the reference
 769         // object removed from the list of discovered references during
 770         // reference processing.
 771         cg = find_intrinsic(method(), false);
 772       }
 773       if (cg == NULL) {
 774         float past_uses = method()->interpreter_invocation_count();
 775         float expected_uses = past_uses;
 776         cg = CallGenerator::for_inline(method(), expected_uses);
 777       }
 778     }
 779     if (failing())  return;
 780     if (cg == NULL) {
 781       record_method_not_compilable_all_tiers("cannot parse method");
 782       return;
 783     }
 784     JVMState* jvms = build_start_state(start(), tf());
 785     if ((jvms = cg->generate(jvms)) == NULL) {
 786       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 787         record_method_not_compilable("method parse failed");
 788       }
 789       return;
 790     }
 791     GraphKit kit(jvms);
 792 
 793     if (!kit.stopped()) {
 794       // Accept return values, and transfer control we know not where.
 795       // This is done by a special, unique ReturnNode bound to root.
 796       return_values(kit.jvms());
 797     }
 798 
 799     if (kit.has_exceptions()) {
 800       // Any exceptions that escape from this call must be rethrown
 801       // to whatever caller is dynamically above us on the stack.
 802       // This is done by a special, unique RethrowNode bound to root.
 803       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 804     }
 805 
 806     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 807 
 808     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 809       inline_string_calls(true);
 810     }
 811 
 812     if (failing())  return;
 813 
 814     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 815 
 816     // Remove clutter produced by parsing.
 817     if (!failing()) {
 818       ResourceMark rm;
 819       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 820     }
 821   }
 822 
 823   // Note:  Large methods are capped off in do_one_bytecode().
 824   if (failing())  return;
 825 
 826   // After parsing, node notes are no longer automagic.
 827   // They must be propagated by register_new_node_with_optimizer(),
 828   // clone(), or the like.
 829   set_default_node_notes(NULL);
 830 
 831   for (;;) {
 832     int successes = Inline_Warm();
 833     if (failing())  return;
 834     if (successes == 0)  break;
 835   }
 836 
 837   // Drain the list.
 838   Finish_Warm();
 839 #ifndef PRODUCT
 840   if (_printer && _printer->should_print(1)) {
 841     _printer->print_inlining();
 842   }
 843 #endif
 844 
 845   if (failing())  return;
 846   NOT_PRODUCT( verify_graph_edges(); )
 847 
 848   // Now optimize
 849   Optimize();
 850   if (failing())  return;
 851   NOT_PRODUCT( verify_graph_edges(); )
 852 
 853 #ifndef PRODUCT
 854   if (PrintIdeal) {
 855     ttyLocker ttyl;  // keep the following output all in one block
 856     // This output goes directly to the tty, not the compiler log.
 857     // To enable tools to match it up with the compilation activity,
 858     // be sure to tag this tty output with the compile ID.
 859     if (xtty != NULL) {
 860       xtty->head("ideal compile_id='%d'%s", compile_id(),
 861                  is_osr_compilation()    ? " compile_kind='osr'" :
 862                  "");
 863     }
 864     root()->dump(9999);
 865     if (xtty != NULL) {
 866       xtty->tail("ideal");
 867     }
 868   }
 869 #endif
 870 
 871   NOT_PRODUCT( verify_barriers(); )
 872 
 873   // Dump compilation data to replay it.
 874   if (directive->DumpReplayOption) {
 875     env()->dump_replay_data(_compile_id);
 876   }
 877   if (directive->DumpInlineOption && (ilt() != NULL)) {
 878     env()->dump_inline_data(_compile_id);
 879   }
 880 
 881   // Now that we know the size of all the monitors we can add a fixed slot
 882   // for the original deopt pc.
 883 
 884   _orig_pc_slot =  fixed_slots();
 885   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 886   set_fixed_slots(next_slot);
 887 
 888   // Compute when to use implicit null checks. Used by matching trap based
 889   // nodes and NullCheck optimization.
 890   set_allowed_deopt_reasons();
 891 
 892   // Now generate code
 893   Code_Gen();
 894   if (failing())  return;
 895 
 896   // Check if we want to skip execution of all compiled code.
 897   {
 898 #ifndef PRODUCT
 899     if (OptoNoExecute) {
 900       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 901       return;
 902     }
 903 #endif
 904     TracePhase tp("install_code", &timers[_t_registerMethod]);
 905 
 906     if (is_osr_compilation()) {
 907       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 908       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 909     } else {
 910       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 911       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 912     }
 913 
 914     env()->register_method(_method, _entry_bci,
 915                            &_code_offsets,
 916                            _orig_pc_slot_offset_in_bytes,
 917                            code_buffer(),
 918                            frame_size_in_words(), _oop_map_set,
 919                            &_handler_table, &_inc_table,
 920                            compiler,
 921                            has_unsafe_access(),
 922                            SharedRuntime::is_wide_vector(max_vector_size()),
 923                            rtm_state()
 924                            );
 925 
 926     if (log() != NULL) // Print code cache state into compiler log
 927       log()->code_cache_state();
 928   }
 929 }
 930 
 931 //------------------------------Compile----------------------------------------
 932 // Compile a runtime stub
 933 Compile::Compile( ciEnv* ci_env,
 934                   TypeFunc_generator generator,
 935                   address stub_function,
 936                   const char *stub_name,
 937                   int is_fancy_jump,
 938                   bool pass_tls,
 939                   bool save_arg_registers,
 940                   bool return_pc,
 941                   DirectiveSet* directive)
 942   : Phase(Compiler),
 943     _env(ci_env),
 944     _directive(directive),
 945     _log(ci_env->log()),
 946     _compile_id(0),
 947     _save_argument_registers(save_arg_registers),
 948     _method(NULL),
 949     _stub_name(stub_name),
 950     _stub_function(stub_function),
 951     _stub_entry_point(NULL),
 952     _entry_bci(InvocationEntryBci),
 953     _initial_gvn(NULL),
 954     _for_igvn(NULL),
 955     _warm_calls(NULL),
 956     _orig_pc_slot(0),
 957     _orig_pc_slot_offset_in_bytes(0),
 958     _subsume_loads(true),
 959     _do_escape_analysis(false),
 960     _eliminate_boxing(false),
 961     _failure_reason(NULL),
 962     _code_buffer("Compile::Fill_buffer"),
 963     _has_method_handle_invokes(false),
 964     _mach_constant_base_node(NULL),
 965     _node_bundling_limit(0),
 966     _node_bundling_base(NULL),
 967     _java_calls(0),
 968     _inner_loops(0),
 969 #ifndef PRODUCT
 970     _trace_opto_output(TraceOptoOutput),
 971     _in_dump_cnt(0),
 972     _printer(NULL),
 973 #endif
 974     _comp_arena(mtCompiler),
 975     _node_arena(mtCompiler),
 976     _old_arena(mtCompiler),
 977     _Compile_types(mtCompiler),
 978     _dead_node_list(comp_arena()),
 979     _dead_node_count(0),
 980     _congraph(NULL),
 981     _replay_inline_data(NULL),
 982     _number_of_mh_late_inlines(0),
 983     _inlining_progress(false),
 984     _inlining_incrementally(false),
 985     _print_inlining_list(NULL),
 986     _print_inlining_stream(NULL),
 987     _print_inlining_idx(0),
 988     _print_inlining_output(NULL),
 989     _allowed_reasons(0),
 990     _interpreter_frame_size(0),
 991     _max_node_limit(MaxNodeLimit) {
 992   C = this;
 993 
 994   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
 995   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
 996 
 997 #ifndef PRODUCT
 998   set_print_assembly(PrintFrameConverterAssembly);
 999   set_parsed_irreducible_loop(false);
1000 #endif
1001   set_has_irreducible_loop(false); // no loops
1002 
1003   CompileWrapper cw(this);
1004   Init(/*AliasLevel=*/ 0);
1005   init_tf((*generator)());
1006 
1007   {
1008     // The following is a dummy for the sake of GraphKit::gen_stub
1009     Unique_Node_List for_igvn(comp_arena());
1010     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1011     PhaseGVN gvn(Thread::current()->resource_area(),255);
1012     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1013     gvn.transform_no_reclaim(top());
1014 
1015     GraphKit kit;
1016     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1017   }
1018 
1019   NOT_PRODUCT( verify_graph_edges(); )
1020   Code_Gen();
1021   if (failing())  return;
1022 
1023 
1024   // Entry point will be accessed using compile->stub_entry_point();
1025   if (code_buffer() == NULL) {
1026     Matcher::soft_match_failure();
1027   } else {
1028     if (PrintAssembly && (WizardMode || Verbose))
1029       tty->print_cr("### Stub::%s", stub_name);
1030 
1031     if (!failing()) {
1032       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1033 
1034       // Make the NMethod
1035       // For now we mark the frame as never safe for profile stackwalking
1036       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1037                                                       code_buffer(),
1038                                                       CodeOffsets::frame_never_safe,
1039                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1040                                                       frame_size_in_words(),
1041                                                       _oop_map_set,
1042                                                       save_arg_registers);
1043       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1044 
1045       _stub_entry_point = rs->entry_point();
1046     }
1047   }
1048 }
1049 
1050 //------------------------------Init-------------------------------------------
1051 // Prepare for a single compilation
1052 void Compile::Init(int aliaslevel) {
1053   _unique  = 0;
1054   _regalloc = NULL;
1055 
1056   _tf      = NULL;  // filled in later
1057   _top     = NULL;  // cached later
1058   _matcher = NULL;  // filled in later
1059   _cfg     = NULL;  // filled in later
1060 
1061   set_24_bit_selection_and_mode(Use24BitFP, false);
1062 
1063   _node_note_array = NULL;
1064   _default_node_notes = NULL;
1065   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1066 
1067   _immutable_memory = NULL; // filled in at first inquiry
1068 
1069   // Globally visible Nodes
1070   // First set TOP to NULL to give safe behavior during creation of RootNode
1071   set_cached_top_node(NULL);
1072   set_root(new RootNode());
1073   // Now that you have a Root to point to, create the real TOP
1074   set_cached_top_node( new ConNode(Type::TOP) );
1075   set_recent_alloc(NULL, NULL);
1076 
1077   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1078   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1079   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1080   env()->set_dependencies(new Dependencies(env()));
1081 
1082   _fixed_slots = 0;
1083   set_has_split_ifs(false);
1084   set_has_loops(has_method() && method()->has_loops()); // first approximation
1085   set_has_stringbuilder(false);
1086   set_has_boxed_value(false);
1087   _trap_can_recompile = false;  // no traps emitted yet
1088   _major_progress = true; // start out assuming good things will happen
1089   set_has_unsafe_access(false);
1090   set_max_vector_size(0);
1091   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1092   set_decompile_count(0);
1093 
1094   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1095   set_num_loop_opts(LoopOptsCount);
1096   set_do_inlining(Inline);
1097   set_max_inline_size(MaxInlineSize);
1098   set_freq_inline_size(FreqInlineSize);
1099   set_do_scheduling(OptoScheduling);
1100   set_do_count_invocations(false);
1101   set_do_method_data_update(false);
1102 
1103   set_do_vector_loop(false);
1104 
1105   if (AllowVectorizeOnDemand) {
1106     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1107       set_do_vector_loop(true);
1108       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());})
1109     } else if (has_method() && method()->name() != 0 &&
1110                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1111       set_do_vector_loop(true);
1112     }
1113   }
1114   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1115   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());})
1116 
1117   set_age_code(has_method() && method()->profile_aging());
1118   set_rtm_state(NoRTM); // No RTM lock eliding by default
1119   _max_node_limit = _directive->MaxNodeLimitOption;
1120 
1121 #if INCLUDE_RTM_OPT
1122   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1123     int rtm_state = method()->method_data()->rtm_state();
1124     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1125       // Don't generate RTM lock eliding code.
1126       set_rtm_state(NoRTM);
1127     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1128       // Generate RTM lock eliding code without abort ratio calculation code.
1129       set_rtm_state(UseRTM);
1130     } else if (UseRTMDeopt) {
1131       // Generate RTM lock eliding code and include abort ratio calculation
1132       // code if UseRTMDeopt is on.
1133       set_rtm_state(ProfileRTM);
1134     }
1135   }
1136 #endif
1137   if (debug_info()->recording_non_safepoints()) {
1138     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1139                         (comp_arena(), 8, 0, NULL));
1140     set_default_node_notes(Node_Notes::make(this));
1141   }
1142 
1143   // // -- Initialize types before each compile --
1144   // // Update cached type information
1145   // if( _method && _method->constants() )
1146   //   Type::update_loaded_types(_method, _method->constants());
1147 
1148   // Init alias_type map.
1149   if (!_do_escape_analysis && aliaslevel == 3)
1150     aliaslevel = 2;  // No unique types without escape analysis
1151   _AliasLevel = aliaslevel;
1152   const int grow_ats = 16;
1153   _max_alias_types = grow_ats;
1154   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1155   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1156   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1157   {
1158     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1159   }
1160   // Initialize the first few types.
1161   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1162   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1163   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1164   _num_alias_types = AliasIdxRaw+1;
1165   // Zero out the alias type cache.
1166   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1167   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1168   probe_alias_cache(NULL)->_index = AliasIdxTop;
1169 
1170   _intrinsics = NULL;
1171   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1172   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1173   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1174   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1175   register_library_intrinsics();
1176 }
1177 
1178 //---------------------------init_start----------------------------------------
1179 // Install the StartNode on this compile object.
1180 void Compile::init_start(StartNode* s) {
1181   if (failing())
1182     return; // already failing
1183   assert(s == start(), "");
1184 }
1185 
1186 /**
1187  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1188  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1189  * the ideal graph.
1190  */
1191 StartNode* Compile::start() const {
1192   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1193   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1194     Node* start = root()->fast_out(i);
1195     if (start->is_Start()) {
1196       return start->as_Start();
1197     }
1198   }
1199   fatal("Did not find Start node!");
1200   return NULL;
1201 }
1202 
1203 //-------------------------------immutable_memory-------------------------------------
1204 // Access immutable memory
1205 Node* Compile::immutable_memory() {
1206   if (_immutable_memory != NULL) {
1207     return _immutable_memory;
1208   }
1209   StartNode* s = start();
1210   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1211     Node *p = s->fast_out(i);
1212     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1213       _immutable_memory = p;
1214       return _immutable_memory;
1215     }
1216   }
1217   ShouldNotReachHere();
1218   return NULL;
1219 }
1220 
1221 //----------------------set_cached_top_node------------------------------------
1222 // Install the cached top node, and make sure Node::is_top works correctly.
1223 void Compile::set_cached_top_node(Node* tn) {
1224   if (tn != NULL)  verify_top(tn);
1225   Node* old_top = _top;
1226   _top = tn;
1227   // Calling Node::setup_is_top allows the nodes the chance to adjust
1228   // their _out arrays.
1229   if (_top != NULL)     _top->setup_is_top();
1230   if (old_top != NULL)  old_top->setup_is_top();
1231   assert(_top == NULL || top()->is_top(), "");
1232 }
1233 
1234 #ifdef ASSERT
1235 uint Compile::count_live_nodes_by_graph_walk() {
1236   Unique_Node_List useful(comp_arena());
1237   // Get useful node list by walking the graph.
1238   identify_useful_nodes(useful);
1239   return useful.size();
1240 }
1241 
1242 void Compile::print_missing_nodes() {
1243 
1244   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1245   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1246     return;
1247   }
1248 
1249   // This is an expensive function. It is executed only when the user
1250   // specifies VerifyIdealNodeCount option or otherwise knows the
1251   // additional work that needs to be done to identify reachable nodes
1252   // by walking the flow graph and find the missing ones using
1253   // _dead_node_list.
1254 
1255   Unique_Node_List useful(comp_arena());
1256   // Get useful node list by walking the graph.
1257   identify_useful_nodes(useful);
1258 
1259   uint l_nodes = C->live_nodes();
1260   uint l_nodes_by_walk = useful.size();
1261 
1262   if (l_nodes != l_nodes_by_walk) {
1263     if (_log != NULL) {
1264       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1265       _log->stamp();
1266       _log->end_head();
1267     }
1268     VectorSet& useful_member_set = useful.member_set();
1269     int last_idx = l_nodes_by_walk;
1270     for (int i = 0; i < last_idx; i++) {
1271       if (useful_member_set.test(i)) {
1272         if (_dead_node_list.test(i)) {
1273           if (_log != NULL) {
1274             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1275           }
1276           if (PrintIdealNodeCount) {
1277             // Print the log message to tty
1278               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1279               useful.at(i)->dump();
1280           }
1281         }
1282       }
1283       else if (! _dead_node_list.test(i)) {
1284         if (_log != NULL) {
1285           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1286         }
1287         if (PrintIdealNodeCount) {
1288           // Print the log message to tty
1289           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1290         }
1291       }
1292     }
1293     if (_log != NULL) {
1294       _log->tail("mismatched_nodes");
1295     }
1296   }
1297 }
1298 void Compile::record_modified_node(Node* n) {
1299   if (_modified_nodes != NULL && !_inlining_incrementally &&
1300       n->outcnt() != 0 && !n->is_Con()) {
1301     _modified_nodes->push(n);
1302   }
1303 }
1304 
1305 void Compile::remove_modified_node(Node* n) {
1306   if (_modified_nodes != NULL) {
1307     _modified_nodes->remove(n);
1308   }
1309 }
1310 #endif
1311 
1312 #ifndef PRODUCT
1313 void Compile::verify_top(Node* tn) const {
1314   if (tn != NULL) {
1315     assert(tn->is_Con(), "top node must be a constant");
1316     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1317     assert(tn->in(0) != NULL, "must have live top node");
1318   }
1319 }
1320 #endif
1321 
1322 
1323 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1324 
1325 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1326   guarantee(arr != NULL, "");
1327   int num_blocks = arr->length();
1328   if (grow_by < num_blocks)  grow_by = num_blocks;
1329   int num_notes = grow_by * _node_notes_block_size;
1330   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1331   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1332   while (num_notes > 0) {
1333     arr->append(notes);
1334     notes     += _node_notes_block_size;
1335     num_notes -= _node_notes_block_size;
1336   }
1337   assert(num_notes == 0, "exact multiple, please");
1338 }
1339 
1340 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1341   if (source == NULL || dest == NULL)  return false;
1342 
1343   if (dest->is_Con())
1344     return false;               // Do not push debug info onto constants.
1345 
1346 #ifdef ASSERT
1347   // Leave a bread crumb trail pointing to the original node:
1348   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1349     dest->set_debug_orig(source);
1350   }
1351 #endif
1352 
1353   if (node_note_array() == NULL)
1354     return false;               // Not collecting any notes now.
1355 
1356   // This is a copy onto a pre-existing node, which may already have notes.
1357   // If both nodes have notes, do not overwrite any pre-existing notes.
1358   Node_Notes* source_notes = node_notes_at(source->_idx);
1359   if (source_notes == NULL || source_notes->is_clear())  return false;
1360   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1361   if (dest_notes == NULL || dest_notes->is_clear()) {
1362     return set_node_notes_at(dest->_idx, source_notes);
1363   }
1364 
1365   Node_Notes merged_notes = (*source_notes);
1366   // The order of operations here ensures that dest notes will win...
1367   merged_notes.update_from(dest_notes);
1368   return set_node_notes_at(dest->_idx, &merged_notes);
1369 }
1370 
1371 
1372 //--------------------------allow_range_check_smearing-------------------------
1373 // Gating condition for coalescing similar range checks.
1374 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1375 // single covering check that is at least as strong as any of them.
1376 // If the optimization succeeds, the simplified (strengthened) range check
1377 // will always succeed.  If it fails, we will deopt, and then give up
1378 // on the optimization.
1379 bool Compile::allow_range_check_smearing() const {
1380   // If this method has already thrown a range-check,
1381   // assume it was because we already tried range smearing
1382   // and it failed.
1383   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1384   return !already_trapped;
1385 }
1386 
1387 
1388 //------------------------------flatten_alias_type-----------------------------
1389 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1390   int offset = tj->offset();
1391   TypePtr::PTR ptr = tj->ptr();
1392 
1393   // Known instance (scalarizable allocation) alias only with itself.
1394   bool is_known_inst = tj->isa_oopptr() != NULL &&
1395                        tj->is_oopptr()->is_known_instance();
1396 
1397   // Process weird unsafe references.
1398   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1399     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1400     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1401     tj = TypeOopPtr::BOTTOM;
1402     ptr = tj->ptr();
1403     offset = tj->offset();
1404   }
1405 
1406   // Array pointers need some flattening
1407   const TypeAryPtr *ta = tj->isa_aryptr();
1408   if (ta && ta->is_stable()) {
1409     // Erase stability property for alias analysis.
1410     tj = ta = ta->cast_to_stable(false);
1411   }
1412   if( ta && is_known_inst ) {
1413     if ( offset != Type::OffsetBot &&
1414          offset > arrayOopDesc::length_offset_in_bytes() ) {
1415       offset = Type::OffsetBot; // Flatten constant access into array body only
1416       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, Type::Offset(offset), ta->field_offset(), ta->instance_id());
1417     }
1418   } else if( ta && _AliasLevel >= 2 ) {
1419     // For arrays indexed by constant indices, we flatten the alias
1420     // space to include all of the array body.  Only the header, klass
1421     // and array length can be accessed un-aliased.
1422     // For flattened value type array, each field has its own slice so
1423     // we must include the field offset.
1424     if( offset != Type::OffsetBot ) {
1425       if( ta->const_oop() ) { // MethodData* or Method*
1426         offset = Type::OffsetBot;   // Flatten constant access into array body
1427         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1428       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1429         // range is OK as-is.
1430         tj = ta = TypeAryPtr::RANGE;
1431       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1432         tj = TypeInstPtr::KLASS; // all klass loads look alike
1433         ta = TypeAryPtr::RANGE; // generic ignored junk
1434         ptr = TypePtr::BotPTR;
1435       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1436         tj = TypeInstPtr::MARK;
1437         ta = TypeAryPtr::RANGE; // generic ignored junk
1438         ptr = TypePtr::BotPTR;
1439       } else {                  // Random constant offset into array body
1440         offset = Type::OffsetBot;   // Flatten constant access into array body
1441         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1442       }
1443     }
1444     // Arrays of fixed size alias with arrays of unknown size.
1445     if (ta->size() != TypeInt::POS) {
1446       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1447       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,Type::Offset(offset), ta->field_offset());
1448     }
1449     // Arrays of known objects become arrays of unknown objects.
1450     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1451       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1452       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1453     }
1454     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1455       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1456       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,Type::Offset(offset), ta->field_offset());
1457     }
1458     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1459     // cannot be distinguished by bytecode alone.
1460     if (ta->elem() == TypeInt::BOOL) {
1461       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1462       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1463       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,Type::Offset(offset), ta->field_offset());
1464     }
1465     // During the 2nd round of IterGVN, NotNull castings are removed.
1466     // Make sure the Bottom and NotNull variants alias the same.
1467     // Also, make sure exact and non-exact variants alias the same.
1468     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1469       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,Type::Offset(offset), ta->field_offset());
1470     }
1471   }
1472 
1473   // Oop pointers need some flattening
1474   const TypeInstPtr *to = tj->isa_instptr();
1475   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1476     ciInstanceKlass *k = to->klass()->as_instance_klass();
1477     if( ptr == TypePtr::Constant ) {
1478       if (to->klass() != ciEnv::current()->Class_klass() ||
1479           offset < k->size_helper() * wordSize) {
1480         // No constant oop pointers (such as Strings); they alias with
1481         // unknown strings.
1482         assert(!is_known_inst, "not scalarizable allocation");
1483         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1484       }
1485     } else if( is_known_inst ) {
1486       tj = to; // Keep NotNull and klass_is_exact for instance type
1487     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1488       // During the 2nd round of IterGVN, NotNull castings are removed.
1489       // Make sure the Bottom and NotNull variants alias the same.
1490       // Also, make sure exact and non-exact variants alias the same.
1491       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,Type::Offset(offset));
1492     }
1493     if (to->speculative() != NULL) {
1494       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),Type::Offset(to->offset()), to->instance_id());
1495     }
1496     // Canonicalize the holder of this field
1497     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1498       // First handle header references such as a LoadKlassNode, even if the
1499       // object's klass is unloaded at compile time (4965979).
1500       if (!is_known_inst) { // Do it only for non-instance types
1501         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, Type::Offset(offset));
1502       }
1503     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1504       // Static fields are in the space above the normal instance
1505       // fields in the java.lang.Class instance.
1506       if (to->klass() != ciEnv::current()->Class_klass()) {
1507         to = NULL;
1508         tj = TypeOopPtr::BOTTOM;
1509         offset = tj->offset();
1510       }
1511     } else {
1512       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1513       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1514         if( is_known_inst ) {
1515           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, Type::Offset(offset), to->instance_id());
1516         } else {
1517           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, Type::Offset(offset));
1518         }
1519       }
1520     }
1521   }
1522 
1523   // Klass pointers to object array klasses need some flattening
1524   const TypeKlassPtr *tk = tj->isa_klassptr();
1525   if( tk ) {
1526     // If we are referencing a field within a Klass, we need
1527     // to assume the worst case of an Object.  Both exact and
1528     // inexact types must flatten to the same alias class so
1529     // use NotNull as the PTR.
1530     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1531 
1532       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1533                                    TypeKlassPtr::OBJECT->klass(),
1534                                    Type::Offset(offset));
1535     }
1536 
1537     ciKlass* klass = tk->klass();
1538     if( klass->is_obj_array_klass() ) {
1539       ciKlass* k = TypeAryPtr::OOPS->klass();
1540       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1541         k = TypeInstPtr::BOTTOM->klass();
1542       tj = tk = TypeKlassPtr::make(TypePtr::NotNull, k, Type::Offset(offset));
1543     }
1544 
1545     // Check for precise loads from the primary supertype array and force them
1546     // to the supertype cache alias index.  Check for generic array loads from
1547     // the primary supertype array and also force them to the supertype cache
1548     // alias index.  Since the same load can reach both, we need to merge
1549     // these 2 disparate memories into the same alias class.  Since the
1550     // primary supertype array is read-only, there's no chance of confusion
1551     // where we bypass an array load and an array store.
1552     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1553     if (offset == Type::OffsetBot ||
1554         (offset >= primary_supers_offset &&
1555          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1556         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1557       offset = in_bytes(Klass::secondary_super_cache_offset());
1558       tj = tk = TypeKlassPtr::make(TypePtr::NotNull, tk->klass(), Type::Offset(offset));
1559     }
1560   }
1561 
1562   // Flatten all Raw pointers together.
1563   if (tj->base() == Type::RawPtr)
1564     tj = TypeRawPtr::BOTTOM;
1565 
1566   if (tj->base() == Type::AnyPtr)
1567     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1568 
1569   // Flatten all to bottom for now
1570   switch( _AliasLevel ) {
1571   case 0:
1572     tj = TypePtr::BOTTOM;
1573     break;
1574   case 1:                       // Flatten to: oop, static, field or array
1575     switch (tj->base()) {
1576     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1577     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1578     case Type::AryPtr:   // do not distinguish arrays at all
1579     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1580     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1581     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1582     default: ShouldNotReachHere();
1583     }
1584     break;
1585   case 2:                       // No collapsing at level 2; keep all splits
1586   case 3:                       // No collapsing at level 3; keep all splits
1587     break;
1588   default:
1589     Unimplemented();
1590   }
1591 
1592   offset = tj->offset();
1593   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1594 
1595   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1596           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1597           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1598           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1599           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1600           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1601           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr)  ,
1602           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1603   assert( tj->ptr() != TypePtr::TopPTR &&
1604           tj->ptr() != TypePtr::AnyNull &&
1605           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1606 //    assert( tj->ptr() != TypePtr::Constant ||
1607 //            tj->base() == Type::RawPtr ||
1608 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1609 
1610   return tj;
1611 }
1612 
1613 void Compile::AliasType::Init(int i, const TypePtr* at) {
1614   _index = i;
1615   _adr_type = at;
1616   _field = NULL;
1617   _element = NULL;
1618   _is_rewritable = true; // default
1619   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1620   if (atoop != NULL && atoop->is_known_instance()) {
1621     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1622     _general_index = Compile::current()->get_alias_index(gt);
1623   } else {
1624     _general_index = 0;
1625   }
1626 }
1627 
1628 //---------------------------------print_on------------------------------------
1629 #ifndef PRODUCT
1630 void Compile::AliasType::print_on(outputStream* st) {
1631   if (index() < 10)
1632         st->print("@ <%d> ", index());
1633   else  st->print("@ <%d>",  index());
1634   st->print(is_rewritable() ? "   " : " RO");
1635   int offset = adr_type()->offset();
1636   if (offset == Type::OffsetBot)
1637         st->print(" +any");
1638   else  st->print(" +%-3d", offset);
1639   st->print(" in ");
1640   adr_type()->dump_on(st);
1641   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1642   if (field() != NULL && tjp) {
1643     if (tjp->klass()  != field()->holder() ||
1644         tjp->offset() != field()->offset_in_bytes()) {
1645       st->print(" != ");
1646       field()->print();
1647       st->print(" ***");
1648     }
1649   }
1650 }
1651 
1652 void print_alias_types() {
1653   Compile* C = Compile::current();
1654   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1655   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1656     C->alias_type(idx)->print_on(tty);
1657     tty->cr();
1658   }
1659 }
1660 #endif
1661 
1662 
1663 //----------------------------probe_alias_cache--------------------------------
1664 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1665   intptr_t key = (intptr_t) adr_type;
1666   key ^= key >> logAliasCacheSize;
1667   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1668 }
1669 
1670 
1671 //-----------------------------grow_alias_types--------------------------------
1672 void Compile::grow_alias_types() {
1673   const int old_ats  = _max_alias_types; // how many before?
1674   const int new_ats  = old_ats;          // how many more?
1675   const int grow_ats = old_ats+new_ats;  // how many now?
1676   _max_alias_types = grow_ats;
1677   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1678   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1679   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1680   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1681 }
1682 
1683 
1684 //--------------------------------find_alias_type------------------------------
1685 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1686   if (_AliasLevel == 0)
1687     return alias_type(AliasIdxBot);
1688 
1689   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1690   if (ace->_adr_type == adr_type) {
1691     return alias_type(ace->_index);
1692   }
1693 
1694   // Handle special cases.
1695   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1696   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1697 
1698   // Do it the slow way.
1699   const TypePtr* flat = flatten_alias_type(adr_type);
1700 
1701 #ifdef ASSERT
1702   assert(flat == flatten_alias_type(flat), "idempotent");
1703   assert(flat != TypePtr::BOTTOM,     "cannot alias-analyze an untyped ptr");
1704   if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1705     const TypeOopPtr* foop = flat->is_oopptr();
1706     // Scalarizable allocations have exact klass always.
1707     bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1708     const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1709     assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1710   }
1711   assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1712 #endif
1713 
1714   int idx = AliasIdxTop;
1715   for (int i = 0; i < num_alias_types(); i++) {
1716     if (alias_type(i)->adr_type() == flat) {
1717       idx = i;
1718       break;
1719     }
1720   }
1721 
1722   if (idx == AliasIdxTop) {
1723     if (no_create)  return NULL;
1724     // Grow the array if necessary.
1725     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1726     // Add a new alias type.
1727     idx = _num_alias_types++;
1728     _alias_types[idx]->Init(idx, flat);
1729     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1730     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1731     if (flat->isa_instptr()) {
1732       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1733           && flat->is_instptr()->klass() == env()->Class_klass())
1734         alias_type(idx)->set_rewritable(false);
1735     }
1736     if (flat->isa_aryptr()) {
1737 #ifdef ASSERT
1738       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1739       // (T_BYTE has the weakest alignment and size restrictions...)
1740       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1741 #endif
1742       if (flat->offset() == TypePtr::OffsetBot) {
1743         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1744       }
1745     }
1746     if (flat->isa_klassptr()) {
1747       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1748         alias_type(idx)->set_rewritable(false);
1749       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1750         alias_type(idx)->set_rewritable(false);
1751       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1752         alias_type(idx)->set_rewritable(false);
1753       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1754         alias_type(idx)->set_rewritable(false);
1755     }
1756     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1757     // but the base pointer type is not distinctive enough to identify
1758     // references into JavaThread.)
1759 
1760     // Check for final fields.
1761     const TypeInstPtr* tinst = flat->isa_instptr();
1762     const TypeValueTypePtr* vtptr = flat->isa_valuetypeptr();
1763     ciField* field = NULL;
1764     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1765       if (tinst->const_oop() != NULL &&
1766           tinst->klass() == ciEnv::current()->Class_klass() &&
1767           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1768         // static field
1769         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1770         field = k->get_field_by_offset(tinst->offset(), true);
1771       } else {
1772         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1773         field = k->get_field_by_offset(tinst->offset(), false);
1774       }
1775     } else if (vtptr) {
1776       // Value type field
1777       ciValueKlass* vk = vtptr->klass()->as_value_klass();
1778       field = vk->get_field_by_offset(vtptr->offset(), false);
1779     }
1780     assert(field == NULL ||
1781            original_field == NULL ||
1782            (field->holder() == original_field->holder() &&
1783             field->offset() == original_field->offset() &&
1784             field->is_static() == original_field->is_static()), "wrong field?");
1785     // Set field() and is_rewritable() attributes.
1786     if (field != NULL)  alias_type(idx)->set_field(field);
1787   }
1788 
1789   // Fill the cache for next time.
1790   ace->_adr_type = adr_type;
1791   ace->_index    = idx;
1792   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1793 
1794   // Might as well try to fill the cache for the flattened version, too.
1795   AliasCacheEntry* face = probe_alias_cache(flat);
1796   if (face->_adr_type == NULL) {
1797     face->_adr_type = flat;
1798     face->_index    = idx;
1799     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1800   }
1801 
1802   return alias_type(idx);
1803 }
1804 
1805 
1806 Compile::AliasType* Compile::alias_type(ciField* field) {
1807   const TypeOopPtr* t;
1808   if (field->is_static())
1809     t = TypeInstPtr::make(field->holder()->java_mirror());
1810   else
1811     t = TypeOopPtr::make_from_klass_raw(field->holder());
1812   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1813   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1814   return atp;
1815 }
1816 
1817 
1818 //------------------------------have_alias_type--------------------------------
1819 bool Compile::have_alias_type(const TypePtr* adr_type) {
1820   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1821   if (ace->_adr_type == adr_type) {
1822     return true;
1823   }
1824 
1825   // Handle special cases.
1826   if (adr_type == NULL)             return true;
1827   if (adr_type == TypePtr::BOTTOM)  return true;
1828 
1829   return find_alias_type(adr_type, true, NULL) != NULL;
1830 }
1831 
1832 //-----------------------------must_alias--------------------------------------
1833 // True if all values of the given address type are in the given alias category.
1834 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1835   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1836   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1837   if (alias_idx == AliasIdxTop)         return false; // the empty category
1838   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1839 
1840   // the only remaining possible overlap is identity
1841   int adr_idx = get_alias_index(adr_type);
1842   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1843   assert(adr_idx == alias_idx ||
1844          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1845           && adr_type                       != TypeOopPtr::BOTTOM),
1846          "should not be testing for overlap with an unsafe pointer");
1847   return adr_idx == alias_idx;
1848 }
1849 
1850 //------------------------------can_alias--------------------------------------
1851 // True if any values of the given address type are in the given alias category.
1852 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1853   if (alias_idx == AliasIdxTop)         return false; // the empty category
1854   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1855   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1856   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1857 
1858   // the only remaining possible overlap is identity
1859   int adr_idx = get_alias_index(adr_type);
1860   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1861   return adr_idx == alias_idx;
1862 }
1863 
1864 
1865 
1866 //---------------------------pop_warm_call-------------------------------------
1867 WarmCallInfo* Compile::pop_warm_call() {
1868   WarmCallInfo* wci = _warm_calls;
1869   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1870   return wci;
1871 }
1872 
1873 //----------------------------Inline_Warm--------------------------------------
1874 int Compile::Inline_Warm() {
1875   // If there is room, try to inline some more warm call sites.
1876   // %%% Do a graph index compaction pass when we think we're out of space?
1877   if (!InlineWarmCalls)  return 0;
1878 
1879   int calls_made_hot = 0;
1880   int room_to_grow   = NodeCountInliningCutoff - unique();
1881   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1882   int amount_grown   = 0;
1883   WarmCallInfo* call;
1884   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1885     int est_size = (int)call->size();
1886     if (est_size > (room_to_grow - amount_grown)) {
1887       // This one won't fit anyway.  Get rid of it.
1888       call->make_cold();
1889       continue;
1890     }
1891     call->make_hot();
1892     calls_made_hot++;
1893     amount_grown   += est_size;
1894     amount_to_grow -= est_size;
1895   }
1896 
1897   if (calls_made_hot > 0)  set_major_progress();
1898   return calls_made_hot;
1899 }
1900 
1901 
1902 //----------------------------Finish_Warm--------------------------------------
1903 void Compile::Finish_Warm() {
1904   if (!InlineWarmCalls)  return;
1905   if (failing())  return;
1906   if (warm_calls() == NULL)  return;
1907 
1908   // Clean up loose ends, if we are out of space for inlining.
1909   WarmCallInfo* call;
1910   while ((call = pop_warm_call()) != NULL) {
1911     call->make_cold();
1912   }
1913 }
1914 
1915 //---------------------cleanup_loop_predicates-----------------------
1916 // Remove the opaque nodes that protect the predicates so that all unused
1917 // checks and uncommon_traps will be eliminated from the ideal graph
1918 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1919   if (predicate_count()==0) return;
1920   for (int i = predicate_count(); i > 0; i--) {
1921     Node * n = predicate_opaque1_node(i-1);
1922     assert(n->Opcode() == Op_Opaque1, "must be");
1923     igvn.replace_node(n, n->in(1));
1924   }
1925   assert(predicate_count()==0, "should be clean!");
1926 }
1927 
1928 void Compile::add_range_check_cast(Node* n) {
1929   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1930   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1931   _range_check_casts->append(n);
1932 }
1933 
1934 // Remove all range check dependent CastIINodes.
1935 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1936   for (int i = range_check_cast_count(); i > 0; i--) {
1937     Node* cast = range_check_cast_node(i-1);
1938     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1939     igvn.replace_node(cast, cast->in(1));
1940   }
1941   assert(range_check_cast_count() == 0, "should be empty");
1942 }
1943 
1944 // StringOpts and late inlining of string methods
1945 void Compile::inline_string_calls(bool parse_time) {
1946   {
1947     // remove useless nodes to make the usage analysis simpler
1948     ResourceMark rm;
1949     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1950   }
1951 
1952   {
1953     ResourceMark rm;
1954     print_method(PHASE_BEFORE_STRINGOPTS, 3);
1955     PhaseStringOpts pso(initial_gvn(), for_igvn());
1956     print_method(PHASE_AFTER_STRINGOPTS, 3);
1957   }
1958 
1959   // now inline anything that we skipped the first time around
1960   if (!parse_time) {
1961     _late_inlines_pos = _late_inlines.length();
1962   }
1963 
1964   while (_string_late_inlines.length() > 0) {
1965     CallGenerator* cg = _string_late_inlines.pop();
1966     cg->do_late_inline();
1967     if (failing())  return;
1968   }
1969   _string_late_inlines.trunc_to(0);
1970 }
1971 
1972 // Late inlining of boxing methods
1973 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1974   if (_boxing_late_inlines.length() > 0) {
1975     assert(has_boxed_value(), "inconsistent");
1976 
1977     PhaseGVN* gvn = initial_gvn();
1978     set_inlining_incrementally(true);
1979 
1980     assert( igvn._worklist.size() == 0, "should be done with igvn" );
1981     for_igvn()->clear();
1982     gvn->replace_with(&igvn);
1983 
1984     _late_inlines_pos = _late_inlines.length();
1985 
1986     while (_boxing_late_inlines.length() > 0) {
1987       CallGenerator* cg = _boxing_late_inlines.pop();
1988       cg->do_late_inline();
1989       if (failing())  return;
1990     }
1991     _boxing_late_inlines.trunc_to(0);
1992 
1993     {
1994       ResourceMark rm;
1995       PhaseRemoveUseless pru(gvn, for_igvn());
1996     }
1997 
1998     igvn = PhaseIterGVN(gvn);
1999     igvn.optimize();
2000 
2001     set_inlining_progress(false);
2002     set_inlining_incrementally(false);
2003   }
2004 }
2005 
2006 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2007   assert(IncrementalInline, "incremental inlining should be on");
2008   PhaseGVN* gvn = initial_gvn();
2009 
2010   set_inlining_progress(false);
2011   for_igvn()->clear();
2012   gvn->replace_with(&igvn);
2013 
2014   {
2015     TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2016     int i = 0;
2017     for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2018       CallGenerator* cg = _late_inlines.at(i);
2019       _late_inlines_pos = i+1;
2020       cg->do_late_inline();
2021       if (failing())  return;
2022     }
2023     int j = 0;
2024     for (; i < _late_inlines.length(); i++, j++) {
2025       _late_inlines.at_put(j, _late_inlines.at(i));
2026     }
2027     _late_inlines.trunc_to(j);
2028   }
2029 
2030   {
2031     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2032     ResourceMark rm;
2033     PhaseRemoveUseless pru(gvn, for_igvn());
2034   }
2035 
2036   {
2037     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2038     igvn = PhaseIterGVN(gvn);
2039   }
2040 }
2041 
2042 // Perform incremental inlining until bound on number of live nodes is reached
2043 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2044   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2045 
2046   PhaseGVN* gvn = initial_gvn();
2047 
2048   set_inlining_incrementally(true);
2049   set_inlining_progress(true);
2050   uint low_live_nodes = 0;
2051 
2052   while(inlining_progress() && _late_inlines.length() > 0) {
2053 
2054     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2055       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2056         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2057         // PhaseIdealLoop is expensive so we only try it once we are
2058         // out of live nodes and we only try it again if the previous
2059         // helped got the number of nodes down significantly
2060         PhaseIdealLoop ideal_loop( igvn, false, true );
2061         if (failing())  return;
2062         low_live_nodes = live_nodes();
2063         _major_progress = true;
2064       }
2065 
2066       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2067         break;
2068       }
2069     }
2070 
2071     inline_incrementally_one(igvn);
2072 
2073     if (failing())  return;
2074 
2075     {
2076       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2077       igvn.optimize();
2078     }
2079 
2080     if (failing())  return;
2081   }
2082 
2083   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2084 
2085   if (_string_late_inlines.length() > 0) {
2086     assert(has_stringbuilder(), "inconsistent");
2087     for_igvn()->clear();
2088     initial_gvn()->replace_with(&igvn);
2089 
2090     inline_string_calls(false);
2091 
2092     if (failing())  return;
2093 
2094     {
2095       TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2096       ResourceMark rm;
2097       PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2098     }
2099 
2100     {
2101       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2102       igvn = PhaseIterGVN(gvn);
2103       igvn.optimize();
2104     }
2105   }
2106 
2107   set_inlining_incrementally(false);
2108 }
2109 
2110 
2111 //------------------------------Optimize---------------------------------------
2112 // Given a graph, optimize it.
2113 void Compile::Optimize() {
2114   TracePhase tp("optimizer", &timers[_t_optimizer]);
2115 
2116 #ifndef PRODUCT
2117   if (_directive->BreakAtCompileOption) {
2118     BREAKPOINT;
2119   }
2120 
2121 #endif
2122 
2123   ResourceMark rm;
2124   int          loop_opts_cnt;
2125 
2126   print_inlining_reinit();
2127 
2128   NOT_PRODUCT( verify_graph_edges(); )
2129 
2130   print_method(PHASE_AFTER_PARSING);
2131 
2132  {
2133   // Iterative Global Value Numbering, including ideal transforms
2134   // Initialize IterGVN with types and values from parse-time GVN
2135   PhaseIterGVN igvn(initial_gvn());
2136 #ifdef ASSERT
2137   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2138 #endif
2139   {
2140     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2141     igvn.optimize();
2142   }
2143 
2144   print_method(PHASE_ITER_GVN1, 2);
2145 
2146   if (failing())  return;
2147 
2148   inline_incrementally(igvn);
2149 
2150   print_method(PHASE_INCREMENTAL_INLINE, 2);
2151 
2152   if (failing())  return;
2153 
2154   if (eliminate_boxing()) {
2155     // Inline valueOf() methods now.
2156     inline_boxing_calls(igvn);
2157 
2158     if (AlwaysIncrementalInline) {
2159       inline_incrementally(igvn);
2160     }
2161 
2162     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2163 
2164     if (failing())  return;
2165   }
2166 
2167   // Remove the speculative part of types and clean up the graph from
2168   // the extra CastPP nodes whose only purpose is to carry them. Do
2169   // that early so that optimizations are not disrupted by the extra
2170   // CastPP nodes.
2171   remove_speculative_types(igvn);
2172 
2173   // No more new expensive nodes will be added to the list from here
2174   // so keep only the actual candidates for optimizations.
2175   cleanup_expensive_nodes(igvn);
2176 
2177   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2178     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2179     initial_gvn()->replace_with(&igvn);
2180     for_igvn()->clear();
2181     Unique_Node_List new_worklist(C->comp_arena());
2182     {
2183       ResourceMark rm;
2184       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2185     }
2186     set_for_igvn(&new_worklist);
2187     igvn = PhaseIterGVN(initial_gvn());
2188     igvn.optimize();
2189   }
2190 
2191   // Perform escape analysis
2192   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2193     if (has_loops()) {
2194       // Cleanup graph (remove dead nodes).
2195       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2196       PhaseIdealLoop ideal_loop( igvn, false, true );
2197       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2198       if (failing())  return;
2199     }
2200     ConnectionGraph::do_analysis(this, &igvn);
2201 
2202     if (failing())  return;
2203 
2204     // Optimize out fields loads from scalar replaceable allocations.
2205     igvn.optimize();
2206     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2207 
2208     if (failing())  return;
2209 
2210     if (congraph() != NULL && macro_count() > 0) {
2211       TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2212       PhaseMacroExpand mexp(igvn);
2213       mexp.eliminate_macro_nodes();
2214       igvn.set_delay_transform(false);
2215 
2216       igvn.optimize();
2217       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2218 
2219       if (failing())  return;
2220     }
2221   }
2222 
2223   // Loop transforms on the ideal graph.  Range Check Elimination,
2224   // peeling, unrolling, etc.
2225 
2226   // Set loop opts counter
2227   loop_opts_cnt = num_loop_opts();
2228   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2229     {
2230       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2231       PhaseIdealLoop ideal_loop( igvn, true );
2232       loop_opts_cnt--;
2233       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2234       if (failing())  return;
2235     }
2236     // Loop opts pass if partial peeling occurred in previous pass
2237     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2238       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2239       PhaseIdealLoop ideal_loop( igvn, false );
2240       loop_opts_cnt--;
2241       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2242       if (failing())  return;
2243     }
2244     // Loop opts pass for loop-unrolling before CCP
2245     if(major_progress() && (loop_opts_cnt > 0)) {
2246       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2247       PhaseIdealLoop ideal_loop( igvn, false );
2248       loop_opts_cnt--;
2249       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2250     }
2251     if (!failing()) {
2252       // Verify that last round of loop opts produced a valid graph
2253       TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2254       PhaseIdealLoop::verify(igvn);
2255     }
2256   }
2257   if (failing())  return;
2258 
2259   // Conditional Constant Propagation;
2260   PhaseCCP ccp( &igvn );
2261   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2262   {
2263     TracePhase tp("ccp", &timers[_t_ccp]);
2264     ccp.do_transform();
2265   }
2266   print_method(PHASE_CPP1, 2);
2267 
2268   assert( true, "Break here to ccp.dump_old2new_map()");
2269 
2270   // Iterative Global Value Numbering, including ideal transforms
2271   {
2272     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2273     igvn = ccp;
2274     igvn.optimize();
2275   }
2276 
2277   print_method(PHASE_ITER_GVN2, 2);
2278 
2279   if (failing())  return;
2280 
2281   // Loop transforms on the ideal graph.  Range Check Elimination,
2282   // peeling, unrolling, etc.
2283   if(loop_opts_cnt > 0) {
2284     debug_only( int cnt = 0; );
2285     while(major_progress() && (loop_opts_cnt > 0)) {
2286       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2287       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2288       PhaseIdealLoop ideal_loop( igvn, true);
2289       loop_opts_cnt--;
2290       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2291       if (failing())  return;
2292     }
2293   }
2294   // Ensure that major progress is now clear
2295   C->clear_major_progress();
2296 
2297   {
2298     // Verify that all previous optimizations produced a valid graph
2299     // at least to this point, even if no loop optimizations were done.
2300     TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2301     PhaseIdealLoop::verify(igvn);
2302   }
2303 
2304   if (range_check_cast_count() > 0) {
2305     // No more loop optimizations. Remove all range check dependent CastIINodes.
2306     C->remove_range_check_casts(igvn);
2307     igvn.optimize();
2308   }
2309 
2310   {
2311     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2312     PhaseMacroExpand  mex(igvn);
2313     if (mex.expand_macro_nodes()) {
2314       assert(failing(), "must bail out w/ explicit message");
2315       return;
2316     }
2317   }
2318 
2319   DEBUG_ONLY( _modified_nodes = NULL; )
2320  } // (End scope of igvn; run destructor if necessary for asserts.)
2321 
2322  process_print_inlining();
2323  // A method with only infinite loops has no edges entering loops from root
2324  {
2325    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2326    if (final_graph_reshaping()) {
2327      assert(failing(), "must bail out w/ explicit message");
2328      return;
2329    }
2330  }
2331 
2332  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2333 }
2334 
2335 // Fixme remove
2336 static void check_for_value_node(Node &n, void* C) {
2337   if (n.is_ValueType()) {
2338 #ifdef ASSERT
2339     ((Compile*)C)->method()->print_short_name();
2340     tty->print_cr("");
2341     n.dump(-1);
2342     assert(false, "Unable to match ValueTypeNode");
2343 #endif
2344     ((Compile*)C)->record_failure("Unable to match ValueTypeNode");
2345   }
2346 }
2347 
2348 //------------------------------Code_Gen---------------------------------------
2349 // Given a graph, generate code for it
2350 void Compile::Code_Gen() {
2351   // FIXME remove
2352   root()->walk(Node::nop, check_for_value_node, this);
2353 
2354   if (failing()) {
2355     return;
2356   }
2357 
2358   // Perform instruction selection.  You might think we could reclaim Matcher
2359   // memory PDQ, but actually the Matcher is used in generating spill code.
2360   // Internals of the Matcher (including some VectorSets) must remain live
2361   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2362   // set a bit in reclaimed memory.
2363 
2364   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2365   // nodes.  Mapping is only valid at the root of each matched subtree.
2366   NOT_PRODUCT( verify_graph_edges(); )
2367 
2368   Matcher matcher;
2369   _matcher = &matcher;
2370   {
2371     TracePhase tp("matcher", &timers[_t_matcher]);
2372     matcher.match();
2373   }
2374   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2375   // nodes.  Mapping is only valid at the root of each matched subtree.
2376   NOT_PRODUCT( verify_graph_edges(); )
2377 
2378   // If you have too many nodes, or if matching has failed, bail out
2379   check_node_count(0, "out of nodes matching instructions");
2380   if (failing()) {
2381     return;
2382   }
2383 
2384   // Build a proper-looking CFG
2385   PhaseCFG cfg(node_arena(), root(), matcher);
2386   _cfg = &cfg;
2387   {
2388     TracePhase tp("scheduler", &timers[_t_scheduler]);
2389     bool success = cfg.do_global_code_motion();
2390     if (!success) {
2391       return;
2392     }
2393 
2394     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2395     NOT_PRODUCT( verify_graph_edges(); )
2396     debug_only( cfg.verify(); )
2397   }
2398 
2399   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2400   _regalloc = &regalloc;
2401   {
2402     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2403     // Perform register allocation.  After Chaitin, use-def chains are
2404     // no longer accurate (at spill code) and so must be ignored.
2405     // Node->LRG->reg mappings are still accurate.
2406     _regalloc->Register_Allocate();
2407 
2408     // Bail out if the allocator builds too many nodes
2409     if (failing()) {
2410       return;
2411     }
2412   }
2413 
2414   // Prior to register allocation we kept empty basic blocks in case the
2415   // the allocator needed a place to spill.  After register allocation we
2416   // are not adding any new instructions.  If any basic block is empty, we
2417   // can now safely remove it.
2418   {
2419     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2420     cfg.remove_empty_blocks();
2421     if (do_freq_based_layout()) {
2422       PhaseBlockLayout layout(cfg);
2423     } else {
2424       cfg.set_loop_alignment();
2425     }
2426     cfg.fixup_flow();
2427   }
2428 
2429   // Apply peephole optimizations
2430   if( OptoPeephole ) {
2431     TracePhase tp("peephole", &timers[_t_peephole]);
2432     PhasePeephole peep( _regalloc, cfg);
2433     peep.do_transform();
2434   }
2435 
2436   // Do late expand if CPU requires this.
2437   if (Matcher::require_postalloc_expand) {
2438     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2439     cfg.postalloc_expand(_regalloc);
2440   }
2441 
2442   // Convert Nodes to instruction bits in a buffer
2443   {
2444     TraceTime tp("output", &timers[_t_output], CITime);
2445     Output();
2446   }
2447 
2448   print_method(PHASE_FINAL_CODE);
2449 
2450   // He's dead, Jim.
2451   _cfg     = (PhaseCFG*)0xdeadbeef;
2452   _regalloc = (PhaseChaitin*)0xdeadbeef;
2453 }
2454 
2455 
2456 //------------------------------dump_asm---------------------------------------
2457 // Dump formatted assembly
2458 #ifndef PRODUCT
2459 void Compile::dump_asm(int *pcs, uint pc_limit) {
2460   bool cut_short = false;
2461   tty->print_cr("#");
2462   tty->print("#  ");  _tf->dump();  tty->cr();
2463   tty->print_cr("#");
2464 
2465   // For all blocks
2466   int pc = 0x0;                 // Program counter
2467   char starts_bundle = ' ';
2468   _regalloc->dump_frame();
2469 
2470   Node *n = NULL;
2471   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2472     if (VMThread::should_terminate()) {
2473       cut_short = true;
2474       break;
2475     }
2476     Block* block = _cfg->get_block(i);
2477     if (block->is_connector() && !Verbose) {
2478       continue;
2479     }
2480     n = block->head();
2481     if (pcs && n->_idx < pc_limit) {
2482       tty->print("%3.3x   ", pcs[n->_idx]);
2483     } else {
2484       tty->print("      ");
2485     }
2486     block->dump_head(_cfg);
2487     if (block->is_connector()) {
2488       tty->print_cr("        # Empty connector block");
2489     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2490       tty->print_cr("        # Block is sole successor of call");
2491     }
2492 
2493     // For all instructions
2494     Node *delay = NULL;
2495     for (uint j = 0; j < block->number_of_nodes(); j++) {
2496       if (VMThread::should_terminate()) {
2497         cut_short = true;
2498         break;
2499       }
2500       n = block->get_node(j);
2501       if (valid_bundle_info(n)) {
2502         Bundle* bundle = node_bundling(n);
2503         if (bundle->used_in_unconditional_delay()) {
2504           delay = n;
2505           continue;
2506         }
2507         if (bundle->starts_bundle()) {
2508           starts_bundle = '+';
2509         }
2510       }
2511 
2512       if (WizardMode) {
2513         n->dump();
2514       }
2515 
2516       if( !n->is_Region() &&    // Dont print in the Assembly
2517           !n->is_Phi() &&       // a few noisely useless nodes
2518           !n->is_Proj() &&
2519           !n->is_MachTemp() &&
2520           !n->is_SafePointScalarObject() &&
2521           !n->is_Catch() &&     // Would be nice to print exception table targets
2522           !n->is_MergeMem() &&  // Not very interesting
2523           !n->is_top() &&       // Debug info table constants
2524           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2525           ) {
2526         if (pcs && n->_idx < pc_limit)
2527           tty->print("%3.3x", pcs[n->_idx]);
2528         else
2529           tty->print("   ");
2530         tty->print(" %c ", starts_bundle);
2531         starts_bundle = ' ';
2532         tty->print("\t");
2533         n->format(_regalloc, tty);
2534         tty->cr();
2535       }
2536 
2537       // If we have an instruction with a delay slot, and have seen a delay,
2538       // then back up and print it
2539       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2540         assert(delay != NULL, "no unconditional delay instruction");
2541         if (WizardMode) delay->dump();
2542 
2543         if (node_bundling(delay)->starts_bundle())
2544           starts_bundle = '+';
2545         if (pcs && n->_idx < pc_limit)
2546           tty->print("%3.3x", pcs[n->_idx]);
2547         else
2548           tty->print("   ");
2549         tty->print(" %c ", starts_bundle);
2550         starts_bundle = ' ';
2551         tty->print("\t");
2552         delay->format(_regalloc, tty);
2553         tty->cr();
2554         delay = NULL;
2555       }
2556 
2557       // Dump the exception table as well
2558       if( n->is_Catch() && (Verbose || WizardMode) ) {
2559         // Print the exception table for this offset
2560         _handler_table.print_subtable_for(pc);
2561       }
2562     }
2563 
2564     if (pcs && n->_idx < pc_limit)
2565       tty->print_cr("%3.3x", pcs[n->_idx]);
2566     else
2567       tty->cr();
2568 
2569     assert(cut_short || delay == NULL, "no unconditional delay branch");
2570 
2571   } // End of per-block dump
2572   tty->cr();
2573 
2574   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2575 }
2576 #endif
2577 
2578 //------------------------------Final_Reshape_Counts---------------------------
2579 // This class defines counters to help identify when a method
2580 // may/must be executed using hardware with only 24-bit precision.
2581 struct Final_Reshape_Counts : public StackObj {
2582   int  _call_count;             // count non-inlined 'common' calls
2583   int  _float_count;            // count float ops requiring 24-bit precision
2584   int  _double_count;           // count double ops requiring more precision
2585   int  _java_call_count;        // count non-inlined 'java' calls
2586   int  _inner_loop_count;       // count loops which need alignment
2587   VectorSet _visited;           // Visitation flags
2588   Node_List _tests;             // Set of IfNodes & PCTableNodes
2589 
2590   Final_Reshape_Counts() :
2591     _call_count(0), _float_count(0), _double_count(0),
2592     _java_call_count(0), _inner_loop_count(0),
2593     _visited( Thread::current()->resource_area() ) { }
2594 
2595   void inc_call_count  () { _call_count  ++; }
2596   void inc_float_count () { _float_count ++; }
2597   void inc_double_count() { _double_count++; }
2598   void inc_java_call_count() { _java_call_count++; }
2599   void inc_inner_loop_count() { _inner_loop_count++; }
2600 
2601   int  get_call_count  () const { return _call_count  ; }
2602   int  get_float_count () const { return _float_count ; }
2603   int  get_double_count() const { return _double_count; }
2604   int  get_java_call_count() const { return _java_call_count; }
2605   int  get_inner_loop_count() const { return _inner_loop_count; }
2606 };
2607 
2608 #ifdef ASSERT
2609 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2610   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2611   // Make sure the offset goes inside the instance layout.
2612   return k->contains_field_offset(tp->offset());
2613   // Note that OffsetBot and OffsetTop are very negative.
2614 }
2615 #endif
2616 
2617 // Eliminate trivially redundant StoreCMs and accumulate their
2618 // precedence edges.
2619 void Compile::eliminate_redundant_card_marks(Node* n) {
2620   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2621   if (n->in(MemNode::Address)->outcnt() > 1) {
2622     // There are multiple users of the same address so it might be
2623     // possible to eliminate some of the StoreCMs
2624     Node* mem = n->in(MemNode::Memory);
2625     Node* adr = n->in(MemNode::Address);
2626     Node* val = n->in(MemNode::ValueIn);
2627     Node* prev = n;
2628     bool done = false;
2629     // Walk the chain of StoreCMs eliminating ones that match.  As
2630     // long as it's a chain of single users then the optimization is
2631     // safe.  Eliminating partially redundant StoreCMs would require
2632     // cloning copies down the other paths.
2633     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2634       if (adr == mem->in(MemNode::Address) &&
2635           val == mem->in(MemNode::ValueIn)) {
2636         // redundant StoreCM
2637         if (mem->req() > MemNode::OopStore) {
2638           // Hasn't been processed by this code yet.
2639           n->add_prec(mem->in(MemNode::OopStore));
2640         } else {
2641           // Already converted to precedence edge
2642           for (uint i = mem->req(); i < mem->len(); i++) {
2643             // Accumulate any precedence edges
2644             if (mem->in(i) != NULL) {
2645               n->add_prec(mem->in(i));
2646             }
2647           }
2648           // Everything above this point has been processed.
2649           done = true;
2650         }
2651         // Eliminate the previous StoreCM
2652         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2653         assert(mem->outcnt() == 0, "should be dead");
2654         mem->disconnect_inputs(NULL, this);
2655       } else {
2656         prev = mem;
2657       }
2658       mem = prev->in(MemNode::Memory);
2659     }
2660   }
2661 }
2662 
2663 //------------------------------final_graph_reshaping_impl----------------------
2664 // Implement items 1-5 from final_graph_reshaping below.
2665 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2666 
2667   if ( n->outcnt() == 0 ) return; // dead node
2668   uint nop = n->Opcode();
2669 
2670   // Check for 2-input instruction with "last use" on right input.
2671   // Swap to left input.  Implements item (2).
2672   if( n->req() == 3 &&          // two-input instruction
2673       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2674       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2675       n->in(2)->outcnt() == 1 &&// right use IS a last use
2676       !n->in(2)->is_Con() ) {   // right use is not a constant
2677     // Check for commutative opcode
2678     switch( nop ) {
2679     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2680     case Op_MaxI:  case Op_MinI:
2681     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2682     case Op_AndL:  case Op_XorL:  case Op_OrL:
2683     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2684       // Move "last use" input to left by swapping inputs
2685       n->swap_edges(1, 2);
2686       break;
2687     }
2688     default:
2689       break;
2690     }
2691   }
2692 
2693 #ifdef ASSERT
2694   if( n->is_Mem() ) {
2695     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2696     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2697             // oop will be recorded in oop map if load crosses safepoint
2698             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2699                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2700             "raw memory operations should have control edge");
2701   }
2702 #endif
2703   // Count FPU ops and common calls, implements item (3)
2704   switch( nop ) {
2705   // Count all float operations that may use FPU
2706   case Op_AddF:
2707   case Op_SubF:
2708   case Op_MulF:
2709   case Op_DivF:
2710   case Op_NegF:
2711   case Op_ModF:
2712   case Op_ConvI2F:
2713   case Op_ConF:
2714   case Op_CmpF:
2715   case Op_CmpF3:
2716   // case Op_ConvL2F: // longs are split into 32-bit halves
2717     frc.inc_float_count();
2718     break;
2719 
2720   case Op_ConvF2D:
2721   case Op_ConvD2F:
2722     frc.inc_float_count();
2723     frc.inc_double_count();
2724     break;
2725 
2726   // Count all double operations that may use FPU
2727   case Op_AddD:
2728   case Op_SubD:
2729   case Op_MulD:
2730   case Op_DivD:
2731   case Op_NegD:
2732   case Op_ModD:
2733   case Op_ConvI2D:
2734   case Op_ConvD2I:
2735   // case Op_ConvL2D: // handled by leaf call
2736   // case Op_ConvD2L: // handled by leaf call
2737   case Op_ConD:
2738   case Op_CmpD:
2739   case Op_CmpD3:
2740     frc.inc_double_count();
2741     break;
2742   case Op_Opaque1:              // Remove Opaque Nodes before matching
2743   case Op_Opaque2:              // Remove Opaque Nodes before matching
2744   case Op_Opaque3:
2745     n->subsume_by(n->in(1), this);
2746     break;
2747   case Op_CallStaticJava:
2748   case Op_CallJava:
2749   case Op_CallDynamicJava:
2750     frc.inc_java_call_count(); // Count java call site;
2751   case Op_CallRuntime:
2752   case Op_CallLeaf:
2753   case Op_CallLeafNoFP: {
2754     assert( n->is_Call(), "" );
2755     CallNode *call = n->as_Call();
2756     // Count call sites where the FP mode bit would have to be flipped.
2757     // Do not count uncommon runtime calls:
2758     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2759     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2760     if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2761       frc.inc_call_count();   // Count the call site
2762     } else {                  // See if uncommon argument is shared
2763       Node *n = call->in(TypeFunc::Parms);
2764       int nop = n->Opcode();
2765       // Clone shared simple arguments to uncommon calls, item (1).
2766       if( n->outcnt() > 1 &&
2767           !n->is_Proj() &&
2768           nop != Op_CreateEx &&
2769           nop != Op_CheckCastPP &&
2770           nop != Op_DecodeN &&
2771           nop != Op_DecodeNKlass &&
2772           !n->is_Mem() ) {
2773         Node *x = n->clone();
2774         call->set_req( TypeFunc::Parms, x );
2775       }
2776     }
2777     break;
2778   }
2779 
2780   case Op_StoreD:
2781   case Op_LoadD:
2782   case Op_LoadD_unaligned:
2783     frc.inc_double_count();
2784     goto handle_mem;
2785   case Op_StoreF:
2786   case Op_LoadF:
2787     frc.inc_float_count();
2788     goto handle_mem;
2789 
2790   case Op_StoreCM:
2791     {
2792       // Convert OopStore dependence into precedence edge
2793       Node* prec = n->in(MemNode::OopStore);
2794       n->del_req(MemNode::OopStore);
2795       n->add_prec(prec);
2796       eliminate_redundant_card_marks(n);
2797     }
2798 
2799     // fall through
2800 
2801   case Op_StoreB:
2802   case Op_StoreC:
2803   case Op_StorePConditional:
2804   case Op_StoreI:
2805   case Op_StoreL:
2806   case Op_StoreIConditional:
2807   case Op_StoreLConditional:
2808   case Op_CompareAndSwapI:
2809   case Op_CompareAndSwapL:
2810   case Op_CompareAndSwapP:
2811   case Op_CompareAndSwapN:
2812   case Op_WeakCompareAndSwapI:
2813   case Op_WeakCompareAndSwapL:
2814   case Op_WeakCompareAndSwapP:
2815   case Op_WeakCompareAndSwapN:
2816   case Op_CompareAndExchangeI:
2817   case Op_CompareAndExchangeL:
2818   case Op_CompareAndExchangeP:
2819   case Op_CompareAndExchangeN:
2820   case Op_GetAndAddI:
2821   case Op_GetAndAddL:
2822   case Op_GetAndSetI:
2823   case Op_GetAndSetL:
2824   case Op_GetAndSetP:
2825   case Op_GetAndSetN:
2826   case Op_StoreP:
2827   case Op_StoreN:
2828   case Op_StoreNKlass:
2829   case Op_LoadB:
2830   case Op_LoadUB:
2831   case Op_LoadUS:
2832   case Op_LoadI:
2833   case Op_LoadKlass:
2834   case Op_LoadNKlass:
2835   case Op_LoadL:
2836   case Op_LoadL_unaligned:
2837   case Op_LoadPLocked:
2838   case Op_LoadP:
2839   case Op_LoadN:
2840   case Op_LoadRange:
2841   case Op_LoadS: {
2842   handle_mem:
2843 #ifdef ASSERT
2844     if( VerifyOptoOopOffsets ) {
2845       assert( n->is_Mem(), "" );
2846       MemNode *mem  = (MemNode*)n;
2847       // Check to see if address types have grounded out somehow.
2848       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2849       assert( !tp || oop_offset_is_sane(tp), "" );
2850     }
2851 #endif
2852     break;
2853   }
2854 
2855   case Op_AddP: {               // Assert sane base pointers
2856     Node *addp = n->in(AddPNode::Address);
2857     assert( !addp->is_AddP() ||
2858             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2859             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2860             "Base pointers must match" );
2861 #ifdef _LP64
2862     if ((UseCompressedOops || UseCompressedClassPointers) &&
2863         addp->Opcode() == Op_ConP &&
2864         addp == n->in(AddPNode::Base) &&
2865         n->in(AddPNode::Offset)->is_Con()) {
2866       // Use addressing with narrow klass to load with offset on x86.
2867       // On sparc loading 32-bits constant and decoding it have less
2868       // instructions (4) then load 64-bits constant (7).
2869       // Do this transformation here since IGVN will convert ConN back to ConP.
2870       const Type* t = addp->bottom_type();
2871       if (t->isa_oopptr() || t->isa_klassptr()) {
2872         Node* nn = NULL;
2873 
2874         int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2875 
2876         // Look for existing ConN node of the same exact type.
2877         Node* r  = root();
2878         uint cnt = r->outcnt();
2879         for (uint i = 0; i < cnt; i++) {
2880           Node* m = r->raw_out(i);
2881           if (m!= NULL && m->Opcode() == op &&
2882               m->bottom_type()->make_ptr() == t) {
2883             nn = m;
2884             break;
2885           }
2886         }
2887         if (nn != NULL) {
2888           // Decode a narrow oop to match address
2889           // [R12 + narrow_oop_reg<<3 + offset]
2890           if (t->isa_oopptr()) {
2891             nn = new DecodeNNode(nn, t);
2892           } else {
2893             nn = new DecodeNKlassNode(nn, t);
2894           }
2895           n->set_req(AddPNode::Base, nn);
2896           n->set_req(AddPNode::Address, nn);
2897           if (addp->outcnt() == 0) {
2898             addp->disconnect_inputs(NULL, this);
2899           }
2900         }
2901       }
2902     }
2903 #endif
2904     break;
2905   }
2906 
2907   case Op_CastPP: {
2908     // Remove CastPP nodes to gain more freedom during scheduling but
2909     // keep the dependency they encode as control or precedence edges
2910     // (if control is set already) on memory operations. Some CastPP
2911     // nodes don't have a control (don't carry a dependency): skip
2912     // those.
2913     if (n->in(0) != NULL) {
2914       ResourceMark rm;
2915       Unique_Node_List wq;
2916       wq.push(n);
2917       for (uint next = 0; next < wq.size(); ++next) {
2918         Node *m = wq.at(next);
2919         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2920           Node* use = m->fast_out(i);
2921           if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
2922             use->ensure_control_or_add_prec(n->in(0));
2923           } else {
2924             switch(use->Opcode()) {
2925             case Op_AddP:
2926             case Op_DecodeN:
2927             case Op_DecodeNKlass:
2928             case Op_CheckCastPP:
2929             case Op_CastPP:
2930               wq.push(use);
2931               break;
2932             }
2933           }
2934         }
2935       }
2936     }
2937     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
2938     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2939       Node* in1 = n->in(1);
2940       const Type* t = n->bottom_type();
2941       Node* new_in1 = in1->clone();
2942       new_in1->as_DecodeN()->set_type(t);
2943 
2944       if (!Matcher::narrow_oop_use_complex_address()) {
2945         //
2946         // x86, ARM and friends can handle 2 adds in addressing mode
2947         // and Matcher can fold a DecodeN node into address by using
2948         // a narrow oop directly and do implicit NULL check in address:
2949         //
2950         // [R12 + narrow_oop_reg<<3 + offset]
2951         // NullCheck narrow_oop_reg
2952         //
2953         // On other platforms (Sparc) we have to keep new DecodeN node and
2954         // use it to do implicit NULL check in address:
2955         //
2956         // decode_not_null narrow_oop_reg, base_reg
2957         // [base_reg + offset]
2958         // NullCheck base_reg
2959         //
2960         // Pin the new DecodeN node to non-null path on these platform (Sparc)
2961         // to keep the information to which NULL check the new DecodeN node
2962         // corresponds to use it as value in implicit_null_check().
2963         //
2964         new_in1->set_req(0, n->in(0));
2965       }
2966 
2967       n->subsume_by(new_in1, this);
2968       if (in1->outcnt() == 0) {
2969         in1->disconnect_inputs(NULL, this);
2970       }
2971     } else {
2972       n->subsume_by(n->in(1), this);
2973       if (n->outcnt() == 0) {
2974         n->disconnect_inputs(NULL, this);
2975       }
2976     }
2977     break;
2978   }
2979 #ifdef _LP64
2980   case Op_CmpP:
2981     // Do this transformation here to preserve CmpPNode::sub() and
2982     // other TypePtr related Ideal optimizations (for example, ptr nullness).
2983     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2984       Node* in1 = n->in(1);
2985       Node* in2 = n->in(2);
2986       if (!in1->is_DecodeNarrowPtr()) {
2987         in2 = in1;
2988         in1 = n->in(2);
2989       }
2990       assert(in1->is_DecodeNarrowPtr(), "sanity");
2991 
2992       Node* new_in2 = NULL;
2993       if (in2->is_DecodeNarrowPtr()) {
2994         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2995         new_in2 = in2->in(1);
2996       } else if (in2->Opcode() == Op_ConP) {
2997         const Type* t = in2->bottom_type();
2998         if (t == TypePtr::NULL_PTR) {
2999           assert(in1->is_DecodeN(), "compare klass to null?");
3000           // Don't convert CmpP null check into CmpN if compressed
3001           // oops implicit null check is not generated.
3002           // This will allow to generate normal oop implicit null check.
3003           if (Matcher::gen_narrow_oop_implicit_null_checks())
3004             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3005           //
3006           // This transformation together with CastPP transformation above
3007           // will generated code for implicit NULL checks for compressed oops.
3008           //
3009           // The original code after Optimize()
3010           //
3011           //    LoadN memory, narrow_oop_reg
3012           //    decode narrow_oop_reg, base_reg
3013           //    CmpP base_reg, NULL
3014           //    CastPP base_reg // NotNull
3015           //    Load [base_reg + offset], val_reg
3016           //
3017           // after these transformations will be
3018           //
3019           //    LoadN memory, narrow_oop_reg
3020           //    CmpN narrow_oop_reg, NULL
3021           //    decode_not_null narrow_oop_reg, base_reg
3022           //    Load [base_reg + offset], val_reg
3023           //
3024           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3025           // since narrow oops can be used in debug info now (see the code in
3026           // final_graph_reshaping_walk()).
3027           //
3028           // At the end the code will be matched to
3029           // on x86:
3030           //
3031           //    Load_narrow_oop memory, narrow_oop_reg
3032           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3033           //    NullCheck narrow_oop_reg
3034           //
3035           // and on sparc:
3036           //
3037           //    Load_narrow_oop memory, narrow_oop_reg
3038           //    decode_not_null narrow_oop_reg, base_reg
3039           //    Load [base_reg + offset], val_reg
3040           //    NullCheck base_reg
3041           //
3042         } else if (t->isa_oopptr()) {
3043           new_in2 = ConNode::make(t->make_narrowoop());
3044         } else if (t->isa_klassptr()) {
3045           new_in2 = ConNode::make(t->make_narrowklass());
3046         }
3047       }
3048       if (new_in2 != NULL) {
3049         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3050         n->subsume_by(cmpN, this);
3051         if (in1->outcnt() == 0) {
3052           in1->disconnect_inputs(NULL, this);
3053         }
3054         if (in2->outcnt() == 0) {
3055           in2->disconnect_inputs(NULL, this);
3056         }
3057       }
3058     }
3059     break;
3060 
3061   case Op_DecodeN:
3062   case Op_DecodeNKlass:
3063     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3064     // DecodeN could be pinned when it can't be fold into
3065     // an address expression, see the code for Op_CastPP above.
3066     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3067     break;
3068 
3069   case Op_EncodeP:
3070   case Op_EncodePKlass: {
3071     Node* in1 = n->in(1);
3072     if (in1->is_DecodeNarrowPtr()) {
3073       n->subsume_by(in1->in(1), this);
3074     } else if (in1->Opcode() == Op_ConP) {
3075       const Type* t = in1->bottom_type();
3076       if (t == TypePtr::NULL_PTR) {
3077         assert(t->isa_oopptr(), "null klass?");
3078         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3079       } else if (t->isa_oopptr()) {
3080         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3081       } else if (t->isa_klassptr()) {
3082         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3083       }
3084     }
3085     if (in1->outcnt() == 0) {
3086       in1->disconnect_inputs(NULL, this);
3087     }
3088     break;
3089   }
3090 
3091   case Op_Proj: {
3092     if (OptimizeStringConcat) {
3093       ProjNode* p = n->as_Proj();
3094       if (p->_is_io_use) {
3095         // Separate projections were used for the exception path which
3096         // are normally removed by a late inline.  If it wasn't inlined
3097         // then they will hang around and should just be replaced with
3098         // the original one.
3099         Node* proj = NULL;
3100         // Replace with just one
3101         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3102           Node *use = i.get();
3103           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3104             proj = use;
3105             break;
3106           }
3107         }
3108         assert(proj != NULL, "must be found");
3109         p->subsume_by(proj, this);
3110       }
3111     }
3112     break;
3113   }
3114 
3115   case Op_Phi:
3116     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3117       // The EncodeP optimization may create Phi with the same edges
3118       // for all paths. It is not handled well by Register Allocator.
3119       Node* unique_in = n->in(1);
3120       assert(unique_in != NULL, "");
3121       uint cnt = n->req();
3122       for (uint i = 2; i < cnt; i++) {
3123         Node* m = n->in(i);
3124         assert(m != NULL, "");
3125         if (unique_in != m)
3126           unique_in = NULL;
3127       }
3128       if (unique_in != NULL) {
3129         n->subsume_by(unique_in, this);
3130       }
3131     }
3132     break;
3133 
3134 #endif
3135 
3136 #ifdef ASSERT
3137   case Op_CastII:
3138     // Verify that all range check dependent CastII nodes were removed.
3139     if (n->isa_CastII()->has_range_check()) {
3140       n->dump(3);
3141       assert(false, "Range check dependent CastII node was not removed");
3142     }
3143     break;
3144 #endif
3145 
3146   case Op_ModI:
3147     if (UseDivMod) {
3148       // Check if a%b and a/b both exist
3149       Node* d = n->find_similar(Op_DivI);
3150       if (d) {
3151         // Replace them with a fused divmod if supported
3152         if (Matcher::has_match_rule(Op_DivModI)) {
3153           DivModINode* divmod = DivModINode::make(n);
3154           d->subsume_by(divmod->div_proj(), this);
3155           n->subsume_by(divmod->mod_proj(), this);
3156         } else {
3157           // replace a%b with a-((a/b)*b)
3158           Node* mult = new MulINode(d, d->in(2));
3159           Node* sub  = new SubINode(d->in(1), mult);
3160           n->subsume_by(sub, this);
3161         }
3162       }
3163     }
3164     break;
3165 
3166   case Op_ModL:
3167     if (UseDivMod) {
3168       // Check if a%b and a/b both exist
3169       Node* d = n->find_similar(Op_DivL);
3170       if (d) {
3171         // Replace them with a fused divmod if supported
3172         if (Matcher::has_match_rule(Op_DivModL)) {
3173           DivModLNode* divmod = DivModLNode::make(n);
3174           d->subsume_by(divmod->div_proj(), this);
3175           n->subsume_by(divmod->mod_proj(), this);
3176         } else {
3177           // replace a%b with a-((a/b)*b)
3178           Node* mult = new MulLNode(d, d->in(2));
3179           Node* sub  = new SubLNode(d->in(1), mult);
3180           n->subsume_by(sub, this);
3181         }
3182       }
3183     }
3184     break;
3185 
3186   case Op_LoadVector:
3187   case Op_StoreVector:
3188     break;
3189 
3190   case Op_AddReductionVI:
3191   case Op_AddReductionVL:
3192   case Op_AddReductionVF:
3193   case Op_AddReductionVD:
3194   case Op_MulReductionVI:
3195   case Op_MulReductionVL:
3196   case Op_MulReductionVF:
3197   case Op_MulReductionVD:
3198     break;
3199 
3200   case Op_PackB:
3201   case Op_PackS:
3202   case Op_PackI:
3203   case Op_PackF:
3204   case Op_PackL:
3205   case Op_PackD:
3206     if (n->req()-1 > 2) {
3207       // Replace many operand PackNodes with a binary tree for matching
3208       PackNode* p = (PackNode*) n;
3209       Node* btp = p->binary_tree_pack(1, n->req());
3210       n->subsume_by(btp, this);
3211     }
3212     break;
3213   case Op_Loop:
3214   case Op_CountedLoop:
3215     if (n->as_Loop()->is_inner_loop()) {
3216       frc.inc_inner_loop_count();
3217     }
3218     break;
3219   case Op_LShiftI:
3220   case Op_RShiftI:
3221   case Op_URShiftI:
3222   case Op_LShiftL:
3223   case Op_RShiftL:
3224   case Op_URShiftL:
3225     if (Matcher::need_masked_shift_count) {
3226       // The cpu's shift instructions don't restrict the count to the
3227       // lower 5/6 bits. We need to do the masking ourselves.
3228       Node* in2 = n->in(2);
3229       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3230       const TypeInt* t = in2->find_int_type();
3231       if (t != NULL && t->is_con()) {
3232         juint shift = t->get_con();
3233         if (shift > mask) { // Unsigned cmp
3234           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3235         }
3236       } else {
3237         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3238           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3239           n->set_req(2, shift);
3240         }
3241       }
3242       if (in2->outcnt() == 0) { // Remove dead node
3243         in2->disconnect_inputs(NULL, this);
3244       }
3245     }
3246     break;
3247   case Op_MemBarStoreStore:
3248   case Op_MemBarRelease:
3249     // Break the link with AllocateNode: it is no longer useful and
3250     // confuses register allocation.
3251     if (n->req() > MemBarNode::Precedent) {
3252       n->set_req(MemBarNode::Precedent, top());
3253     }
3254     break;
3255   case Op_RangeCheck: {
3256     RangeCheckNode* rc = n->as_RangeCheck();
3257     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3258     n->subsume_by(iff, this);
3259     frc._tests.push(iff);
3260     break;
3261   }
3262   case Op_ValueType: {
3263     ValueTypeNode* vt = n->as_ValueType();
3264     vt->make_scalar_in_safepoints(this);
3265     if (vt->outcnt() == 0) {
3266       vt->disconnect_inputs(NULL, this);
3267     }
3268     break;
3269   }
3270   default:
3271     assert( !n->is_Call(), "" );
3272     assert( !n->is_Mem(), "" );
3273     assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3274     break;
3275   }
3276 
3277   // Collect CFG split points
3278   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3279     frc._tests.push(n);
3280   }
3281 }
3282 
3283 //------------------------------final_graph_reshaping_walk---------------------
3284 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3285 // requires that the walk visits a node's inputs before visiting the node.
3286 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3287   ResourceArea *area = Thread::current()->resource_area();
3288   Unique_Node_List sfpt(area);
3289 
3290   frc._visited.set(root->_idx); // first, mark node as visited
3291   uint cnt = root->req();
3292   Node *n = root;
3293   uint  i = 0;
3294   while (true) {
3295     if (i < cnt) {
3296       // Place all non-visited non-null inputs onto stack
3297       Node* m = n->in(i);
3298       ++i;
3299       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3300         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3301           // compute worst case interpreter size in case of a deoptimization
3302           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3303 
3304           sfpt.push(m);
3305         }
3306         cnt = m->req();
3307         nstack.push(n, i); // put on stack parent and next input's index
3308         n = m;
3309         i = 0;
3310       }
3311     } else {
3312       // Now do post-visit work
3313       final_graph_reshaping_impl( n, frc );
3314       if (nstack.is_empty())
3315         break;             // finished
3316       n = nstack.node();   // Get node from stack
3317       cnt = n->req();
3318       i = nstack.index();
3319       nstack.pop();        // Shift to the next node on stack
3320     }
3321   }
3322 
3323   // Skip next transformation if compressed oops are not used.
3324   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3325       (!UseCompressedOops && !UseCompressedClassPointers))
3326     return;
3327 
3328   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3329   // It could be done for an uncommon traps or any safepoints/calls
3330   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3331   while (sfpt.size() > 0) {
3332     n = sfpt.pop();
3333     JVMState *jvms = n->as_SafePoint()->jvms();
3334     assert(jvms != NULL, "sanity");
3335     int start = jvms->debug_start();
3336     int end   = n->req();
3337     bool is_uncommon = (n->is_CallStaticJava() &&
3338                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3339     for (int j = start; j < end; j++) {
3340       Node* in = n->in(j);
3341       if (in->is_DecodeNarrowPtr()) {
3342         bool safe_to_skip = true;
3343         if (!is_uncommon ) {
3344           // Is it safe to skip?
3345           for (uint i = 0; i < in->outcnt(); i++) {
3346             Node* u = in->raw_out(i);
3347             if (!u->is_SafePoint() ||
3348                  u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3349               safe_to_skip = false;
3350             }
3351           }
3352         }
3353         if (safe_to_skip) {
3354           n->set_req(j, in->in(1));
3355         }
3356         if (in->outcnt() == 0) {
3357           in->disconnect_inputs(NULL, this);
3358         }
3359       }
3360     }
3361   }
3362 }
3363 
3364 //------------------------------final_graph_reshaping--------------------------
3365 // Final Graph Reshaping.
3366 //
3367 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3368 //     and not commoned up and forced early.  Must come after regular
3369 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3370 //     inputs to Loop Phis; these will be split by the allocator anyways.
3371 //     Remove Opaque nodes.
3372 // (2) Move last-uses by commutative operations to the left input to encourage
3373 //     Intel update-in-place two-address operations and better register usage
3374 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3375 //     calls canonicalizing them back.
3376 // (3) Count the number of double-precision FP ops, single-precision FP ops
3377 //     and call sites.  On Intel, we can get correct rounding either by
3378 //     forcing singles to memory (requires extra stores and loads after each
3379 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3380 //     clearing the mode bit around call sites).  The mode bit is only used
3381 //     if the relative frequency of single FP ops to calls is low enough.
3382 //     This is a key transform for SPEC mpeg_audio.
3383 // (4) Detect infinite loops; blobs of code reachable from above but not
3384 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3385 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3386 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3387 //     Detection is by looking for IfNodes where only 1 projection is
3388 //     reachable from below or CatchNodes missing some targets.
3389 // (5) Assert for insane oop offsets in debug mode.
3390 
3391 bool Compile::final_graph_reshaping() {
3392   // an infinite loop may have been eliminated by the optimizer,
3393   // in which case the graph will be empty.
3394   if (root()->req() == 1) {
3395     record_method_not_compilable("trivial infinite loop");
3396     return true;
3397   }
3398 
3399   // Expensive nodes have their control input set to prevent the GVN
3400   // from freely commoning them. There's no GVN beyond this point so
3401   // no need to keep the control input. We want the expensive nodes to
3402   // be freely moved to the least frequent code path by gcm.
3403   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3404   for (int i = 0; i < expensive_count(); i++) {
3405     _expensive_nodes->at(i)->set_req(0, NULL);
3406   }
3407 
3408   Final_Reshape_Counts frc;
3409 
3410   // Visit everybody reachable!
3411   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3412   Node_Stack nstack(live_nodes() >> 1);
3413   final_graph_reshaping_walk(nstack, root(), frc);
3414 
3415   // Check for unreachable (from below) code (i.e., infinite loops).
3416   for( uint i = 0; i < frc._tests.size(); i++ ) {
3417     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3418     // Get number of CFG targets.
3419     // Note that PCTables include exception targets after calls.
3420     uint required_outcnt = n->required_outcnt();
3421     if (n->outcnt() != required_outcnt) {
3422       // Check for a few special cases.  Rethrow Nodes never take the
3423       // 'fall-thru' path, so expected kids is 1 less.
3424       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3425         if (n->in(0)->in(0)->is_Call()) {
3426           CallNode *call = n->in(0)->in(0)->as_Call();
3427           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3428             required_outcnt--;      // Rethrow always has 1 less kid
3429           } else if (call->req() > TypeFunc::Parms &&
3430                      call->is_CallDynamicJava()) {
3431             // Check for null receiver. In such case, the optimizer has
3432             // detected that the virtual call will always result in a null
3433             // pointer exception. The fall-through projection of this CatchNode
3434             // will not be populated.
3435             Node *arg0 = call->in(TypeFunc::Parms);
3436             if (arg0->is_Type() &&
3437                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3438               required_outcnt--;
3439             }
3440           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3441                      call->req() > TypeFunc::Parms+1 &&
3442                      call->is_CallStaticJava()) {
3443             // Check for negative array length. In such case, the optimizer has
3444             // detected that the allocation attempt will always result in an
3445             // exception. There is no fall-through projection of this CatchNode .
3446             Node *arg1 = call->in(TypeFunc::Parms+1);
3447             if (arg1->is_Type() &&
3448                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3449               required_outcnt--;
3450             }
3451           }
3452         }
3453       }
3454       // Recheck with a better notion of 'required_outcnt'
3455       if (n->outcnt() != required_outcnt) {
3456         record_method_not_compilable("malformed control flow");
3457         return true;            // Not all targets reachable!
3458       }
3459     }
3460     // Check that I actually visited all kids.  Unreached kids
3461     // must be infinite loops.
3462     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3463       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3464         record_method_not_compilable("infinite loop");
3465         return true;            // Found unvisited kid; must be unreach
3466       }
3467   }
3468 
3469   // If original bytecodes contained a mixture of floats and doubles
3470   // check if the optimizer has made it homogenous, item (3).
3471   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3472       frc.get_float_count() > 32 &&
3473       frc.get_double_count() == 0 &&
3474       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3475     set_24_bit_selection_and_mode( false,  true );
3476   }
3477 
3478   set_java_calls(frc.get_java_call_count());
3479   set_inner_loops(frc.get_inner_loop_count());
3480 
3481   // No infinite loops, no reason to bail out.
3482   return false;
3483 }
3484 
3485 //-----------------------------too_many_traps----------------------------------
3486 // Report if there are too many traps at the current method and bci.
3487 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3488 bool Compile::too_many_traps(ciMethod* method,
3489                              int bci,
3490                              Deoptimization::DeoptReason reason) {
3491   ciMethodData* md = method->method_data();
3492   if (md->is_empty()) {
3493     // Assume the trap has not occurred, or that it occurred only
3494     // because of a transient condition during start-up in the interpreter.
3495     return false;
3496   }
3497   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3498   if (md->has_trap_at(bci, m, reason) != 0) {
3499     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3500     // Also, if there are multiple reasons, or if there is no per-BCI record,
3501     // assume the worst.
3502     if (log())
3503       log()->elem("observe trap='%s' count='%d'",
3504                   Deoptimization::trap_reason_name(reason),
3505                   md->trap_count(reason));
3506     return true;
3507   } else {
3508     // Ignore method/bci and see if there have been too many globally.
3509     return too_many_traps(reason, md);
3510   }
3511 }
3512 
3513 // Less-accurate variant which does not require a method and bci.
3514 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3515                              ciMethodData* logmd) {
3516   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3517     // Too many traps globally.
3518     // Note that we use cumulative trap_count, not just md->trap_count.
3519     if (log()) {
3520       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3521       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3522                   Deoptimization::trap_reason_name(reason),
3523                   mcount, trap_count(reason));
3524     }
3525     return true;
3526   } else {
3527     // The coast is clear.
3528     return false;
3529   }
3530 }
3531 
3532 //--------------------------too_many_recompiles--------------------------------
3533 // Report if there are too many recompiles at the current method and bci.
3534 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3535 // Is not eager to return true, since this will cause the compiler to use
3536 // Action_none for a trap point, to avoid too many recompilations.
3537 bool Compile::too_many_recompiles(ciMethod* method,
3538                                   int bci,
3539                                   Deoptimization::DeoptReason reason) {
3540   ciMethodData* md = method->method_data();
3541   if (md->is_empty()) {
3542     // Assume the trap has not occurred, or that it occurred only
3543     // because of a transient condition during start-up in the interpreter.
3544     return false;
3545   }
3546   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3547   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3548   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3549   Deoptimization::DeoptReason per_bc_reason
3550     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3551   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3552   if ((per_bc_reason == Deoptimization::Reason_none
3553        || md->has_trap_at(bci, m, reason) != 0)
3554       // The trap frequency measure we care about is the recompile count:
3555       && md->trap_recompiled_at(bci, m)
3556       && md->overflow_recompile_count() >= bc_cutoff) {
3557     // Do not emit a trap here if it has already caused recompilations.
3558     // Also, if there are multiple reasons, or if there is no per-BCI record,
3559     // assume the worst.
3560     if (log())
3561       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3562                   Deoptimization::trap_reason_name(reason),
3563                   md->trap_count(reason),
3564                   md->overflow_recompile_count());
3565     return true;
3566   } else if (trap_count(reason) != 0
3567              && decompile_count() >= m_cutoff) {
3568     // Too many recompiles globally, and we have seen this sort of trap.
3569     // Use cumulative decompile_count, not just md->decompile_count.
3570     if (log())
3571       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3572                   Deoptimization::trap_reason_name(reason),
3573                   md->trap_count(reason), trap_count(reason),
3574                   md->decompile_count(), decompile_count());
3575     return true;
3576   } else {
3577     // The coast is clear.
3578     return false;
3579   }
3580 }
3581 
3582 // Compute when not to trap. Used by matching trap based nodes and
3583 // NullCheck optimization.
3584 void Compile::set_allowed_deopt_reasons() {
3585   _allowed_reasons = 0;
3586   if (is_method_compilation()) {
3587     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3588       assert(rs < BitsPerInt, "recode bit map");
3589       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3590         _allowed_reasons |= nth_bit(rs);
3591       }
3592     }
3593   }
3594 }
3595 
3596 #ifndef PRODUCT
3597 //------------------------------verify_graph_edges---------------------------
3598 // Walk the Graph and verify that there is a one-to-one correspondence
3599 // between Use-Def edges and Def-Use edges in the graph.
3600 void Compile::verify_graph_edges(bool no_dead_code) {
3601   if (VerifyGraphEdges) {
3602     ResourceArea *area = Thread::current()->resource_area();
3603     Unique_Node_List visited(area);
3604     // Call recursive graph walk to check edges
3605     _root->verify_edges(visited);
3606     if (no_dead_code) {
3607       // Now make sure that no visited node is used by an unvisited node.
3608       bool dead_nodes = false;
3609       Unique_Node_List checked(area);
3610       while (visited.size() > 0) {
3611         Node* n = visited.pop();
3612         checked.push(n);
3613         for (uint i = 0; i < n->outcnt(); i++) {
3614           Node* use = n->raw_out(i);
3615           if (checked.member(use))  continue;  // already checked
3616           if (visited.member(use))  continue;  // already in the graph
3617           if (use->is_Con())        continue;  // a dead ConNode is OK
3618           // At this point, we have found a dead node which is DU-reachable.
3619           if (!dead_nodes) {
3620             tty->print_cr("*** Dead nodes reachable via DU edges:");
3621             dead_nodes = true;
3622           }
3623           use->dump(2);
3624           tty->print_cr("---");
3625           checked.push(use);  // No repeats; pretend it is now checked.
3626         }
3627       }
3628       assert(!dead_nodes, "using nodes must be reachable from root");
3629     }
3630   }
3631 }
3632 
3633 // Verify GC barriers consistency
3634 // Currently supported:
3635 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3636 void Compile::verify_barriers() {
3637   if (UseG1GC) {
3638     // Verify G1 pre-barriers
3639     const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active());
3640 
3641     ResourceArea *area = Thread::current()->resource_area();
3642     Unique_Node_List visited(area);
3643     Node_List worklist(area);
3644     // We're going to walk control flow backwards starting from the Root
3645     worklist.push(_root);
3646     while (worklist.size() > 0) {
3647       Node* x = worklist.pop();
3648       if (x == NULL || x == top()) continue;
3649       if (visited.member(x)) {
3650         continue;
3651       } else {
3652         visited.push(x);
3653       }
3654 
3655       if (x->is_Region()) {
3656         for (uint i = 1; i < x->req(); i++) {
3657           worklist.push(x->in(i));
3658         }
3659       } else {
3660         worklist.push(x->in(0));
3661         // We are looking for the pattern:
3662         //                            /->ThreadLocal
3663         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3664         //              \->ConI(0)
3665         // We want to verify that the If and the LoadB have the same control
3666         // See GraphKit::g1_write_barrier_pre()
3667         if (x->is_If()) {
3668           IfNode *iff = x->as_If();
3669           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3670             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3671             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3672                 && cmp->in(1)->is_Load()) {
3673               LoadNode* load = cmp->in(1)->as_Load();
3674               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3675                   && load->in(2)->in(3)->is_Con()
3676                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3677 
3678                 Node* if_ctrl = iff->in(0);
3679                 Node* load_ctrl = load->in(0);
3680 
3681                 if (if_ctrl != load_ctrl) {
3682                   // Skip possible CProj->NeverBranch in infinite loops
3683                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3684                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3685                     if_ctrl = if_ctrl->in(0)->in(0);
3686                   }
3687                 }
3688                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3689               }
3690             }
3691           }
3692         }
3693       }
3694     }
3695   }
3696 }
3697 
3698 #endif
3699 
3700 // The Compile object keeps track of failure reasons separately from the ciEnv.
3701 // This is required because there is not quite a 1-1 relation between the
3702 // ciEnv and its compilation task and the Compile object.  Note that one
3703 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3704 // to backtrack and retry without subsuming loads.  Other than this backtracking
3705 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3706 // by the logic in C2Compiler.
3707 void Compile::record_failure(const char* reason) {
3708   if (log() != NULL) {
3709     log()->elem("failure reason='%s' phase='compile'", reason);
3710   }
3711   if (_failure_reason == NULL) {
3712     // Record the first failure reason.
3713     _failure_reason = reason;
3714   }
3715 
3716   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3717     C->print_method(PHASE_FAILURE);
3718   }
3719   _root = NULL;  // flush the graph, too
3720 }
3721 
3722 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3723   : TraceTime(name, accumulator, CITime, CITimeVerbose),
3724     _phase_name(name), _dolog(CITimeVerbose)
3725 {
3726   if (_dolog) {
3727     C = Compile::current();
3728     _log = C->log();
3729   } else {
3730     C = NULL;
3731     _log = NULL;
3732   }
3733   if (_log != NULL) {
3734     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3735     _log->stamp();
3736     _log->end_head();
3737   }
3738 }
3739 
3740 Compile::TracePhase::~TracePhase() {
3741 
3742   C = Compile::current();
3743   if (_dolog) {
3744     _log = C->log();
3745   } else {
3746     _log = NULL;
3747   }
3748 
3749 #ifdef ASSERT
3750   if (PrintIdealNodeCount) {
3751     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3752                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3753   }
3754 
3755   if (VerifyIdealNodeCount) {
3756     Compile::current()->print_missing_nodes();
3757   }
3758 #endif
3759 
3760   if (_log != NULL) {
3761     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3762   }
3763 }
3764 
3765 //=============================================================================
3766 // Two Constant's are equal when the type and the value are equal.
3767 bool Compile::Constant::operator==(const Constant& other) {
3768   if (type()          != other.type()         )  return false;
3769   if (can_be_reused() != other.can_be_reused())  return false;
3770   // For floating point values we compare the bit pattern.
3771   switch (type()) {
3772   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3773   case T_LONG:
3774   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
3775   case T_OBJECT:
3776   case T_ADDRESS: return (_v._value.l == other._v._value.l);
3777   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
3778   case T_METADATA: return (_v._metadata == other._v._metadata);
3779   default: ShouldNotReachHere();
3780   }
3781   return false;
3782 }
3783 
3784 static int type_to_size_in_bytes(BasicType t) {
3785   switch (t) {
3786   case T_LONG:    return sizeof(jlong  );
3787   case T_FLOAT:   return sizeof(jfloat );
3788   case T_DOUBLE:  return sizeof(jdouble);
3789   case T_METADATA: return sizeof(Metadata*);
3790     // We use T_VOID as marker for jump-table entries (labels) which
3791     // need an internal word relocation.
3792   case T_VOID:
3793   case T_ADDRESS:
3794   case T_OBJECT:  return sizeof(jobject);
3795   }
3796 
3797   ShouldNotReachHere();
3798   return -1;
3799 }
3800 
3801 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3802   // sort descending
3803   if (a->freq() > b->freq())  return -1;
3804   if (a->freq() < b->freq())  return  1;
3805   return 0;
3806 }
3807 
3808 void Compile::ConstantTable::calculate_offsets_and_size() {
3809   // First, sort the array by frequencies.
3810   _constants.sort(qsort_comparator);
3811 
3812 #ifdef ASSERT
3813   // Make sure all jump-table entries were sorted to the end of the
3814   // array (they have a negative frequency).
3815   bool found_void = false;
3816   for (int i = 0; i < _constants.length(); i++) {
3817     Constant con = _constants.at(i);
3818     if (con.type() == T_VOID)
3819       found_void = true;  // jump-tables
3820     else
3821       assert(!found_void, "wrong sorting");
3822   }
3823 #endif
3824 
3825   int offset = 0;
3826   for (int i = 0; i < _constants.length(); i++) {
3827     Constant* con = _constants.adr_at(i);
3828 
3829     // Align offset for type.
3830     int typesize = type_to_size_in_bytes(con->type());
3831     offset = align_size_up(offset, typesize);
3832     con->set_offset(offset);   // set constant's offset
3833 
3834     if (con->type() == T_VOID) {
3835       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3836       offset = offset + typesize * n->outcnt();  // expand jump-table
3837     } else {
3838       offset = offset + typesize;
3839     }
3840   }
3841 
3842   // Align size up to the next section start (which is insts; see
3843   // CodeBuffer::align_at_start).
3844   assert(_size == -1, "already set?");
3845   _size = align_size_up(offset, CodeEntryAlignment);
3846 }
3847 
3848 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3849   MacroAssembler _masm(&cb);
3850   for (int i = 0; i < _constants.length(); i++) {
3851     Constant con = _constants.at(i);
3852     address constant_addr = NULL;
3853     switch (con.type()) {
3854     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
3855     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3856     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3857     case T_OBJECT: {
3858       jobject obj = con.get_jobject();
3859       int oop_index = _masm.oop_recorder()->find_index(obj);
3860       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3861       break;
3862     }
3863     case T_ADDRESS: {
3864       address addr = (address) con.get_jobject();
3865       constant_addr = _masm.address_constant(addr);
3866       break;
3867     }
3868     // We use T_VOID as marker for jump-table entries (labels) which
3869     // need an internal word relocation.
3870     case T_VOID: {
3871       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3872       // Fill the jump-table with a dummy word.  The real value is
3873       // filled in later in fill_jump_table.
3874       address dummy = (address) n;
3875       constant_addr = _masm.address_constant(dummy);
3876       // Expand jump-table
3877       for (uint i = 1; i < n->outcnt(); i++) {
3878         address temp_addr = _masm.address_constant(dummy + i);
3879         assert(temp_addr, "consts section too small");
3880       }
3881       break;
3882     }
3883     case T_METADATA: {
3884       Metadata* obj = con.get_metadata();
3885       int metadata_index = _masm.oop_recorder()->find_index(obj);
3886       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3887       break;
3888     }
3889     default: ShouldNotReachHere();
3890     }
3891     assert(constant_addr, "consts section too small");
3892     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3893             "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
3894   }
3895 }
3896 
3897 int Compile::ConstantTable::find_offset(Constant& con) const {
3898   int idx = _constants.find(con);
3899   assert(idx != -1, "constant must be in constant table");
3900   int offset = _constants.at(idx).offset();
3901   assert(offset != -1, "constant table not emitted yet?");
3902   return offset;
3903 }
3904 
3905 void Compile::ConstantTable::add(Constant& con) {
3906   if (con.can_be_reused()) {
3907     int idx = _constants.find(con);
3908     if (idx != -1 && _constants.at(idx).can_be_reused()) {
3909       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
3910       return;
3911     }
3912   }
3913   (void) _constants.append(con);
3914 }
3915 
3916 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3917   Block* b = Compile::current()->cfg()->get_block_for_node(n);
3918   Constant con(type, value, b->_freq);
3919   add(con);
3920   return con;
3921 }
3922 
3923 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3924   Constant con(metadata);
3925   add(con);
3926   return con;
3927 }
3928 
3929 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3930   jvalue value;
3931   BasicType type = oper->type()->basic_type();
3932   switch (type) {
3933   case T_LONG:    value.j = oper->constantL(); break;
3934   case T_FLOAT:   value.f = oper->constantF(); break;
3935   case T_DOUBLE:  value.d = oper->constantD(); break;
3936   case T_OBJECT:
3937   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3938   case T_METADATA: return add((Metadata*)oper->constant()); break;
3939   default: guarantee(false, "unhandled type: %s", type2name(type));
3940   }
3941   return add(n, type, value);
3942 }
3943 
3944 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3945   jvalue value;
3946   // We can use the node pointer here to identify the right jump-table
3947   // as this method is called from Compile::Fill_buffer right before
3948   // the MachNodes are emitted and the jump-table is filled (means the
3949   // MachNode pointers do not change anymore).
3950   value.l = (jobject) n;
3951   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
3952   add(con);
3953   return con;
3954 }
3955 
3956 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3957   // If called from Compile::scratch_emit_size do nothing.
3958   if (Compile::current()->in_scratch_emit_size())  return;
3959 
3960   assert(labels.is_nonempty(), "must be");
3961   assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
3962 
3963   // Since MachConstantNode::constant_offset() also contains
3964   // table_base_offset() we need to subtract the table_base_offset()
3965   // to get the plain offset into the constant table.
3966   int offset = n->constant_offset() - table_base_offset();
3967 
3968   MacroAssembler _masm(&cb);
3969   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3970 
3971   for (uint i = 0; i < n->outcnt(); i++) {
3972     address* constant_addr = &jump_table_base[i];
3973     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));
3974     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3975     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3976   }
3977 }
3978 
3979 //----------------------------static_subtype_check-----------------------------
3980 // Shortcut important common cases when superklass is exact:
3981 // (0) superklass is java.lang.Object (can occur in reflective code)
3982 // (1) subklass is already limited to a subtype of superklass => always ok
3983 // (2) subklass does not overlap with superklass => always fail
3984 // (3) superklass has NO subtypes and we can check with a simple compare.
3985 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
3986   if (StressReflectiveCode) {
3987     return SSC_full_test;       // Let caller generate the general case.
3988   }
3989 
3990   if (superk == env()->Object_klass()) {
3991     return SSC_always_true;     // (0) this test cannot fail
3992   }
3993 
3994   ciType* superelem = superk;
3995   if (superelem->is_array_klass())
3996     superelem = superelem->as_array_klass()->base_element_type();
3997 
3998   if (!subk->is_interface()) {  // cannot trust static interface types yet
3999     if (subk->is_subtype_of(superk)) {
4000       return SSC_always_true;   // (1) false path dead; no dynamic test needed
4001     }
4002     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4003         !superk->is_subtype_of(subk)) {
4004       return SSC_always_false;
4005     }
4006   }
4007 
4008   // If casting to an instance klass, it must have no subtypes
4009   if (superk->is_interface()) {
4010     // Cannot trust interfaces yet.
4011     // %%% S.B. superk->nof_implementors() == 1
4012   } else if (superelem->is_instance_klass()) {
4013     ciInstanceKlass* ik = superelem->as_instance_klass();
4014     if (!ik->has_subklass() && !ik->is_interface()) {
4015       if (!ik->is_final()) {
4016         // Add a dependency if there is a chance of a later subclass.
4017         dependencies()->assert_leaf_type(ik);
4018       }
4019       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
4020     }
4021   } else {
4022     // A primitive array type has no subtypes.
4023     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
4024   }
4025 
4026   return SSC_full_test;
4027 }
4028 
4029 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4030 #ifdef _LP64
4031   // The scaled index operand to AddP must be a clean 64-bit value.
4032   // Java allows a 32-bit int to be incremented to a negative
4033   // value, which appears in a 64-bit register as a large
4034   // positive number.  Using that large positive number as an
4035   // operand in pointer arithmetic has bad consequences.
4036   // On the other hand, 32-bit overflow is rare, and the possibility
4037   // can often be excluded, if we annotate the ConvI2L node with
4038   // a type assertion that its value is known to be a small positive
4039   // number.  (The prior range check has ensured this.)
4040   // This assertion is used by ConvI2LNode::Ideal.
4041   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4042   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4043   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4044   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4045 #endif
4046   return idx;
4047 }
4048 
4049 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4050 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4051   if (ctrl != NULL) {
4052     // Express control dependency by a CastII node with a narrow type.
4053     value = new CastIINode(value, itype, false, true /* range check dependency */);
4054     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4055     // node from floating above the range check during loop optimizations. Otherwise, the
4056     // ConvI2L node may be eliminated independently of the range check, causing the data path
4057     // to become TOP while the control path is still there (although it's unreachable).
4058     value->set_req(0, ctrl);
4059     // Save CastII node to remove it after loop optimizations.
4060     phase->C->add_range_check_cast(value);
4061     value = phase->transform(value);
4062   }
4063   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4064   return phase->transform(new ConvI2LNode(value, ltype));
4065 }
4066 
4067 // The message about the current inlining is accumulated in
4068 // _print_inlining_stream and transfered into the _print_inlining_list
4069 // once we know whether inlining succeeds or not. For regular
4070 // inlining, messages are appended to the buffer pointed by
4071 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4072 // a new buffer is added after _print_inlining_idx in the list. This
4073 // way we can update the inlining message for late inlining call site
4074 // when the inlining is attempted again.
4075 void Compile::print_inlining_init() {
4076   if (print_inlining() || print_intrinsics()) {
4077     _print_inlining_stream = new stringStream();
4078     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4079   }
4080 }
4081 
4082 void Compile::print_inlining_reinit() {
4083   if (print_inlining() || print_intrinsics()) {
4084     // Re allocate buffer when we change ResourceMark
4085     _print_inlining_stream = new stringStream();
4086   }
4087 }
4088 
4089 void Compile::print_inlining_reset() {
4090   _print_inlining_stream->reset();
4091 }
4092 
4093 void Compile::print_inlining_commit() {
4094   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4095   // Transfer the message from _print_inlining_stream to the current
4096   // _print_inlining_list buffer and clear _print_inlining_stream.
4097   _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4098   print_inlining_reset();
4099 }
4100 
4101 void Compile::print_inlining_push() {
4102   // Add new buffer to the _print_inlining_list at current position
4103   _print_inlining_idx++;
4104   _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4105 }
4106 
4107 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4108   return _print_inlining_list->at(_print_inlining_idx);
4109 }
4110 
4111 void Compile::print_inlining_update(CallGenerator* cg) {
4112   if (print_inlining() || print_intrinsics()) {
4113     if (!cg->is_late_inline()) {
4114       if (print_inlining_current().cg() != NULL) {
4115         print_inlining_push();
4116       }
4117       print_inlining_commit();
4118     } else {
4119       if (print_inlining_current().cg() != cg &&
4120           (print_inlining_current().cg() != NULL ||
4121            print_inlining_current().ss()->size() != 0)) {
4122         print_inlining_push();
4123       }
4124       print_inlining_commit();
4125       print_inlining_current().set_cg(cg);
4126     }
4127   }
4128 }
4129 
4130 void Compile::print_inlining_move_to(CallGenerator* cg) {
4131   // We resume inlining at a late inlining call site. Locate the
4132   // corresponding inlining buffer so that we can update it.
4133   if (print_inlining()) {
4134     for (int i = 0; i < _print_inlining_list->length(); i++) {
4135       if (_print_inlining_list->adr_at(i)->cg() == cg) {
4136         _print_inlining_idx = i;
4137         return;
4138       }
4139     }
4140     ShouldNotReachHere();
4141   }
4142 }
4143 
4144 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4145   if (print_inlining()) {
4146     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4147     assert(print_inlining_current().cg() == cg, "wrong entry");
4148     // replace message with new message
4149     _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4150     print_inlining_commit();
4151     print_inlining_current().set_cg(cg);
4152   }
4153 }
4154 
4155 void Compile::print_inlining_assert_ready() {
4156   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4157 }
4158 
4159 void Compile::process_print_inlining() {
4160   bool do_print_inlining = print_inlining() || print_intrinsics();
4161   if (do_print_inlining || log() != NULL) {
4162     // Print inlining message for candidates that we couldn't inline
4163     // for lack of space
4164     for (int i = 0; i < _late_inlines.length(); i++) {
4165       CallGenerator* cg = _late_inlines.at(i);
4166       if (!cg->is_mh_late_inline()) {
4167         const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4168         if (do_print_inlining) {
4169           cg->print_inlining_late(msg);
4170         }
4171         log_late_inline_failure(cg, msg);
4172       }
4173     }
4174   }
4175   if (do_print_inlining) {
4176     ResourceMark rm;
4177     stringStream ss;
4178     for (int i = 0; i < _print_inlining_list->length(); i++) {
4179       ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4180     }
4181     size_t end = ss.size();
4182     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4183     strncpy(_print_inlining_output, ss.base(), end+1);
4184     _print_inlining_output[end] = 0;
4185   }
4186 }
4187 
4188 void Compile::dump_print_inlining() {
4189   if (_print_inlining_output != NULL) {
4190     tty->print_raw(_print_inlining_output);
4191   }
4192 }
4193 
4194 void Compile::log_late_inline(CallGenerator* cg) {
4195   if (log() != NULL) {
4196     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4197                 cg->unique_id());
4198     JVMState* p = cg->call_node()->jvms();
4199     while (p != NULL) {
4200       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4201       p = p->caller();
4202     }
4203     log()->tail("late_inline");
4204   }
4205 }
4206 
4207 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4208   log_late_inline(cg);
4209   if (log() != NULL) {
4210     log()->inline_fail(msg);
4211   }
4212 }
4213 
4214 void Compile::log_inline_id(CallGenerator* cg) {
4215   if (log() != NULL) {
4216     // The LogCompilation tool needs a unique way to identify late
4217     // inline call sites. This id must be unique for this call site in
4218     // this compilation. Try to have it unique across compilations as
4219     // well because it can be convenient when grepping through the log
4220     // file.
4221     // Distinguish OSR compilations from others in case CICountOSR is
4222     // on.
4223     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4224     cg->set_unique_id(id);
4225     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4226   }
4227 }
4228 
4229 void Compile::log_inline_failure(const char* msg) {
4230   if (C->log() != NULL) {
4231     C->log()->inline_fail(msg);
4232   }
4233 }
4234 
4235 
4236 // Dump inlining replay data to the stream.
4237 // Don't change thread state and acquire any locks.
4238 void Compile::dump_inline_data(outputStream* out) {
4239   InlineTree* inl_tree = ilt();
4240   if (inl_tree != NULL) {
4241     out->print(" inline %d", inl_tree->count());
4242     inl_tree->dump_replay_data(out);
4243   }
4244 }
4245 
4246 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4247   if (n1->Opcode() < n2->Opcode())      return -1;
4248   else if (n1->Opcode() > n2->Opcode()) return 1;
4249 
4250   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4251   for (uint i = 1; i < n1->req(); i++) {
4252     if (n1->in(i) < n2->in(i))      return -1;
4253     else if (n1->in(i) > n2->in(i)) return 1;
4254   }
4255 
4256   return 0;
4257 }
4258 
4259 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4260   Node* n1 = *n1p;
4261   Node* n2 = *n2p;
4262 
4263   return cmp_expensive_nodes(n1, n2);
4264 }
4265 
4266 void Compile::sort_expensive_nodes() {
4267   if (!expensive_nodes_sorted()) {
4268     _expensive_nodes->sort(cmp_expensive_nodes);
4269   }
4270 }
4271 
4272 bool Compile::expensive_nodes_sorted() const {
4273   for (int i = 1; i < _expensive_nodes->length(); i++) {
4274     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4275       return false;
4276     }
4277   }
4278   return true;
4279 }
4280 
4281 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4282   if (_expensive_nodes->length() == 0) {
4283     return false;
4284   }
4285 
4286   assert(OptimizeExpensiveOps, "optimization off?");
4287 
4288   // Take this opportunity to remove dead nodes from the list
4289   int j = 0;
4290   for (int i = 0; i < _expensive_nodes->length(); i++) {
4291     Node* n = _expensive_nodes->at(i);
4292     if (!n->is_unreachable(igvn)) {
4293       assert(n->is_expensive(), "should be expensive");
4294       _expensive_nodes->at_put(j, n);
4295       j++;
4296     }
4297   }
4298   _expensive_nodes->trunc_to(j);
4299 
4300   // Then sort the list so that similar nodes are next to each other
4301   // and check for at least two nodes of identical kind with same data
4302   // inputs.
4303   sort_expensive_nodes();
4304 
4305   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4306     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4307       return true;
4308     }
4309   }
4310 
4311   return false;
4312 }
4313 
4314 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4315   if (_expensive_nodes->length() == 0) {
4316     return;
4317   }
4318 
4319   assert(OptimizeExpensiveOps, "optimization off?");
4320 
4321   // Sort to bring similar nodes next to each other and clear the
4322   // control input of nodes for which there's only a single copy.
4323   sort_expensive_nodes();
4324 
4325   int j = 0;
4326   int identical = 0;
4327   int i = 0;
4328   bool modified = false;
4329   for (; i < _expensive_nodes->length()-1; i++) {
4330     assert(j <= i, "can't write beyond current index");
4331     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4332       identical++;
4333       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4334       continue;
4335     }
4336     if (identical > 0) {
4337       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4338       identical = 0;
4339     } else {
4340       Node* n = _expensive_nodes->at(i);
4341       igvn.replace_input_of(n, 0, NULL);
4342       igvn.hash_insert(n);
4343       modified = true;
4344     }
4345   }
4346   if (identical > 0) {
4347     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4348   } else if (_expensive_nodes->length() >= 1) {
4349     Node* n = _expensive_nodes->at(i);
4350     igvn.replace_input_of(n, 0, NULL);
4351     igvn.hash_insert(n);
4352     modified = true;
4353   }
4354   _expensive_nodes->trunc_to(j);
4355   if (modified) {
4356     igvn.optimize();
4357   }
4358 }
4359 
4360 void Compile::add_expensive_node(Node * n) {
4361   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4362   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4363   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4364   if (OptimizeExpensiveOps) {
4365     _expensive_nodes->append(n);
4366   } else {
4367     // Clear control input and let IGVN optimize expensive nodes if
4368     // OptimizeExpensiveOps is off.
4369     n->set_req(0, NULL);
4370   }
4371 }
4372 
4373 /**
4374  * Remove the speculative part of types and clean up the graph
4375  */
4376 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4377   if (UseTypeSpeculation) {
4378     Unique_Node_List worklist;
4379     worklist.push(root());
4380     int modified = 0;
4381     // Go over all type nodes that carry a speculative type, drop the
4382     // speculative part of the type and enqueue the node for an igvn
4383     // which may optimize it out.
4384     for (uint next = 0; next < worklist.size(); ++next) {
4385       Node *n  = worklist.at(next);
4386       if (n->is_Type()) {
4387         TypeNode* tn = n->as_Type();
4388         const Type* t = tn->type();
4389         const Type* t_no_spec = t->remove_speculative();
4390         if (t_no_spec != t) {
4391           bool in_hash = igvn.hash_delete(n);
4392           assert(in_hash, "node should be in igvn hash table");
4393           tn->set_type(t_no_spec);
4394           igvn.hash_insert(n);
4395           igvn._worklist.push(n); // give it a chance to go away
4396           modified++;
4397         }
4398       }
4399       uint max = n->len();
4400       for( uint i = 0; i < max; ++i ) {
4401         Node *m = n->in(i);
4402         if (not_a_node(m))  continue;
4403         worklist.push(m);
4404       }
4405     }
4406     // Drop the speculative part of all types in the igvn's type table
4407     igvn.remove_speculative_types();
4408     if (modified > 0) {
4409       igvn.optimize();
4410     }
4411 #ifdef ASSERT
4412     // Verify that after the IGVN is over no speculative type has resurfaced
4413     worklist.clear();
4414     worklist.push(root());
4415     for (uint next = 0; next < worklist.size(); ++next) {
4416       Node *n  = worklist.at(next);
4417       const Type* t = igvn.type_or_null(n);
4418       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4419       if (n->is_Type()) {
4420         t = n->as_Type()->type();
4421         assert(t == t->remove_speculative(), "no more speculative types");
4422       }
4423       uint max = n->len();
4424       for( uint i = 0; i < max; ++i ) {
4425         Node *m = n->in(i);
4426         if (not_a_node(m))  continue;
4427         worklist.push(m);
4428       }
4429     }
4430     igvn.check_no_speculative_types();
4431 #endif
4432   }
4433 }
4434 
4435 // Auxiliary method to support randomized stressing/fuzzing.
4436 //
4437 // This method can be called the arbitrary number of times, with current count
4438 // as the argument. The logic allows selecting a single candidate from the
4439 // running list of candidates as follows:
4440 //    int count = 0;
4441 //    Cand* selected = null;
4442 //    while(cand = cand->next()) {
4443 //      if (randomized_select(++count)) {
4444 //        selected = cand;
4445 //      }
4446 //    }
4447 //
4448 // Including count equalizes the chances any candidate is "selected".
4449 // This is useful when we don't have the complete list of candidates to choose
4450 // from uniformly. In this case, we need to adjust the randomicity of the
4451 // selection, or else we will end up biasing the selection towards the latter
4452 // candidates.
4453 //
4454 // Quick back-envelope calculation shows that for the list of n candidates
4455 // the equal probability for the candidate to persist as "best" can be
4456 // achieved by replacing it with "next" k-th candidate with the probability
4457 // of 1/k. It can be easily shown that by the end of the run, the
4458 // probability for any candidate is converged to 1/n, thus giving the
4459 // uniform distribution among all the candidates.
4460 //
4461 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4462 #define RANDOMIZED_DOMAIN_POW 29
4463 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4464 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4465 bool Compile::randomized_select(int count) {
4466   assert(count > 0, "only positive");
4467   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4468 }
4469 
4470 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4471 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4472 
4473 void NodeCloneInfo::dump() const {
4474   tty->print(" {%d:%d} ", idx(), gen());
4475 }
4476 
4477 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4478   uint64_t val = value(old->_idx);
4479   NodeCloneInfo cio(val);
4480   assert(val != 0, "old node should be in the map");
4481   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4482   insert(nnn->_idx, cin.get());
4483 #ifndef PRODUCT
4484   if (is_debug()) {
4485     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4486   }
4487 #endif
4488 }
4489 
4490 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4491   NodeCloneInfo cio(value(old->_idx));
4492   if (cio.get() == 0) {
4493     cio.set(old->_idx, 0);
4494     insert(old->_idx, cio.get());
4495 #ifndef PRODUCT
4496     if (is_debug()) {
4497       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4498     }
4499 #endif
4500   }
4501   clone(old, nnn, gen);
4502 }
4503 
4504 int CloneMap::max_gen() const {
4505   int g = 0;
4506   DictI di(_dict);
4507   for(; di.test(); ++di) {
4508     int t = gen(di._key);
4509     if (g < t) {
4510       g = t;
4511 #ifndef PRODUCT
4512       if (is_debug()) {
4513         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4514       }
4515 #endif
4516     }
4517   }
4518   return g;
4519 }
4520 
4521 void CloneMap::dump(node_idx_t key) const {
4522   uint64_t val = value(key);
4523   if (val != 0) {
4524     NodeCloneInfo ni(val);
4525     ni.dump();
4526   }
4527 }