1 /*
   2  * Copyright (c) 1997, 2016, 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 "memory/resourceArea.hpp"
  37 #include "opto/addnode.hpp"
  38 #include "opto/block.hpp"
  39 #include "opto/c2compiler.hpp"
  40 #include "opto/callGenerator.hpp"
  41 #include "opto/callnode.hpp"
  42 #include "opto/castnode.hpp"
  43 #include "opto/cfgnode.hpp"
  44 #include "opto/chaitin.hpp"
  45 #include "opto/compile.hpp"
  46 #include "opto/connode.hpp"
  47 #include "opto/convertnode.hpp"
  48 #include "opto/divnode.hpp"
  49 #include "opto/escape.hpp"
  50 #include "opto/idealGraphPrinter.hpp"
  51 #include "opto/loopnode.hpp"
  52 #include "opto/machnode.hpp"
  53 #include "opto/macro.hpp"
  54 #include "opto/matcher.hpp"
  55 #include "opto/mathexactnode.hpp"
  56 #include "opto/memnode.hpp"
  57 #include "opto/mulnode.hpp"
  58 #include "opto/narrowptrnode.hpp"
  59 #include "opto/node.hpp"
  60 #include "opto/opcodes.hpp"
  61 #include "opto/output.hpp"
  62 #include "opto/parse.hpp"
  63 #include "opto/phaseX.hpp"
  64 #include "opto/rootnode.hpp"
  65 #include "opto/runtime.hpp"
  66 #include "opto/stringopts.hpp"
  67 #include "opto/type.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/align.hpp"
  75 #include "utilities/copy.hpp"
  76 
  77 
  78 // -------------------- Compile::mach_constant_base_node -----------------------
  79 // Constant table base node singleton.
  80 MachConstantBaseNode* Compile::mach_constant_base_node() {
  81   if (_mach_constant_base_node == NULL) {
  82     _mach_constant_base_node = new MachConstantBaseNode();
  83     _mach_constant_base_node->add_req(C->root());
  84   }
  85   return _mach_constant_base_node;
  86 }
  87 
  88 
  89 /// Support for intrinsics.
  90 
  91 // Return the index at which m must be inserted (or already exists).
  92 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
  93 class IntrinsicDescPair {
  94  private:
  95   ciMethod* _m;
  96   bool _is_virtual;
  97  public:
  98   IntrinsicDescPair(ciMethod* m, bool is_virtual) : _m(m), _is_virtual(is_virtual) {}
  99   static int compare(IntrinsicDescPair* const& key, CallGenerator* const& elt) {
 100     ciMethod* m= elt->method();
 101     ciMethod* key_m = key->_m;
 102     if (key_m < m)      return -1;
 103     else if (key_m > m) return 1;
 104     else {
 105       bool is_virtual = elt->is_virtual();
 106       bool key_virtual = key->_is_virtual;
 107       if (key_virtual < is_virtual)      return -1;
 108       else if (key_virtual > is_virtual) return 1;
 109       else                               return 0;
 110     }
 111   }
 112 };
 113 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual, bool& found) {
 114 #ifdef ASSERT
 115   for (int i = 1; i < _intrinsics->length(); i++) {
 116     CallGenerator* cg1 = _intrinsics->at(i-1);
 117     CallGenerator* cg2 = _intrinsics->at(i);
 118     assert(cg1->method() != cg2->method()
 119            ? cg1->method()     < cg2->method()
 120            : cg1->is_virtual() < cg2->is_virtual(),
 121            "compiler intrinsics list must stay sorted");
 122   }
 123 #endif
 124   IntrinsicDescPair pair(m, is_virtual);
 125   return _intrinsics->find_sorted<IntrinsicDescPair*, IntrinsicDescPair::compare>(&pair, found);
 126 }
 127 
 128 void Compile::register_intrinsic(CallGenerator* cg) {
 129   if (_intrinsics == NULL) {
 130     _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 131   }
 132   int len = _intrinsics->length();
 133   bool found = false;
 134   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual(), found);
 135   assert(!found, "registering twice");
 136   _intrinsics->insert_before(index, cg);
 137   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 138 }
 139 
 140 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 141   assert(m->is_loaded(), "don't try this on unloaded methods");
 142   if (_intrinsics != NULL) {
 143     bool found = false;
 144     int index = intrinsic_insertion_index(m, is_virtual, found);
 145      if (found) {
 146       return _intrinsics->at(index);
 147     }
 148   }
 149   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 150   if (m->intrinsic_id() != vmIntrinsics::_none &&
 151       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 152     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 153     if (cg != NULL) {
 154       // Save it for next time:
 155       register_intrinsic(cg);
 156       return cg;
 157     } else {
 158       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 159     }
 160   }
 161   return NULL;
 162 }
 163 
 164 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 165 // in library_call.cpp.
 166 
 167 
 168 #ifndef PRODUCT
 169 // statistics gathering...
 170 
 171 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 172 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 173 
 174 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 175   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 176   int oflags = _intrinsic_hist_flags[id];
 177   assert(flags != 0, "what happened?");
 178   if (is_virtual) {
 179     flags |= _intrinsic_virtual;
 180   }
 181   bool changed = (flags != oflags);
 182   if ((flags & _intrinsic_worked) != 0) {
 183     juint count = (_intrinsic_hist_count[id] += 1);
 184     if (count == 1) {
 185       changed = true;           // first time
 186     }
 187     // increment the overall count also:
 188     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 189   }
 190   if (changed) {
 191     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 192       // Something changed about the intrinsic's virtuality.
 193       if ((flags & _intrinsic_virtual) != 0) {
 194         // This is the first use of this intrinsic as a virtual call.
 195         if (oflags != 0) {
 196           // We already saw it as a non-virtual, so note both cases.
 197           flags |= _intrinsic_both;
 198         }
 199       } else if ((oflags & _intrinsic_both) == 0) {
 200         // This is the first use of this intrinsic as a non-virtual
 201         flags |= _intrinsic_both;
 202       }
 203     }
 204     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 205   }
 206   // update the overall flags also:
 207   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 208   return changed;
 209 }
 210 
 211 static char* format_flags(int flags, char* buf) {
 212   buf[0] = 0;
 213   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 214   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 215   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 216   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 217   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 218   if (buf[0] == 0)  strcat(buf, ",");
 219   assert(buf[0] == ',', "must be");
 220   return &buf[1];
 221 }
 222 
 223 void Compile::print_intrinsic_statistics() {
 224   char flagsbuf[100];
 225   ttyLocker ttyl;
 226   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 227   tty->print_cr("Compiler intrinsic usage:");
 228   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 229   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 230   #define PRINT_STAT_LINE(name, c, f) \
 231     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 232   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 233     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 234     int   flags = _intrinsic_hist_flags[id];
 235     juint count = _intrinsic_hist_count[id];
 236     if ((flags | count) != 0) {
 237       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 238     }
 239   }
 240   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 241   if (xtty != NULL)  xtty->tail("statistics");
 242 }
 243 
 244 void Compile::print_statistics() {
 245   { ttyLocker ttyl;
 246     if (xtty != NULL)  xtty->head("statistics type='opto'");
 247     Parse::print_statistics();
 248     PhaseCCP::print_statistics();
 249     PhaseRegAlloc::print_statistics();
 250     Scheduling::print_statistics();
 251     PhasePeephole::print_statistics();
 252     PhaseIdealLoop::print_statistics();
 253     if (xtty != NULL)  xtty->tail("statistics");
 254   }
 255   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 256     // put this under its own <statistics> element.
 257     print_intrinsic_statistics();
 258   }
 259 }
 260 #endif //PRODUCT
 261 
 262 // Support for bundling info
 263 Bundle* Compile::node_bundling(const Node *n) {
 264   assert(valid_bundle_info(n), "oob");
 265   return &_node_bundling_base[n->_idx];
 266 }
 267 
 268 bool Compile::valid_bundle_info(const Node *n) {
 269   return (_node_bundling_limit > n->_idx);
 270 }
 271 
 272 
 273 void Compile::gvn_replace_by(Node* n, Node* nn) {
 274   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 275     Node* use = n->last_out(i);
 276     bool is_in_table = initial_gvn()->hash_delete(use);
 277     uint uses_found = 0;
 278     for (uint j = 0; j < use->len(); j++) {
 279       if (use->in(j) == n) {
 280         if (j < use->req())
 281           use->set_req(j, nn);
 282         else
 283           use->set_prec(j, nn);
 284         uses_found++;
 285       }
 286     }
 287     if (is_in_table) {
 288       // reinsert into table
 289       initial_gvn()->hash_find_insert(use);
 290     }
 291     record_for_igvn(use);
 292     i -= uses_found;    // we deleted 1 or more copies of this edge
 293   }
 294 }
 295 
 296 
 297 static inline bool not_a_node(const Node* n) {
 298   if (n == NULL)                   return true;
 299   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 300   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 301   return false;
 302 }
 303 
 304 // Identify all nodes that are reachable from below, useful.
 305 // Use breadth-first pass that records state in a Unique_Node_List,
 306 // recursive traversal is slower.
 307 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 308   int estimated_worklist_size = live_nodes();
 309   useful.map( estimated_worklist_size, NULL );  // preallocate space
 310 
 311   // Initialize worklist
 312   if (root() != NULL)     { useful.push(root()); }
 313   // If 'top' is cached, declare it useful to preserve cached node
 314   if( cached_top_node() ) { useful.push(cached_top_node()); }
 315 
 316   // Push all useful nodes onto the list, breadthfirst
 317   for( uint next = 0; next < useful.size(); ++next ) {
 318     assert( next < unique(), "Unique useful nodes < total nodes");
 319     Node *n  = useful.at(next);
 320     uint max = n->len();
 321     for( uint i = 0; i < max; ++i ) {
 322       Node *m = n->in(i);
 323       if (not_a_node(m))  continue;
 324       useful.push(m);
 325     }
 326   }
 327 }
 328 
 329 // Update dead_node_list with any missing dead nodes using useful
 330 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 331 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 332   uint max_idx = unique();
 333   VectorSet& useful_node_set = useful.member_set();
 334 
 335   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 336     // If node with index node_idx is not in useful set,
 337     // mark it as dead in dead node list.
 338     if (! useful_node_set.test(node_idx) ) {
 339       record_dead_node(node_idx);
 340     }
 341   }
 342 }
 343 
 344 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 345   int shift = 0;
 346   for (int i = 0; i < inlines->length(); i++) {
 347     CallGenerator* cg = inlines->at(i);
 348     CallNode* call = cg->call_node();
 349     if (shift > 0) {
 350       inlines->at_put(i-shift, cg);
 351     }
 352     if (!useful.member(call)) {
 353       shift++;
 354     }
 355   }
 356   inlines->trunc_to(inlines->length()-shift);
 357 }
 358 
 359 // Disconnect all useless nodes by disconnecting those at the boundary.
 360 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 361   uint next = 0;
 362   while (next < useful.size()) {
 363     Node *n = useful.at(next++);
 364     if (n->is_SafePoint()) {
 365       // We're done with a parsing phase. Replaced nodes are not valid
 366       // beyond that point.
 367       n->as_SafePoint()->delete_replaced_nodes();
 368     }
 369     // Use raw traversal of out edges since this code removes out edges
 370     int max = n->outcnt();
 371     for (int j = 0; j < max; ++j) {
 372       Node* child = n->raw_out(j);
 373       if (! useful.member(child)) {
 374         assert(!child->is_top() || child != top(),
 375                "If top is cached in Compile object it is in useful list");
 376         // Only need to remove this out-edge to the useless node
 377         n->raw_del_out(j);
 378         --j;
 379         --max;
 380       }
 381     }
 382     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 383       record_for_igvn(n->unique_out());
 384     }
 385   }
 386   // Remove useless macro and predicate opaq nodes
 387   for (int i = C->macro_count()-1; i >= 0; i--) {
 388     Node* n = C->macro_node(i);
 389     if (!useful.member(n)) {
 390       remove_macro_node(n);
 391     }
 392   }
 393   // Remove useless CastII nodes with range check dependency
 394   for (int i = range_check_cast_count() - 1; i >= 0; i--) {
 395     Node* cast = range_check_cast_node(i);
 396     if (!useful.member(cast)) {
 397       remove_range_check_cast(cast);
 398     }
 399   }
 400   // Remove useless expensive node
 401   for (int i = C->expensive_count()-1; i >= 0; i--) {
 402     Node* n = C->expensive_node(i);
 403     if (!useful.member(n)) {
 404       remove_expensive_node(n);
 405     }
 406   }
 407   // clean up the late inline lists
 408   remove_useless_late_inlines(&_string_late_inlines, useful);
 409   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 410   remove_useless_late_inlines(&_late_inlines, useful);
 411   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 412 }
 413 
 414 //------------------------------frame_size_in_words-----------------------------
 415 // frame_slots in units of words
 416 int Compile::frame_size_in_words() const {
 417   // shift is 0 in LP32 and 1 in LP64
 418   const int shift = (LogBytesPerWord - LogBytesPerInt);
 419   int words = _frame_slots >> shift;
 420   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 421   return words;
 422 }
 423 
 424 // To bang the stack of this compiled method we use the stack size
 425 // that the interpreter would need in case of a deoptimization. This
 426 // removes the need to bang the stack in the deoptimization blob which
 427 // in turn simplifies stack overflow handling.
 428 int Compile::bang_size_in_bytes() const {
 429   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
 430 }
 431 
 432 // ============================================================================
 433 //------------------------------CompileWrapper---------------------------------
 434 class CompileWrapper : public StackObj {
 435   Compile *const _compile;
 436  public:
 437   CompileWrapper(Compile* compile);
 438 
 439   ~CompileWrapper();
 440 };
 441 
 442 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 443   // the Compile* pointer is stored in the current ciEnv:
 444   ciEnv* env = compile->env();
 445   assert(env == ciEnv::current(), "must already be a ciEnv active");
 446   assert(env->compiler_data() == NULL, "compile already active?");
 447   env->set_compiler_data(compile);
 448   assert(compile == Compile::current(), "sanity");
 449 
 450   compile->set_type_dict(NULL);
 451   compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
 452   compile->clone_map().set_clone_idx(0);
 453   compile->set_type_hwm(NULL);
 454   compile->set_type_last_size(0);
 455   compile->set_last_tf(NULL, NULL);
 456   compile->set_indexSet_arena(NULL);
 457   compile->set_indexSet_free_block_list(NULL);
 458   compile->init_type_arena();
 459   Type::Initialize(compile);
 460   _compile->set_scratch_buffer_blob(NULL);
 461   _compile->begin_method();
 462   _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
 463 }
 464 CompileWrapper::~CompileWrapper() {
 465   _compile->end_method();
 466   if (_compile->scratch_buffer_blob() != NULL)
 467     BufferBlob::free(_compile->scratch_buffer_blob());
 468   _compile->env()->set_compiler_data(NULL);
 469 }
 470 
 471 
 472 //----------------------------print_compile_messages---------------------------
 473 void Compile::print_compile_messages() {
 474 #ifndef PRODUCT
 475   // Check if recompiling
 476   if (_subsume_loads == false && PrintOpto) {
 477     // Recompiling without allowing machine instructions to subsume loads
 478     tty->print_cr("*********************************************************");
 479     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 480     tty->print_cr("*********************************************************");
 481   }
 482   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 483     // Recompiling without escape analysis
 484     tty->print_cr("*********************************************************");
 485     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 486     tty->print_cr("*********************************************************");
 487   }
 488   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 489     // Recompiling without boxing elimination
 490     tty->print_cr("*********************************************************");
 491     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 492     tty->print_cr("*********************************************************");
 493   }
 494   if (C->directive()->BreakAtCompileOption) {
 495     // Open the debugger when compiling this method.
 496     tty->print("### Breaking when compiling: ");
 497     method()->print_short_name();
 498     tty->cr();
 499     BREAKPOINT;
 500   }
 501 
 502   if( PrintOpto ) {
 503     if (is_osr_compilation()) {
 504       tty->print("[OSR]%3d", _compile_id);
 505     } else {
 506       tty->print("%3d", _compile_id);
 507     }
 508   }
 509 #endif
 510 }
 511 
 512 
 513 //-----------------------init_scratch_buffer_blob------------------------------
 514 // Construct a temporary BufferBlob and cache it for this compile.
 515 void Compile::init_scratch_buffer_blob(int const_size) {
 516   // If there is already a scratch buffer blob allocated and the
 517   // constant section is big enough, use it.  Otherwise free the
 518   // current and allocate a new one.
 519   BufferBlob* blob = scratch_buffer_blob();
 520   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 521     // Use the current blob.
 522   } else {
 523     if (blob != NULL) {
 524       BufferBlob::free(blob);
 525     }
 526 
 527     ResourceMark rm;
 528     _scratch_const_size = const_size;
 529     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
 530     blob = BufferBlob::create("Compile::scratch_buffer", size);
 531     // Record the buffer blob for next time.
 532     set_scratch_buffer_blob(blob);
 533     // Have we run out of code space?
 534     if (scratch_buffer_blob() == NULL) {
 535       // Let CompilerBroker disable further compilations.
 536       record_failure("Not enough space for scratch buffer in CodeCache");
 537       return;
 538     }
 539   }
 540 
 541   // Initialize the relocation buffers
 542   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 543   set_scratch_locs_memory(locs_buf);
 544 }
 545 
 546 
 547 //-----------------------scratch_emit_size-------------------------------------
 548 // Helper function that computes size by emitting code
 549 uint Compile::scratch_emit_size(const Node* n) {
 550   // Start scratch_emit_size section.
 551   set_in_scratch_emit_size(true);
 552 
 553   // Emit into a trash buffer and count bytes emitted.
 554   // This is a pretty expensive way to compute a size,
 555   // but it works well enough if seldom used.
 556   // All common fixed-size instructions are given a size
 557   // method by the AD file.
 558   // Note that the scratch buffer blob and locs memory are
 559   // allocated at the beginning of the compile task, and
 560   // may be shared by several calls to scratch_emit_size.
 561   // The allocation of the scratch buffer blob is particularly
 562   // expensive, since it has to grab the code cache lock.
 563   BufferBlob* blob = this->scratch_buffer_blob();
 564   assert(blob != NULL, "Initialize BufferBlob at start");
 565   assert(blob->size() > MAX_inst_size, "sanity");
 566   relocInfo* locs_buf = scratch_locs_memory();
 567   address blob_begin = blob->content_begin();
 568   address blob_end   = (address)locs_buf;
 569   assert(blob->contains(blob_end), "sanity");
 570   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 571   buf.initialize_consts_size(_scratch_const_size);
 572   buf.initialize_stubs_size(MAX_stubs_size);
 573   assert(locs_buf != NULL, "sanity");
 574   int lsize = MAX_locs_size / 3;
 575   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 576   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 577   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 578   // Mark as scratch buffer.
 579   buf.consts()->set_scratch_emit();
 580   buf.insts()->set_scratch_emit();
 581   buf.stubs()->set_scratch_emit();
 582 
 583   // Do the emission.
 584 
 585   Label fakeL; // Fake label for branch instructions.
 586   Label*   saveL = NULL;
 587   uint save_bnum = 0;
 588   bool is_branch = n->is_MachBranch();
 589   if (is_branch) {
 590     MacroAssembler masm(&buf);
 591     masm.bind(fakeL);
 592     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 593     n->as_MachBranch()->label_set(&fakeL, 0);
 594   }
 595   n->emit(buf, this->regalloc());
 596 
 597   // Emitting into the scratch buffer should not fail
 598   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
 599 
 600   if (is_branch) // Restore label.
 601     n->as_MachBranch()->label_set(saveL, save_bnum);
 602 
 603   // End scratch_emit_size section.
 604   set_in_scratch_emit_size(false);
 605 
 606   return buf.insts_size();
 607 }
 608 
 609 
 610 // ============================================================================
 611 //------------------------------Compile standard-------------------------------
 612 debug_only( int Compile::_debug_idx = 100000; )
 613 
 614 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 615 // the continuation bci for on stack replacement.
 616 
 617 
 618 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 619                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
 620                 : Phase(Compiler),
 621                   _env(ci_env),
 622                   _directive(directive),
 623                   _log(ci_env->log()),
 624                   _compile_id(ci_env->compile_id()),
 625                   _save_argument_registers(false),
 626                   _stub_name(NULL),
 627                   _stub_function(NULL),
 628                   _stub_entry_point(NULL),
 629                   _method(target),
 630                   _entry_bci(osr_bci),
 631                   _initial_gvn(NULL),
 632                   _for_igvn(NULL),
 633                   _warm_calls(NULL),
 634                   _subsume_loads(subsume_loads),
 635                   _do_escape_analysis(do_escape_analysis),
 636                   _eliminate_boxing(eliminate_boxing),
 637                   _failure_reason(NULL),
 638                   _code_buffer("Compile::Fill_buffer"),
 639                   _orig_pc_slot(0),
 640                   _orig_pc_slot_offset_in_bytes(0),
 641                   _has_method_handle_invokes(false),
 642                   _mach_constant_base_node(NULL),
 643                   _node_bundling_limit(0),
 644                   _node_bundling_base(NULL),
 645                   _java_calls(0),
 646                   _inner_loops(0),
 647                   _scratch_const_size(-1),
 648                   _in_scratch_emit_size(false),
 649                   _dead_node_list(comp_arena()),
 650                   _dead_node_count(0),
 651 #ifndef PRODUCT
 652                   _trace_opto_output(directive->TraceOptoOutputOption),
 653                   _in_dump_cnt(0),
 654                   _printer(IdealGraphPrinter::printer()),
 655 #endif
 656                   _congraph(NULL),
 657                   _comp_arena(mtCompiler),
 658                   _node_arena(mtCompiler),
 659                   _old_arena(mtCompiler),
 660                   _Compile_types(mtCompiler),
 661                   _replay_inline_data(NULL),
 662                   _late_inlines(comp_arena(), 2, 0, NULL),
 663                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 664                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 665                   _late_inlines_pos(0),
 666                   _number_of_mh_late_inlines(0),
 667                   _inlining_progress(false),
 668                   _inlining_incrementally(false),
 669                   _print_inlining_list(NULL),
 670                   _print_inlining_stream(NULL),
 671                   _print_inlining_idx(0),
 672                   _print_inlining_output(NULL),
 673                   _interpreter_frame_size(0),
 674                   _max_node_limit(MaxNodeLimit),
 675                   _has_reserved_stack_access(target->has_reserved_stack_access()) {
 676   C = this;
 677 #ifndef PRODUCT
 678   if (_printer != NULL) {
 679     _printer->set_compile(this);
 680   }
 681 #endif
 682   CompileWrapper cw(this);
 683 
 684   if (CITimeVerbose) {
 685     tty->print(" ");
 686     target->holder()->name()->print();
 687     tty->print(".");
 688     target->print_short_name();
 689     tty->print("  ");
 690   }
 691   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 692   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 693 
 694 #ifndef PRODUCT
 695   bool print_opto_assembly = directive->PrintOptoAssemblyOption;
 696   if (!print_opto_assembly) {
 697     bool print_assembly = directive->PrintAssemblyOption;
 698     if (print_assembly && !Disassembler::can_decode()) {
 699       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 700       print_opto_assembly = true;
 701     }
 702   }
 703   set_print_assembly(print_opto_assembly);
 704   set_parsed_irreducible_loop(false);
 705 
 706   if (directive->ReplayInlineOption) {
 707     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 708   }
 709 #endif
 710   set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
 711   set_print_intrinsics(directive->PrintIntrinsicsOption);
 712   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 713 
 714   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 715     // Make sure the method being compiled gets its own MDO,
 716     // so we can at least track the decompile_count().
 717     // Need MDO to record RTM code generation state.
 718     method()->ensure_method_data();
 719   }
 720 
 721   Init(::AliasLevel);
 722 
 723 
 724   print_compile_messages();
 725 
 726   _ilt = InlineTree::build_inline_tree_root();
 727 
 728   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 729   assert(num_alias_types() >= AliasIdxRaw, "");
 730 
 731 #define MINIMUM_NODE_HASH  1023
 732   // Node list that Iterative GVN will start with
 733   Unique_Node_List for_igvn(comp_arena());
 734   set_for_igvn(&for_igvn);
 735 
 736   // GVN that will be run immediately on new nodes
 737   uint estimated_size = method()->code_size()*4+64;
 738   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 739   PhaseGVN gvn(node_arena(), estimated_size);
 740   set_initial_gvn(&gvn);
 741 
 742   print_inlining_init();
 743   { // Scope for timing the parser
 744     TracePhase tp("parse", &timers[_t_parser]);
 745 
 746     // Put top into the hash table ASAP.
 747     initial_gvn()->transform_no_reclaim(top());
 748 
 749     // Set up tf(), start(), and find a CallGenerator.
 750     CallGenerator* cg = NULL;
 751     if (is_osr_compilation()) {
 752       const TypeTuple *domain = StartOSRNode::osr_domain();
 753       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 754       init_tf(TypeFunc::make(domain, range));
 755       StartNode* s = new StartOSRNode(root(), domain);
 756       initial_gvn()->set_type_bottom(s);
 757       init_start(s);
 758       cg = CallGenerator::for_osr(method(), entry_bci());
 759     } else {
 760       // Normal case.
 761       init_tf(TypeFunc::make(method()));
 762       StartNode* s = new StartNode(root(), tf()->domain());
 763       initial_gvn()->set_type_bottom(s);
 764       init_start(s);
 765       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
 766         // With java.lang.ref.reference.get() we must go through the
 767         // intrinsic when G1 is enabled - even when get() is the root
 768         // method of the compile - so that, if necessary, the value in
 769         // the referent field of the reference object gets recorded by
 770         // the pre-barrier code.
 771         // Specifically, if G1 is enabled, the value in the referent
 772         // field is recorded by the G1 SATB pre barrier. This will
 773         // result in the referent being marked live and the reference
 774         // object removed from the list of discovered references during
 775         // reference processing.
 776         cg = find_intrinsic(method(), false);
 777       }
 778       if (cg == NULL) {
 779         float past_uses = method()->interpreter_invocation_count();
 780         float expected_uses = past_uses;
 781         cg = CallGenerator::for_inline(method(), expected_uses);
 782       }
 783     }
 784     if (failing())  return;
 785     if (cg == NULL) {
 786       record_method_not_compilable("cannot parse method");
 787       return;
 788     }
 789     JVMState* jvms = build_start_state(start(), tf());
 790     if ((jvms = cg->generate(jvms)) == NULL) {
 791       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 792         record_method_not_compilable("method parse failed");
 793       }
 794       return;
 795     }
 796     GraphKit kit(jvms);
 797 
 798     if (!kit.stopped()) {
 799       // Accept return values, and transfer control we know not where.
 800       // This is done by a special, unique ReturnNode bound to root.
 801       return_values(kit.jvms());
 802     }
 803 
 804     if (kit.has_exceptions()) {
 805       // Any exceptions that escape from this call must be rethrown
 806       // to whatever caller is dynamically above us on the stack.
 807       // This is done by a special, unique RethrowNode bound to root.
 808       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 809     }
 810 
 811     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 812 
 813     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 814       inline_string_calls(true);
 815     }
 816 
 817     if (failing())  return;
 818 
 819     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 820 
 821     // Remove clutter produced by parsing.
 822     if (!failing()) {
 823       ResourceMark rm;
 824       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 825     }
 826   }
 827 
 828   // Note:  Large methods are capped off in do_one_bytecode().
 829   if (failing())  return;
 830 
 831   // After parsing, node notes are no longer automagic.
 832   // They must be propagated by register_new_node_with_optimizer(),
 833   // clone(), or the like.
 834   set_default_node_notes(NULL);
 835 
 836   for (;;) {
 837     int successes = Inline_Warm();
 838     if (failing())  return;
 839     if (successes == 0)  break;
 840   }
 841 
 842   // Drain the list.
 843   Finish_Warm();
 844 #ifndef PRODUCT
 845   if (_printer && _printer->should_print(1)) {
 846     _printer->print_inlining();
 847   }
 848 #endif
 849 
 850   if (failing())  return;
 851   NOT_PRODUCT( verify_graph_edges(); )
 852 
 853   // Now optimize
 854   Optimize();
 855   if (failing())  return;
 856   NOT_PRODUCT( verify_graph_edges(); )
 857 
 858 #ifndef PRODUCT
 859   if (PrintIdeal) {
 860     ttyLocker ttyl;  // keep the following output all in one block
 861     // This output goes directly to the tty, not the compiler log.
 862     // To enable tools to match it up with the compilation activity,
 863     // be sure to tag this tty output with the compile ID.
 864     if (xtty != NULL) {
 865       xtty->head("ideal compile_id='%d'%s", compile_id(),
 866                  is_osr_compilation()    ? " compile_kind='osr'" :
 867                  "");
 868     }
 869     root()->dump(9999);
 870     if (xtty != NULL) {
 871       xtty->tail("ideal");
 872     }
 873   }
 874 #endif
 875 
 876   NOT_PRODUCT( verify_barriers(); )
 877 
 878   // Dump compilation data to replay it.
 879   if (directive->DumpReplayOption) {
 880     env()->dump_replay_data(_compile_id);
 881   }
 882   if (directive->DumpInlineOption && (ilt() != NULL)) {
 883     env()->dump_inline_data(_compile_id);
 884   }
 885 
 886   // Now that we know the size of all the monitors we can add a fixed slot
 887   // for the original deopt pc.
 888 
 889   _orig_pc_slot =  fixed_slots();
 890   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 891   set_fixed_slots(next_slot);
 892 
 893   // Compute when to use implicit null checks. Used by matching trap based
 894   // nodes and NullCheck optimization.
 895   set_allowed_deopt_reasons();
 896 
 897   // Now generate code
 898   Code_Gen();
 899   if (failing())  return;
 900 
 901   // Check if we want to skip execution of all compiled code.
 902   {
 903 #ifndef PRODUCT
 904     if (OptoNoExecute) {
 905       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 906       return;
 907     }
 908 #endif
 909     TracePhase tp("install_code", &timers[_t_registerMethod]);
 910 
 911     if (is_osr_compilation()) {
 912       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 913       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 914     } else {
 915       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 916       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 917     }
 918 
 919     env()->register_method(_method, _entry_bci,
 920                            &_code_offsets,
 921                            _orig_pc_slot_offset_in_bytes,
 922                            code_buffer(),
 923                            frame_size_in_words(), _oop_map_set,
 924                            &_handler_table, &_inc_table,
 925                            compiler,
 926                            has_unsafe_access(),
 927                            SharedRuntime::is_wide_vector(max_vector_size()),
 928                            rtm_state()
 929                            );
 930 
 931     if (log() != NULL) // Print code cache state into compiler log
 932       log()->code_cache_state();
 933   }
 934 }
 935 
 936 //------------------------------Compile----------------------------------------
 937 // Compile a runtime stub
 938 Compile::Compile( ciEnv* ci_env,
 939                   TypeFunc_generator generator,
 940                   address stub_function,
 941                   const char *stub_name,
 942                   int is_fancy_jump,
 943                   bool pass_tls,
 944                   bool save_arg_registers,
 945                   bool return_pc,
 946                   DirectiveSet* directive)
 947   : Phase(Compiler),
 948     _env(ci_env),
 949     _directive(directive),
 950     _log(ci_env->log()),
 951     _compile_id(0),
 952     _save_argument_registers(save_arg_registers),
 953     _method(NULL),
 954     _stub_name(stub_name),
 955     _stub_function(stub_function),
 956     _stub_entry_point(NULL),
 957     _entry_bci(InvocationEntryBci),
 958     _initial_gvn(NULL),
 959     _for_igvn(NULL),
 960     _warm_calls(NULL),
 961     _orig_pc_slot(0),
 962     _orig_pc_slot_offset_in_bytes(0),
 963     _subsume_loads(true),
 964     _do_escape_analysis(false),
 965     _eliminate_boxing(false),
 966     _failure_reason(NULL),
 967     _code_buffer("Compile::Fill_buffer"),
 968     _has_method_handle_invokes(false),
 969     _mach_constant_base_node(NULL),
 970     _node_bundling_limit(0),
 971     _node_bundling_base(NULL),
 972     _java_calls(0),
 973     _inner_loops(0),
 974 #ifndef PRODUCT
 975     _trace_opto_output(directive->TraceOptoOutputOption),
 976     _in_dump_cnt(0),
 977     _printer(NULL),
 978 #endif
 979     _comp_arena(mtCompiler),
 980     _node_arena(mtCompiler),
 981     _old_arena(mtCompiler),
 982     _Compile_types(mtCompiler),
 983     _dead_node_list(comp_arena()),
 984     _dead_node_count(0),
 985     _congraph(NULL),
 986     _replay_inline_data(NULL),
 987     _number_of_mh_late_inlines(0),
 988     _inlining_progress(false),
 989     _inlining_incrementally(false),
 990     _print_inlining_list(NULL),
 991     _print_inlining_stream(NULL),
 992     _print_inlining_idx(0),
 993     _print_inlining_output(NULL),
 994     _allowed_reasons(0),
 995     _interpreter_frame_size(0),
 996     _max_node_limit(MaxNodeLimit) {
 997   C = this;
 998 
 999   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1000   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1001 
1002 #ifndef PRODUCT
1003   set_print_assembly(PrintFrameConverterAssembly);
1004   set_parsed_irreducible_loop(false);
1005 #endif
1006   set_has_irreducible_loop(false); // no loops
1007 
1008   CompileWrapper cw(this);
1009   Init(/*AliasLevel=*/ 0);
1010   init_tf((*generator)());
1011 
1012   {
1013     // The following is a dummy for the sake of GraphKit::gen_stub
1014     Unique_Node_List for_igvn(comp_arena());
1015     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1016     PhaseGVN gvn(Thread::current()->resource_area(),255);
1017     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1018     gvn.transform_no_reclaim(top());
1019 
1020     GraphKit kit;
1021     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1022   }
1023 
1024   NOT_PRODUCT( verify_graph_edges(); )
1025   Code_Gen();
1026   if (failing())  return;
1027 
1028 
1029   // Entry point will be accessed using compile->stub_entry_point();
1030   if (code_buffer() == NULL) {
1031     Matcher::soft_match_failure();
1032   } else {
1033     if (PrintAssembly && (WizardMode || Verbose))
1034       tty->print_cr("### Stub::%s", stub_name);
1035 
1036     if (!failing()) {
1037       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1038 
1039       // Make the NMethod
1040       // For now we mark the frame as never safe for profile stackwalking
1041       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1042                                                       code_buffer(),
1043                                                       CodeOffsets::frame_never_safe,
1044                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1045                                                       frame_size_in_words(),
1046                                                       _oop_map_set,
1047                                                       save_arg_registers);
1048       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1049 
1050       _stub_entry_point = rs->entry_point();
1051     }
1052   }
1053 }
1054 
1055 //------------------------------Init-------------------------------------------
1056 // Prepare for a single compilation
1057 void Compile::Init(int aliaslevel) {
1058   _unique  = 0;
1059   _regalloc = NULL;
1060 
1061   _tf      = NULL;  // filled in later
1062   _top     = NULL;  // cached later
1063   _matcher = NULL;  // filled in later
1064   _cfg     = NULL;  // filled in later
1065 
1066   set_24_bit_selection_and_mode(Use24BitFP, false);
1067 
1068   _node_note_array = NULL;
1069   _default_node_notes = NULL;
1070   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1071 
1072   _immutable_memory = NULL; // filled in at first inquiry
1073 
1074   // Globally visible Nodes
1075   // First set TOP to NULL to give safe behavior during creation of RootNode
1076   set_cached_top_node(NULL);
1077   set_root(new RootNode());
1078   // Now that you have a Root to point to, create the real TOP
1079   set_cached_top_node( new ConNode(Type::TOP) );
1080   set_recent_alloc(NULL, NULL);
1081 
1082   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1083   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1084   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1085   env()->set_dependencies(new Dependencies(env()));
1086 
1087   _fixed_slots = 0;
1088   set_has_split_ifs(false);
1089   set_has_loops(has_method() && method()->has_loops()); // first approximation
1090   set_has_stringbuilder(false);
1091   set_has_boxed_value(false);
1092   _trap_can_recompile = false;  // no traps emitted yet
1093   _major_progress = true; // start out assuming good things will happen
1094   set_has_unsafe_access(false);
1095   set_max_vector_size(0);
1096   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1097   set_decompile_count(0);
1098 
1099   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1100   set_num_loop_opts(LoopOptsCount);
1101   set_do_inlining(Inline);
1102   set_max_inline_size(MaxInlineSize);
1103   set_freq_inline_size(FreqInlineSize);
1104   set_do_scheduling(OptoScheduling);
1105   set_do_count_invocations(false);
1106   set_do_method_data_update(false);
1107 
1108   set_do_vector_loop(false);
1109 
1110   if (AllowVectorizeOnDemand) {
1111     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1112       set_do_vector_loop(true);
1113       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());})
1114     } else if (has_method() && method()->name() != 0 &&
1115                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1116       set_do_vector_loop(true);
1117     }
1118   }
1119   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1120   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());})
1121 
1122   set_age_code(has_method() && method()->profile_aging());
1123   set_rtm_state(NoRTM); // No RTM lock eliding by default
1124   _max_node_limit = _directive->MaxNodeLimitOption;
1125 
1126 #if INCLUDE_RTM_OPT
1127   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1128     int rtm_state = method()->method_data()->rtm_state();
1129     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1130       // Don't generate RTM lock eliding code.
1131       set_rtm_state(NoRTM);
1132     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1133       // Generate RTM lock eliding code without abort ratio calculation code.
1134       set_rtm_state(UseRTM);
1135     } else if (UseRTMDeopt) {
1136       // Generate RTM lock eliding code and include abort ratio calculation
1137       // code if UseRTMDeopt is on.
1138       set_rtm_state(ProfileRTM);
1139     }
1140   }
1141 #endif
1142   if (debug_info()->recording_non_safepoints()) {
1143     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1144                         (comp_arena(), 8, 0, NULL));
1145     set_default_node_notes(Node_Notes::make(this));
1146   }
1147 
1148   // // -- Initialize types before each compile --
1149   // // Update cached type information
1150   // if( _method && _method->constants() )
1151   //   Type::update_loaded_types(_method, _method->constants());
1152 
1153   // Init alias_type map.
1154   if (!_do_escape_analysis && aliaslevel == 3)
1155     aliaslevel = 2;  // No unique types without escape analysis
1156   _AliasLevel = aliaslevel;
1157   const int grow_ats = 16;
1158   _max_alias_types = grow_ats;
1159   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1160   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1161   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1162   {
1163     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1164   }
1165   // Initialize the first few types.
1166   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1167   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1168   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1169   _num_alias_types = AliasIdxRaw+1;
1170   // Zero out the alias type cache.
1171   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1172   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1173   probe_alias_cache(NULL)->_index = AliasIdxTop;
1174 
1175   _intrinsics = NULL;
1176   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1177   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1178   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1179   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1180   register_library_intrinsics();
1181 }
1182 
1183 //---------------------------init_start----------------------------------------
1184 // Install the StartNode on this compile object.
1185 void Compile::init_start(StartNode* s) {
1186   if (failing())
1187     return; // already failing
1188   assert(s == start(), "");
1189 }
1190 
1191 /**
1192  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1193  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1194  * the ideal graph.
1195  */
1196 StartNode* Compile::start() const {
1197   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1198   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1199     Node* start = root()->fast_out(i);
1200     if (start->is_Start()) {
1201       return start->as_Start();
1202     }
1203   }
1204   fatal("Did not find Start node!");
1205   return NULL;
1206 }
1207 
1208 //-------------------------------immutable_memory-------------------------------------
1209 // Access immutable memory
1210 Node* Compile::immutable_memory() {
1211   if (_immutable_memory != NULL) {
1212     return _immutable_memory;
1213   }
1214   StartNode* s = start();
1215   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1216     Node *p = s->fast_out(i);
1217     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1218       _immutable_memory = p;
1219       return _immutable_memory;
1220     }
1221   }
1222   ShouldNotReachHere();
1223   return NULL;
1224 }
1225 
1226 //----------------------set_cached_top_node------------------------------------
1227 // Install the cached top node, and make sure Node::is_top works correctly.
1228 void Compile::set_cached_top_node(Node* tn) {
1229   if (tn != NULL)  verify_top(tn);
1230   Node* old_top = _top;
1231   _top = tn;
1232   // Calling Node::setup_is_top allows the nodes the chance to adjust
1233   // their _out arrays.
1234   if (_top != NULL)     _top->setup_is_top();
1235   if (old_top != NULL)  old_top->setup_is_top();
1236   assert(_top == NULL || top()->is_top(), "");
1237 }
1238 
1239 #ifdef ASSERT
1240 uint Compile::count_live_nodes_by_graph_walk() {
1241   Unique_Node_List useful(comp_arena());
1242   // Get useful node list by walking the graph.
1243   identify_useful_nodes(useful);
1244   return useful.size();
1245 }
1246 
1247 void Compile::print_missing_nodes() {
1248 
1249   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1250   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1251     return;
1252   }
1253 
1254   // This is an expensive function. It is executed only when the user
1255   // specifies VerifyIdealNodeCount option or otherwise knows the
1256   // additional work that needs to be done to identify reachable nodes
1257   // by walking the flow graph and find the missing ones using
1258   // _dead_node_list.
1259 
1260   Unique_Node_List useful(comp_arena());
1261   // Get useful node list by walking the graph.
1262   identify_useful_nodes(useful);
1263 
1264   uint l_nodes = C->live_nodes();
1265   uint l_nodes_by_walk = useful.size();
1266 
1267   if (l_nodes != l_nodes_by_walk) {
1268     if (_log != NULL) {
1269       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1270       _log->stamp();
1271       _log->end_head();
1272     }
1273     VectorSet& useful_member_set = useful.member_set();
1274     int last_idx = l_nodes_by_walk;
1275     for (int i = 0; i < last_idx; i++) {
1276       if (useful_member_set.test(i)) {
1277         if (_dead_node_list.test(i)) {
1278           if (_log != NULL) {
1279             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1280           }
1281           if (PrintIdealNodeCount) {
1282             // Print the log message to tty
1283               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1284               useful.at(i)->dump();
1285           }
1286         }
1287       }
1288       else if (! _dead_node_list.test(i)) {
1289         if (_log != NULL) {
1290           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1291         }
1292         if (PrintIdealNodeCount) {
1293           // Print the log message to tty
1294           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1295         }
1296       }
1297     }
1298     if (_log != NULL) {
1299       _log->tail("mismatched_nodes");
1300     }
1301   }
1302 }
1303 void Compile::record_modified_node(Node* n) {
1304   if (_modified_nodes != NULL && !_inlining_incrementally &&
1305       n->outcnt() != 0 && !n->is_Con()) {
1306     _modified_nodes->push(n);
1307   }
1308 }
1309 
1310 void Compile::remove_modified_node(Node* n) {
1311   if (_modified_nodes != NULL) {
1312     _modified_nodes->remove(n);
1313   }
1314 }
1315 #endif
1316 
1317 #ifndef PRODUCT
1318 void Compile::verify_top(Node* tn) const {
1319   if (tn != NULL) {
1320     assert(tn->is_Con(), "top node must be a constant");
1321     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1322     assert(tn->in(0) != NULL, "must have live top node");
1323   }
1324 }
1325 #endif
1326 
1327 
1328 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1329 
1330 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1331   guarantee(arr != NULL, "");
1332   int num_blocks = arr->length();
1333   if (grow_by < num_blocks)  grow_by = num_blocks;
1334   int num_notes = grow_by * _node_notes_block_size;
1335   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1336   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1337   while (num_notes > 0) {
1338     arr->append(notes);
1339     notes     += _node_notes_block_size;
1340     num_notes -= _node_notes_block_size;
1341   }
1342   assert(num_notes == 0, "exact multiple, please");
1343 }
1344 
1345 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1346   if (source == NULL || dest == NULL)  return false;
1347 
1348   if (dest->is_Con())
1349     return false;               // Do not push debug info onto constants.
1350 
1351 #ifdef ASSERT
1352   // Leave a bread crumb trail pointing to the original node:
1353   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1354     dest->set_debug_orig(source);
1355   }
1356 #endif
1357 
1358   if (node_note_array() == NULL)
1359     return false;               // Not collecting any notes now.
1360 
1361   // This is a copy onto a pre-existing node, which may already have notes.
1362   // If both nodes have notes, do not overwrite any pre-existing notes.
1363   Node_Notes* source_notes = node_notes_at(source->_idx);
1364   if (source_notes == NULL || source_notes->is_clear())  return false;
1365   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1366   if (dest_notes == NULL || dest_notes->is_clear()) {
1367     return set_node_notes_at(dest->_idx, source_notes);
1368   }
1369 
1370   Node_Notes merged_notes = (*source_notes);
1371   // The order of operations here ensures that dest notes will win...
1372   merged_notes.update_from(dest_notes);
1373   return set_node_notes_at(dest->_idx, &merged_notes);
1374 }
1375 
1376 
1377 //--------------------------allow_range_check_smearing-------------------------
1378 // Gating condition for coalescing similar range checks.
1379 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1380 // single covering check that is at least as strong as any of them.
1381 // If the optimization succeeds, the simplified (strengthened) range check
1382 // will always succeed.  If it fails, we will deopt, and then give up
1383 // on the optimization.
1384 bool Compile::allow_range_check_smearing() const {
1385   // If this method has already thrown a range-check,
1386   // assume it was because we already tried range smearing
1387   // and it failed.
1388   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1389   return !already_trapped;
1390 }
1391 
1392 
1393 //------------------------------flatten_alias_type-----------------------------
1394 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1395   int offset = tj->offset();
1396   TypePtr::PTR ptr = tj->ptr();
1397 
1398   // Known instance (scalarizable allocation) alias only with itself.
1399   bool is_known_inst = tj->isa_oopptr() != NULL &&
1400                        tj->is_oopptr()->is_known_instance();
1401 
1402   // Process weird unsafe references.
1403   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1404     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1405     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1406     tj = TypeOopPtr::BOTTOM;
1407     ptr = tj->ptr();
1408     offset = tj->offset();
1409   }
1410 
1411   // Array pointers need some flattening
1412   const TypeAryPtr *ta = tj->isa_aryptr();
1413   if (ta && ta->is_stable()) {
1414     // Erase stability property for alias analysis.
1415     tj = ta = ta->cast_to_stable(false);
1416   }
1417   if( ta && is_known_inst ) {
1418     if ( offset != Type::OffsetBot &&
1419          offset > arrayOopDesc::length_offset_in_bytes() ) {
1420       offset = Type::OffsetBot; // Flatten constant access into array body only
1421       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1422     }
1423   } else if( ta && _AliasLevel >= 2 ) {
1424     // For arrays indexed by constant indices, we flatten the alias
1425     // space to include all of the array body.  Only the header, klass
1426     // and array length can be accessed un-aliased.
1427     if( offset != Type::OffsetBot ) {
1428       if( ta->const_oop() ) { // MethodData* or Method*
1429         offset = Type::OffsetBot;   // Flatten constant access into array body
1430         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1431       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1432         // range is OK as-is.
1433         tj = ta = TypeAryPtr::RANGE;
1434       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1435         tj = TypeInstPtr::KLASS; // all klass loads look alike
1436         ta = TypeAryPtr::RANGE; // generic ignored junk
1437         ptr = TypePtr::BotPTR;
1438       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1439         tj = TypeInstPtr::MARK;
1440         ta = TypeAryPtr::RANGE; // generic ignored junk
1441         ptr = TypePtr::BotPTR;
1442       } else {                  // Random constant offset into array body
1443         offset = Type::OffsetBot;   // Flatten constant access into array body
1444         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1445       }
1446     }
1447     // Arrays of fixed size alias with arrays of unknown size.
1448     if (ta->size() != TypeInt::POS) {
1449       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1450       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1451     }
1452     // Arrays of known objects become arrays of unknown objects.
1453     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1454       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1455       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1456     }
1457     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1458       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1459       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1460     }
1461     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1462     // cannot be distinguished by bytecode alone.
1463     if (ta->elem() == TypeInt::BOOL) {
1464       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1465       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1466       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1467     }
1468     // During the 2nd round of IterGVN, NotNull castings are removed.
1469     // Make sure the Bottom and NotNull variants alias the same.
1470     // Also, make sure exact and non-exact variants alias the same.
1471     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1472       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1473     }
1474   }
1475 
1476   // Oop pointers need some flattening
1477   const TypeInstPtr *to = tj->isa_instptr();
1478   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1479     ciInstanceKlass *k = to->klass()->as_instance_klass();
1480     if( ptr == TypePtr::Constant ) {
1481       if (to->klass() != ciEnv::current()->Class_klass() ||
1482           offset < k->size_helper() * wordSize) {
1483         // No constant oop pointers (such as Strings); they alias with
1484         // unknown strings.
1485         assert(!is_known_inst, "not scalarizable allocation");
1486         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1487       }
1488     } else if( is_known_inst ) {
1489       tj = to; // Keep NotNull and klass_is_exact for instance type
1490     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1491       // During the 2nd round of IterGVN, NotNull castings are removed.
1492       // Make sure the Bottom and NotNull variants alias the same.
1493       // Also, make sure exact and non-exact variants alias the same.
1494       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1495     }
1496     if (to->speculative() != NULL) {
1497       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1498     }
1499     // Canonicalize the holder of this field
1500     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1501       // First handle header references such as a LoadKlassNode, even if the
1502       // object's klass is unloaded at compile time (4965979).
1503       if (!is_known_inst) { // Do it only for non-instance types
1504         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1505       }
1506     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1507       // Static fields are in the space above the normal instance
1508       // fields in the java.lang.Class instance.
1509       if (to->klass() != ciEnv::current()->Class_klass()) {
1510         to = NULL;
1511         tj = TypeOopPtr::BOTTOM;
1512         offset = tj->offset();
1513       }
1514     } else {
1515       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1516       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1517         if( is_known_inst ) {
1518           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1519         } else {
1520           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1521         }
1522       }
1523     }
1524   }
1525 
1526   // Klass pointers to object array klasses need some flattening
1527   const TypeKlassPtr *tk = tj->isa_klassptr();
1528   if( tk ) {
1529     // If we are referencing a field within a Klass, we need
1530     // to assume the worst case of an Object.  Both exact and
1531     // inexact types must flatten to the same alias class so
1532     // use NotNull as the PTR.
1533     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1534 
1535       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1536                                    TypeKlassPtr::OBJECT->klass(),
1537                                    offset);
1538     }
1539 
1540     ciKlass* klass = tk->klass();
1541     if( klass->is_obj_array_klass() ) {
1542       ciKlass* k = TypeAryPtr::OOPS->klass();
1543       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1544         k = TypeInstPtr::BOTTOM->klass();
1545       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1546     }
1547 
1548     // Check for precise loads from the primary supertype array and force them
1549     // to the supertype cache alias index.  Check for generic array loads from
1550     // the primary supertype array and also force them to the supertype cache
1551     // alias index.  Since the same load can reach both, we need to merge
1552     // these 2 disparate memories into the same alias class.  Since the
1553     // primary supertype array is read-only, there's no chance of confusion
1554     // where we bypass an array load and an array store.
1555     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1556     if (offset == Type::OffsetBot ||
1557         (offset >= primary_supers_offset &&
1558          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1559         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1560       offset = in_bytes(Klass::secondary_super_cache_offset());
1561       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1562     }
1563   }
1564 
1565   // Flatten all Raw pointers together.
1566   if (tj->base() == Type::RawPtr)
1567     tj = TypeRawPtr::BOTTOM;
1568 
1569   if (tj->base() == Type::AnyPtr)
1570     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1571 
1572   // Flatten all to bottom for now
1573   switch( _AliasLevel ) {
1574   case 0:
1575     tj = TypePtr::BOTTOM;
1576     break;
1577   case 1:                       // Flatten to: oop, static, field or array
1578     switch (tj->base()) {
1579     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1580     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1581     case Type::AryPtr:   // do not distinguish arrays at all
1582     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1583     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1584     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1585     default: ShouldNotReachHere();
1586     }
1587     break;
1588   case 2:                       // No collapsing at level 2; keep all splits
1589   case 3:                       // No collapsing at level 3; keep all splits
1590     break;
1591   default:
1592     Unimplemented();
1593   }
1594 
1595   offset = tj->offset();
1596   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1597 
1598   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1599           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1600           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1601           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1602           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1603           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1604           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr)  ,
1605           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1606   assert( tj->ptr() != TypePtr::TopPTR &&
1607           tj->ptr() != TypePtr::AnyNull &&
1608           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1609 //    assert( tj->ptr() != TypePtr::Constant ||
1610 //            tj->base() == Type::RawPtr ||
1611 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1612 
1613   return tj;
1614 }
1615 
1616 void Compile::AliasType::Init(int i, const TypePtr* at) {
1617   _index = i;
1618   _adr_type = at;
1619   _field = NULL;
1620   _element = NULL;
1621   _is_rewritable = true; // default
1622   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1623   if (atoop != NULL && atoop->is_known_instance()) {
1624     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1625     _general_index = Compile::current()->get_alias_index(gt);
1626   } else {
1627     _general_index = 0;
1628   }
1629 }
1630 
1631 BasicType Compile::AliasType::basic_type() const {
1632   if (element() != NULL) {
1633     const Type* element = adr_type()->is_aryptr()->elem();
1634     return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1635   } if (field() != NULL) {
1636     return field()->layout_type();
1637   } else {
1638     return T_ILLEGAL; // unknown
1639   }
1640 }
1641 
1642 //---------------------------------print_on------------------------------------
1643 #ifndef PRODUCT
1644 void Compile::AliasType::print_on(outputStream* st) {
1645   if (index() < 10)
1646         st->print("@ <%d> ", index());
1647   else  st->print("@ <%d>",  index());
1648   st->print(is_rewritable() ? "   " : " RO");
1649   int offset = adr_type()->offset();
1650   if (offset == Type::OffsetBot)
1651         st->print(" +any");
1652   else  st->print(" +%-3d", offset);
1653   st->print(" in ");
1654   adr_type()->dump_on(st);
1655   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1656   if (field() != NULL && tjp) {
1657     if (tjp->klass()  != field()->holder() ||
1658         tjp->offset() != field()->offset_in_bytes()) {
1659       st->print(" != ");
1660       field()->print();
1661       st->print(" ***");
1662     }
1663   }
1664 }
1665 
1666 void print_alias_types() {
1667   Compile* C = Compile::current();
1668   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1669   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1670     C->alias_type(idx)->print_on(tty);
1671     tty->cr();
1672   }
1673 }
1674 #endif
1675 
1676 
1677 //----------------------------probe_alias_cache--------------------------------
1678 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1679   intptr_t key = (intptr_t) adr_type;
1680   key ^= key >> logAliasCacheSize;
1681   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1682 }
1683 
1684 
1685 //-----------------------------grow_alias_types--------------------------------
1686 void Compile::grow_alias_types() {
1687   const int old_ats  = _max_alias_types; // how many before?
1688   const int new_ats  = old_ats;          // how many more?
1689   const int grow_ats = old_ats+new_ats;  // how many now?
1690   _max_alias_types = grow_ats;
1691   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1692   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1693   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1694   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1695 }
1696 
1697 
1698 //--------------------------------find_alias_type------------------------------
1699 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1700   if (_AliasLevel == 0)
1701     return alias_type(AliasIdxBot);
1702 
1703   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1704   if (ace->_adr_type == adr_type) {
1705     return alias_type(ace->_index);
1706   }
1707 
1708   // Handle special cases.
1709   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1710   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1711 
1712   // Do it the slow way.
1713   const TypePtr* flat = flatten_alias_type(adr_type);
1714 
1715 #ifdef ASSERT
1716   {
1717     ResourceMark rm;
1718     assert(flat == flatten_alias_type(flat), "not idempotent: adr_type = %s; flat = %s => %s",
1719            Type::str(adr_type), Type::str(flat), Type::str(flatten_alias_type(flat)));
1720     assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr: adr_type = %s",
1721            Type::str(adr_type));
1722     if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1723       const TypeOopPtr* foop = flat->is_oopptr();
1724       // Scalarizable allocations have exact klass always.
1725       bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1726       const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1727       assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type: foop = %s; xoop = %s",
1728              Type::str(foop), Type::str(xoop));
1729     }
1730   }
1731 #endif
1732 
1733   int idx = AliasIdxTop;
1734   for (int i = 0; i < num_alias_types(); i++) {
1735     if (alias_type(i)->adr_type() == flat) {
1736       idx = i;
1737       break;
1738     }
1739   }
1740 
1741   if (idx == AliasIdxTop) {
1742     if (no_create)  return NULL;
1743     // Grow the array if necessary.
1744     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1745     // Add a new alias type.
1746     idx = _num_alias_types++;
1747     _alias_types[idx]->Init(idx, flat);
1748     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1749     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1750     if (flat->isa_instptr()) {
1751       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1752           && flat->is_instptr()->klass() == env()->Class_klass())
1753         alias_type(idx)->set_rewritable(false);
1754     }
1755     if (flat->isa_aryptr()) {
1756 #ifdef ASSERT
1757       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1758       // (T_BYTE has the weakest alignment and size restrictions...)
1759       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1760 #endif
1761       if (flat->offset() == TypePtr::OffsetBot) {
1762         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1763       }
1764     }
1765     if (flat->isa_klassptr()) {
1766       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1767         alias_type(idx)->set_rewritable(false);
1768       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1769         alias_type(idx)->set_rewritable(false);
1770       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1771         alias_type(idx)->set_rewritable(false);
1772       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1773         alias_type(idx)->set_rewritable(false);
1774     }
1775     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1776     // but the base pointer type is not distinctive enough to identify
1777     // references into JavaThread.)
1778 
1779     // Check for final fields.
1780     const TypeInstPtr* tinst = flat->isa_instptr();
1781     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1782       ciField* field;
1783       if (tinst->const_oop() != NULL &&
1784           tinst->klass() == ciEnv::current()->Class_klass() &&
1785           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1786         // static field
1787         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1788         field = k->get_field_by_offset(tinst->offset(), true);
1789       } else {
1790         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1791         field = k->get_field_by_offset(tinst->offset(), false);
1792       }
1793       assert(field == NULL ||
1794              original_field == NULL ||
1795              (field->holder() == original_field->holder() &&
1796               field->offset() == original_field->offset() &&
1797               field->is_static() == original_field->is_static()), "wrong field?");
1798       // Set field() and is_rewritable() attributes.
1799       if (field != NULL)  alias_type(idx)->set_field(field);
1800     }
1801   }
1802 
1803   // Fill the cache for next time.
1804   ace->_adr_type = adr_type;
1805   ace->_index    = idx;
1806   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1807 
1808   // Might as well try to fill the cache for the flattened version, too.
1809   AliasCacheEntry* face = probe_alias_cache(flat);
1810   if (face->_adr_type == NULL) {
1811     face->_adr_type = flat;
1812     face->_index    = idx;
1813     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1814   }
1815 
1816   return alias_type(idx);
1817 }
1818 
1819 
1820 Compile::AliasType* Compile::alias_type(ciField* field) {
1821   const TypeOopPtr* t;
1822   if (field->is_static())
1823     t = TypeInstPtr::make(field->holder()->java_mirror());
1824   else
1825     t = TypeOopPtr::make_from_klass_raw(field->holder());
1826   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1827   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1828   return atp;
1829 }
1830 
1831 
1832 //------------------------------have_alias_type--------------------------------
1833 bool Compile::have_alias_type(const TypePtr* adr_type) {
1834   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1835   if (ace->_adr_type == adr_type) {
1836     return true;
1837   }
1838 
1839   // Handle special cases.
1840   if (adr_type == NULL)             return true;
1841   if (adr_type == TypePtr::BOTTOM)  return true;
1842 
1843   return find_alias_type(adr_type, true, NULL) != NULL;
1844 }
1845 
1846 //-----------------------------must_alias--------------------------------------
1847 // True if all values of the given address type are in the given alias category.
1848 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1849   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1850   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1851   if (alias_idx == AliasIdxTop)         return false; // the empty category
1852   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1853 
1854   // the only remaining possible overlap is identity
1855   int adr_idx = get_alias_index(adr_type);
1856   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1857   assert(adr_idx == alias_idx ||
1858          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1859           && adr_type                       != TypeOopPtr::BOTTOM),
1860          "should not be testing for overlap with an unsafe pointer");
1861   return adr_idx == alias_idx;
1862 }
1863 
1864 //------------------------------can_alias--------------------------------------
1865 // True if any values of the given address type are in the given alias category.
1866 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1867   if (alias_idx == AliasIdxTop)         return false; // the empty category
1868   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1869   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1870   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1871 
1872   // the only remaining possible overlap is identity
1873   int adr_idx = get_alias_index(adr_type);
1874   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1875   return adr_idx == alias_idx;
1876 }
1877 
1878 
1879 
1880 //---------------------------pop_warm_call-------------------------------------
1881 WarmCallInfo* Compile::pop_warm_call() {
1882   WarmCallInfo* wci = _warm_calls;
1883   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1884   return wci;
1885 }
1886 
1887 //----------------------------Inline_Warm--------------------------------------
1888 int Compile::Inline_Warm() {
1889   // If there is room, try to inline some more warm call sites.
1890   // %%% Do a graph index compaction pass when we think we're out of space?
1891   if (!InlineWarmCalls)  return 0;
1892 
1893   int calls_made_hot = 0;
1894   int room_to_grow   = NodeCountInliningCutoff - unique();
1895   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1896   int amount_grown   = 0;
1897   WarmCallInfo* call;
1898   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1899     int est_size = (int)call->size();
1900     if (est_size > (room_to_grow - amount_grown)) {
1901       // This one won't fit anyway.  Get rid of it.
1902       call->make_cold();
1903       continue;
1904     }
1905     call->make_hot();
1906     calls_made_hot++;
1907     amount_grown   += est_size;
1908     amount_to_grow -= est_size;
1909   }
1910 
1911   if (calls_made_hot > 0)  set_major_progress();
1912   return calls_made_hot;
1913 }
1914 
1915 
1916 //----------------------------Finish_Warm--------------------------------------
1917 void Compile::Finish_Warm() {
1918   if (!InlineWarmCalls)  return;
1919   if (failing())  return;
1920   if (warm_calls() == NULL)  return;
1921 
1922   // Clean up loose ends, if we are out of space for inlining.
1923   WarmCallInfo* call;
1924   while ((call = pop_warm_call()) != NULL) {
1925     call->make_cold();
1926   }
1927 }
1928 
1929 //---------------------cleanup_loop_predicates-----------------------
1930 // Remove the opaque nodes that protect the predicates so that all unused
1931 // checks and uncommon_traps will be eliminated from the ideal graph
1932 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1933   if (predicate_count()==0) return;
1934   for (int i = predicate_count(); i > 0; i--) {
1935     Node * n = predicate_opaque1_node(i-1);
1936     assert(n->Opcode() == Op_Opaque1, "must be");
1937     igvn.replace_node(n, n->in(1));
1938   }
1939   assert(predicate_count()==0, "should be clean!");
1940 }
1941 
1942 void Compile::add_range_check_cast(Node* n) {
1943   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1944   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1945   _range_check_casts->append(n);
1946 }
1947 
1948 // Remove all range check dependent CastIINodes.
1949 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1950   for (int i = range_check_cast_count(); i > 0; i--) {
1951     Node* cast = range_check_cast_node(i-1);
1952     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1953     igvn.replace_node(cast, cast->in(1));
1954   }
1955   assert(range_check_cast_count() == 0, "should be empty");
1956 }
1957 
1958 // StringOpts and late inlining of string methods
1959 void Compile::inline_string_calls(bool parse_time) {
1960   {
1961     // remove useless nodes to make the usage analysis simpler
1962     ResourceMark rm;
1963     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1964   }
1965 
1966   {
1967     ResourceMark rm;
1968     print_method(PHASE_BEFORE_STRINGOPTS, 3);
1969     PhaseStringOpts pso(initial_gvn(), for_igvn());
1970     print_method(PHASE_AFTER_STRINGOPTS, 3);
1971   }
1972 
1973   // now inline anything that we skipped the first time around
1974   if (!parse_time) {
1975     _late_inlines_pos = _late_inlines.length();
1976   }
1977 
1978   while (_string_late_inlines.length() > 0) {
1979     CallGenerator* cg = _string_late_inlines.pop();
1980     cg->do_late_inline();
1981     if (failing())  return;
1982   }
1983   _string_late_inlines.trunc_to(0);
1984 }
1985 
1986 // Late inlining of boxing methods
1987 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1988   if (_boxing_late_inlines.length() > 0) {
1989     assert(has_boxed_value(), "inconsistent");
1990 
1991     PhaseGVN* gvn = initial_gvn();
1992     set_inlining_incrementally(true);
1993 
1994     assert( igvn._worklist.size() == 0, "should be done with igvn" );
1995     for_igvn()->clear();
1996     gvn->replace_with(&igvn);
1997 
1998     _late_inlines_pos = _late_inlines.length();
1999 
2000     while (_boxing_late_inlines.length() > 0) {
2001       CallGenerator* cg = _boxing_late_inlines.pop();
2002       cg->do_late_inline();
2003       if (failing())  return;
2004     }
2005     _boxing_late_inlines.trunc_to(0);
2006 
2007     {
2008       ResourceMark rm;
2009       PhaseRemoveUseless pru(gvn, for_igvn());
2010     }
2011 
2012     igvn = PhaseIterGVN(gvn);
2013     igvn.optimize();
2014 
2015     set_inlining_progress(false);
2016     set_inlining_incrementally(false);
2017   }
2018 }
2019 
2020 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2021   assert(IncrementalInline, "incremental inlining should be on");
2022   PhaseGVN* gvn = initial_gvn();
2023 
2024   set_inlining_progress(false);
2025   for_igvn()->clear();
2026   gvn->replace_with(&igvn);
2027 
2028   {
2029     TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2030     int i = 0;
2031     for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2032       CallGenerator* cg = _late_inlines.at(i);
2033       _late_inlines_pos = i+1;
2034       cg->do_late_inline();
2035       if (failing())  return;
2036     }
2037     int j = 0;
2038     for (; i < _late_inlines.length(); i++, j++) {
2039       _late_inlines.at_put(j, _late_inlines.at(i));
2040     }
2041     _late_inlines.trunc_to(j);
2042   }
2043 
2044   {
2045     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2046     ResourceMark rm;
2047     PhaseRemoveUseless pru(gvn, for_igvn());
2048   }
2049 
2050   {
2051     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2052     igvn = PhaseIterGVN(gvn);
2053   }
2054 }
2055 
2056 // Perform incremental inlining until bound on number of live nodes is reached
2057 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2058   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2059 
2060   PhaseGVN* gvn = initial_gvn();
2061 
2062   set_inlining_incrementally(true);
2063   set_inlining_progress(true);
2064   uint low_live_nodes = 0;
2065 
2066   while(inlining_progress() && _late_inlines.length() > 0) {
2067 
2068     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2069       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2070         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2071         // PhaseIdealLoop is expensive so we only try it once we are
2072         // out of live nodes and we only try it again if the previous
2073         // helped got the number of nodes down significantly
2074         PhaseIdealLoop ideal_loop( igvn, false, true );
2075         if (failing())  return;
2076         low_live_nodes = live_nodes();
2077         _major_progress = true;
2078       }
2079 
2080       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2081         break;
2082       }
2083     }
2084 
2085     inline_incrementally_one(igvn);
2086 
2087     if (failing())  return;
2088 
2089     {
2090       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2091       igvn.optimize();
2092     }
2093 
2094     if (failing())  return;
2095   }
2096 
2097   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2098 
2099   if (_string_late_inlines.length() > 0) {
2100     assert(has_stringbuilder(), "inconsistent");
2101     for_igvn()->clear();
2102     initial_gvn()->replace_with(&igvn);
2103 
2104     inline_string_calls(false);
2105 
2106     if (failing())  return;
2107 
2108     {
2109       TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2110       ResourceMark rm;
2111       PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2112     }
2113 
2114     {
2115       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2116       igvn = PhaseIterGVN(gvn);
2117       igvn.optimize();
2118     }
2119   }
2120 
2121   set_inlining_incrementally(false);
2122 }
2123 
2124 
2125 //------------------------------Optimize---------------------------------------
2126 // Given a graph, optimize it.
2127 void Compile::Optimize() {
2128   TracePhase tp("optimizer", &timers[_t_optimizer]);
2129 
2130 #ifndef PRODUCT
2131   if (_directive->BreakAtCompileOption) {
2132     BREAKPOINT;
2133   }
2134 
2135 #endif
2136 
2137   ResourceMark rm;
2138   int          loop_opts_cnt;
2139 
2140   print_inlining_reinit();
2141 
2142   NOT_PRODUCT( verify_graph_edges(); )
2143 
2144   print_method(PHASE_AFTER_PARSING);
2145 
2146  {
2147   // Iterative Global Value Numbering, including ideal transforms
2148   // Initialize IterGVN with types and values from parse-time GVN
2149   PhaseIterGVN igvn(initial_gvn());
2150 #ifdef ASSERT
2151   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2152 #endif
2153   {
2154     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2155     igvn.optimize();
2156   }
2157 
2158   print_method(PHASE_ITER_GVN1, 2);
2159 
2160   if (failing())  return;
2161 
2162   inline_incrementally(igvn);
2163 
2164   print_method(PHASE_INCREMENTAL_INLINE, 2);
2165 
2166   if (failing())  return;
2167 
2168   if (eliminate_boxing()) {
2169     // Inline valueOf() methods now.
2170     inline_boxing_calls(igvn);
2171 
2172     if (AlwaysIncrementalInline) {
2173       inline_incrementally(igvn);
2174     }
2175 
2176     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2177 
2178     if (failing())  return;
2179   }
2180 
2181   // Remove the speculative part of types and clean up the graph from
2182   // the extra CastPP nodes whose only purpose is to carry them. Do
2183   // that early so that optimizations are not disrupted by the extra
2184   // CastPP nodes.
2185   remove_speculative_types(igvn);
2186 
2187   // No more new expensive nodes will be added to the list from here
2188   // so keep only the actual candidates for optimizations.
2189   cleanup_expensive_nodes(igvn);
2190 
2191   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2192     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2193     initial_gvn()->replace_with(&igvn);
2194     for_igvn()->clear();
2195     Unique_Node_List new_worklist(C->comp_arena());
2196     {
2197       ResourceMark rm;
2198       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2199     }
2200     set_for_igvn(&new_worklist);
2201     igvn = PhaseIterGVN(initial_gvn());
2202     igvn.optimize();
2203   }
2204 
2205   // Perform escape analysis
2206   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2207     if (has_loops()) {
2208       // Cleanup graph (remove dead nodes).
2209       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2210       PhaseIdealLoop ideal_loop( igvn, false, true );
2211       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2212       if (failing())  return;
2213     }
2214     ConnectionGraph::do_analysis(this, &igvn);
2215 
2216     if (failing())  return;
2217 
2218     // Optimize out fields loads from scalar replaceable allocations.
2219     igvn.optimize();
2220     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2221 
2222     if (failing())  return;
2223 
2224     if (congraph() != NULL && macro_count() > 0) {
2225       TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2226       PhaseMacroExpand mexp(igvn);
2227       mexp.eliminate_macro_nodes();
2228       igvn.set_delay_transform(false);
2229 
2230       igvn.optimize();
2231       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2232 
2233       if (failing())  return;
2234     }
2235   }
2236 
2237   // Loop transforms on the ideal graph.  Range Check Elimination,
2238   // peeling, unrolling, etc.
2239 
2240   // Set loop opts counter
2241   loop_opts_cnt = num_loop_opts();
2242   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2243     {
2244       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2245       PhaseIdealLoop ideal_loop( igvn, true );
2246       loop_opts_cnt--;
2247       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2248       if (failing())  return;
2249     }
2250     // Loop opts pass if partial peeling occurred in previous pass
2251     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2252       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2253       PhaseIdealLoop ideal_loop( igvn, false );
2254       loop_opts_cnt--;
2255       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2256       if (failing())  return;
2257     }
2258     // Loop opts pass for loop-unrolling before CCP
2259     if(major_progress() && (loop_opts_cnt > 0)) {
2260       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2261       PhaseIdealLoop ideal_loop( igvn, false );
2262       loop_opts_cnt--;
2263       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2264     }
2265     if (!failing()) {
2266       // Verify that last round of loop opts produced a valid graph
2267       TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2268       PhaseIdealLoop::verify(igvn);
2269     }
2270   }
2271   if (failing())  return;
2272 
2273   // Conditional Constant Propagation;
2274   PhaseCCP ccp( &igvn );
2275   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2276   {
2277     TracePhase tp("ccp", &timers[_t_ccp]);
2278     ccp.do_transform();
2279   }
2280   print_method(PHASE_CPP1, 2);
2281 
2282   assert( true, "Break here to ccp.dump_old2new_map()");
2283 
2284   // Iterative Global Value Numbering, including ideal transforms
2285   {
2286     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2287     igvn = ccp;
2288     igvn.optimize();
2289   }
2290 
2291   print_method(PHASE_ITER_GVN2, 2);
2292 
2293   if (failing())  return;
2294 
2295   // Loop transforms on the ideal graph.  Range Check Elimination,
2296   // peeling, unrolling, etc.
2297   if(loop_opts_cnt > 0) {
2298     debug_only( int cnt = 0; );
2299     while(major_progress() && (loop_opts_cnt > 0)) {
2300       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2301       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2302       PhaseIdealLoop ideal_loop( igvn, true);
2303       loop_opts_cnt--;
2304       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2305       if (failing())  return;
2306     }
2307   }
2308   // Ensure that major progress is now clear
2309   C->clear_major_progress();
2310 
2311   {
2312     // Verify that all previous optimizations produced a valid graph
2313     // at least to this point, even if no loop optimizations were done.
2314     TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2315     PhaseIdealLoop::verify(igvn);
2316   }
2317 
2318   if (range_check_cast_count() > 0) {
2319     // No more loop optimizations. Remove all range check dependent CastIINodes.
2320     C->remove_range_check_casts(igvn);
2321     igvn.optimize();
2322   }
2323 
2324   {
2325     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2326     PhaseMacroExpand  mex(igvn);
2327     if (mex.expand_macro_nodes()) {
2328       assert(failing(), "must bail out w/ explicit message");
2329       return;
2330     }
2331   }
2332 
2333   DEBUG_ONLY( _modified_nodes = NULL; )
2334  } // (End scope of igvn; run destructor if necessary for asserts.)
2335 
2336  process_print_inlining();
2337  // A method with only infinite loops has no edges entering loops from root
2338  {
2339    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2340    if (final_graph_reshaping()) {
2341      assert(failing(), "must bail out w/ explicit message");
2342      return;
2343    }
2344  }
2345 
2346  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2347 }
2348 
2349 
2350 //------------------------------Code_Gen---------------------------------------
2351 // Given a graph, generate code for it
2352 void Compile::Code_Gen() {
2353   if (failing()) {
2354     return;
2355   }
2356 
2357   // Perform instruction selection.  You might think we could reclaim Matcher
2358   // memory PDQ, but actually the Matcher is used in generating spill code.
2359   // Internals of the Matcher (including some VectorSets) must remain live
2360   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2361   // set a bit in reclaimed memory.
2362 
2363   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2364   // nodes.  Mapping is only valid at the root of each matched subtree.
2365   NOT_PRODUCT( verify_graph_edges(); )
2366 
2367   Matcher matcher;
2368   _matcher = &matcher;
2369   {
2370     TracePhase tp("matcher", &timers[_t_matcher]);
2371     matcher.match();
2372   }
2373   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2374   // nodes.  Mapping is only valid at the root of each matched subtree.
2375   NOT_PRODUCT( verify_graph_edges(); )
2376 
2377   // If you have too many nodes, or if matching has failed, bail out
2378   check_node_count(0, "out of nodes matching instructions");
2379   if (failing()) {
2380     return;
2381   }
2382 
2383   // Build a proper-looking CFG
2384   PhaseCFG cfg(node_arena(), root(), matcher);
2385   _cfg = &cfg;
2386   {
2387     TracePhase tp("scheduler", &timers[_t_scheduler]);
2388     bool success = cfg.do_global_code_motion();
2389     if (!success) {
2390       return;
2391     }
2392 
2393     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2394     NOT_PRODUCT( verify_graph_edges(); )
2395     debug_only( cfg.verify(); )
2396   }
2397 
2398   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2399   _regalloc = &regalloc;
2400   {
2401     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2402     // Perform register allocation.  After Chaitin, use-def chains are
2403     // no longer accurate (at spill code) and so must be ignored.
2404     // Node->LRG->reg mappings are still accurate.
2405     _regalloc->Register_Allocate();
2406 
2407     // Bail out if the allocator builds too many nodes
2408     if (failing()) {
2409       return;
2410     }
2411   }
2412 
2413   // Prior to register allocation we kept empty basic blocks in case the
2414   // the allocator needed a place to spill.  After register allocation we
2415   // are not adding any new instructions.  If any basic block is empty, we
2416   // can now safely remove it.
2417   {
2418     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2419     cfg.remove_empty_blocks();
2420     if (do_freq_based_layout()) {
2421       PhaseBlockLayout layout(cfg);
2422     } else {
2423       cfg.set_loop_alignment();
2424     }
2425     cfg.fixup_flow();
2426   }
2427 
2428   // Apply peephole optimizations
2429   if( OptoPeephole ) {
2430     TracePhase tp("peephole", &timers[_t_peephole]);
2431     PhasePeephole peep( _regalloc, cfg);
2432     peep.do_transform();
2433   }
2434 
2435   // Do late expand if CPU requires this.
2436   if (Matcher::require_postalloc_expand) {
2437     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2438     cfg.postalloc_expand(_regalloc);
2439   }
2440 
2441   // Convert Nodes to instruction bits in a buffer
2442   {
2443     TraceTime tp("output", &timers[_t_output], CITime);
2444     Output();
2445   }
2446 
2447   print_method(PHASE_FINAL_CODE);
2448 
2449   // He's dead, Jim.
2450   _cfg     = (PhaseCFG*)0xdeadbeef;
2451   _regalloc = (PhaseChaitin*)0xdeadbeef;
2452 }
2453 
2454 
2455 //------------------------------dump_asm---------------------------------------
2456 // Dump formatted assembly
2457 #ifndef PRODUCT
2458 void Compile::dump_asm(int *pcs, uint pc_limit) {
2459   bool cut_short = false;
2460   tty->print_cr("#");
2461   tty->print("#  ");  _tf->dump();  tty->cr();
2462   tty->print_cr("#");
2463 
2464   // For all blocks
2465   int pc = 0x0;                 // Program counter
2466   char starts_bundle = ' ';
2467   _regalloc->dump_frame();
2468 
2469   Node *n = NULL;
2470   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2471     if (VMThread::should_terminate()) {
2472       cut_short = true;
2473       break;
2474     }
2475     Block* block = _cfg->get_block(i);
2476     if (block->is_connector() && !Verbose) {
2477       continue;
2478     }
2479     n = block->head();
2480     if (pcs && n->_idx < pc_limit) {
2481       tty->print("%3.3x   ", pcs[n->_idx]);
2482     } else {
2483       tty->print("      ");
2484     }
2485     block->dump_head(_cfg);
2486     if (block->is_connector()) {
2487       tty->print_cr("        # Empty connector block");
2488     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2489       tty->print_cr("        # Block is sole successor of call");
2490     }
2491 
2492     // For all instructions
2493     Node *delay = NULL;
2494     for (uint j = 0; j < block->number_of_nodes(); j++) {
2495       if (VMThread::should_terminate()) {
2496         cut_short = true;
2497         break;
2498       }
2499       n = block->get_node(j);
2500       if (valid_bundle_info(n)) {
2501         Bundle* bundle = node_bundling(n);
2502         if (bundle->used_in_unconditional_delay()) {
2503           delay = n;
2504           continue;
2505         }
2506         if (bundle->starts_bundle()) {
2507           starts_bundle = '+';
2508         }
2509       }
2510 
2511       if (WizardMode) {
2512         n->dump();
2513       }
2514 
2515       if( !n->is_Region() &&    // Dont print in the Assembly
2516           !n->is_Phi() &&       // a few noisely useless nodes
2517           !n->is_Proj() &&
2518           !n->is_MachTemp() &&
2519           !n->is_SafePointScalarObject() &&
2520           !n->is_Catch() &&     // Would be nice to print exception table targets
2521           !n->is_MergeMem() &&  // Not very interesting
2522           !n->is_top() &&       // Debug info table constants
2523           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2524           ) {
2525         if (pcs && n->_idx < pc_limit)
2526           tty->print("%3.3x", pcs[n->_idx]);
2527         else
2528           tty->print("   ");
2529         tty->print(" %c ", starts_bundle);
2530         starts_bundle = ' ';
2531         tty->print("\t");
2532         n->format(_regalloc, tty);
2533         tty->cr();
2534       }
2535 
2536       // If we have an instruction with a delay slot, and have seen a delay,
2537       // then back up and print it
2538       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2539         assert(delay != NULL, "no unconditional delay instruction");
2540         if (WizardMode) delay->dump();
2541 
2542         if (node_bundling(delay)->starts_bundle())
2543           starts_bundle = '+';
2544         if (pcs && n->_idx < pc_limit)
2545           tty->print("%3.3x", pcs[n->_idx]);
2546         else
2547           tty->print("   ");
2548         tty->print(" %c ", starts_bundle);
2549         starts_bundle = ' ';
2550         tty->print("\t");
2551         delay->format(_regalloc, tty);
2552         tty->cr();
2553         delay = NULL;
2554       }
2555 
2556       // Dump the exception table as well
2557       if( n->is_Catch() && (Verbose || WizardMode) ) {
2558         // Print the exception table for this offset
2559         _handler_table.print_subtable_for(pc);
2560       }
2561     }
2562 
2563     if (pcs && n->_idx < pc_limit)
2564       tty->print_cr("%3.3x", pcs[n->_idx]);
2565     else
2566       tty->cr();
2567 
2568     assert(cut_short || delay == NULL, "no unconditional delay branch");
2569 
2570   } // End of per-block dump
2571   tty->cr();
2572 
2573   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2574 }
2575 #endif
2576 
2577 //------------------------------Final_Reshape_Counts---------------------------
2578 // This class defines counters to help identify when a method
2579 // may/must be executed using hardware with only 24-bit precision.
2580 struct Final_Reshape_Counts : public StackObj {
2581   int  _call_count;             // count non-inlined 'common' calls
2582   int  _float_count;            // count float ops requiring 24-bit precision
2583   int  _double_count;           // count double ops requiring more precision
2584   int  _java_call_count;        // count non-inlined 'java' calls
2585   int  _inner_loop_count;       // count loops which need alignment
2586   VectorSet _visited;           // Visitation flags
2587   Node_List _tests;             // Set of IfNodes & PCTableNodes
2588 
2589   Final_Reshape_Counts() :
2590     _call_count(0), _float_count(0), _double_count(0),
2591     _java_call_count(0), _inner_loop_count(0),
2592     _visited( Thread::current()->resource_area() ) { }
2593 
2594   void inc_call_count  () { _call_count  ++; }
2595   void inc_float_count () { _float_count ++; }
2596   void inc_double_count() { _double_count++; }
2597   void inc_java_call_count() { _java_call_count++; }
2598   void inc_inner_loop_count() { _inner_loop_count++; }
2599 
2600   int  get_call_count  () const { return _call_count  ; }
2601   int  get_float_count () const { return _float_count ; }
2602   int  get_double_count() const { return _double_count; }
2603   int  get_java_call_count() const { return _java_call_count; }
2604   int  get_inner_loop_count() const { return _inner_loop_count; }
2605 };
2606 
2607 #ifdef ASSERT
2608 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2609   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2610   // Make sure the offset goes inside the instance layout.
2611   return k->contains_field_offset(tp->offset());
2612   // Note that OffsetBot and OffsetTop are very negative.
2613 }
2614 #endif
2615 
2616 // Eliminate trivially redundant StoreCMs and accumulate their
2617 // precedence edges.
2618 void Compile::eliminate_redundant_card_marks(Node* n) {
2619   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2620   if (n->in(MemNode::Address)->outcnt() > 1) {
2621     // There are multiple users of the same address so it might be
2622     // possible to eliminate some of the StoreCMs
2623     Node* mem = n->in(MemNode::Memory);
2624     Node* adr = n->in(MemNode::Address);
2625     Node* val = n->in(MemNode::ValueIn);
2626     Node* prev = n;
2627     bool done = false;
2628     // Walk the chain of StoreCMs eliminating ones that match.  As
2629     // long as it's a chain of single users then the optimization is
2630     // safe.  Eliminating partially redundant StoreCMs would require
2631     // cloning copies down the other paths.
2632     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2633       if (adr == mem->in(MemNode::Address) &&
2634           val == mem->in(MemNode::ValueIn)) {
2635         // redundant StoreCM
2636         if (mem->req() > MemNode::OopStore) {
2637           // Hasn't been processed by this code yet.
2638           n->add_prec(mem->in(MemNode::OopStore));
2639         } else {
2640           // Already converted to precedence edge
2641           for (uint i = mem->req(); i < mem->len(); i++) {
2642             // Accumulate any precedence edges
2643             if (mem->in(i) != NULL) {
2644               n->add_prec(mem->in(i));
2645             }
2646           }
2647           // Everything above this point has been processed.
2648           done = true;
2649         }
2650         // Eliminate the previous StoreCM
2651         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2652         assert(mem->outcnt() == 0, "should be dead");
2653         mem->disconnect_inputs(NULL, this);
2654       } else {
2655         prev = mem;
2656       }
2657       mem = prev->in(MemNode::Memory);
2658     }
2659   }
2660 }
2661 
2662 //------------------------------final_graph_reshaping_impl----------------------
2663 // Implement items 1-5 from final_graph_reshaping below.
2664 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2665 
2666   if ( n->outcnt() == 0 ) return; // dead node
2667   uint nop = n->Opcode();
2668 
2669   // Check for 2-input instruction with "last use" on right input.
2670   // Swap to left input.  Implements item (2).
2671   if( n->req() == 3 &&          // two-input instruction
2672       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2673       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2674       n->in(2)->outcnt() == 1 &&// right use IS a last use
2675       !n->in(2)->is_Con() ) {   // right use is not a constant
2676     // Check for commutative opcode
2677     switch( nop ) {
2678     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2679     case Op_MaxI:  case Op_MinI:
2680     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2681     case Op_AndL:  case Op_XorL:  case Op_OrL:
2682     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2683       // Move "last use" input to left by swapping inputs
2684       n->swap_edges(1, 2);
2685       break;
2686     }
2687     default:
2688       break;
2689     }
2690   }
2691 
2692 #ifdef ASSERT
2693   if( n->is_Mem() ) {
2694     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2695     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2696             // oop will be recorded in oop map if load crosses safepoint
2697             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2698                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2699             "raw memory operations should have control edge");
2700   }
2701 #endif
2702   // Count FPU ops and common calls, implements item (3)
2703   switch( nop ) {
2704   // Count all float operations that may use FPU
2705   case Op_AddF:
2706   case Op_SubF:
2707   case Op_MulF:
2708   case Op_DivF:
2709   case Op_NegF:
2710   case Op_ModF:
2711   case Op_ConvI2F:
2712   case Op_ConF:
2713   case Op_CmpF:
2714   case Op_CmpF3:
2715   // case Op_ConvL2F: // longs are split into 32-bit halves
2716     frc.inc_float_count();
2717     break;
2718 
2719   case Op_ConvF2D:
2720   case Op_ConvD2F:
2721     frc.inc_float_count();
2722     frc.inc_double_count();
2723     break;
2724 
2725   // Count all double operations that may use FPU
2726   case Op_AddD:
2727   case Op_SubD:
2728   case Op_MulD:
2729   case Op_DivD:
2730   case Op_NegD:
2731   case Op_ModD:
2732   case Op_ConvI2D:
2733   case Op_ConvD2I:
2734   // case Op_ConvL2D: // handled by leaf call
2735   // case Op_ConvD2L: // handled by leaf call
2736   case Op_ConD:
2737   case Op_CmpD:
2738   case Op_CmpD3:
2739     frc.inc_double_count();
2740     break;
2741   case Op_Opaque1:              // Remove Opaque Nodes before matching
2742   case Op_Opaque2:              // Remove Opaque Nodes before matching
2743   case Op_Opaque3:
2744     n->subsume_by(n->in(1), this);
2745     break;
2746   case Op_CallStaticJava:
2747   case Op_CallJava:
2748   case Op_CallDynamicJava:
2749     frc.inc_java_call_count(); // Count java call site;
2750   case Op_CallRuntime:
2751   case Op_CallLeaf:
2752   case Op_CallLeafNoFP: {
2753     assert( n->is_Call(), "" );
2754     CallNode *call = n->as_Call();
2755     // Count call sites where the FP mode bit would have to be flipped.
2756     // Do not count uncommon runtime calls:
2757     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2758     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2759     if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2760       frc.inc_call_count();   // Count the call site
2761     } else {                  // See if uncommon argument is shared
2762       Node *n = call->in(TypeFunc::Parms);
2763       int nop = n->Opcode();
2764       // Clone shared simple arguments to uncommon calls, item (1).
2765       if( n->outcnt() > 1 &&
2766           !n->is_Proj() &&
2767           nop != Op_CreateEx &&
2768           nop != Op_CheckCastPP &&
2769           nop != Op_DecodeN &&
2770           nop != Op_DecodeNKlass &&
2771           !n->is_Mem() ) {
2772         Node *x = n->clone();
2773         call->set_req( TypeFunc::Parms, x );
2774       }
2775     }
2776     break;
2777   }
2778 
2779   case Op_StoreD:
2780   case Op_LoadD:
2781   case Op_LoadD_unaligned:
2782     frc.inc_double_count();
2783     goto handle_mem;
2784   case Op_StoreF:
2785   case Op_LoadF:
2786     frc.inc_float_count();
2787     goto handle_mem;
2788 
2789   case Op_StoreCM:
2790     {
2791       // Convert OopStore dependence into precedence edge
2792       Node* prec = n->in(MemNode::OopStore);
2793       n->del_req(MemNode::OopStore);
2794       n->add_prec(prec);
2795       eliminate_redundant_card_marks(n);
2796     }
2797 
2798     // fall through
2799 
2800   case Op_StoreB:
2801   case Op_StoreC:
2802   case Op_StorePConditional:
2803   case Op_StoreI:
2804   case Op_StoreL:
2805   case Op_StoreIConditional:
2806   case Op_StoreLConditional:
2807   case Op_CompareAndSwapB:
2808   case Op_CompareAndSwapS:
2809   case Op_CompareAndSwapI:
2810   case Op_CompareAndSwapL:
2811   case Op_CompareAndSwapP:
2812   case Op_CompareAndSwapN:
2813   case Op_WeakCompareAndSwapB:
2814   case Op_WeakCompareAndSwapS:
2815   case Op_WeakCompareAndSwapI:
2816   case Op_WeakCompareAndSwapL:
2817   case Op_WeakCompareAndSwapP:
2818   case Op_WeakCompareAndSwapN:
2819   case Op_CompareAndExchangeB:
2820   case Op_CompareAndExchangeS:
2821   case Op_CompareAndExchangeI:
2822   case Op_CompareAndExchangeL:
2823   case Op_CompareAndExchangeP:
2824   case Op_CompareAndExchangeN:
2825   case Op_GetAndAddS:
2826   case Op_GetAndAddB:
2827   case Op_GetAndAddI:
2828   case Op_GetAndAddL:
2829   case Op_GetAndSetS:
2830   case Op_GetAndSetB:
2831   case Op_GetAndSetI:
2832   case Op_GetAndSetL:
2833   case Op_GetAndSetP:
2834   case Op_GetAndSetN:
2835   case Op_StoreP:
2836   case Op_StoreN:
2837   case Op_StoreNKlass:
2838   case Op_LoadB:
2839   case Op_LoadUB:
2840   case Op_LoadUS:
2841   case Op_LoadI:
2842   case Op_LoadKlass:
2843   case Op_LoadNKlass:
2844   case Op_LoadL:
2845   case Op_LoadL_unaligned:
2846   case Op_LoadPLocked:
2847   case Op_LoadP:
2848   case Op_LoadN:
2849   case Op_LoadRange:
2850   case Op_LoadS: {
2851   handle_mem:
2852 #ifdef ASSERT
2853     if( VerifyOptoOopOffsets ) {
2854       assert( n->is_Mem(), "" );
2855       MemNode *mem  = (MemNode*)n;
2856       // Check to see if address types have grounded out somehow.
2857       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2858       assert( !tp || oop_offset_is_sane(tp), "" );
2859     }
2860 #endif
2861     break;
2862   }
2863 
2864   case Op_AddP: {               // Assert sane base pointers
2865     Node *addp = n->in(AddPNode::Address);
2866     assert( !addp->is_AddP() ||
2867             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2868             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2869             "Base pointers must match (addp %u)", addp->_idx );
2870 #ifdef _LP64
2871     if ((UseCompressedOops || UseCompressedClassPointers) &&
2872         addp->Opcode() == Op_ConP &&
2873         addp == n->in(AddPNode::Base) &&
2874         n->in(AddPNode::Offset)->is_Con()) {
2875       // If the transformation of ConP to ConN+DecodeN is beneficial depends
2876       // on the platform and on the compressed oops mode.
2877       // Use addressing with narrow klass to load with offset on x86.
2878       // Some platforms can use the constant pool to load ConP.
2879       // Do this transformation here since IGVN will convert ConN back to ConP.
2880       const Type* t = addp->bottom_type();
2881       bool is_oop   = t->isa_oopptr() != NULL;
2882       bool is_klass = t->isa_klassptr() != NULL;
2883 
2884       if ((is_oop   && Matcher::const_oop_prefer_decode()  ) ||
2885           (is_klass && Matcher::const_klass_prefer_decode())) {
2886         Node* nn = NULL;
2887 
2888         int op = is_oop ? Op_ConN : Op_ConNKlass;
2889 
2890         // Look for existing ConN node of the same exact type.
2891         Node* r  = root();
2892         uint cnt = r->outcnt();
2893         for (uint i = 0; i < cnt; i++) {
2894           Node* m = r->raw_out(i);
2895           if (m!= NULL && m->Opcode() == op &&
2896               m->bottom_type()->make_ptr() == t) {
2897             nn = m;
2898             break;
2899           }
2900         }
2901         if (nn != NULL) {
2902           // Decode a narrow oop to match address
2903           // [R12 + narrow_oop_reg<<3 + offset]
2904           if (is_oop) {
2905             nn = new DecodeNNode(nn, t);
2906           } else {
2907             nn = new DecodeNKlassNode(nn, t);
2908           }
2909           // Check for succeeding AddP which uses the same Base.
2910           // Otherwise we will run into the assertion above when visiting that guy.
2911           for (uint i = 0; i < n->outcnt(); ++i) {
2912             Node *out_i = n->raw_out(i);
2913             if (out_i && out_i->is_AddP() && out_i->in(AddPNode::Base) == addp) {
2914               out_i->set_req(AddPNode::Base, nn);
2915 #ifdef ASSERT
2916               for (uint j = 0; j < out_i->outcnt(); ++j) {
2917                 Node *out_j = out_i->raw_out(j);
2918                 assert(out_j == NULL || !out_j->is_AddP() || out_j->in(AddPNode::Base) != addp,
2919                        "more than 2 AddP nodes in a chain (out_j %u)", out_j->_idx);
2920               }
2921 #endif
2922             }
2923           }
2924           n->set_req(AddPNode::Base, nn);
2925           n->set_req(AddPNode::Address, nn);
2926           if (addp->outcnt() == 0) {
2927             addp->disconnect_inputs(NULL, this);
2928           }
2929         }
2930       }
2931     }
2932 #endif
2933     // platform dependent reshaping of the address expression
2934     reshape_address(n->as_AddP());
2935     break;
2936   }
2937 
2938   case Op_CastPP: {
2939     // Remove CastPP nodes to gain more freedom during scheduling but
2940     // keep the dependency they encode as control or precedence edges
2941     // (if control is set already) on memory operations. Some CastPP
2942     // nodes don't have a control (don't carry a dependency): skip
2943     // those.
2944     if (n->in(0) != NULL) {
2945       ResourceMark rm;
2946       Unique_Node_List wq;
2947       wq.push(n);
2948       for (uint next = 0; next < wq.size(); ++next) {
2949         Node *m = wq.at(next);
2950         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2951           Node* use = m->fast_out(i);
2952           if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
2953             use->ensure_control_or_add_prec(n->in(0));
2954           } else {
2955             switch(use->Opcode()) {
2956             case Op_AddP:
2957             case Op_DecodeN:
2958             case Op_DecodeNKlass:
2959             case Op_CheckCastPP:
2960             case Op_CastPP:
2961               wq.push(use);
2962               break;
2963             }
2964           }
2965         }
2966       }
2967     }
2968     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
2969     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2970       Node* in1 = n->in(1);
2971       const Type* t = n->bottom_type();
2972       Node* new_in1 = in1->clone();
2973       new_in1->as_DecodeN()->set_type(t);
2974 
2975       if (!Matcher::narrow_oop_use_complex_address()) {
2976         //
2977         // x86, ARM and friends can handle 2 adds in addressing mode
2978         // and Matcher can fold a DecodeN node into address by using
2979         // a narrow oop directly and do implicit NULL check in address:
2980         //
2981         // [R12 + narrow_oop_reg<<3 + offset]
2982         // NullCheck narrow_oop_reg
2983         //
2984         // On other platforms (Sparc) we have to keep new DecodeN node and
2985         // use it to do implicit NULL check in address:
2986         //
2987         // decode_not_null narrow_oop_reg, base_reg
2988         // [base_reg + offset]
2989         // NullCheck base_reg
2990         //
2991         // Pin the new DecodeN node to non-null path on these platform (Sparc)
2992         // to keep the information to which NULL check the new DecodeN node
2993         // corresponds to use it as value in implicit_null_check().
2994         //
2995         new_in1->set_req(0, n->in(0));
2996       }
2997 
2998       n->subsume_by(new_in1, this);
2999       if (in1->outcnt() == 0) {
3000         in1->disconnect_inputs(NULL, this);
3001       }
3002     } else {
3003       n->subsume_by(n->in(1), this);
3004       if (n->outcnt() == 0) {
3005         n->disconnect_inputs(NULL, this);
3006       }
3007     }
3008     break;
3009   }
3010 #ifdef _LP64
3011   case Op_CmpP:
3012     // Do this transformation here to preserve CmpPNode::sub() and
3013     // other TypePtr related Ideal optimizations (for example, ptr nullness).
3014     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
3015       Node* in1 = n->in(1);
3016       Node* in2 = n->in(2);
3017       if (!in1->is_DecodeNarrowPtr()) {
3018         in2 = in1;
3019         in1 = n->in(2);
3020       }
3021       assert(in1->is_DecodeNarrowPtr(), "sanity");
3022 
3023       Node* new_in2 = NULL;
3024       if (in2->is_DecodeNarrowPtr()) {
3025         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
3026         new_in2 = in2->in(1);
3027       } else if (in2->Opcode() == Op_ConP) {
3028         const Type* t = in2->bottom_type();
3029         if (t == TypePtr::NULL_PTR) {
3030           assert(in1->is_DecodeN(), "compare klass to null?");
3031           // Don't convert CmpP null check into CmpN if compressed
3032           // oops implicit null check is not generated.
3033           // This will allow to generate normal oop implicit null check.
3034           if (Matcher::gen_narrow_oop_implicit_null_checks())
3035             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
3036           //
3037           // This transformation together with CastPP transformation above
3038           // will generated code for implicit NULL checks for compressed oops.
3039           //
3040           // The original code after Optimize()
3041           //
3042           //    LoadN memory, narrow_oop_reg
3043           //    decode narrow_oop_reg, base_reg
3044           //    CmpP base_reg, NULL
3045           //    CastPP base_reg // NotNull
3046           //    Load [base_reg + offset], val_reg
3047           //
3048           // after these transformations will be
3049           //
3050           //    LoadN memory, narrow_oop_reg
3051           //    CmpN narrow_oop_reg, NULL
3052           //    decode_not_null narrow_oop_reg, base_reg
3053           //    Load [base_reg + offset], val_reg
3054           //
3055           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3056           // since narrow oops can be used in debug info now (see the code in
3057           // final_graph_reshaping_walk()).
3058           //
3059           // At the end the code will be matched to
3060           // on x86:
3061           //
3062           //    Load_narrow_oop memory, narrow_oop_reg
3063           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3064           //    NullCheck narrow_oop_reg
3065           //
3066           // and on sparc:
3067           //
3068           //    Load_narrow_oop memory, narrow_oop_reg
3069           //    decode_not_null narrow_oop_reg, base_reg
3070           //    Load [base_reg + offset], val_reg
3071           //    NullCheck base_reg
3072           //
3073         } else if (t->isa_oopptr()) {
3074           new_in2 = ConNode::make(t->make_narrowoop());
3075         } else if (t->isa_klassptr()) {
3076           new_in2 = ConNode::make(t->make_narrowklass());
3077         }
3078       }
3079       if (new_in2 != NULL) {
3080         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3081         n->subsume_by(cmpN, this);
3082         if (in1->outcnt() == 0) {
3083           in1->disconnect_inputs(NULL, this);
3084         }
3085         if (in2->outcnt() == 0) {
3086           in2->disconnect_inputs(NULL, this);
3087         }
3088       }
3089     }
3090     break;
3091 
3092   case Op_DecodeN:
3093   case Op_DecodeNKlass:
3094     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3095     // DecodeN could be pinned when it can't be fold into
3096     // an address expression, see the code for Op_CastPP above.
3097     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3098     break;
3099 
3100   case Op_EncodeP:
3101   case Op_EncodePKlass: {
3102     Node* in1 = n->in(1);
3103     if (in1->is_DecodeNarrowPtr()) {
3104       n->subsume_by(in1->in(1), this);
3105     } else if (in1->Opcode() == Op_ConP) {
3106       const Type* t = in1->bottom_type();
3107       if (t == TypePtr::NULL_PTR) {
3108         assert(t->isa_oopptr(), "null klass?");
3109         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3110       } else if (t->isa_oopptr()) {
3111         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3112       } else if (t->isa_klassptr()) {
3113         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3114       }
3115     }
3116     if (in1->outcnt() == 0) {
3117       in1->disconnect_inputs(NULL, this);
3118     }
3119     break;
3120   }
3121 
3122   case Op_Proj: {
3123     if (OptimizeStringConcat) {
3124       ProjNode* p = n->as_Proj();
3125       if (p->_is_io_use) {
3126         // Separate projections were used for the exception path which
3127         // are normally removed by a late inline.  If it wasn't inlined
3128         // then they will hang around and should just be replaced with
3129         // the original one.
3130         Node* proj = NULL;
3131         // Replace with just one
3132         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3133           Node *use = i.get();
3134           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3135             proj = use;
3136             break;
3137           }
3138         }
3139         assert(proj != NULL, "must be found");
3140         p->subsume_by(proj, this);
3141       }
3142     }
3143     break;
3144   }
3145 
3146   case Op_Phi:
3147     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3148       // The EncodeP optimization may create Phi with the same edges
3149       // for all paths. It is not handled well by Register Allocator.
3150       Node* unique_in = n->in(1);
3151       assert(unique_in != NULL, "");
3152       uint cnt = n->req();
3153       for (uint i = 2; i < cnt; i++) {
3154         Node* m = n->in(i);
3155         assert(m != NULL, "");
3156         if (unique_in != m)
3157           unique_in = NULL;
3158       }
3159       if (unique_in != NULL) {
3160         n->subsume_by(unique_in, this);
3161       }
3162     }
3163     break;
3164 
3165 #endif
3166 
3167 #ifdef ASSERT
3168   case Op_CastII:
3169     // Verify that all range check dependent CastII nodes were removed.
3170     if (n->isa_CastII()->has_range_check()) {
3171       n->dump(3);
3172       assert(false, "Range check dependent CastII node was not removed");
3173     }
3174     break;
3175 #endif
3176 
3177   case Op_ModI:
3178     if (UseDivMod) {
3179       // Check if a%b and a/b both exist
3180       Node* d = n->find_similar(Op_DivI);
3181       if (d) {
3182         // Replace them with a fused divmod if supported
3183         if (Matcher::has_match_rule(Op_DivModI)) {
3184           DivModINode* divmod = DivModINode::make(n);
3185           d->subsume_by(divmod->div_proj(), this);
3186           n->subsume_by(divmod->mod_proj(), this);
3187         } else {
3188           // replace a%b with a-((a/b)*b)
3189           Node* mult = new MulINode(d, d->in(2));
3190           Node* sub  = new SubINode(d->in(1), mult);
3191           n->subsume_by(sub, this);
3192         }
3193       }
3194     }
3195     break;
3196 
3197   case Op_ModL:
3198     if (UseDivMod) {
3199       // Check if a%b and a/b both exist
3200       Node* d = n->find_similar(Op_DivL);
3201       if (d) {
3202         // Replace them with a fused divmod if supported
3203         if (Matcher::has_match_rule(Op_DivModL)) {
3204           DivModLNode* divmod = DivModLNode::make(n);
3205           d->subsume_by(divmod->div_proj(), this);
3206           n->subsume_by(divmod->mod_proj(), this);
3207         } else {
3208           // replace a%b with a-((a/b)*b)
3209           Node* mult = new MulLNode(d, d->in(2));
3210           Node* sub  = new SubLNode(d->in(1), mult);
3211           n->subsume_by(sub, this);
3212         }
3213       }
3214     }
3215     break;
3216 
3217   case Op_LoadVector:
3218   case Op_StoreVector:
3219     break;
3220 
3221   case Op_AddReductionVI:
3222   case Op_AddReductionVL:
3223   case Op_AddReductionVF:
3224   case Op_AddReductionVD:
3225   case Op_MulReductionVI:
3226   case Op_MulReductionVL:
3227   case Op_MulReductionVF:
3228   case Op_MulReductionVD:
3229     break;
3230 
3231   case Op_PackB:
3232   case Op_PackS:
3233   case Op_PackI:
3234   case Op_PackF:
3235   case Op_PackL:
3236   case Op_PackD:
3237     if (n->req()-1 > 2) {
3238       // Replace many operand PackNodes with a binary tree for matching
3239       PackNode* p = (PackNode*) n;
3240       Node* btp = p->binary_tree_pack(1, n->req());
3241       n->subsume_by(btp, this);
3242     }
3243     break;
3244   case Op_Loop:
3245   case Op_CountedLoop:
3246     if (n->as_Loop()->is_inner_loop()) {
3247       frc.inc_inner_loop_count();
3248     }
3249     break;
3250   case Op_LShiftI:
3251   case Op_RShiftI:
3252   case Op_URShiftI:
3253   case Op_LShiftL:
3254   case Op_RShiftL:
3255   case Op_URShiftL:
3256     if (Matcher::need_masked_shift_count) {
3257       // The cpu's shift instructions don't restrict the count to the
3258       // lower 5/6 bits. We need to do the masking ourselves.
3259       Node* in2 = n->in(2);
3260       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3261       const TypeInt* t = in2->find_int_type();
3262       if (t != NULL && t->is_con()) {
3263         juint shift = t->get_con();
3264         if (shift > mask) { // Unsigned cmp
3265           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3266         }
3267       } else {
3268         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3269           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3270           n->set_req(2, shift);
3271         }
3272       }
3273       if (in2->outcnt() == 0) { // Remove dead node
3274         in2->disconnect_inputs(NULL, this);
3275       }
3276     }
3277     break;
3278   case Op_MemBarStoreStore:
3279   case Op_MemBarRelease:
3280     // Break the link with AllocateNode: it is no longer useful and
3281     // confuses register allocation.
3282     if (n->req() > MemBarNode::Precedent) {
3283       n->set_req(MemBarNode::Precedent, top());
3284     }
3285     break;
3286   case Op_RangeCheck: {
3287     RangeCheckNode* rc = n->as_RangeCheck();
3288     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3289     n->subsume_by(iff, this);
3290     frc._tests.push(iff);
3291     break;
3292   }
3293   case Op_ConvI2L: {
3294     if (!Matcher::convi2l_type_required) {
3295       // Code generation on some platforms doesn't need accurate
3296       // ConvI2L types. Widening the type can help remove redundant
3297       // address computations.
3298       n->as_Type()->set_type(TypeLong::INT);
3299       ResourceMark rm;
3300       Node_List wq;
3301       wq.push(n);
3302       for (uint next = 0; next < wq.size(); next++) {
3303         Node *m = wq.at(next);
3304 
3305         for(;;) {
3306           // Loop over all nodes with identical inputs edges as m
3307           Node* k = m->find_similar(m->Opcode());
3308           if (k == NULL) {
3309             break;
3310           }
3311           // Push their uses so we get a chance to remove node made
3312           // redundant
3313           for (DUIterator_Fast imax, i = k->fast_outs(imax); i < imax; i++) {
3314             Node* u = k->fast_out(i);
3315             assert(!wq.contains(u), "shouldn't process one node several times");
3316             if (u->Opcode() == Op_LShiftL ||
3317                 u->Opcode() == Op_AddL ||
3318                 u->Opcode() == Op_SubL ||
3319                 u->Opcode() == Op_AddP) {
3320               wq.push(u);
3321             }
3322           }
3323           // Replace all nodes with identical edges as m with m
3324           k->subsume_by(m, this);
3325         }
3326       }
3327     }
3328     break;
3329   }
3330   default:
3331     assert( !n->is_Call(), "" );
3332     assert( !n->is_Mem(), "" );
3333     assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3334     break;
3335   }
3336 
3337   // Collect CFG split points
3338   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3339     frc._tests.push(n);
3340   }
3341 }
3342 
3343 //------------------------------final_graph_reshaping_walk---------------------
3344 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3345 // requires that the walk visits a node's inputs before visiting the node.
3346 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3347   ResourceArea *area = Thread::current()->resource_area();
3348   Unique_Node_List sfpt(area);
3349 
3350   frc._visited.set(root->_idx); // first, mark node as visited
3351   uint cnt = root->req();
3352   Node *n = root;
3353   uint  i = 0;
3354   while (true) {
3355     if (i < cnt) {
3356       // Place all non-visited non-null inputs onto stack
3357       Node* m = n->in(i);
3358       ++i;
3359       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3360         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3361           // compute worst case interpreter size in case of a deoptimization
3362           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3363 
3364           sfpt.push(m);
3365         }
3366         cnt = m->req();
3367         nstack.push(n, i); // put on stack parent and next input's index
3368         n = m;
3369         i = 0;
3370       }
3371     } else {
3372       // Now do post-visit work
3373       final_graph_reshaping_impl( n, frc );
3374       if (nstack.is_empty())
3375         break;             // finished
3376       n = nstack.node();   // Get node from stack
3377       cnt = n->req();
3378       i = nstack.index();
3379       nstack.pop();        // Shift to the next node on stack
3380     }
3381   }
3382 
3383   // Skip next transformation if compressed oops are not used.
3384   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3385       (!UseCompressedOops && !UseCompressedClassPointers))
3386     return;
3387 
3388   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3389   // It could be done for an uncommon traps or any safepoints/calls
3390   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3391   while (sfpt.size() > 0) {
3392     n = sfpt.pop();
3393     JVMState *jvms = n->as_SafePoint()->jvms();
3394     assert(jvms != NULL, "sanity");
3395     int start = jvms->debug_start();
3396     int end   = n->req();
3397     bool is_uncommon = (n->is_CallStaticJava() &&
3398                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3399     for (int j = start; j < end; j++) {
3400       Node* in = n->in(j);
3401       if (in->is_DecodeNarrowPtr()) {
3402         bool safe_to_skip = true;
3403         if (!is_uncommon ) {
3404           // Is it safe to skip?
3405           for (uint i = 0; i < in->outcnt(); i++) {
3406             Node* u = in->raw_out(i);
3407             if (!u->is_SafePoint() ||
3408                  u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3409               safe_to_skip = false;
3410             }
3411           }
3412         }
3413         if (safe_to_skip) {
3414           n->set_req(j, in->in(1));
3415         }
3416         if (in->outcnt() == 0) {
3417           in->disconnect_inputs(NULL, this);
3418         }
3419       }
3420     }
3421   }
3422 }
3423 
3424 //------------------------------final_graph_reshaping--------------------------
3425 // Final Graph Reshaping.
3426 //
3427 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3428 //     and not commoned up and forced early.  Must come after regular
3429 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3430 //     inputs to Loop Phis; these will be split by the allocator anyways.
3431 //     Remove Opaque nodes.
3432 // (2) Move last-uses by commutative operations to the left input to encourage
3433 //     Intel update-in-place two-address operations and better register usage
3434 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3435 //     calls canonicalizing them back.
3436 // (3) Count the number of double-precision FP ops, single-precision FP ops
3437 //     and call sites.  On Intel, we can get correct rounding either by
3438 //     forcing singles to memory (requires extra stores and loads after each
3439 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3440 //     clearing the mode bit around call sites).  The mode bit is only used
3441 //     if the relative frequency of single FP ops to calls is low enough.
3442 //     This is a key transform for SPEC mpeg_audio.
3443 // (4) Detect infinite loops; blobs of code reachable from above but not
3444 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3445 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3446 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3447 //     Detection is by looking for IfNodes where only 1 projection is
3448 //     reachable from below or CatchNodes missing some targets.
3449 // (5) Assert for insane oop offsets in debug mode.
3450 
3451 bool Compile::final_graph_reshaping() {
3452   // an infinite loop may have been eliminated by the optimizer,
3453   // in which case the graph will be empty.
3454   if (root()->req() == 1) {
3455     record_method_not_compilable("trivial infinite loop");
3456     return true;
3457   }
3458 
3459   // Expensive nodes have their control input set to prevent the GVN
3460   // from freely commoning them. There's no GVN beyond this point so
3461   // no need to keep the control input. We want the expensive nodes to
3462   // be freely moved to the least frequent code path by gcm.
3463   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3464   for (int i = 0; i < expensive_count(); i++) {
3465     _expensive_nodes->at(i)->set_req(0, NULL);
3466   }
3467 
3468   Final_Reshape_Counts frc;
3469 
3470   // Visit everybody reachable!
3471   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3472   Node_Stack nstack(live_nodes() >> 1);
3473   final_graph_reshaping_walk(nstack, root(), frc);
3474 
3475   // Check for unreachable (from below) code (i.e., infinite loops).
3476   for( uint i = 0; i < frc._tests.size(); i++ ) {
3477     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3478     // Get number of CFG targets.
3479     // Note that PCTables include exception targets after calls.
3480     uint required_outcnt = n->required_outcnt();
3481     if (n->outcnt() != required_outcnt) {
3482       // Check for a few special cases.  Rethrow Nodes never take the
3483       // 'fall-thru' path, so expected kids is 1 less.
3484       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3485         if (n->in(0)->in(0)->is_Call()) {
3486           CallNode *call = n->in(0)->in(0)->as_Call();
3487           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3488             required_outcnt--;      // Rethrow always has 1 less kid
3489           } else if (call->req() > TypeFunc::Parms &&
3490                      call->is_CallDynamicJava()) {
3491             // Check for null receiver. In such case, the optimizer has
3492             // detected that the virtual call will always result in a null
3493             // pointer exception. The fall-through projection of this CatchNode
3494             // will not be populated.
3495             Node *arg0 = call->in(TypeFunc::Parms);
3496             if (arg0->is_Type() &&
3497                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3498               required_outcnt--;
3499             }
3500           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3501                      call->req() > TypeFunc::Parms+1 &&
3502                      call->is_CallStaticJava()) {
3503             // Check for negative array length. In such case, the optimizer has
3504             // detected that the allocation attempt will always result in an
3505             // exception. There is no fall-through projection of this CatchNode .
3506             Node *arg1 = call->in(TypeFunc::Parms+1);
3507             if (arg1->is_Type() &&
3508                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3509               required_outcnt--;
3510             }
3511           }
3512         }
3513       }
3514       // Recheck with a better notion of 'required_outcnt'
3515       if (n->outcnt() != required_outcnt) {
3516         record_method_not_compilable("malformed control flow");
3517         return true;            // Not all targets reachable!
3518       }
3519     }
3520     // Check that I actually visited all kids.  Unreached kids
3521     // must be infinite loops.
3522     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3523       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3524         record_method_not_compilable("infinite loop");
3525         return true;            // Found unvisited kid; must be unreach
3526       }
3527   }
3528 
3529   // If original bytecodes contained a mixture of floats and doubles
3530   // check if the optimizer has made it homogenous, item (3).
3531   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3532       frc.get_float_count() > 32 &&
3533       frc.get_double_count() == 0 &&
3534       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3535     set_24_bit_selection_and_mode( false,  true );
3536   }
3537 
3538   set_java_calls(frc.get_java_call_count());
3539   set_inner_loops(frc.get_inner_loop_count());
3540 
3541   // No infinite loops, no reason to bail out.
3542   return false;
3543 }
3544 
3545 //-----------------------------too_many_traps----------------------------------
3546 // Report if there are too many traps at the current method and bci.
3547 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3548 bool Compile::too_many_traps(ciMethod* method,
3549                              int bci,
3550                              Deoptimization::DeoptReason reason) {
3551   ciMethodData* md = method->method_data();
3552   if (md->is_empty()) {
3553     // Assume the trap has not occurred, or that it occurred only
3554     // because of a transient condition during start-up in the interpreter.
3555     return false;
3556   }
3557   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3558   if (md->has_trap_at(bci, m, reason) != 0) {
3559     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3560     // Also, if there are multiple reasons, or if there is no per-BCI record,
3561     // assume the worst.
3562     if (log())
3563       log()->elem("observe trap='%s' count='%d'",
3564                   Deoptimization::trap_reason_name(reason),
3565                   md->trap_count(reason));
3566     return true;
3567   } else {
3568     // Ignore method/bci and see if there have been too many globally.
3569     return too_many_traps(reason, md);
3570   }
3571 }
3572 
3573 // Less-accurate variant which does not require a method and bci.
3574 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3575                              ciMethodData* logmd) {
3576   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3577     // Too many traps globally.
3578     // Note that we use cumulative trap_count, not just md->trap_count.
3579     if (log()) {
3580       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3581       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3582                   Deoptimization::trap_reason_name(reason),
3583                   mcount, trap_count(reason));
3584     }
3585     return true;
3586   } else {
3587     // The coast is clear.
3588     return false;
3589   }
3590 }
3591 
3592 //--------------------------too_many_recompiles--------------------------------
3593 // Report if there are too many recompiles at the current method and bci.
3594 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3595 // Is not eager to return true, since this will cause the compiler to use
3596 // Action_none for a trap point, to avoid too many recompilations.
3597 bool Compile::too_many_recompiles(ciMethod* method,
3598                                   int bci,
3599                                   Deoptimization::DeoptReason reason) {
3600   ciMethodData* md = method->method_data();
3601   if (md->is_empty()) {
3602     // Assume the trap has not occurred, or that it occurred only
3603     // because of a transient condition during start-up in the interpreter.
3604     return false;
3605   }
3606   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3607   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3608   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3609   Deoptimization::DeoptReason per_bc_reason
3610     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3611   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3612   if ((per_bc_reason == Deoptimization::Reason_none
3613        || md->has_trap_at(bci, m, reason) != 0)
3614       // The trap frequency measure we care about is the recompile count:
3615       && md->trap_recompiled_at(bci, m)
3616       && md->overflow_recompile_count() >= bc_cutoff) {
3617     // Do not emit a trap here if it has already caused recompilations.
3618     // Also, if there are multiple reasons, or if there is no per-BCI record,
3619     // assume the worst.
3620     if (log())
3621       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3622                   Deoptimization::trap_reason_name(reason),
3623                   md->trap_count(reason),
3624                   md->overflow_recompile_count());
3625     return true;
3626   } else if (trap_count(reason) != 0
3627              && decompile_count() >= m_cutoff) {
3628     // Too many recompiles globally, and we have seen this sort of trap.
3629     // Use cumulative decompile_count, not just md->decompile_count.
3630     if (log())
3631       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3632                   Deoptimization::trap_reason_name(reason),
3633                   md->trap_count(reason), trap_count(reason),
3634                   md->decompile_count(), decompile_count());
3635     return true;
3636   } else {
3637     // The coast is clear.
3638     return false;
3639   }
3640 }
3641 
3642 // Compute when not to trap. Used by matching trap based nodes and
3643 // NullCheck optimization.
3644 void Compile::set_allowed_deopt_reasons() {
3645   _allowed_reasons = 0;
3646   if (is_method_compilation()) {
3647     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3648       assert(rs < BitsPerInt, "recode bit map");
3649       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3650         _allowed_reasons |= nth_bit(rs);
3651       }
3652     }
3653   }
3654 }
3655 
3656 #ifndef PRODUCT
3657 //------------------------------verify_graph_edges---------------------------
3658 // Walk the Graph and verify that there is a one-to-one correspondence
3659 // between Use-Def edges and Def-Use edges in the graph.
3660 void Compile::verify_graph_edges(bool no_dead_code) {
3661   if (VerifyGraphEdges) {
3662     ResourceArea *area = Thread::current()->resource_area();
3663     Unique_Node_List visited(area);
3664     // Call recursive graph walk to check edges
3665     _root->verify_edges(visited);
3666     if (no_dead_code) {
3667       // Now make sure that no visited node is used by an unvisited node.
3668       bool dead_nodes = false;
3669       Unique_Node_List checked(area);
3670       while (visited.size() > 0) {
3671         Node* n = visited.pop();
3672         checked.push(n);
3673         for (uint i = 0; i < n->outcnt(); i++) {
3674           Node* use = n->raw_out(i);
3675           if (checked.member(use))  continue;  // already checked
3676           if (visited.member(use))  continue;  // already in the graph
3677           if (use->is_Con())        continue;  // a dead ConNode is OK
3678           // At this point, we have found a dead node which is DU-reachable.
3679           if (!dead_nodes) {
3680             tty->print_cr("*** Dead nodes reachable via DU edges:");
3681             dead_nodes = true;
3682           }
3683           use->dump(2);
3684           tty->print_cr("---");
3685           checked.push(use);  // No repeats; pretend it is now checked.
3686         }
3687       }
3688       assert(!dead_nodes, "using nodes must be reachable from root");
3689     }
3690   }
3691 }
3692 
3693 // Verify GC barriers consistency
3694 // Currently supported:
3695 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3696 void Compile::verify_barriers() {
3697   if (UseG1GC) {
3698     // Verify G1 pre-barriers
3699     const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active());
3700 
3701     ResourceArea *area = Thread::current()->resource_area();
3702     Unique_Node_List visited(area);
3703     Node_List worklist(area);
3704     // We're going to walk control flow backwards starting from the Root
3705     worklist.push(_root);
3706     while (worklist.size() > 0) {
3707       Node* x = worklist.pop();
3708       if (x == NULL || x == top()) continue;
3709       if (visited.member(x)) {
3710         continue;
3711       } else {
3712         visited.push(x);
3713       }
3714 
3715       if (x->is_Region()) {
3716         for (uint i = 1; i < x->req(); i++) {
3717           worklist.push(x->in(i));
3718         }
3719       } else {
3720         worklist.push(x->in(0));
3721         // We are looking for the pattern:
3722         //                            /->ThreadLocal
3723         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3724         //              \->ConI(0)
3725         // We want to verify that the If and the LoadB have the same control
3726         // See GraphKit::g1_write_barrier_pre()
3727         if (x->is_If()) {
3728           IfNode *iff = x->as_If();
3729           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3730             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3731             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3732                 && cmp->in(1)->is_Load()) {
3733               LoadNode* load = cmp->in(1)->as_Load();
3734               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3735                   && load->in(2)->in(3)->is_Con()
3736                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3737 
3738                 Node* if_ctrl = iff->in(0);
3739                 Node* load_ctrl = load->in(0);
3740 
3741                 if (if_ctrl != load_ctrl) {
3742                   // Skip possible CProj->NeverBranch in infinite loops
3743                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3744                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3745                     if_ctrl = if_ctrl->in(0)->in(0);
3746                   }
3747                 }
3748                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3749               }
3750             }
3751           }
3752         }
3753       }
3754     }
3755   }
3756 }
3757 
3758 #endif
3759 
3760 // The Compile object keeps track of failure reasons separately from the ciEnv.
3761 // This is required because there is not quite a 1-1 relation between the
3762 // ciEnv and its compilation task and the Compile object.  Note that one
3763 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3764 // to backtrack and retry without subsuming loads.  Other than this backtracking
3765 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3766 // by the logic in C2Compiler.
3767 void Compile::record_failure(const char* reason) {
3768   if (log() != NULL) {
3769     log()->elem("failure reason='%s' phase='compile'", reason);
3770   }
3771   if (_failure_reason == NULL) {
3772     // Record the first failure reason.
3773     _failure_reason = reason;
3774   }
3775 
3776   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3777     C->print_method(PHASE_FAILURE);
3778   }
3779   _root = NULL;  // flush the graph, too
3780 }
3781 
3782 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3783   : TraceTime(name, accumulator, CITime, CITimeVerbose),
3784     _phase_name(name), _dolog(CITimeVerbose)
3785 {
3786   if (_dolog) {
3787     C = Compile::current();
3788     _log = C->log();
3789   } else {
3790     C = NULL;
3791     _log = NULL;
3792   }
3793   if (_log != NULL) {
3794     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3795     _log->stamp();
3796     _log->end_head();
3797   }
3798 }
3799 
3800 Compile::TracePhase::~TracePhase() {
3801 
3802   C = Compile::current();
3803   if (_dolog) {
3804     _log = C->log();
3805   } else {
3806     _log = NULL;
3807   }
3808 
3809 #ifdef ASSERT
3810   if (PrintIdealNodeCount) {
3811     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3812                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3813   }
3814 
3815   if (VerifyIdealNodeCount) {
3816     Compile::current()->print_missing_nodes();
3817   }
3818 #endif
3819 
3820   if (_log != NULL) {
3821     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3822   }
3823 }
3824 
3825 //=============================================================================
3826 // Two Constant's are equal when the type and the value are equal.
3827 bool Compile::Constant::operator==(const Constant& other) {
3828   if (type()          != other.type()         )  return false;
3829   if (can_be_reused() != other.can_be_reused())  return false;
3830   // For floating point values we compare the bit pattern.
3831   switch (type()) {
3832   case T_INT:
3833   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3834   case T_LONG:
3835   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
3836   case T_OBJECT:
3837   case T_ADDRESS: return (_v._value.l == other._v._value.l);
3838   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
3839   case T_METADATA: return (_v._metadata == other._v._metadata);
3840   default: ShouldNotReachHere();
3841   }
3842   return false;
3843 }
3844 
3845 static int type_to_size_in_bytes(BasicType t) {
3846   switch (t) {
3847   case T_INT:     return sizeof(jint   );
3848   case T_LONG:    return sizeof(jlong  );
3849   case T_FLOAT:   return sizeof(jfloat );
3850   case T_DOUBLE:  return sizeof(jdouble);
3851   case T_METADATA: return sizeof(Metadata*);
3852     // We use T_VOID as marker for jump-table entries (labels) which
3853     // need an internal word relocation.
3854   case T_VOID:
3855   case T_ADDRESS:
3856   case T_OBJECT:  return sizeof(jobject);
3857   }
3858 
3859   ShouldNotReachHere();
3860   return -1;
3861 }
3862 
3863 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3864   // sort descending
3865   if (a->freq() > b->freq())  return -1;
3866   if (a->freq() < b->freq())  return  1;
3867   return 0;
3868 }
3869 
3870 void Compile::ConstantTable::calculate_offsets_and_size() {
3871   // First, sort the array by frequencies.
3872   _constants.sort(qsort_comparator);
3873 
3874 #ifdef ASSERT
3875   // Make sure all jump-table entries were sorted to the end of the
3876   // array (they have a negative frequency).
3877   bool found_void = false;
3878   for (int i = 0; i < _constants.length(); i++) {
3879     Constant con = _constants.at(i);
3880     if (con.type() == T_VOID)
3881       found_void = true;  // jump-tables
3882     else
3883       assert(!found_void, "wrong sorting");
3884   }
3885 #endif
3886 
3887   int offset = 0;
3888   for (int i = 0; i < _constants.length(); i++) {
3889     Constant* con = _constants.adr_at(i);
3890 
3891     // Align offset for type.
3892     int typesize = type_to_size_in_bytes(con->type());
3893     offset = align_up(offset, typesize);
3894     con->set_offset(offset);   // set constant's offset
3895 
3896     if (con->type() == T_VOID) {
3897       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3898       offset = offset + typesize * n->outcnt();  // expand jump-table
3899     } else {
3900       offset = offset + typesize;
3901     }
3902   }
3903 
3904   // Align size up to the next section start (which is insts; see
3905   // CodeBuffer::align_at_start).
3906   assert(_size == -1, "already set?");
3907   _size = align_up(offset, (int)CodeEntryAlignment);
3908 }
3909 
3910 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3911   MacroAssembler _masm(&cb);
3912   for (int i = 0; i < _constants.length(); i++) {
3913     Constant con = _constants.at(i);
3914     address constant_addr = NULL;
3915     switch (con.type()) {
3916     case T_INT:    constant_addr = _masm.int_constant(   con.get_jint()   ); break;
3917     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
3918     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3919     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3920     case T_OBJECT: {
3921       jobject obj = con.get_jobject();
3922       int oop_index = _masm.oop_recorder()->find_index(obj);
3923       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3924       break;
3925     }
3926     case T_ADDRESS: {
3927       address addr = (address) con.get_jobject();
3928       constant_addr = _masm.address_constant(addr);
3929       break;
3930     }
3931     // We use T_VOID as marker for jump-table entries (labels) which
3932     // need an internal word relocation.
3933     case T_VOID: {
3934       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3935       // Fill the jump-table with a dummy word.  The real value is
3936       // filled in later in fill_jump_table.
3937       address dummy = (address) n;
3938       constant_addr = _masm.address_constant(dummy);
3939       // Expand jump-table
3940       for (uint i = 1; i < n->outcnt(); i++) {
3941         address temp_addr = _masm.address_constant(dummy + i);
3942         assert(temp_addr, "consts section too small");
3943       }
3944       break;
3945     }
3946     case T_METADATA: {
3947       Metadata* obj = con.get_metadata();
3948       int metadata_index = _masm.oop_recorder()->find_index(obj);
3949       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3950       break;
3951     }
3952     default: ShouldNotReachHere();
3953     }
3954     assert(constant_addr, "consts section too small");
3955     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3956             "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
3957   }
3958 }
3959 
3960 int Compile::ConstantTable::find_offset(Constant& con) const {
3961   int idx = _constants.find(con);
3962   assert(idx != -1, "constant must be in constant table");
3963   int offset = _constants.at(idx).offset();
3964   assert(offset != -1, "constant table not emitted yet?");
3965   return offset;
3966 }
3967 
3968 void Compile::ConstantTable::add(Constant& con) {
3969   if (con.can_be_reused()) {
3970     int idx = _constants.find(con);
3971     if (idx != -1 && _constants.at(idx).can_be_reused()) {
3972       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
3973       return;
3974     }
3975   }
3976   (void) _constants.append(con);
3977 }
3978 
3979 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3980   Block* b = Compile::current()->cfg()->get_block_for_node(n);
3981   Constant con(type, value, b->_freq);
3982   add(con);
3983   return con;
3984 }
3985 
3986 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3987   Constant con(metadata);
3988   add(con);
3989   return con;
3990 }
3991 
3992 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3993   jvalue value;
3994   BasicType type = oper->type()->basic_type();
3995   switch (type) {
3996   case T_LONG:    value.j = oper->constantL(); break;
3997   case T_FLOAT:   value.f = oper->constantF(); break;
3998   case T_DOUBLE:  value.d = oper->constantD(); break;
3999   case T_OBJECT:
4000   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
4001   case T_METADATA: return add((Metadata*)oper->constant()); break;
4002   default: guarantee(false, "unhandled type: %s", type2name(type));
4003   }
4004   return add(n, type, value);
4005 }
4006 
4007 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
4008   jvalue value;
4009   // We can use the node pointer here to identify the right jump-table
4010   // as this method is called from Compile::Fill_buffer right before
4011   // the MachNodes are emitted and the jump-table is filled (means the
4012   // MachNode pointers do not change anymore).
4013   value.l = (jobject) n;
4014   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
4015   add(con);
4016   return con;
4017 }
4018 
4019 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
4020   // If called from Compile::scratch_emit_size do nothing.
4021   if (Compile::current()->in_scratch_emit_size())  return;
4022 
4023   assert(labels.is_nonempty(), "must be");
4024   assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
4025 
4026   // Since MachConstantNode::constant_offset() also contains
4027   // table_base_offset() we need to subtract the table_base_offset()
4028   // to get the plain offset into the constant table.
4029   int offset = n->constant_offset() - table_base_offset();
4030 
4031   MacroAssembler _masm(&cb);
4032   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
4033 
4034   for (uint i = 0; i < n->outcnt(); i++) {
4035     address* constant_addr = &jump_table_base[i];
4036     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));
4037     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
4038     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
4039   }
4040 }
4041 
4042 //----------------------------static_subtype_check-----------------------------
4043 // Shortcut important common cases when superklass is exact:
4044 // (0) superklass is java.lang.Object (can occur in reflective code)
4045 // (1) subklass is already limited to a subtype of superklass => always ok
4046 // (2) subklass does not overlap with superklass => always fail
4047 // (3) superklass has NO subtypes and we can check with a simple compare.
4048 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
4049   if (StressReflectiveCode) {
4050     return SSC_full_test;       // Let caller generate the general case.
4051   }
4052 
4053   if (superk == env()->Object_klass()) {
4054     return SSC_always_true;     // (0) this test cannot fail
4055   }
4056 
4057   ciType* superelem = superk;
4058   if (superelem->is_array_klass())
4059     superelem = superelem->as_array_klass()->base_element_type();
4060 
4061   if (!subk->is_interface()) {  // cannot trust static interface types yet
4062     if (subk->is_subtype_of(superk)) {
4063       return SSC_always_true;   // (1) false path dead; no dynamic test needed
4064     }
4065     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
4066         !superk->is_subtype_of(subk)) {
4067       return SSC_always_false;
4068     }
4069   }
4070 
4071   // If casting to an instance klass, it must have no subtypes
4072   if (superk->is_interface()) {
4073     // Cannot trust interfaces yet.
4074     // %%% S.B. superk->nof_implementors() == 1
4075   } else if (superelem->is_instance_klass()) {
4076     ciInstanceKlass* ik = superelem->as_instance_klass();
4077     if (!ik->has_subklass() && !ik->is_interface()) {
4078       if (!ik->is_final()) {
4079         // Add a dependency if there is a chance of a later subclass.
4080         dependencies()->assert_leaf_type(ik);
4081       }
4082       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
4083     }
4084   } else {
4085     // A primitive array type has no subtypes.
4086     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
4087   }
4088 
4089   return SSC_full_test;
4090 }
4091 
4092 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4093 #ifdef _LP64
4094   // The scaled index operand to AddP must be a clean 64-bit value.
4095   // Java allows a 32-bit int to be incremented to a negative
4096   // value, which appears in a 64-bit register as a large
4097   // positive number.  Using that large positive number as an
4098   // operand in pointer arithmetic has bad consequences.
4099   // On the other hand, 32-bit overflow is rare, and the possibility
4100   // can often be excluded, if we annotate the ConvI2L node with
4101   // a type assertion that its value is known to be a small positive
4102   // number.  (The prior range check has ensured this.)
4103   // This assertion is used by ConvI2LNode::Ideal.
4104   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4105   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4106   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4107   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4108 #endif
4109   return idx;
4110 }
4111 
4112 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4113 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4114   if (ctrl != NULL) {
4115     // Express control dependency by a CastII node with a narrow type.
4116     value = new CastIINode(value, itype, false, true /* range check dependency */);
4117     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4118     // node from floating above the range check during loop optimizations. Otherwise, the
4119     // ConvI2L node may be eliminated independently of the range check, causing the data path
4120     // to become TOP while the control path is still there (although it's unreachable).
4121     value->set_req(0, ctrl);
4122     // Save CastII node to remove it after loop optimizations.
4123     phase->C->add_range_check_cast(value);
4124     value = phase->transform(value);
4125   }
4126   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4127   return phase->transform(new ConvI2LNode(value, ltype));
4128 }
4129 
4130 // The message about the current inlining is accumulated in
4131 // _print_inlining_stream and transfered into the _print_inlining_list
4132 // once we know whether inlining succeeds or not. For regular
4133 // inlining, messages are appended to the buffer pointed by
4134 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4135 // a new buffer is added after _print_inlining_idx in the list. This
4136 // way we can update the inlining message for late inlining call site
4137 // when the inlining is attempted again.
4138 void Compile::print_inlining_init() {
4139   if (print_inlining() || print_intrinsics()) {
4140     _print_inlining_stream = new stringStream();
4141     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4142   }
4143 }
4144 
4145 void Compile::print_inlining_reinit() {
4146   if (print_inlining() || print_intrinsics()) {
4147     // Re allocate buffer when we change ResourceMark
4148     _print_inlining_stream = new stringStream();
4149   }
4150 }
4151 
4152 void Compile::print_inlining_reset() {
4153   _print_inlining_stream->reset();
4154 }
4155 
4156 void Compile::print_inlining_commit() {
4157   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4158   // Transfer the message from _print_inlining_stream to the current
4159   // _print_inlining_list buffer and clear _print_inlining_stream.
4160   _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4161   print_inlining_reset();
4162 }
4163 
4164 void Compile::print_inlining_push() {
4165   // Add new buffer to the _print_inlining_list at current position
4166   _print_inlining_idx++;
4167   _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4168 }
4169 
4170 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4171   return _print_inlining_list->at(_print_inlining_idx);
4172 }
4173 
4174 void Compile::print_inlining_update(CallGenerator* cg) {
4175   if (print_inlining() || print_intrinsics()) {
4176     if (!cg->is_late_inline()) {
4177       if (print_inlining_current().cg() != NULL) {
4178         print_inlining_push();
4179       }
4180       print_inlining_commit();
4181     } else {
4182       if (print_inlining_current().cg() != cg &&
4183           (print_inlining_current().cg() != NULL ||
4184            print_inlining_current().ss()->size() != 0)) {
4185         print_inlining_push();
4186       }
4187       print_inlining_commit();
4188       print_inlining_current().set_cg(cg);
4189     }
4190   }
4191 }
4192 
4193 void Compile::print_inlining_move_to(CallGenerator* cg) {
4194   // We resume inlining at a late inlining call site. Locate the
4195   // corresponding inlining buffer so that we can update it.
4196   if (print_inlining()) {
4197     for (int i = 0; i < _print_inlining_list->length(); i++) {
4198       if (_print_inlining_list->adr_at(i)->cg() == cg) {
4199         _print_inlining_idx = i;
4200         return;
4201       }
4202     }
4203     ShouldNotReachHere();
4204   }
4205 }
4206 
4207 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4208   if (print_inlining()) {
4209     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4210     assert(print_inlining_current().cg() == cg, "wrong entry");
4211     // replace message with new message
4212     _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4213     print_inlining_commit();
4214     print_inlining_current().set_cg(cg);
4215   }
4216 }
4217 
4218 void Compile::print_inlining_assert_ready() {
4219   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4220 }
4221 
4222 void Compile::process_print_inlining() {
4223   bool do_print_inlining = print_inlining() || print_intrinsics();
4224   if (do_print_inlining || log() != NULL) {
4225     // Print inlining message for candidates that we couldn't inline
4226     // for lack of space
4227     for (int i = 0; i < _late_inlines.length(); i++) {
4228       CallGenerator* cg = _late_inlines.at(i);
4229       if (!cg->is_mh_late_inline()) {
4230         const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4231         if (do_print_inlining) {
4232           cg->print_inlining_late(msg);
4233         }
4234         log_late_inline_failure(cg, msg);
4235       }
4236     }
4237   }
4238   if (do_print_inlining) {
4239     ResourceMark rm;
4240     stringStream ss;
4241     for (int i = 0; i < _print_inlining_list->length(); i++) {
4242       ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4243     }
4244     size_t end = ss.size();
4245     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4246     strncpy(_print_inlining_output, ss.base(), end+1);
4247     _print_inlining_output[end] = 0;
4248   }
4249 }
4250 
4251 void Compile::dump_print_inlining() {
4252   if (_print_inlining_output != NULL) {
4253     tty->print_raw(_print_inlining_output);
4254   }
4255 }
4256 
4257 void Compile::log_late_inline(CallGenerator* cg) {
4258   if (log() != NULL) {
4259     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4260                 cg->unique_id());
4261     JVMState* p = cg->call_node()->jvms();
4262     while (p != NULL) {
4263       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4264       p = p->caller();
4265     }
4266     log()->tail("late_inline");
4267   }
4268 }
4269 
4270 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4271   log_late_inline(cg);
4272   if (log() != NULL) {
4273     log()->inline_fail(msg);
4274   }
4275 }
4276 
4277 void Compile::log_inline_id(CallGenerator* cg) {
4278   if (log() != NULL) {
4279     // The LogCompilation tool needs a unique way to identify late
4280     // inline call sites. This id must be unique for this call site in
4281     // this compilation. Try to have it unique across compilations as
4282     // well because it can be convenient when grepping through the log
4283     // file.
4284     // Distinguish OSR compilations from others in case CICountOSR is
4285     // on.
4286     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4287     cg->set_unique_id(id);
4288     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4289   }
4290 }
4291 
4292 void Compile::log_inline_failure(const char* msg) {
4293   if (C->log() != NULL) {
4294     C->log()->inline_fail(msg);
4295   }
4296 }
4297 
4298 
4299 // Dump inlining replay data to the stream.
4300 // Don't change thread state and acquire any locks.
4301 void Compile::dump_inline_data(outputStream* out) {
4302   InlineTree* inl_tree = ilt();
4303   if (inl_tree != NULL) {
4304     out->print(" inline %d", inl_tree->count());
4305     inl_tree->dump_replay_data(out);
4306   }
4307 }
4308 
4309 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4310   if (n1->Opcode() < n2->Opcode())      return -1;
4311   else if (n1->Opcode() > n2->Opcode()) return 1;
4312 
4313   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4314   for (uint i = 1; i < n1->req(); i++) {
4315     if (n1->in(i) < n2->in(i))      return -1;
4316     else if (n1->in(i) > n2->in(i)) return 1;
4317   }
4318 
4319   return 0;
4320 }
4321 
4322 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4323   Node* n1 = *n1p;
4324   Node* n2 = *n2p;
4325 
4326   return cmp_expensive_nodes(n1, n2);
4327 }
4328 
4329 void Compile::sort_expensive_nodes() {
4330   if (!expensive_nodes_sorted()) {
4331     _expensive_nodes->sort(cmp_expensive_nodes);
4332   }
4333 }
4334 
4335 bool Compile::expensive_nodes_sorted() const {
4336   for (int i = 1; i < _expensive_nodes->length(); i++) {
4337     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4338       return false;
4339     }
4340   }
4341   return true;
4342 }
4343 
4344 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4345   if (_expensive_nodes->length() == 0) {
4346     return false;
4347   }
4348 
4349   assert(OptimizeExpensiveOps, "optimization off?");
4350 
4351   // Take this opportunity to remove dead nodes from the list
4352   int j = 0;
4353   for (int i = 0; i < _expensive_nodes->length(); i++) {
4354     Node* n = _expensive_nodes->at(i);
4355     if (!n->is_unreachable(igvn)) {
4356       assert(n->is_expensive(), "should be expensive");
4357       _expensive_nodes->at_put(j, n);
4358       j++;
4359     }
4360   }
4361   _expensive_nodes->trunc_to(j);
4362 
4363   // Then sort the list so that similar nodes are next to each other
4364   // and check for at least two nodes of identical kind with same data
4365   // inputs.
4366   sort_expensive_nodes();
4367 
4368   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4369     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4370       return true;
4371     }
4372   }
4373 
4374   return false;
4375 }
4376 
4377 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4378   if (_expensive_nodes->length() == 0) {
4379     return;
4380   }
4381 
4382   assert(OptimizeExpensiveOps, "optimization off?");
4383 
4384   // Sort to bring similar nodes next to each other and clear the
4385   // control input of nodes for which there's only a single copy.
4386   sort_expensive_nodes();
4387 
4388   int j = 0;
4389   int identical = 0;
4390   int i = 0;
4391   bool modified = false;
4392   for (; i < _expensive_nodes->length()-1; i++) {
4393     assert(j <= i, "can't write beyond current index");
4394     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4395       identical++;
4396       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4397       continue;
4398     }
4399     if (identical > 0) {
4400       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4401       identical = 0;
4402     } else {
4403       Node* n = _expensive_nodes->at(i);
4404       igvn.replace_input_of(n, 0, NULL);
4405       igvn.hash_insert(n);
4406       modified = true;
4407     }
4408   }
4409   if (identical > 0) {
4410     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4411   } else if (_expensive_nodes->length() >= 1) {
4412     Node* n = _expensive_nodes->at(i);
4413     igvn.replace_input_of(n, 0, NULL);
4414     igvn.hash_insert(n);
4415     modified = true;
4416   }
4417   _expensive_nodes->trunc_to(j);
4418   if (modified) {
4419     igvn.optimize();
4420   }
4421 }
4422 
4423 void Compile::add_expensive_node(Node * n) {
4424   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4425   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4426   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4427   if (OptimizeExpensiveOps) {
4428     _expensive_nodes->append(n);
4429   } else {
4430     // Clear control input and let IGVN optimize expensive nodes if
4431     // OptimizeExpensiveOps is off.
4432     n->set_req(0, NULL);
4433   }
4434 }
4435 
4436 /**
4437  * Remove the speculative part of types and clean up the graph
4438  */
4439 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4440   if (UseTypeSpeculation) {
4441     Unique_Node_List worklist;
4442     worklist.push(root());
4443     int modified = 0;
4444     // Go over all type nodes that carry a speculative type, drop the
4445     // speculative part of the type and enqueue the node for an igvn
4446     // which may optimize it out.
4447     for (uint next = 0; next < worklist.size(); ++next) {
4448       Node *n  = worklist.at(next);
4449       if (n->is_Type()) {
4450         TypeNode* tn = n->as_Type();
4451         const Type* t = tn->type();
4452         const Type* t_no_spec = t->remove_speculative();
4453         if (t_no_spec != t) {
4454           bool in_hash = igvn.hash_delete(n);
4455           assert(in_hash, "node should be in igvn hash table");
4456           tn->set_type(t_no_spec);
4457           igvn.hash_insert(n);
4458           igvn._worklist.push(n); // give it a chance to go away
4459           modified++;
4460         }
4461       }
4462       uint max = n->len();
4463       for( uint i = 0; i < max; ++i ) {
4464         Node *m = n->in(i);
4465         if (not_a_node(m))  continue;
4466         worklist.push(m);
4467       }
4468     }
4469     // Drop the speculative part of all types in the igvn's type table
4470     igvn.remove_speculative_types();
4471     if (modified > 0) {
4472       igvn.optimize();
4473     }
4474 #ifdef ASSERT
4475     // Verify that after the IGVN is over no speculative type has resurfaced
4476     worklist.clear();
4477     worklist.push(root());
4478     for (uint next = 0; next < worklist.size(); ++next) {
4479       Node *n  = worklist.at(next);
4480       const Type* t = igvn.type_or_null(n);
4481       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4482       if (n->is_Type()) {
4483         t = n->as_Type()->type();
4484         assert(t == t->remove_speculative(), "no more speculative types");
4485       }
4486       uint max = n->len();
4487       for( uint i = 0; i < max; ++i ) {
4488         Node *m = n->in(i);
4489         if (not_a_node(m))  continue;
4490         worklist.push(m);
4491       }
4492     }
4493     igvn.check_no_speculative_types();
4494 #endif
4495   }
4496 }
4497 
4498 // Auxiliary method to support randomized stressing/fuzzing.
4499 //
4500 // This method can be called the arbitrary number of times, with current count
4501 // as the argument. The logic allows selecting a single candidate from the
4502 // running list of candidates as follows:
4503 //    int count = 0;
4504 //    Cand* selected = null;
4505 //    while(cand = cand->next()) {
4506 //      if (randomized_select(++count)) {
4507 //        selected = cand;
4508 //      }
4509 //    }
4510 //
4511 // Including count equalizes the chances any candidate is "selected".
4512 // This is useful when we don't have the complete list of candidates to choose
4513 // from uniformly. In this case, we need to adjust the randomicity of the
4514 // selection, or else we will end up biasing the selection towards the latter
4515 // candidates.
4516 //
4517 // Quick back-envelope calculation shows that for the list of n candidates
4518 // the equal probability for the candidate to persist as "best" can be
4519 // achieved by replacing it with "next" k-th candidate with the probability
4520 // of 1/k. It can be easily shown that by the end of the run, the
4521 // probability for any candidate is converged to 1/n, thus giving the
4522 // uniform distribution among all the candidates.
4523 //
4524 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4525 #define RANDOMIZED_DOMAIN_POW 29
4526 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4527 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4528 bool Compile::randomized_select(int count) {
4529   assert(count > 0, "only positive");
4530   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4531 }
4532 
4533 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4534 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4535 
4536 void NodeCloneInfo::dump() const {
4537   tty->print(" {%d:%d} ", idx(), gen());
4538 }
4539 
4540 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4541   uint64_t val = value(old->_idx);
4542   NodeCloneInfo cio(val);
4543   assert(val != 0, "old node should be in the map");
4544   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4545   insert(nnn->_idx, cin.get());
4546 #ifndef PRODUCT
4547   if (is_debug()) {
4548     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4549   }
4550 #endif
4551 }
4552 
4553 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4554   NodeCloneInfo cio(value(old->_idx));
4555   if (cio.get() == 0) {
4556     cio.set(old->_idx, 0);
4557     insert(old->_idx, cio.get());
4558 #ifndef PRODUCT
4559     if (is_debug()) {
4560       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4561     }
4562 #endif
4563   }
4564   clone(old, nnn, gen);
4565 }
4566 
4567 int CloneMap::max_gen() const {
4568   int g = 0;
4569   DictI di(_dict);
4570   for(; di.test(); ++di) {
4571     int t = gen(di._key);
4572     if (g < t) {
4573       g = t;
4574 #ifndef PRODUCT
4575       if (is_debug()) {
4576         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4577       }
4578 #endif
4579     }
4580   }
4581   return g;
4582 }
4583 
4584 void CloneMap::dump(node_idx_t key) const {
4585   uint64_t val = value(key);
4586   if (val != 0) {
4587     NodeCloneInfo ni(val);
4588     ni.dump();
4589   }
4590 }