1 /*
   2  * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.hpp"
  27 #include "asm/macroAssembler.inline.hpp"
  28 #include "ci/ciReplay.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/exceptionHandlerTable.hpp"
  31 #include "code/nmethod.hpp"
  32 #include "compiler/compileBroker.hpp"
  33 #include "compiler/compileLog.hpp"
  34 #include "compiler/disassembler.hpp"
  35 #include "compiler/oopMap.hpp"
  36 #include "opto/addnode.hpp"
  37 #include "opto/block.hpp"
  38 #include "opto/c2compiler.hpp"
  39 #include "opto/callGenerator.hpp"
  40 #include "opto/callnode.hpp"
  41 #include "opto/castnode.hpp"
  42 #include "opto/cfgnode.hpp"
  43 #include "opto/chaitin.hpp"
  44 #include "opto/compile.hpp"
  45 #include "opto/connode.hpp"
  46 #include "opto/convertnode.hpp"
  47 #include "opto/divnode.hpp"
  48 #include "opto/escape.hpp"
  49 #include "opto/idealGraphPrinter.hpp"
  50 #include "opto/loopnode.hpp"
  51 #include "opto/machnode.hpp"
  52 #include "opto/macro.hpp"
  53 #include "opto/matcher.hpp"
  54 #include "opto/mathexactnode.hpp"
  55 #include "opto/memnode.hpp"
  56 #include "opto/mulnode.hpp"
  57 #include "opto/narrowptrnode.hpp"
  58 #include "opto/node.hpp"
  59 #include "opto/opcodes.hpp"
  60 #include "opto/output.hpp"
  61 #include "opto/parse.hpp"
  62 #include "opto/phaseX.hpp"
  63 #include "opto/rootnode.hpp"
  64 #include "opto/runtime.hpp"
  65 #include "opto/stringopts.hpp"
  66 #include "opto/type.hpp"
  67 #include "opto/vectornode.hpp"
  68 #include "runtime/arguments.hpp"
  69 #include "runtime/sharedRuntime.hpp"
  70 #include "runtime/signature.hpp"
  71 #include "runtime/stubRoutines.hpp"
  72 #include "runtime/timer.hpp"
  73 #include "utilities/copy.hpp"
  74 
  75 
  76 // -------------------- Compile::mach_constant_base_node -----------------------
  77 // Constant table base node singleton.
  78 MachConstantBaseNode* Compile::mach_constant_base_node() {
  79   if (_mach_constant_base_node == NULL) {
  80     _mach_constant_base_node = new MachConstantBaseNode();
  81     _mach_constant_base_node->add_req(C->root());
  82   }
  83   return _mach_constant_base_node;
  84 }
  85 
  86 
  87 /// Support for intrinsics.
  88 
  89 // Return the index at which m must be inserted (or already exists).
  90 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
  91 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
  92 #ifdef ASSERT
  93   for (int i = 1; i < _intrinsics->length(); i++) {
  94     CallGenerator* cg1 = _intrinsics->at(i-1);
  95     CallGenerator* cg2 = _intrinsics->at(i);
  96     assert(cg1->method() != cg2->method()
  97            ? cg1->method()     < cg2->method()
  98            : cg1->is_virtual() < cg2->is_virtual(),
  99            "compiler intrinsics list must stay sorted");
 100   }
 101 #endif
 102   // Binary search sorted list, in decreasing intervals [lo, hi].
 103   int lo = 0, hi = _intrinsics->length()-1;
 104   while (lo <= hi) {
 105     int mid = (uint)(hi + lo) / 2;
 106     ciMethod* mid_m = _intrinsics->at(mid)->method();
 107     if (m < mid_m) {
 108       hi = mid-1;
 109     } else if (m > mid_m) {
 110       lo = mid+1;
 111     } else {
 112       // look at minor sort key
 113       bool mid_virt = _intrinsics->at(mid)->is_virtual();
 114       if (is_virtual < mid_virt) {
 115         hi = mid-1;
 116       } else if (is_virtual > mid_virt) {
 117         lo = mid+1;
 118       } else {
 119         return mid;  // exact match
 120       }
 121     }
 122   }
 123   return lo;  // inexact match
 124 }
 125 
 126 void Compile::register_intrinsic(CallGenerator* cg) {
 127   if (_intrinsics == NULL) {
 128     _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 129   }
 130   // This code is stolen from ciObjectFactory::insert.
 131   // Really, GrowableArray should have methods for
 132   // insert_at, remove_at, and binary_search.
 133   int len = _intrinsics->length();
 134   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
 135   if (index == len) {
 136     _intrinsics->append(cg);
 137   } else {
 138 #ifdef ASSERT
 139     CallGenerator* oldcg = _intrinsics->at(index);
 140     assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
 141 #endif
 142     _intrinsics->append(_intrinsics->at(len-1));
 143     int pos;
 144     for (pos = len-2; pos >= index; pos--) {
 145       _intrinsics->at_put(pos+1,_intrinsics->at(pos));
 146     }
 147     _intrinsics->at_put(index, cg);
 148   }
 149   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 150 }
 151 
 152 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 153   assert(m->is_loaded(), "don't try this on unloaded methods");
 154   if (_intrinsics != NULL) {
 155     int index = intrinsic_insertion_index(m, is_virtual);
 156     if (index < _intrinsics->length()
 157         && _intrinsics->at(index)->method() == m
 158         && _intrinsics->at(index)->is_virtual() == is_virtual) {
 159       return _intrinsics->at(index);
 160     }
 161   }
 162   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 163   if (m->intrinsic_id() != vmIntrinsics::_none &&
 164       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 165     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 166     if (cg != NULL) {
 167       // Save it for next time:
 168       register_intrinsic(cg);
 169       return cg;
 170     } else {
 171       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 172     }
 173   }
 174   return NULL;
 175 }
 176 
 177 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 178 // in library_call.cpp.
 179 
 180 
 181 #ifndef PRODUCT
 182 // statistics gathering...
 183 
 184 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 185 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 186 
 187 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 188   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 189   int oflags = _intrinsic_hist_flags[id];
 190   assert(flags != 0, "what happened?");
 191   if (is_virtual) {
 192     flags |= _intrinsic_virtual;
 193   }
 194   bool changed = (flags != oflags);
 195   if ((flags & _intrinsic_worked) != 0) {
 196     juint count = (_intrinsic_hist_count[id] += 1);
 197     if (count == 1) {
 198       changed = true;           // first time
 199     }
 200     // increment the overall count also:
 201     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 202   }
 203   if (changed) {
 204     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 205       // Something changed about the intrinsic's virtuality.
 206       if ((flags & _intrinsic_virtual) != 0) {
 207         // This is the first use of this intrinsic as a virtual call.
 208         if (oflags != 0) {
 209           // We already saw it as a non-virtual, so note both cases.
 210           flags |= _intrinsic_both;
 211         }
 212       } else if ((oflags & _intrinsic_both) == 0) {
 213         // This is the first use of this intrinsic as a non-virtual
 214         flags |= _intrinsic_both;
 215       }
 216     }
 217     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 218   }
 219   // update the overall flags also:
 220   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 221   return changed;
 222 }
 223 
 224 static char* format_flags(int flags, char* buf) {
 225   buf[0] = 0;
 226   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 227   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 228   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 229   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 230   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 231   if (buf[0] == 0)  strcat(buf, ",");
 232   assert(buf[0] == ',', "must be");
 233   return &buf[1];
 234 }
 235 
 236 void Compile::print_intrinsic_statistics() {
 237   char flagsbuf[100];
 238   ttyLocker ttyl;
 239   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 240   tty->print_cr("Compiler intrinsic usage:");
 241   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 242   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 243   #define PRINT_STAT_LINE(name, c, f) \
 244     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 245   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 246     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 247     int   flags = _intrinsic_hist_flags[id];
 248     juint count = _intrinsic_hist_count[id];
 249     if ((flags | count) != 0) {
 250       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 251     }
 252   }
 253   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 254   if (xtty != NULL)  xtty->tail("statistics");
 255 }
 256 
 257 void Compile::print_statistics() {
 258   { ttyLocker ttyl;
 259     if (xtty != NULL)  xtty->head("statistics type='opto'");
 260     Parse::print_statistics();
 261     PhaseCCP::print_statistics();
 262     PhaseRegAlloc::print_statistics();
 263     Scheduling::print_statistics();
 264     PhasePeephole::print_statistics();
 265     PhaseIdealLoop::print_statistics();
 266     if (xtty != NULL)  xtty->tail("statistics");
 267   }
 268   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 269     // put this under its own <statistics> element.
 270     print_intrinsic_statistics();
 271   }
 272 }
 273 #endif //PRODUCT
 274 
 275 // Support for bundling info
 276 Bundle* Compile::node_bundling(const Node *n) {
 277   assert(valid_bundle_info(n), "oob");
 278   return &_node_bundling_base[n->_idx];
 279 }
 280 
 281 bool Compile::valid_bundle_info(const Node *n) {
 282   return (_node_bundling_limit > n->_idx);
 283 }
 284 
 285 
 286 void Compile::gvn_replace_by(Node* n, Node* nn) {
 287   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 288     Node* use = n->last_out(i);
 289     bool is_in_table = initial_gvn()->hash_delete(use);
 290     uint uses_found = 0;
 291     for (uint j = 0; j < use->len(); j++) {
 292       if (use->in(j) == n) {
 293         if (j < use->req())
 294           use->set_req(j, nn);
 295         else
 296           use->set_prec(j, nn);
 297         uses_found++;
 298       }
 299     }
 300     if (is_in_table) {
 301       // reinsert into table
 302       initial_gvn()->hash_find_insert(use);
 303     }
 304     record_for_igvn(use);
 305     i -= uses_found;    // we deleted 1 or more copies of this edge
 306   }
 307 }
 308 
 309 
 310 static inline bool not_a_node(const Node* n) {
 311   if (n == NULL)                   return true;
 312   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 313   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 314   return false;
 315 }
 316 
 317 // Identify all nodes that are reachable from below, useful.
 318 // Use breadth-first pass that records state in a Unique_Node_List,
 319 // recursive traversal is slower.
 320 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 321   int estimated_worklist_size = live_nodes();
 322   useful.map( estimated_worklist_size, NULL );  // preallocate space
 323 
 324   // Initialize worklist
 325   if (root() != NULL)     { useful.push(root()); }
 326   // If 'top' is cached, declare it useful to preserve cached node
 327   if( cached_top_node() ) { useful.push(cached_top_node()); }
 328 
 329   // Push all useful nodes onto the list, breadthfirst
 330   for( uint next = 0; next < useful.size(); ++next ) {
 331     assert( next < unique(), "Unique useful nodes < total nodes");
 332     Node *n  = useful.at(next);
 333     uint max = n->len();
 334     for( uint i = 0; i < max; ++i ) {
 335       Node *m = n->in(i);
 336       if (not_a_node(m))  continue;
 337       useful.push(m);
 338     }
 339   }
 340 }
 341 
 342 // Update dead_node_list with any missing dead nodes using useful
 343 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 344 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 345   uint max_idx = unique();
 346   VectorSet& useful_node_set = useful.member_set();
 347 
 348   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 349     // If node with index node_idx is not in useful set,
 350     // mark it as dead in dead node list.
 351     if (! useful_node_set.test(node_idx) ) {
 352       record_dead_node(node_idx);
 353     }
 354   }
 355 }
 356 
 357 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 358   int shift = 0;
 359   for (int i = 0; i < inlines->length(); i++) {
 360     CallGenerator* cg = inlines->at(i);
 361     CallNode* call = cg->call_node();
 362     if (shift > 0) {
 363       inlines->at_put(i-shift, cg);
 364     }
 365     if (!useful.member(call)) {
 366       shift++;
 367     }
 368   }
 369   inlines->trunc_to(inlines->length()-shift);
 370 }
 371 
 372 // Disconnect all useless nodes by disconnecting those at the boundary.
 373 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 374   uint next = 0;
 375   while (next < useful.size()) {
 376     Node *n = useful.at(next++);
 377     if (n->is_SafePoint()) {
 378       // We're done with a parsing phase. Replaced nodes are not valid
 379       // beyond that point.
 380       n->as_SafePoint()->delete_replaced_nodes();
 381     }
 382     // Use raw traversal of out edges since this code removes out edges
 383     int max = n->outcnt();
 384     for (int j = 0; j < max; ++j) {
 385       Node* child = n->raw_out(j);
 386       if (! useful.member(child)) {
 387         assert(!child->is_top() || child != top(),
 388                "If top is cached in Compile object it is in useful list");
 389         // Only need to remove this out-edge to the useless node
 390         n->raw_del_out(j);
 391         --j;
 392         --max;
 393       }
 394     }
 395     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 396       record_for_igvn(n->unique_out());
 397     }
 398   }
 399   // Remove useless macro and predicate opaq nodes
 400   for (int i = C->macro_count()-1; i >= 0; i--) {
 401     Node* n = C->macro_node(i);
 402     if (!useful.member(n)) {
 403       remove_macro_node(n);
 404     }
 405   }
 406   // Remove useless CastII nodes with range check dependency
 407   for (int i = range_check_cast_count() - 1; i >= 0; i--) {
 408     Node* cast = range_check_cast_node(i);
 409     if (!useful.member(cast)) {
 410       remove_range_check_cast(cast);
 411     }
 412   }
 413   // Remove useless expensive node
 414   for (int i = C->expensive_count()-1; i >= 0; i--) {
 415     Node* n = C->expensive_node(i);
 416     if (!useful.member(n)) {
 417       remove_expensive_node(n);
 418     }
 419   }
 420   // clean up the late inline lists
 421   remove_useless_late_inlines(&_string_late_inlines, useful);
 422   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 423   remove_useless_late_inlines(&_late_inlines, useful);
 424   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 425 }
 426 
 427 //------------------------------frame_size_in_words-----------------------------
 428 // frame_slots in units of words
 429 int Compile::frame_size_in_words() const {
 430   // shift is 0 in LP32 and 1 in LP64
 431   const int shift = (LogBytesPerWord - LogBytesPerInt);
 432   int words = _frame_slots >> shift;
 433   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 434   return words;
 435 }
 436 
 437 // To bang the stack of this compiled method we use the stack size
 438 // that the interpreter would need in case of a deoptimization. This
 439 // removes the need to bang the stack in the deoptimization blob which
 440 // in turn simplifies stack overflow handling.
 441 int Compile::bang_size_in_bytes() const {
 442   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
 443 }
 444 
 445 // ============================================================================
 446 //------------------------------CompileWrapper---------------------------------
 447 class CompileWrapper : public StackObj {
 448   Compile *const _compile;
 449  public:
 450   CompileWrapper(Compile* compile);
 451 
 452   ~CompileWrapper();
 453 };
 454 
 455 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 456   // the Compile* pointer is stored in the current ciEnv:
 457   ciEnv* env = compile->env();
 458   assert(env == ciEnv::current(), "must already be a ciEnv active");
 459   assert(env->compiler_data() == NULL, "compile already active?");
 460   env->set_compiler_data(compile);
 461   assert(compile == Compile::current(), "sanity");
 462 
 463   compile->set_type_dict(NULL);
 464   compile->set_clone_map(new Dict(cmpkey, hashkey, _compile->comp_arena()));
 465   compile->clone_map().set_clone_idx(0);
 466   compile->set_type_hwm(NULL);
 467   compile->set_type_last_size(0);
 468   compile->set_last_tf(NULL, NULL);
 469   compile->set_indexSet_arena(NULL);
 470   compile->set_indexSet_free_block_list(NULL);
 471   compile->init_type_arena();
 472   Type::Initialize(compile);
 473   _compile->set_scratch_buffer_blob(NULL);
 474   _compile->begin_method();
 475   _compile->clone_map().set_debug(_compile->has_method() && _compile->directive()->CloneMapDebugOption);
 476 }
 477 CompileWrapper::~CompileWrapper() {
 478   _compile->end_method();
 479   if (_compile->scratch_buffer_blob() != NULL)
 480     BufferBlob::free(_compile->scratch_buffer_blob());
 481   _compile->env()->set_compiler_data(NULL);
 482 }
 483 
 484 
 485 //----------------------------print_compile_messages---------------------------
 486 void Compile::print_compile_messages() {
 487 #ifndef PRODUCT
 488   // Check if recompiling
 489   if (_subsume_loads == false && PrintOpto) {
 490     // Recompiling without allowing machine instructions to subsume loads
 491     tty->print_cr("*********************************************************");
 492     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 493     tty->print_cr("*********************************************************");
 494   }
 495   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 496     // Recompiling without escape analysis
 497     tty->print_cr("*********************************************************");
 498     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 499     tty->print_cr("*********************************************************");
 500   }
 501   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 502     // Recompiling without boxing elimination
 503     tty->print_cr("*********************************************************");
 504     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 505     tty->print_cr("*********************************************************");
 506   }
 507   if (C->directive()->BreakAtCompileOption) {
 508     // Open the debugger when compiling this method.
 509     tty->print("### Breaking when compiling: ");
 510     method()->print_short_name();
 511     tty->cr();
 512     BREAKPOINT;
 513   }
 514 
 515   if( PrintOpto ) {
 516     if (is_osr_compilation()) {
 517       tty->print("[OSR]%3d", _compile_id);
 518     } else {
 519       tty->print("%3d", _compile_id);
 520     }
 521   }
 522 #endif
 523 }
 524 
 525 
 526 //-----------------------init_scratch_buffer_blob------------------------------
 527 // Construct a temporary BufferBlob and cache it for this compile.
 528 void Compile::init_scratch_buffer_blob(int const_size) {
 529   // If there is already a scratch buffer blob allocated and the
 530   // constant section is big enough, use it.  Otherwise free the
 531   // current and allocate a new one.
 532   BufferBlob* blob = scratch_buffer_blob();
 533   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 534     // Use the current blob.
 535   } else {
 536     if (blob != NULL) {
 537       BufferBlob::free(blob);
 538     }
 539 
 540     ResourceMark rm;
 541     _scratch_const_size = const_size;
 542     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
 543     blob = BufferBlob::create("Compile::scratch_buffer", size);
 544     // Record the buffer blob for next time.
 545     set_scratch_buffer_blob(blob);
 546     // Have we run out of code space?
 547     if (scratch_buffer_blob() == NULL) {
 548       // Let CompilerBroker disable further compilations.
 549       record_failure("Not enough space for scratch buffer in CodeCache");
 550       return;
 551     }
 552   }
 553 
 554   // Initialize the relocation buffers
 555   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 556   set_scratch_locs_memory(locs_buf);
 557 }
 558 
 559 
 560 //-----------------------scratch_emit_size-------------------------------------
 561 // Helper function that computes size by emitting code
 562 uint Compile::scratch_emit_size(const Node* n) {
 563   // Start scratch_emit_size section.
 564   set_in_scratch_emit_size(true);
 565 
 566   // Emit into a trash buffer and count bytes emitted.
 567   // This is a pretty expensive way to compute a size,
 568   // but it works well enough if seldom used.
 569   // All common fixed-size instructions are given a size
 570   // method by the AD file.
 571   // Note that the scratch buffer blob and locs memory are
 572   // allocated at the beginning of the compile task, and
 573   // may be shared by several calls to scratch_emit_size.
 574   // The allocation of the scratch buffer blob is particularly
 575   // expensive, since it has to grab the code cache lock.
 576   BufferBlob* blob = this->scratch_buffer_blob();
 577   assert(blob != NULL, "Initialize BufferBlob at start");
 578   assert(blob->size() > MAX_inst_size, "sanity");
 579   relocInfo* locs_buf = scratch_locs_memory();
 580   address blob_begin = blob->content_begin();
 581   address blob_end   = (address)locs_buf;
 582   assert(blob->content_contains(blob_end), "sanity");
 583   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 584   buf.initialize_consts_size(_scratch_const_size);
 585   buf.initialize_stubs_size(MAX_stubs_size);
 586   assert(locs_buf != NULL, "sanity");
 587   int lsize = MAX_locs_size / 3;
 588   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 589   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 590   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 591 
 592   // Do the emission.
 593 
 594   Label fakeL; // Fake label for branch instructions.
 595   Label*   saveL = NULL;
 596   uint save_bnum = 0;
 597   bool is_branch = n->is_MachBranch();
 598   if (is_branch) {
 599     MacroAssembler masm(&buf);
 600     masm.bind(fakeL);
 601     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 602     n->as_MachBranch()->label_set(&fakeL, 0);
 603   }
 604   n->emit(buf, this->regalloc());
 605 
 606   // Emitting into the scratch buffer should not fail
 607   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
 608 
 609   if (is_branch) // Restore label.
 610     n->as_MachBranch()->label_set(saveL, save_bnum);
 611 
 612   // End scratch_emit_size section.
 613   set_in_scratch_emit_size(false);
 614 
 615   return buf.insts_size();
 616 }
 617 
 618 
 619 // ============================================================================
 620 //------------------------------Compile standard-------------------------------
 621 debug_only( int Compile::_debug_idx = 100000; )
 622 
 623 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 624 // the continuation bci for on stack replacement.
 625 
 626 
 627 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 628                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing, DirectiveSet* directive)
 629                 : Phase(Compiler),
 630                   _env(ci_env),
 631                   _directive(directive),
 632                   _log(ci_env->log()),
 633                   _compile_id(ci_env->compile_id()),
 634                   _save_argument_registers(false),
 635                   _stub_name(NULL),
 636                   _stub_function(NULL),
 637                   _stub_entry_point(NULL),
 638                   _method(target),
 639                   _entry_bci(osr_bci),
 640                   _initial_gvn(NULL),
 641                   _for_igvn(NULL),
 642                   _warm_calls(NULL),
 643                   _subsume_loads(subsume_loads),
 644                   _do_escape_analysis(do_escape_analysis),
 645                   _eliminate_boxing(eliminate_boxing),
 646                   _failure_reason(NULL),
 647                   _code_buffer("Compile::Fill_buffer"),
 648                   _orig_pc_slot(0),
 649                   _orig_pc_slot_offset_in_bytes(0),
 650                   _has_method_handle_invokes(false),
 651                   _mach_constant_base_node(NULL),
 652                   _node_bundling_limit(0),
 653                   _node_bundling_base(NULL),
 654                   _java_calls(0),
 655                   _inner_loops(0),
 656                   _scratch_const_size(-1),
 657                   _in_scratch_emit_size(false),
 658                   _dead_node_list(comp_arena()),
 659                   _dead_node_count(0),
 660 #ifndef PRODUCT
 661                   _trace_opto_output(directive->TraceOptoOutputOption),
 662                   _in_dump_cnt(0),
 663                   _printer(IdealGraphPrinter::printer()),
 664 #endif
 665                   _congraph(NULL),
 666                   _comp_arena(mtCompiler),
 667                   _node_arena(mtCompiler),
 668                   _old_arena(mtCompiler),
 669                   _Compile_types(mtCompiler),
 670                   _replay_inline_data(NULL),
 671                   _late_inlines(comp_arena(), 2, 0, NULL),
 672                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 673                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 674                   _late_inlines_pos(0),
 675                   _number_of_mh_late_inlines(0),
 676                   _inlining_progress(false),
 677                   _inlining_incrementally(false),
 678                   _print_inlining_list(NULL),
 679                   _print_inlining_stream(NULL),
 680                   _print_inlining_idx(0),
 681                   _print_inlining_output(NULL),
 682                   _interpreter_frame_size(0),
 683                   _max_node_limit(MaxNodeLimit),
 684                   _has_reserved_stack_access(target->has_reserved_stack_access()) {
 685   C = this;
 686 #ifndef PRODUCT
 687   if (_printer != NULL) {
 688     _printer->set_compile(this);
 689   }
 690 #endif
 691   CompileWrapper cw(this);
 692 
 693   if (CITimeVerbose) {
 694     tty->print(" ");
 695     target->holder()->name()->print();
 696     tty->print(".");
 697     target->print_short_name();
 698     tty->print("  ");
 699   }
 700   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 701   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 702 
 703 #ifndef PRODUCT
 704   bool print_opto_assembly = directive->PrintOptoAssemblyOption;
 705   if (!print_opto_assembly) {
 706     bool print_assembly = directive->PrintAssemblyOption;
 707     if (print_assembly && !Disassembler::can_decode()) {
 708       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 709       print_opto_assembly = true;
 710     }
 711   }
 712   set_print_assembly(print_opto_assembly);
 713   set_parsed_irreducible_loop(false);
 714 
 715   if (directive->ReplayInlineOption) {
 716     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 717   }
 718 #endif
 719   set_print_inlining(directive->PrintInliningOption || PrintOptoInlining);
 720   set_print_intrinsics(directive->PrintIntrinsicsOption);
 721   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 722 
 723   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 724     // Make sure the method being compiled gets its own MDO,
 725     // so we can at least track the decompile_count().
 726     // Need MDO to record RTM code generation state.
 727     method()->ensure_method_data();
 728   }
 729 
 730   Init(::AliasLevel);
 731 
 732 
 733   print_compile_messages();
 734 
 735   _ilt = InlineTree::build_inline_tree_root();
 736 
 737   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 738   assert(num_alias_types() >= AliasIdxRaw, "");
 739 
 740 #define MINIMUM_NODE_HASH  1023
 741   // Node list that Iterative GVN will start with
 742   Unique_Node_List for_igvn(comp_arena());
 743   set_for_igvn(&for_igvn);
 744 
 745   // GVN that will be run immediately on new nodes
 746   uint estimated_size = method()->code_size()*4+64;
 747   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 748   PhaseGVN gvn(node_arena(), estimated_size);
 749   set_initial_gvn(&gvn);
 750 
 751   print_inlining_init();
 752   { // Scope for timing the parser
 753     TracePhase tp("parse", &timers[_t_parser]);
 754 
 755     // Put top into the hash table ASAP.
 756     initial_gvn()->transform_no_reclaim(top());
 757 
 758     // Set up tf(), start(), and find a CallGenerator.
 759     CallGenerator* cg = NULL;
 760     if (is_osr_compilation()) {
 761       const TypeTuple *domain = StartOSRNode::osr_domain();
 762       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 763       init_tf(TypeFunc::make(domain, range));
 764       StartNode* s = new StartOSRNode(root(), domain);
 765       initial_gvn()->set_type_bottom(s);
 766       init_start(s);
 767       cg = CallGenerator::for_osr(method(), entry_bci());
 768     } else {
 769       // Normal case.
 770       init_tf(TypeFunc::make(method()));
 771       StartNode* s = new StartNode(root(), tf()->domain());
 772       initial_gvn()->set_type_bottom(s);
 773       init_start(s);
 774       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
 775         // With java.lang.ref.reference.get() we must go through the
 776         // intrinsic when G1 is enabled - even when get() is the root
 777         // method of the compile - so that, if necessary, the value in
 778         // the referent field of the reference object gets recorded by
 779         // the pre-barrier code.
 780         // Specifically, if G1 is enabled, the value in the referent
 781         // field is recorded by the G1 SATB pre barrier. This will
 782         // result in the referent being marked live and the reference
 783         // object removed from the list of discovered references during
 784         // reference processing.
 785         cg = find_intrinsic(method(), false);
 786       }
 787       if (cg == NULL) {
 788         float past_uses = method()->interpreter_invocation_count();
 789         float expected_uses = past_uses;
 790         cg = CallGenerator::for_inline(method(), expected_uses);
 791       }
 792     }
 793     if (failing())  return;
 794     if (cg == NULL) {
 795       record_method_not_compilable_all_tiers("cannot parse method");
 796       return;
 797     }
 798     JVMState* jvms = build_start_state(start(), tf());
 799     if ((jvms = cg->generate(jvms)) == NULL) {
 800       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 801         record_method_not_compilable("method parse failed");
 802       }
 803       return;
 804     }
 805     GraphKit kit(jvms);
 806 
 807     if (!kit.stopped()) {
 808       // Accept return values, and transfer control we know not where.
 809       // This is done by a special, unique ReturnNode bound to root.
 810       return_values(kit.jvms());
 811     }
 812 
 813     if (kit.has_exceptions()) {
 814       // Any exceptions that escape from this call must be rethrown
 815       // to whatever caller is dynamically above us on the stack.
 816       // This is done by a special, unique RethrowNode bound to root.
 817       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 818     }
 819 
 820     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 821 
 822     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 823       inline_string_calls(true);
 824     }
 825 
 826     if (failing())  return;
 827 
 828     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 829 
 830     // Remove clutter produced by parsing.
 831     if (!failing()) {
 832       ResourceMark rm;
 833       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 834     }
 835   }
 836 
 837   // Note:  Large methods are capped off in do_one_bytecode().
 838   if (failing())  return;
 839 
 840   // After parsing, node notes are no longer automagic.
 841   // They must be propagated by register_new_node_with_optimizer(),
 842   // clone(), or the like.
 843   set_default_node_notes(NULL);
 844 
 845   for (;;) {
 846     int successes = Inline_Warm();
 847     if (failing())  return;
 848     if (successes == 0)  break;
 849   }
 850 
 851   // Drain the list.
 852   Finish_Warm();
 853 #ifndef PRODUCT
 854   if (_printer && _printer->should_print(1)) {
 855     _printer->print_inlining();
 856   }
 857 #endif
 858 
 859   if (failing())  return;
 860   NOT_PRODUCT( verify_graph_edges(); )
 861 
 862   // Now optimize
 863   Optimize();
 864   if (failing())  return;
 865   NOT_PRODUCT( verify_graph_edges(); )
 866 
 867 #ifndef PRODUCT
 868   if (PrintIdeal) {
 869     ttyLocker ttyl;  // keep the following output all in one block
 870     // This output goes directly to the tty, not the compiler log.
 871     // To enable tools to match it up with the compilation activity,
 872     // be sure to tag this tty output with the compile ID.
 873     if (xtty != NULL) {
 874       xtty->head("ideal compile_id='%d'%s", compile_id(),
 875                  is_osr_compilation()    ? " compile_kind='osr'" :
 876                  "");
 877     }
 878     root()->dump(9999);
 879     if (xtty != NULL) {
 880       xtty->tail("ideal");
 881     }
 882   }
 883 #endif
 884 
 885   NOT_PRODUCT( verify_barriers(); )
 886 
 887   // Dump compilation data to replay it.
 888   if (directive->DumpReplayOption) {
 889     env()->dump_replay_data(_compile_id);
 890   }
 891   if (directive->DumpInlineOption && (ilt() != NULL)) {
 892     env()->dump_inline_data(_compile_id);
 893   }
 894 
 895   // Now that we know the size of all the monitors we can add a fixed slot
 896   // for the original deopt pc.
 897 
 898   _orig_pc_slot =  fixed_slots();
 899   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 900   set_fixed_slots(next_slot);
 901 
 902   // Compute when to use implicit null checks. Used by matching trap based
 903   // nodes and NullCheck optimization.
 904   set_allowed_deopt_reasons();
 905 
 906   // Now generate code
 907   Code_Gen();
 908   if (failing())  return;
 909 
 910   // Check if we want to skip execution of all compiled code.
 911   {
 912 #ifndef PRODUCT
 913     if (OptoNoExecute) {
 914       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 915       return;
 916     }
 917 #endif
 918     TracePhase tp("install_code", &timers[_t_registerMethod]);
 919 
 920     if (is_osr_compilation()) {
 921       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 922       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 923     } else {
 924       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 925       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 926     }
 927 
 928     env()->register_method(_method, _entry_bci,
 929                            &_code_offsets,
 930                            _orig_pc_slot_offset_in_bytes,
 931                            code_buffer(),
 932                            frame_size_in_words(), _oop_map_set,
 933                            &_handler_table, &_inc_table,
 934                            compiler,
 935                            has_unsafe_access(),
 936                            SharedRuntime::is_wide_vector(max_vector_size()),
 937                            rtm_state()
 938                            );
 939 
 940     if (log() != NULL) // Print code cache state into compiler log
 941       log()->code_cache_state();
 942   }
 943 }
 944 
 945 //------------------------------Compile----------------------------------------
 946 // Compile a runtime stub
 947 Compile::Compile( ciEnv* ci_env,
 948                   TypeFunc_generator generator,
 949                   address stub_function,
 950                   const char *stub_name,
 951                   int is_fancy_jump,
 952                   bool pass_tls,
 953                   bool save_arg_registers,
 954                   bool return_pc,
 955                   DirectiveSet* directive)
 956   : Phase(Compiler),
 957     _env(ci_env),
 958     _directive(directive),
 959     _log(ci_env->log()),
 960     _compile_id(0),
 961     _save_argument_registers(save_arg_registers),
 962     _method(NULL),
 963     _stub_name(stub_name),
 964     _stub_function(stub_function),
 965     _stub_entry_point(NULL),
 966     _entry_bci(InvocationEntryBci),
 967     _initial_gvn(NULL),
 968     _for_igvn(NULL),
 969     _warm_calls(NULL),
 970     _orig_pc_slot(0),
 971     _orig_pc_slot_offset_in_bytes(0),
 972     _subsume_loads(true),
 973     _do_escape_analysis(false),
 974     _eliminate_boxing(false),
 975     _failure_reason(NULL),
 976     _code_buffer("Compile::Fill_buffer"),
 977     _has_method_handle_invokes(false),
 978     _mach_constant_base_node(NULL),
 979     _node_bundling_limit(0),
 980     _node_bundling_base(NULL),
 981     _java_calls(0),
 982     _inner_loops(0),
 983 #ifndef PRODUCT
 984     _trace_opto_output(TraceOptoOutput),
 985     _in_dump_cnt(0),
 986     _printer(NULL),
 987 #endif
 988     _comp_arena(mtCompiler),
 989     _node_arena(mtCompiler),
 990     _old_arena(mtCompiler),
 991     _Compile_types(mtCompiler),
 992     _dead_node_list(comp_arena()),
 993     _dead_node_count(0),
 994     _congraph(NULL),
 995     _replay_inline_data(NULL),
 996     _number_of_mh_late_inlines(0),
 997     _inlining_progress(false),
 998     _inlining_incrementally(false),
 999     _print_inlining_list(NULL),
1000     _print_inlining_stream(NULL),
1001     _print_inlining_idx(0),
1002     _print_inlining_output(NULL),
1003     _allowed_reasons(0),
1004     _interpreter_frame_size(0),
1005     _max_node_limit(MaxNodeLimit) {
1006   C = this;
1007 
1008   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
1009   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
1010 
1011 #ifndef PRODUCT
1012   set_print_assembly(PrintFrameConverterAssembly);
1013   set_parsed_irreducible_loop(false);
1014 #endif
1015   set_has_irreducible_loop(false); // no loops
1016 
1017   CompileWrapper cw(this);
1018   Init(/*AliasLevel=*/ 0);
1019   init_tf((*generator)());
1020 
1021   {
1022     // The following is a dummy for the sake of GraphKit::gen_stub
1023     Unique_Node_List for_igvn(comp_arena());
1024     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1025     PhaseGVN gvn(Thread::current()->resource_area(),255);
1026     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1027     gvn.transform_no_reclaim(top());
1028 
1029     GraphKit kit;
1030     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1031   }
1032 
1033   NOT_PRODUCT( verify_graph_edges(); )
1034   Code_Gen();
1035   if (failing())  return;
1036 
1037 
1038   // Entry point will be accessed using compile->stub_entry_point();
1039   if (code_buffer() == NULL) {
1040     Matcher::soft_match_failure();
1041   } else {
1042     if (PrintAssembly && (WizardMode || Verbose))
1043       tty->print_cr("### Stub::%s", stub_name);
1044 
1045     if (!failing()) {
1046       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1047 
1048       // Make the NMethod
1049       // For now we mark the frame as never safe for profile stackwalking
1050       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1051                                                       code_buffer(),
1052                                                       CodeOffsets::frame_never_safe,
1053                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1054                                                       frame_size_in_words(),
1055                                                       _oop_map_set,
1056                                                       save_arg_registers);
1057       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1058 
1059       _stub_entry_point = rs->entry_point();
1060     }
1061   }
1062 }
1063 
1064 //------------------------------Init-------------------------------------------
1065 // Prepare for a single compilation
1066 void Compile::Init(int aliaslevel) {
1067   _unique  = 0;
1068   _regalloc = NULL;
1069 
1070   _tf      = NULL;  // filled in later
1071   _top     = NULL;  // cached later
1072   _matcher = NULL;  // filled in later
1073   _cfg     = NULL;  // filled in later
1074 
1075   set_24_bit_selection_and_mode(Use24BitFP, false);
1076 
1077   _node_note_array = NULL;
1078   _default_node_notes = NULL;
1079   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1080 
1081   _immutable_memory = NULL; // filled in at first inquiry
1082 
1083   // Globally visible Nodes
1084   // First set TOP to NULL to give safe behavior during creation of RootNode
1085   set_cached_top_node(NULL);
1086   set_root(new RootNode());
1087   // Now that you have a Root to point to, create the real TOP
1088   set_cached_top_node( new ConNode(Type::TOP) );
1089   set_recent_alloc(NULL, NULL);
1090 
1091   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1092   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1093   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1094   env()->set_dependencies(new Dependencies(env()));
1095 
1096   _fixed_slots = 0;
1097   set_has_split_ifs(false);
1098   set_has_loops(has_method() && method()->has_loops()); // first approximation
1099   set_has_stringbuilder(false);
1100   set_has_boxed_value(false);
1101   _trap_can_recompile = false;  // no traps emitted yet
1102   _major_progress = true; // start out assuming good things will happen
1103   set_has_unsafe_access(false);
1104   set_max_vector_size(0);
1105   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1106   set_decompile_count(0);
1107 
1108   set_do_freq_based_layout(_directive->BlockLayoutByFrequencyOption);
1109   set_num_loop_opts(LoopOptsCount);
1110   set_do_inlining(Inline);
1111   set_max_inline_size(MaxInlineSize);
1112   set_freq_inline_size(FreqInlineSize);
1113   set_do_scheduling(OptoScheduling);
1114   set_do_count_invocations(false);
1115   set_do_method_data_update(false);
1116 
1117   set_do_vector_loop(false);
1118 
1119   if (AllowVectorizeOnDemand) {
1120     if (has_method() && (_directive->VectorizeOption || _directive->VectorizeDebugOption)) {
1121       set_do_vector_loop(true);
1122       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());})
1123     } else if (has_method() && method()->name() != 0 &&
1124                method()->intrinsic_id() == vmIntrinsics::_forEachRemaining) {
1125       set_do_vector_loop(true);
1126     }
1127   }
1128   set_use_cmove(UseCMoveUnconditionally /* || do_vector_loop()*/); //TODO: consider do_vector_loop() mandate use_cmove unconditionally
1129   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());})
1130 
1131   set_age_code(has_method() && method()->profile_aging());
1132   set_rtm_state(NoRTM); // No RTM lock eliding by default
1133   _max_node_limit = _directive->MaxNodeLimitOption;
1134 
1135 #if INCLUDE_RTM_OPT
1136   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1137     int rtm_state = method()->method_data()->rtm_state();
1138     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1139       // Don't generate RTM lock eliding code.
1140       set_rtm_state(NoRTM);
1141     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1142       // Generate RTM lock eliding code without abort ratio calculation code.
1143       set_rtm_state(UseRTM);
1144     } else if (UseRTMDeopt) {
1145       // Generate RTM lock eliding code and include abort ratio calculation
1146       // code if UseRTMDeopt is on.
1147       set_rtm_state(ProfileRTM);
1148     }
1149   }
1150 #endif
1151   if (debug_info()->recording_non_safepoints()) {
1152     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1153                         (comp_arena(), 8, 0, NULL));
1154     set_default_node_notes(Node_Notes::make(this));
1155   }
1156 
1157   // // -- Initialize types before each compile --
1158   // // Update cached type information
1159   // if( _method && _method->constants() )
1160   //   Type::update_loaded_types(_method, _method->constants());
1161 
1162   // Init alias_type map.
1163   if (!_do_escape_analysis && aliaslevel == 3)
1164     aliaslevel = 2;  // No unique types without escape analysis
1165   _AliasLevel = aliaslevel;
1166   const int grow_ats = 16;
1167   _max_alias_types = grow_ats;
1168   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1169   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1170   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1171   {
1172     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1173   }
1174   // Initialize the first few types.
1175   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1176   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1177   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1178   _num_alias_types = AliasIdxRaw+1;
1179   // Zero out the alias type cache.
1180   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1181   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1182   probe_alias_cache(NULL)->_index = AliasIdxTop;
1183 
1184   _intrinsics = NULL;
1185   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1186   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1187   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1188   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1189   register_library_intrinsics();
1190 }
1191 
1192 //---------------------------init_start----------------------------------------
1193 // Install the StartNode on this compile object.
1194 void Compile::init_start(StartNode* s) {
1195   if (failing())
1196     return; // already failing
1197   assert(s == start(), "");
1198 }
1199 
1200 /**
1201  * Return the 'StartNode'. We must not have a pending failure, since the ideal graph
1202  * can be in an inconsistent state, i.e., we can get segmentation faults when traversing
1203  * the ideal graph.
1204  */
1205 StartNode* Compile::start() const {
1206   assert (!failing(), "Must not have pending failure. Reason is: %s", failure_reason());
1207   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1208     Node* start = root()->fast_out(i);
1209     if (start->is_Start()) {
1210       return start->as_Start();
1211     }
1212   }
1213   fatal("Did not find Start node!");
1214   return NULL;
1215 }
1216 
1217 //-------------------------------immutable_memory-------------------------------------
1218 // Access immutable memory
1219 Node* Compile::immutable_memory() {
1220   if (_immutable_memory != NULL) {
1221     return _immutable_memory;
1222   }
1223   StartNode* s = start();
1224   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1225     Node *p = s->fast_out(i);
1226     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1227       _immutable_memory = p;
1228       return _immutable_memory;
1229     }
1230   }
1231   ShouldNotReachHere();
1232   return NULL;
1233 }
1234 
1235 //----------------------set_cached_top_node------------------------------------
1236 // Install the cached top node, and make sure Node::is_top works correctly.
1237 void Compile::set_cached_top_node(Node* tn) {
1238   if (tn != NULL)  verify_top(tn);
1239   Node* old_top = _top;
1240   _top = tn;
1241   // Calling Node::setup_is_top allows the nodes the chance to adjust
1242   // their _out arrays.
1243   if (_top != NULL)     _top->setup_is_top();
1244   if (old_top != NULL)  old_top->setup_is_top();
1245   assert(_top == NULL || top()->is_top(), "");
1246 }
1247 
1248 #ifdef ASSERT
1249 uint Compile::count_live_nodes_by_graph_walk() {
1250   Unique_Node_List useful(comp_arena());
1251   // Get useful node list by walking the graph.
1252   identify_useful_nodes(useful);
1253   return useful.size();
1254 }
1255 
1256 void Compile::print_missing_nodes() {
1257 
1258   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1259   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1260     return;
1261   }
1262 
1263   // This is an expensive function. It is executed only when the user
1264   // specifies VerifyIdealNodeCount option or otherwise knows the
1265   // additional work that needs to be done to identify reachable nodes
1266   // by walking the flow graph and find the missing ones using
1267   // _dead_node_list.
1268 
1269   Unique_Node_List useful(comp_arena());
1270   // Get useful node list by walking the graph.
1271   identify_useful_nodes(useful);
1272 
1273   uint l_nodes = C->live_nodes();
1274   uint l_nodes_by_walk = useful.size();
1275 
1276   if (l_nodes != l_nodes_by_walk) {
1277     if (_log != NULL) {
1278       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1279       _log->stamp();
1280       _log->end_head();
1281     }
1282     VectorSet& useful_member_set = useful.member_set();
1283     int last_idx = l_nodes_by_walk;
1284     for (int i = 0; i < last_idx; i++) {
1285       if (useful_member_set.test(i)) {
1286         if (_dead_node_list.test(i)) {
1287           if (_log != NULL) {
1288             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1289           }
1290           if (PrintIdealNodeCount) {
1291             // Print the log message to tty
1292               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1293               useful.at(i)->dump();
1294           }
1295         }
1296       }
1297       else if (! _dead_node_list.test(i)) {
1298         if (_log != NULL) {
1299           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1300         }
1301         if (PrintIdealNodeCount) {
1302           // Print the log message to tty
1303           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1304         }
1305       }
1306     }
1307     if (_log != NULL) {
1308       _log->tail("mismatched_nodes");
1309     }
1310   }
1311 }
1312 void Compile::record_modified_node(Node* n) {
1313   if (_modified_nodes != NULL && !_inlining_incrementally &&
1314       n->outcnt() != 0 && !n->is_Con()) {
1315     _modified_nodes->push(n);
1316   }
1317 }
1318 
1319 void Compile::remove_modified_node(Node* n) {
1320   if (_modified_nodes != NULL) {
1321     _modified_nodes->remove(n);
1322   }
1323 }
1324 #endif
1325 
1326 #ifndef PRODUCT
1327 void Compile::verify_top(Node* tn) const {
1328   if (tn != NULL) {
1329     assert(tn->is_Con(), "top node must be a constant");
1330     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1331     assert(tn->in(0) != NULL, "must have live top node");
1332   }
1333 }
1334 #endif
1335 
1336 
1337 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1338 
1339 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1340   guarantee(arr != NULL, "");
1341   int num_blocks = arr->length();
1342   if (grow_by < num_blocks)  grow_by = num_blocks;
1343   int num_notes = grow_by * _node_notes_block_size;
1344   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1345   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1346   while (num_notes > 0) {
1347     arr->append(notes);
1348     notes     += _node_notes_block_size;
1349     num_notes -= _node_notes_block_size;
1350   }
1351   assert(num_notes == 0, "exact multiple, please");
1352 }
1353 
1354 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1355   if (source == NULL || dest == NULL)  return false;
1356 
1357   if (dest->is_Con())
1358     return false;               // Do not push debug info onto constants.
1359 
1360 #ifdef ASSERT
1361   // Leave a bread crumb trail pointing to the original node:
1362   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1363     dest->set_debug_orig(source);
1364   }
1365 #endif
1366 
1367   if (node_note_array() == NULL)
1368     return false;               // Not collecting any notes now.
1369 
1370   // This is a copy onto a pre-existing node, which may already have notes.
1371   // If both nodes have notes, do not overwrite any pre-existing notes.
1372   Node_Notes* source_notes = node_notes_at(source->_idx);
1373   if (source_notes == NULL || source_notes->is_clear())  return false;
1374   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1375   if (dest_notes == NULL || dest_notes->is_clear()) {
1376     return set_node_notes_at(dest->_idx, source_notes);
1377   }
1378 
1379   Node_Notes merged_notes = (*source_notes);
1380   // The order of operations here ensures that dest notes will win...
1381   merged_notes.update_from(dest_notes);
1382   return set_node_notes_at(dest->_idx, &merged_notes);
1383 }
1384 
1385 
1386 //--------------------------allow_range_check_smearing-------------------------
1387 // Gating condition for coalescing similar range checks.
1388 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1389 // single covering check that is at least as strong as any of them.
1390 // If the optimization succeeds, the simplified (strengthened) range check
1391 // will always succeed.  If it fails, we will deopt, and then give up
1392 // on the optimization.
1393 bool Compile::allow_range_check_smearing() const {
1394   // If this method has already thrown a range-check,
1395   // assume it was because we already tried range smearing
1396   // and it failed.
1397   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1398   return !already_trapped;
1399 }
1400 
1401 
1402 //------------------------------flatten_alias_type-----------------------------
1403 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1404   int offset = tj->offset();
1405   TypePtr::PTR ptr = tj->ptr();
1406 
1407   // Known instance (scalarizable allocation) alias only with itself.
1408   bool is_known_inst = tj->isa_oopptr() != NULL &&
1409                        tj->is_oopptr()->is_known_instance();
1410 
1411   // Process weird unsafe references.
1412   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1413     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1414     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1415     tj = TypeOopPtr::BOTTOM;
1416     ptr = tj->ptr();
1417     offset = tj->offset();
1418   }
1419 
1420   // Array pointers need some flattening
1421   const TypeAryPtr *ta = tj->isa_aryptr();
1422   if (ta && ta->is_stable()) {
1423     // Erase stability property for alias analysis.
1424     tj = ta = ta->cast_to_stable(false);
1425   }
1426   if( ta && is_known_inst ) {
1427     if ( offset != Type::OffsetBot &&
1428          offset > arrayOopDesc::length_offset_in_bytes() ) {
1429       offset = Type::OffsetBot; // Flatten constant access into array body only
1430       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1431     }
1432   } else if( ta && _AliasLevel >= 2 ) {
1433     // For arrays indexed by constant indices, we flatten the alias
1434     // space to include all of the array body.  Only the header, klass
1435     // and array length can be accessed un-aliased.
1436     if( offset != Type::OffsetBot ) {
1437       if( ta->const_oop() ) { // MethodData* or Method*
1438         offset = Type::OffsetBot;   // Flatten constant access into array body
1439         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1440       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1441         // range is OK as-is.
1442         tj = ta = TypeAryPtr::RANGE;
1443       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1444         tj = TypeInstPtr::KLASS; // all klass loads look alike
1445         ta = TypeAryPtr::RANGE; // generic ignored junk
1446         ptr = TypePtr::BotPTR;
1447       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1448         tj = TypeInstPtr::MARK;
1449         ta = TypeAryPtr::RANGE; // generic ignored junk
1450         ptr = TypePtr::BotPTR;
1451       } else {                  // Random constant offset into array body
1452         offset = Type::OffsetBot;   // Flatten constant access into array body
1453         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1454       }
1455     }
1456     // Arrays of fixed size alias with arrays of unknown size.
1457     if (ta->size() != TypeInt::POS) {
1458       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1459       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1460     }
1461     // Arrays of known objects become arrays of unknown objects.
1462     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1463       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1464       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1465     }
1466     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1467       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1468       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1469     }
1470     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1471     // cannot be distinguished by bytecode alone.
1472     if (ta->elem() == TypeInt::BOOL) {
1473       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1474       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1475       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1476     }
1477     // During the 2nd round of IterGVN, NotNull castings are removed.
1478     // Make sure the Bottom and NotNull variants alias the same.
1479     // Also, make sure exact and non-exact variants alias the same.
1480     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1481       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1482     }
1483   }
1484 
1485   // Oop pointers need some flattening
1486   const TypeInstPtr *to = tj->isa_instptr();
1487   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1488     ciInstanceKlass *k = to->klass()->as_instance_klass();
1489     if( ptr == TypePtr::Constant ) {
1490       if (to->klass() != ciEnv::current()->Class_klass() ||
1491           offset < k->size_helper() * wordSize) {
1492         // No constant oop pointers (such as Strings); they alias with
1493         // unknown strings.
1494         assert(!is_known_inst, "not scalarizable allocation");
1495         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1496       }
1497     } else if( is_known_inst ) {
1498       tj = to; // Keep NotNull and klass_is_exact for instance type
1499     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1500       // During the 2nd round of IterGVN, NotNull castings are removed.
1501       // Make sure the Bottom and NotNull variants alias the same.
1502       // Also, make sure exact and non-exact variants alias the same.
1503       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1504     }
1505     if (to->speculative() != NULL) {
1506       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1507     }
1508     // Canonicalize the holder of this field
1509     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1510       // First handle header references such as a LoadKlassNode, even if the
1511       // object's klass is unloaded at compile time (4965979).
1512       if (!is_known_inst) { // Do it only for non-instance types
1513         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1514       }
1515     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1516       // Static fields are in the space above the normal instance
1517       // fields in the java.lang.Class instance.
1518       if (to->klass() != ciEnv::current()->Class_klass()) {
1519         to = NULL;
1520         tj = TypeOopPtr::BOTTOM;
1521         offset = tj->offset();
1522       }
1523     } else {
1524       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1525       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1526         if( is_known_inst ) {
1527           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1528         } else {
1529           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1530         }
1531       }
1532     }
1533   }
1534 
1535   // Klass pointers to object array klasses need some flattening
1536   const TypeKlassPtr *tk = tj->isa_klassptr();
1537   if( tk ) {
1538     // If we are referencing a field within a Klass, we need
1539     // to assume the worst case of an Object.  Both exact and
1540     // inexact types must flatten to the same alias class so
1541     // use NotNull as the PTR.
1542     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1543 
1544       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1545                                    TypeKlassPtr::OBJECT->klass(),
1546                                    offset);
1547     }
1548 
1549     ciKlass* klass = tk->klass();
1550     if( klass->is_obj_array_klass() ) {
1551       ciKlass* k = TypeAryPtr::OOPS->klass();
1552       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1553         k = TypeInstPtr::BOTTOM->klass();
1554       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1555     }
1556 
1557     // Check for precise loads from the primary supertype array and force them
1558     // to the supertype cache alias index.  Check for generic array loads from
1559     // the primary supertype array and also force them to the supertype cache
1560     // alias index.  Since the same load can reach both, we need to merge
1561     // these 2 disparate memories into the same alias class.  Since the
1562     // primary supertype array is read-only, there's no chance of confusion
1563     // where we bypass an array load and an array store.
1564     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1565     if (offset == Type::OffsetBot ||
1566         (offset >= primary_supers_offset &&
1567          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1568         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1569       offset = in_bytes(Klass::secondary_super_cache_offset());
1570       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1571     }
1572   }
1573 
1574   // Flatten all Raw pointers together.
1575   if (tj->base() == Type::RawPtr)
1576     tj = TypeRawPtr::BOTTOM;
1577 
1578   if (tj->base() == Type::AnyPtr)
1579     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1580 
1581   // Flatten all to bottom for now
1582   switch( _AliasLevel ) {
1583   case 0:
1584     tj = TypePtr::BOTTOM;
1585     break;
1586   case 1:                       // Flatten to: oop, static, field or array
1587     switch (tj->base()) {
1588     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1589     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1590     case Type::AryPtr:   // do not distinguish arrays at all
1591     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1592     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1593     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1594     default: ShouldNotReachHere();
1595     }
1596     break;
1597   case 2:                       // No collapsing at level 2; keep all splits
1598   case 3:                       // No collapsing at level 3; keep all splits
1599     break;
1600   default:
1601     Unimplemented();
1602   }
1603 
1604   offset = tj->offset();
1605   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1606 
1607   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1608           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1609           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1610           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1611           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1612           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1613           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr)  ,
1614           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1615   assert( tj->ptr() != TypePtr::TopPTR &&
1616           tj->ptr() != TypePtr::AnyNull &&
1617           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1618 //    assert( tj->ptr() != TypePtr::Constant ||
1619 //            tj->base() == Type::RawPtr ||
1620 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1621 
1622   return tj;
1623 }
1624 
1625 void Compile::AliasType::Init(int i, const TypePtr* at) {
1626   _index = i;
1627   _adr_type = at;
1628   _field = NULL;
1629   _element = NULL;
1630   _is_rewritable = true; // default
1631   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1632   if (atoop != NULL && atoop->is_known_instance()) {
1633     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1634     _general_index = Compile::current()->get_alias_index(gt);
1635   } else {
1636     _general_index = 0;
1637   }
1638 }
1639 
1640 //---------------------------------print_on------------------------------------
1641 #ifndef PRODUCT
1642 void Compile::AliasType::print_on(outputStream* st) {
1643   if (index() < 10)
1644         st->print("@ <%d> ", index());
1645   else  st->print("@ <%d>",  index());
1646   st->print(is_rewritable() ? "   " : " RO");
1647   int offset = adr_type()->offset();
1648   if (offset == Type::OffsetBot)
1649         st->print(" +any");
1650   else  st->print(" +%-3d", offset);
1651   st->print(" in ");
1652   adr_type()->dump_on(st);
1653   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1654   if (field() != NULL && tjp) {
1655     if (tjp->klass()  != field()->holder() ||
1656         tjp->offset() != field()->offset_in_bytes()) {
1657       st->print(" != ");
1658       field()->print();
1659       st->print(" ***");
1660     }
1661   }
1662 }
1663 
1664 void print_alias_types() {
1665   Compile* C = Compile::current();
1666   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1667   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1668     C->alias_type(idx)->print_on(tty);
1669     tty->cr();
1670   }
1671 }
1672 #endif
1673 
1674 
1675 //----------------------------probe_alias_cache--------------------------------
1676 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1677   intptr_t key = (intptr_t) adr_type;
1678   key ^= key >> logAliasCacheSize;
1679   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1680 }
1681 
1682 
1683 //-----------------------------grow_alias_types--------------------------------
1684 void Compile::grow_alias_types() {
1685   const int old_ats  = _max_alias_types; // how many before?
1686   const int new_ats  = old_ats;          // how many more?
1687   const int grow_ats = old_ats+new_ats;  // how many now?
1688   _max_alias_types = grow_ats;
1689   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1690   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1691   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1692   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1693 }
1694 
1695 
1696 //--------------------------------find_alias_type------------------------------
1697 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1698   if (_AliasLevel == 0)
1699     return alias_type(AliasIdxBot);
1700 
1701   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1702   if (ace->_adr_type == adr_type) {
1703     return alias_type(ace->_index);
1704   }
1705 
1706   // Handle special cases.
1707   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1708   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1709 
1710   // Do it the slow way.
1711   const TypePtr* flat = flatten_alias_type(adr_type);
1712 
1713 #ifdef ASSERT
1714   assert(flat == flatten_alias_type(flat), "idempotent");
1715   assert(flat != TypePtr::BOTTOM,     "cannot alias-analyze an untyped ptr");
1716   if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1717     const TypeOopPtr* foop = flat->is_oopptr();
1718     // Scalarizable allocations have exact klass always.
1719     bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1720     const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1721     assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1722   }
1723   assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1724 #endif
1725 
1726   int idx = AliasIdxTop;
1727   for (int i = 0; i < num_alias_types(); i++) {
1728     if (alias_type(i)->adr_type() == flat) {
1729       idx = i;
1730       break;
1731     }
1732   }
1733 
1734   if (idx == AliasIdxTop) {
1735     if (no_create)  return NULL;
1736     // Grow the array if necessary.
1737     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1738     // Add a new alias type.
1739     idx = _num_alias_types++;
1740     _alias_types[idx]->Init(idx, flat);
1741     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1742     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1743     if (flat->isa_instptr()) {
1744       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1745           && flat->is_instptr()->klass() == env()->Class_klass())
1746         alias_type(idx)->set_rewritable(false);
1747     }
1748     if (flat->isa_aryptr()) {
1749 #ifdef ASSERT
1750       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1751       // (T_BYTE has the weakest alignment and size restrictions...)
1752       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1753 #endif
1754       if (flat->offset() == TypePtr::OffsetBot) {
1755         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1756       }
1757     }
1758     if (flat->isa_klassptr()) {
1759       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1760         alias_type(idx)->set_rewritable(false);
1761       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1762         alias_type(idx)->set_rewritable(false);
1763       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1764         alias_type(idx)->set_rewritable(false);
1765       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1766         alias_type(idx)->set_rewritable(false);
1767     }
1768     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1769     // but the base pointer type is not distinctive enough to identify
1770     // references into JavaThread.)
1771 
1772     // Check for final fields.
1773     const TypeInstPtr* tinst = flat->isa_instptr();
1774     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1775       ciField* field;
1776       if (tinst->const_oop() != NULL &&
1777           tinst->klass() == ciEnv::current()->Class_klass() &&
1778           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1779         // static field
1780         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1781         field = k->get_field_by_offset(tinst->offset(), true);
1782       } else {
1783         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1784         field = k->get_field_by_offset(tinst->offset(), false);
1785       }
1786       assert(field == NULL ||
1787              original_field == NULL ||
1788              (field->holder() == original_field->holder() &&
1789               field->offset() == original_field->offset() &&
1790               field->is_static() == original_field->is_static()), "wrong field?");
1791       // Set field() and is_rewritable() attributes.
1792       if (field != NULL)  alias_type(idx)->set_field(field);
1793     }
1794   }
1795 
1796   // Fill the cache for next time.
1797   ace->_adr_type = adr_type;
1798   ace->_index    = idx;
1799   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1800 
1801   // Might as well try to fill the cache for the flattened version, too.
1802   AliasCacheEntry* face = probe_alias_cache(flat);
1803   if (face->_adr_type == NULL) {
1804     face->_adr_type = flat;
1805     face->_index    = idx;
1806     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1807   }
1808 
1809   return alias_type(idx);
1810 }
1811 
1812 
1813 Compile::AliasType* Compile::alias_type(ciField* field) {
1814   const TypeOopPtr* t;
1815   if (field->is_static())
1816     t = TypeInstPtr::make(field->holder()->java_mirror());
1817   else
1818     t = TypeOopPtr::make_from_klass_raw(field->holder());
1819   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1820   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1821   return atp;
1822 }
1823 
1824 
1825 //------------------------------have_alias_type--------------------------------
1826 bool Compile::have_alias_type(const TypePtr* adr_type) {
1827   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1828   if (ace->_adr_type == adr_type) {
1829     return true;
1830   }
1831 
1832   // Handle special cases.
1833   if (adr_type == NULL)             return true;
1834   if (adr_type == TypePtr::BOTTOM)  return true;
1835 
1836   return find_alias_type(adr_type, true, NULL) != NULL;
1837 }
1838 
1839 //-----------------------------must_alias--------------------------------------
1840 // True if all values of the given address type are in the given alias category.
1841 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1842   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1843   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1844   if (alias_idx == AliasIdxTop)         return false; // the empty category
1845   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1846 
1847   // the only remaining possible overlap is identity
1848   int adr_idx = get_alias_index(adr_type);
1849   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1850   assert(adr_idx == alias_idx ||
1851          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1852           && adr_type                       != TypeOopPtr::BOTTOM),
1853          "should not be testing for overlap with an unsafe pointer");
1854   return adr_idx == alias_idx;
1855 }
1856 
1857 //------------------------------can_alias--------------------------------------
1858 // True if any values of the given address type are in the given alias category.
1859 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1860   if (alias_idx == AliasIdxTop)         return false; // the empty category
1861   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1862   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1863   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1864 
1865   // the only remaining possible overlap is identity
1866   int adr_idx = get_alias_index(adr_type);
1867   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1868   return adr_idx == alias_idx;
1869 }
1870 
1871 
1872 
1873 //---------------------------pop_warm_call-------------------------------------
1874 WarmCallInfo* Compile::pop_warm_call() {
1875   WarmCallInfo* wci = _warm_calls;
1876   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1877   return wci;
1878 }
1879 
1880 //----------------------------Inline_Warm--------------------------------------
1881 int Compile::Inline_Warm() {
1882   // If there is room, try to inline some more warm call sites.
1883   // %%% Do a graph index compaction pass when we think we're out of space?
1884   if (!InlineWarmCalls)  return 0;
1885 
1886   int calls_made_hot = 0;
1887   int room_to_grow   = NodeCountInliningCutoff - unique();
1888   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1889   int amount_grown   = 0;
1890   WarmCallInfo* call;
1891   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1892     int est_size = (int)call->size();
1893     if (est_size > (room_to_grow - amount_grown)) {
1894       // This one won't fit anyway.  Get rid of it.
1895       call->make_cold();
1896       continue;
1897     }
1898     call->make_hot();
1899     calls_made_hot++;
1900     amount_grown   += est_size;
1901     amount_to_grow -= est_size;
1902   }
1903 
1904   return calls_made_hot;
1905 }
1906 
1907 
1908 //----------------------------Finish_Warm--------------------------------------
1909 void Compile::Finish_Warm() {
1910   if (!InlineWarmCalls)  return;
1911   if (failing())  return;
1912   if (warm_calls() == NULL)  return;
1913 
1914   // Clean up loose ends, if we are out of space for inlining.
1915   WarmCallInfo* call;
1916   while ((call = pop_warm_call()) != NULL) {
1917     call->make_cold();
1918   }
1919 }
1920 
1921 //---------------------cleanup_loop_predicates-----------------------
1922 // Remove the opaque nodes that protect the predicates so that all unused
1923 // checks and uncommon_traps will be eliminated from the ideal graph
1924 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1925   if (predicate_count()==0) return;
1926   for (int i = predicate_count(); i > 0; i--) {
1927     Node * n = predicate_opaque1_node(i-1);
1928     assert(n->Opcode() == Op_Opaque1, "must be");
1929     igvn.replace_node(n, n->in(1));
1930   }
1931   assert(predicate_count()==0, "should be clean!");
1932 }
1933 
1934 void Compile::add_range_check_cast(Node* n) {
1935   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1936   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1937   _range_check_casts->append(n);
1938 }
1939 
1940 // Remove all range check dependent CastIINodes.
1941 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1942   for (int i = range_check_cast_count(); i > 0; i--) {
1943     Node* cast = range_check_cast_node(i-1);
1944     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1945     igvn.replace_node(cast, cast->in(1));
1946   }
1947   assert(range_check_cast_count() == 0, "should be empty");
1948 }
1949 
1950 // StringOpts and late inlining of string methods
1951 void Compile::inline_string_calls(bool parse_time) {
1952   {
1953     // remove useless nodes to make the usage analysis simpler
1954     ResourceMark rm;
1955     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1956   }
1957 
1958   {
1959     ResourceMark rm;
1960     print_method(PHASE_BEFORE_STRINGOPTS, 3);
1961     PhaseStringOpts pso(initial_gvn(), for_igvn());
1962     print_method(PHASE_AFTER_STRINGOPTS, 3);
1963   }
1964 
1965   // now inline anything that we skipped the first time around
1966   if (!parse_time) {
1967     _late_inlines_pos = _late_inlines.length();
1968   }
1969 
1970   while (_string_late_inlines.length() > 0) {
1971     CallGenerator* cg = _string_late_inlines.pop();
1972     cg->do_late_inline();
1973     if (failing())  return;
1974   }
1975   _string_late_inlines.trunc_to(0);
1976 }
1977 
1978 // Late inlining of boxing methods
1979 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1980   if (_boxing_late_inlines.length() > 0) {
1981     assert(has_boxed_value(), "inconsistent");
1982 
1983     PhaseGVN* gvn = initial_gvn();
1984     set_inlining_incrementally(true);
1985 
1986     assert( igvn._worklist.size() == 0, "should be done with igvn" );
1987     for_igvn()->clear();
1988     gvn->replace_with(&igvn);
1989 
1990     _late_inlines_pos = _late_inlines.length();
1991 
1992     while (_boxing_late_inlines.length() > 0) {
1993       CallGenerator* cg = _boxing_late_inlines.pop();
1994       cg->do_late_inline();
1995       if (failing())  return;
1996     }
1997     _boxing_late_inlines.trunc_to(0);
1998 
1999     {
2000       ResourceMark rm;
2001       PhaseRemoveUseless pru(gvn, for_igvn());
2002     }
2003 
2004     igvn = PhaseIterGVN(gvn);
2005     igvn.optimize();
2006 
2007     set_inlining_progress(false);
2008     set_inlining_incrementally(false);
2009   }
2010 }
2011 
2012 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2013   assert(IncrementalInline, "incremental inlining should be on");
2014   PhaseGVN* gvn = initial_gvn();
2015 
2016   set_inlining_progress(false);
2017   for_igvn()->clear();
2018   gvn->replace_with(&igvn);
2019 
2020   {
2021     TracePhase tp("incrementalInline_inline", &timers[_t_incrInline_inline]);
2022     int i = 0;
2023     for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2024       CallGenerator* cg = _late_inlines.at(i);
2025       _late_inlines_pos = i+1;
2026       cg->do_late_inline();
2027       if (failing())  return;
2028     }
2029     int j = 0;
2030     for (; i < _late_inlines.length(); i++, j++) {
2031       _late_inlines.at_put(j, _late_inlines.at(i));
2032     }
2033     _late_inlines.trunc_to(j);
2034   }
2035 
2036   {
2037     TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2038     ResourceMark rm;
2039     PhaseRemoveUseless pru(gvn, for_igvn());
2040   }
2041 
2042   {
2043     TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2044     igvn = PhaseIterGVN(gvn);
2045   }
2046 }
2047 
2048 // Perform incremental inlining until bound on number of live nodes is reached
2049 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2050   TracePhase tp("incrementalInline", &timers[_t_incrInline]);
2051 
2052   PhaseGVN* gvn = initial_gvn();
2053 
2054   set_inlining_incrementally(true);
2055   set_inlining_progress(true);
2056   uint low_live_nodes = 0;
2057 
2058   while(inlining_progress() && _late_inlines.length() > 0) {
2059 
2060     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2061       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2062         TracePhase tp("incrementalInline_ideal", &timers[_t_incrInline_ideal]);
2063         // PhaseIdealLoop is expensive so we only try it once we are
2064         // out of live nodes and we only try it again if the previous
2065         // helped got the number of nodes down significantly
2066         PhaseIdealLoop ideal_loop( igvn, false, true );
2067         if (failing())  return;
2068         low_live_nodes = live_nodes();
2069       }
2070 
2071       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2072         break;
2073       }
2074     }
2075 
2076     inline_incrementally_one(igvn);
2077 
2078     if (failing())  return;
2079 
2080     {
2081       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2082       igvn.optimize();
2083     }
2084 
2085     if (failing())  return;
2086   }
2087 
2088   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2089 
2090   if (_string_late_inlines.length() > 0) {
2091     assert(has_stringbuilder(), "inconsistent");
2092     for_igvn()->clear();
2093     initial_gvn()->replace_with(&igvn);
2094 
2095     inline_string_calls(false);
2096 
2097     if (failing())  return;
2098 
2099     {
2100       TracePhase tp("incrementalInline_pru", &timers[_t_incrInline_pru]);
2101       ResourceMark rm;
2102       PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2103     }
2104 
2105     {
2106       TracePhase tp("incrementalInline_igvn", &timers[_t_incrInline_igvn]);
2107       igvn = PhaseIterGVN(gvn);
2108       igvn.optimize();
2109     }
2110   }
2111 
2112   set_inlining_incrementally(false);
2113 }
2114 
2115 
2116 //------------------------------Optimize---------------------------------------
2117 // Given a graph, optimize it.
2118 void Compile::Optimize() {
2119   TracePhase tp("optimizer", &timers[_t_optimizer]);
2120 
2121 #ifndef PRODUCT
2122   if (_directive->BreakAtCompileOption) {
2123     BREAKPOINT;
2124   }
2125 
2126 #endif
2127 
2128   ResourceMark rm;
2129   int          loop_opts_cnt;
2130 
2131   print_inlining_reinit();
2132 
2133   NOT_PRODUCT( verify_graph_edges(); )
2134 
2135   print_method(PHASE_AFTER_PARSING);
2136 
2137  {
2138   // Iterative Global Value Numbering, including ideal transforms
2139   // Initialize IterGVN with types and values from parse-time GVN
2140   PhaseIterGVN igvn(initial_gvn());
2141 #ifdef ASSERT
2142   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2143 #endif
2144   {
2145     TracePhase tp("iterGVN", &timers[_t_iterGVN]);
2146     igvn.optimize();
2147   }
2148 
2149   print_method(PHASE_ITER_GVN1, 2);
2150 
2151   if (failing())  return;
2152 
2153   inline_incrementally(igvn);
2154 
2155   print_method(PHASE_INCREMENTAL_INLINE, 2);
2156 
2157   if (failing())  return;
2158 
2159   if (eliminate_boxing()) {
2160     // Inline valueOf() methods now.
2161     inline_boxing_calls(igvn);
2162 
2163     if (AlwaysIncrementalInline) {
2164       inline_incrementally(igvn);
2165     }
2166 
2167     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2168 
2169     if (failing())  return;
2170   }
2171 
2172   // Remove the speculative part of types and clean up the graph from
2173   // the extra CastPP nodes whose only purpose is to carry them. Do
2174   // that early so that optimizations are not disrupted by the extra
2175   // CastPP nodes.
2176   remove_speculative_types(igvn);
2177 
2178   // No more new expensive nodes will be added to the list from here
2179   // so keep only the actual candidates for optimizations.
2180   cleanup_expensive_nodes(igvn);
2181 
2182   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2183     Compile::TracePhase tp("", &timers[_t_renumberLive]);
2184     initial_gvn()->replace_with(&igvn);
2185     for_igvn()->clear();
2186     Unique_Node_List new_worklist(C->comp_arena());
2187     {
2188       ResourceMark rm;
2189       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2190     }
2191     set_for_igvn(&new_worklist);
2192     igvn = PhaseIterGVN(initial_gvn());
2193     igvn.optimize();
2194   }
2195 
2196   // Perform escape analysis
2197   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2198     if (has_loops()) {
2199       // Cleanup graph (remove dead nodes).
2200       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2201       PhaseIdealLoop ideal_loop( igvn, false, true );
2202       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2203       if (failing())  return;
2204     }
2205     ConnectionGraph::do_analysis(this, &igvn);
2206 
2207     if (failing())  return;
2208 
2209     // Optimize out fields loads from scalar replaceable allocations.
2210     igvn.optimize();
2211     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2212 
2213     if (failing())  return;
2214 
2215     if (congraph() != NULL && macro_count() > 0) {
2216       TracePhase tp("macroEliminate", &timers[_t_macroEliminate]);
2217       PhaseMacroExpand mexp(igvn);
2218       mexp.eliminate_macro_nodes();
2219       igvn.set_delay_transform(false);
2220 
2221       igvn.optimize();
2222       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2223 
2224       if (failing())  return;
2225     }
2226   }
2227 
2228   // Loop transforms on the ideal graph.  Range Check Elimination,
2229   // peeling, unrolling, etc.
2230 
2231   // Set loop opts counter
2232   loop_opts_cnt = num_loop_opts();
2233   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2234     {
2235       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2236       PhaseIdealLoop ideal_loop( igvn, true );
2237       loop_opts_cnt--;
2238       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2239       if (failing())  return;
2240     }
2241     // Loop opts pass if partial peeling occurred in previous pass
2242     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2243       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2244       PhaseIdealLoop ideal_loop( igvn, false );
2245       loop_opts_cnt--;
2246       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2247       if (failing())  return;
2248     }
2249     // Loop opts pass for loop-unrolling before CCP
2250     if(major_progress() && (loop_opts_cnt > 0)) {
2251       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2252       PhaseIdealLoop ideal_loop( igvn, false );
2253       loop_opts_cnt--;
2254       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2255     }
2256     if (!failing()) {
2257       // Verify that last round of loop opts produced a valid graph
2258       TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2259       PhaseIdealLoop::verify(igvn);
2260     }
2261   }
2262   if (failing())  return;
2263 
2264   // Conditional Constant Propagation;
2265   PhaseCCP ccp( &igvn );
2266   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2267   {
2268     TracePhase tp("ccp", &timers[_t_ccp]);
2269     ccp.do_transform();
2270   }
2271   print_method(PHASE_CPP1, 2);
2272 
2273   assert( true, "Break here to ccp.dump_old2new_map()");
2274 
2275   // Iterative Global Value Numbering, including ideal transforms
2276   {
2277     TracePhase tp("iterGVN2", &timers[_t_iterGVN2]);
2278     igvn = ccp;
2279     igvn.optimize();
2280   }
2281 
2282   print_method(PHASE_ITER_GVN2, 2);
2283 
2284   if (failing())  return;
2285 
2286   // Loop transforms on the ideal graph.  Range Check Elimination,
2287   // peeling, unrolling, etc.
2288   if(loop_opts_cnt > 0) {
2289     debug_only( int cnt = 0; );
2290     while(major_progress() && (loop_opts_cnt > 0)) {
2291       TracePhase tp("idealLoop", &timers[_t_idealLoop]);
2292       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2293       PhaseIdealLoop ideal_loop( igvn, true);
2294       loop_opts_cnt--;
2295       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2296       if (failing())  return;
2297     }
2298   }
2299   // Ensure that major progress is now clear
2300   C->clear_major_progress();
2301 
2302   {
2303     // Verify that all previous optimizations produced a valid graph
2304     // at least to this point, even if no loop optimizations were done.
2305     TracePhase tp("idealLoopVerify", &timers[_t_idealLoopVerify]);
2306     PhaseIdealLoop::verify(igvn);
2307   }
2308 
2309   if (range_check_cast_count() > 0) {
2310     // No more loop optimizations. Remove all range check dependent CastIINodes.
2311     C->remove_range_check_casts(igvn);
2312     igvn.optimize();
2313   }
2314 
2315   {
2316     TracePhase tp("macroExpand", &timers[_t_macroExpand]);
2317     PhaseMacroExpand  mex(igvn);
2318     if (mex.expand_macro_nodes()) {
2319       assert(failing(), "must bail out w/ explicit message");
2320       return;
2321     }
2322   }
2323 
2324   DEBUG_ONLY( _modified_nodes = NULL; )
2325  } // (End scope of igvn; run destructor if necessary for asserts.)
2326 
2327  process_print_inlining();
2328  // A method with only infinite loops has no edges entering loops from root
2329  {
2330    TracePhase tp("graphReshape", &timers[_t_graphReshaping]);
2331    if (final_graph_reshaping()) {
2332      assert(failing(), "must bail out w/ explicit message");
2333      return;
2334    }
2335  }
2336 
2337  print_method(PHASE_OPTIMIZE_FINISHED, 2);
2338 }
2339 
2340 
2341 //------------------------------Code_Gen---------------------------------------
2342 // Given a graph, generate code for it
2343 void Compile::Code_Gen() {
2344   if (failing()) {
2345     return;
2346   }
2347 
2348   // Perform instruction selection.  You might think we could reclaim Matcher
2349   // memory PDQ, but actually the Matcher is used in generating spill code.
2350   // Internals of the Matcher (including some VectorSets) must remain live
2351   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2352   // set a bit in reclaimed memory.
2353 
2354   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2355   // nodes.  Mapping is only valid at the root of each matched subtree.
2356   NOT_PRODUCT( verify_graph_edges(); )
2357 
2358   Matcher matcher;
2359   _matcher = &matcher;
2360   {
2361     TracePhase tp("matcher", &timers[_t_matcher]);
2362     matcher.match();
2363   }
2364   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2365   // nodes.  Mapping is only valid at the root of each matched subtree.
2366   NOT_PRODUCT( verify_graph_edges(); )
2367 
2368   // If you have too many nodes, or if matching has failed, bail out
2369   check_node_count(0, "out of nodes matching instructions");
2370   if (failing()) {
2371     return;
2372   }
2373 
2374   // Build a proper-looking CFG
2375   PhaseCFG cfg(node_arena(), root(), matcher);
2376   _cfg = &cfg;
2377   {
2378     TracePhase tp("scheduler", &timers[_t_scheduler]);
2379     bool success = cfg.do_global_code_motion();
2380     if (!success) {
2381       return;
2382     }
2383 
2384     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2385     NOT_PRODUCT( verify_graph_edges(); )
2386     debug_only( cfg.verify(); )
2387   }
2388 
2389   PhaseChaitin regalloc(unique(), cfg, matcher, false);
2390   _regalloc = &regalloc;
2391   {
2392     TracePhase tp("regalloc", &timers[_t_registerAllocation]);
2393     // Perform register allocation.  After Chaitin, use-def chains are
2394     // no longer accurate (at spill code) and so must be ignored.
2395     // Node->LRG->reg mappings are still accurate.
2396     _regalloc->Register_Allocate();
2397 
2398     // Bail out if the allocator builds too many nodes
2399     if (failing()) {
2400       return;
2401     }
2402   }
2403 
2404   // Prior to register allocation we kept empty basic blocks in case the
2405   // the allocator needed a place to spill.  After register allocation we
2406   // are not adding any new instructions.  If any basic block is empty, we
2407   // can now safely remove it.
2408   {
2409     TracePhase tp("blockOrdering", &timers[_t_blockOrdering]);
2410     cfg.remove_empty_blocks();
2411     if (do_freq_based_layout()) {
2412       PhaseBlockLayout layout(cfg);
2413     } else {
2414       cfg.set_loop_alignment();
2415     }
2416     cfg.fixup_flow();
2417   }
2418 
2419   // Apply peephole optimizations
2420   if( OptoPeephole ) {
2421     TracePhase tp("peephole", &timers[_t_peephole]);
2422     PhasePeephole peep( _regalloc, cfg);
2423     peep.do_transform();
2424   }
2425 
2426   // Do late expand if CPU requires this.
2427   if (Matcher::require_postalloc_expand) {
2428     TracePhase tp("postalloc_expand", &timers[_t_postalloc_expand]);
2429     cfg.postalloc_expand(_regalloc);
2430   }
2431 
2432   // Convert Nodes to instruction bits in a buffer
2433   {
2434     TraceTime tp("output", &timers[_t_output], CITime);
2435     Output();
2436   }
2437 
2438   print_method(PHASE_FINAL_CODE);
2439 
2440   // He's dead, Jim.
2441   _cfg     = (PhaseCFG*)0xdeadbeef;
2442   _regalloc = (PhaseChaitin*)0xdeadbeef;
2443 }
2444 
2445 
2446 //------------------------------dump_asm---------------------------------------
2447 // Dump formatted assembly
2448 #ifndef PRODUCT
2449 void Compile::dump_asm(int *pcs, uint pc_limit) {
2450   bool cut_short = false;
2451   tty->print_cr("#");
2452   tty->print("#  ");  _tf->dump();  tty->cr();
2453   tty->print_cr("#");
2454 
2455   // For all blocks
2456   int pc = 0x0;                 // Program counter
2457   char starts_bundle = ' ';
2458   _regalloc->dump_frame();
2459 
2460   Node *n = NULL;
2461   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2462     if (VMThread::should_terminate()) {
2463       cut_short = true;
2464       break;
2465     }
2466     Block* block = _cfg->get_block(i);
2467     if (block->is_connector() && !Verbose) {
2468       continue;
2469     }
2470     n = block->head();
2471     if (pcs && n->_idx < pc_limit) {
2472       tty->print("%3.3x   ", pcs[n->_idx]);
2473     } else {
2474       tty->print("      ");
2475     }
2476     block->dump_head(_cfg);
2477     if (block->is_connector()) {
2478       tty->print_cr("        # Empty connector block");
2479     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2480       tty->print_cr("        # Block is sole successor of call");
2481     }
2482 
2483     // For all instructions
2484     Node *delay = NULL;
2485     for (uint j = 0; j < block->number_of_nodes(); j++) {
2486       if (VMThread::should_terminate()) {
2487         cut_short = true;
2488         break;
2489       }
2490       n = block->get_node(j);
2491       if (valid_bundle_info(n)) {
2492         Bundle* bundle = node_bundling(n);
2493         if (bundle->used_in_unconditional_delay()) {
2494           delay = n;
2495           continue;
2496         }
2497         if (bundle->starts_bundle()) {
2498           starts_bundle = '+';
2499         }
2500       }
2501 
2502       if (WizardMode) {
2503         n->dump();
2504       }
2505 
2506       if( !n->is_Region() &&    // Dont print in the Assembly
2507           !n->is_Phi() &&       // a few noisely useless nodes
2508           !n->is_Proj() &&
2509           !n->is_MachTemp() &&
2510           !n->is_SafePointScalarObject() &&
2511           !n->is_Catch() &&     // Would be nice to print exception table targets
2512           !n->is_MergeMem() &&  // Not very interesting
2513           !n->is_top() &&       // Debug info table constants
2514           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2515           ) {
2516         if (pcs && n->_idx < pc_limit)
2517           tty->print("%3.3x", pcs[n->_idx]);
2518         else
2519           tty->print("   ");
2520         tty->print(" %c ", starts_bundle);
2521         starts_bundle = ' ';
2522         tty->print("\t");
2523         n->format(_regalloc, tty);
2524         tty->cr();
2525       }
2526 
2527       // If we have an instruction with a delay slot, and have seen a delay,
2528       // then back up and print it
2529       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2530         assert(delay != NULL, "no unconditional delay instruction");
2531         if (WizardMode) delay->dump();
2532 
2533         if (node_bundling(delay)->starts_bundle())
2534           starts_bundle = '+';
2535         if (pcs && n->_idx < pc_limit)
2536           tty->print("%3.3x", pcs[n->_idx]);
2537         else
2538           tty->print("   ");
2539         tty->print(" %c ", starts_bundle);
2540         starts_bundle = ' ';
2541         tty->print("\t");
2542         delay->format(_regalloc, tty);
2543         tty->cr();
2544         delay = NULL;
2545       }
2546 
2547       // Dump the exception table as well
2548       if( n->is_Catch() && (Verbose || WizardMode) ) {
2549         // Print the exception table for this offset
2550         _handler_table.print_subtable_for(pc);
2551       }
2552     }
2553 
2554     if (pcs && n->_idx < pc_limit)
2555       tty->print_cr("%3.3x", pcs[n->_idx]);
2556     else
2557       tty->cr();
2558 
2559     assert(cut_short || delay == NULL, "no unconditional delay branch");
2560 
2561   } // End of per-block dump
2562   tty->cr();
2563 
2564   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2565 }
2566 #endif
2567 
2568 //------------------------------Final_Reshape_Counts---------------------------
2569 // This class defines counters to help identify when a method
2570 // may/must be executed using hardware with only 24-bit precision.
2571 struct Final_Reshape_Counts : public StackObj {
2572   int  _call_count;             // count non-inlined 'common' calls
2573   int  _float_count;            // count float ops requiring 24-bit precision
2574   int  _double_count;           // count double ops requiring more precision
2575   int  _java_call_count;        // count non-inlined 'java' calls
2576   int  _inner_loop_count;       // count loops which need alignment
2577   VectorSet _visited;           // Visitation flags
2578   Node_List _tests;             // Set of IfNodes & PCTableNodes
2579 
2580   Final_Reshape_Counts() :
2581     _call_count(0), _float_count(0), _double_count(0),
2582     _java_call_count(0), _inner_loop_count(0),
2583     _visited( Thread::current()->resource_area() ) { }
2584 
2585   void inc_call_count  () { _call_count  ++; }
2586   void inc_float_count () { _float_count ++; }
2587   void inc_double_count() { _double_count++; }
2588   void inc_java_call_count() { _java_call_count++; }
2589   void inc_inner_loop_count() { _inner_loop_count++; }
2590 
2591   int  get_call_count  () const { return _call_count  ; }
2592   int  get_float_count () const { return _float_count ; }
2593   int  get_double_count() const { return _double_count; }
2594   int  get_java_call_count() const { return _java_call_count; }
2595   int  get_inner_loop_count() const { return _inner_loop_count; }
2596 };
2597 
2598 #ifdef ASSERT
2599 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2600   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2601   // Make sure the offset goes inside the instance layout.
2602   return k->contains_field_offset(tp->offset());
2603   // Note that OffsetBot and OffsetTop are very negative.
2604 }
2605 #endif
2606 
2607 // Eliminate trivially redundant StoreCMs and accumulate their
2608 // precedence edges.
2609 void Compile::eliminate_redundant_card_marks(Node* n) {
2610   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2611   if (n->in(MemNode::Address)->outcnt() > 1) {
2612     // There are multiple users of the same address so it might be
2613     // possible to eliminate some of the StoreCMs
2614     Node* mem = n->in(MemNode::Memory);
2615     Node* adr = n->in(MemNode::Address);
2616     Node* val = n->in(MemNode::ValueIn);
2617     Node* prev = n;
2618     bool done = false;
2619     // Walk the chain of StoreCMs eliminating ones that match.  As
2620     // long as it's a chain of single users then the optimization is
2621     // safe.  Eliminating partially redundant StoreCMs would require
2622     // cloning copies down the other paths.
2623     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2624       if (adr == mem->in(MemNode::Address) &&
2625           val == mem->in(MemNode::ValueIn)) {
2626         // redundant StoreCM
2627         if (mem->req() > MemNode::OopStore) {
2628           // Hasn't been processed by this code yet.
2629           n->add_prec(mem->in(MemNode::OopStore));
2630         } else {
2631           // Already converted to precedence edge
2632           for (uint i = mem->req(); i < mem->len(); i++) {
2633             // Accumulate any precedence edges
2634             if (mem->in(i) != NULL) {
2635               n->add_prec(mem->in(i));
2636             }
2637           }
2638           // Everything above this point has been processed.
2639           done = true;
2640         }
2641         // Eliminate the previous StoreCM
2642         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2643         assert(mem->outcnt() == 0, "should be dead");
2644         mem->disconnect_inputs(NULL, this);
2645       } else {
2646         prev = mem;
2647       }
2648       mem = prev->in(MemNode::Memory);
2649     }
2650   }
2651 }
2652 
2653 //------------------------------final_graph_reshaping_impl----------------------
2654 // Implement items 1-5 from final_graph_reshaping below.
2655 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2656 
2657   if ( n->outcnt() == 0 ) return; // dead node
2658   uint nop = n->Opcode();
2659 
2660   // Check for 2-input instruction with "last use" on right input.
2661   // Swap to left input.  Implements item (2).
2662   if( n->req() == 3 &&          // two-input instruction
2663       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2664       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2665       n->in(2)->outcnt() == 1 &&// right use IS a last use
2666       !n->in(2)->is_Con() ) {   // right use is not a constant
2667     // Check for commutative opcode
2668     switch( nop ) {
2669     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2670     case Op_MaxI:  case Op_MinI:
2671     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2672     case Op_AndL:  case Op_XorL:  case Op_OrL:
2673     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2674       // Move "last use" input to left by swapping inputs
2675       n->swap_edges(1, 2);
2676       break;
2677     }
2678     default:
2679       break;
2680     }
2681   }
2682 
2683 #ifdef ASSERT
2684   if( n->is_Mem() ) {
2685     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2686     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2687             // oop will be recorded in oop map if load crosses safepoint
2688             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2689                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2690             "raw memory operations should have control edge");
2691   }
2692 #endif
2693   // Count FPU ops and common calls, implements item (3)
2694   switch( nop ) {
2695   // Count all float operations that may use FPU
2696   case Op_AddF:
2697   case Op_SubF:
2698   case Op_MulF:
2699   case Op_DivF:
2700   case Op_NegF:
2701   case Op_ModF:
2702   case Op_ConvI2F:
2703   case Op_ConF:
2704   case Op_CmpF:
2705   case Op_CmpF3:
2706   // case Op_ConvL2F: // longs are split into 32-bit halves
2707     frc.inc_float_count();
2708     break;
2709 
2710   case Op_ConvF2D:
2711   case Op_ConvD2F:
2712     frc.inc_float_count();
2713     frc.inc_double_count();
2714     break;
2715 
2716   // Count all double operations that may use FPU
2717   case Op_AddD:
2718   case Op_SubD:
2719   case Op_MulD:
2720   case Op_DivD:
2721   case Op_NegD:
2722   case Op_ModD:
2723   case Op_ConvI2D:
2724   case Op_ConvD2I:
2725   // case Op_ConvL2D: // handled by leaf call
2726   // case Op_ConvD2L: // handled by leaf call
2727   case Op_ConD:
2728   case Op_CmpD:
2729   case Op_CmpD3:
2730     frc.inc_double_count();
2731     break;
2732   case Op_Opaque1:              // Remove Opaque Nodes before matching
2733   case Op_Opaque2:              // Remove Opaque Nodes before matching
2734   case Op_Opaque3:
2735     n->subsume_by(n->in(1), this);
2736     break;
2737   case Op_CallStaticJava:
2738   case Op_CallJava:
2739   case Op_CallDynamicJava:
2740     frc.inc_java_call_count(); // Count java call site;
2741   case Op_CallRuntime:
2742   case Op_CallLeaf:
2743   case Op_CallLeafNoFP: {
2744     assert( n->is_Call(), "" );
2745     CallNode *call = n->as_Call();
2746     // Count call sites where the FP mode bit would have to be flipped.
2747     // Do not count uncommon runtime calls:
2748     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2749     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2750     if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2751       frc.inc_call_count();   // Count the call site
2752     } else {                  // See if uncommon argument is shared
2753       Node *n = call->in(TypeFunc::Parms);
2754       int nop = n->Opcode();
2755       // Clone shared simple arguments to uncommon calls, item (1).
2756       if( n->outcnt() > 1 &&
2757           !n->is_Proj() &&
2758           nop != Op_CreateEx &&
2759           nop != Op_CheckCastPP &&
2760           nop != Op_DecodeN &&
2761           nop != Op_DecodeNKlass &&
2762           !n->is_Mem() ) {
2763         Node *x = n->clone();
2764         call->set_req( TypeFunc::Parms, x );
2765       }
2766     }
2767     break;
2768   }
2769 
2770   case Op_StoreD:
2771   case Op_LoadD:
2772   case Op_LoadD_unaligned:
2773     frc.inc_double_count();
2774     goto handle_mem;
2775   case Op_StoreF:
2776   case Op_LoadF:
2777     frc.inc_float_count();
2778     goto handle_mem;
2779 
2780   case Op_StoreCM:
2781     {
2782       // Convert OopStore dependence into precedence edge
2783       Node* prec = n->in(MemNode::OopStore);
2784       n->del_req(MemNode::OopStore);
2785       n->add_prec(prec);
2786       eliminate_redundant_card_marks(n);
2787     }
2788 
2789     // fall through
2790 
2791   case Op_StoreB:
2792   case Op_StoreC:
2793   case Op_StorePConditional:
2794   case Op_StoreI:
2795   case Op_StoreL:
2796   case Op_StoreIConditional:
2797   case Op_StoreLConditional:
2798   case Op_CompareAndSwapI:
2799   case Op_CompareAndSwapL:
2800   case Op_CompareAndSwapP:
2801   case Op_CompareAndSwapN:
2802   case Op_GetAndAddI:
2803   case Op_GetAndAddL:
2804   case Op_GetAndSetI:
2805   case Op_GetAndSetL:
2806   case Op_GetAndSetP:
2807   case Op_GetAndSetN:
2808   case Op_StoreP:
2809   case Op_StoreN:
2810   case Op_StoreNKlass:
2811   case Op_LoadB:
2812   case Op_LoadUB:
2813   case Op_LoadUS:
2814   case Op_LoadI:
2815   case Op_LoadKlass:
2816   case Op_LoadNKlass:
2817   case Op_LoadL:
2818   case Op_LoadL_unaligned:
2819   case Op_LoadPLocked:
2820   case Op_LoadP:
2821   case Op_LoadN:
2822   case Op_LoadRange:
2823   case Op_LoadS: {
2824   handle_mem:
2825 #ifdef ASSERT
2826     if( VerifyOptoOopOffsets ) {
2827       assert( n->is_Mem(), "" );
2828       MemNode *mem  = (MemNode*)n;
2829       // Check to see if address types have grounded out somehow.
2830       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2831       assert( !tp || oop_offset_is_sane(tp), "" );
2832     }
2833 #endif
2834     break;
2835   }
2836 
2837   case Op_AddP: {               // Assert sane base pointers
2838     Node *addp = n->in(AddPNode::Address);
2839     assert( !addp->is_AddP() ||
2840             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2841             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2842             "Base pointers must match" );
2843 #ifdef _LP64
2844     if ((UseCompressedOops || UseCompressedClassPointers) &&
2845         addp->Opcode() == Op_ConP &&
2846         addp == n->in(AddPNode::Base) &&
2847         n->in(AddPNode::Offset)->is_Con()) {
2848       // Use addressing with narrow klass to load with offset on x86.
2849       // On sparc loading 32-bits constant and decoding it have less
2850       // instructions (4) then load 64-bits constant (7).
2851       // Do this transformation here since IGVN will convert ConN back to ConP.
2852       const Type* t = addp->bottom_type();
2853       if (t->isa_oopptr() || t->isa_klassptr()) {
2854         Node* nn = NULL;
2855 
2856         int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2857 
2858         // Look for existing ConN node of the same exact type.
2859         Node* r  = root();
2860         uint cnt = r->outcnt();
2861         for (uint i = 0; i < cnt; i++) {
2862           Node* m = r->raw_out(i);
2863           if (m!= NULL && m->Opcode() == op &&
2864               m->bottom_type()->make_ptr() == t) {
2865             nn = m;
2866             break;
2867           }
2868         }
2869         if (nn != NULL) {
2870           // Decode a narrow oop to match address
2871           // [R12 + narrow_oop_reg<<3 + offset]
2872           if (t->isa_oopptr()) {
2873             nn = new DecodeNNode(nn, t);
2874           } else {
2875             nn = new DecodeNKlassNode(nn, t);
2876           }
2877           n->set_req(AddPNode::Base, nn);
2878           n->set_req(AddPNode::Address, nn);
2879           if (addp->outcnt() == 0) {
2880             addp->disconnect_inputs(NULL, this);
2881           }
2882         }
2883       }
2884     }
2885 #endif
2886     break;
2887   }
2888 
2889   case Op_CastPP: {
2890     // Remove CastPP nodes to gain more freedom during scheduling but
2891     // keep the dependency they encode as control or precedence edges
2892     // (if control is set already) on memory operations. Some CastPP
2893     // nodes don't have a control (don't carry a dependency): skip
2894     // those.
2895     if (n->in(0) != NULL) {
2896       ResourceMark rm;
2897       Unique_Node_List wq;
2898       wq.push(n);
2899       for (uint next = 0; next < wq.size(); ++next) {
2900         Node *m = wq.at(next);
2901         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2902           Node* use = m->fast_out(i);
2903           if (use->is_Mem() || use->is_EncodeNarrowPtr()) {
2904             use->ensure_control_or_add_prec(n->in(0));
2905           } else {
2906             switch(use->Opcode()) {
2907             case Op_AddP:
2908             case Op_DecodeN:
2909             case Op_DecodeNKlass:
2910             case Op_CheckCastPP:
2911             case Op_CastPP:
2912               wq.push(use);
2913               break;
2914             }
2915           }
2916         }
2917       }
2918     }
2919     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
2920     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2921       Node* in1 = n->in(1);
2922       const Type* t = n->bottom_type();
2923       Node* new_in1 = in1->clone();
2924       new_in1->as_DecodeN()->set_type(t);
2925 
2926       if (!Matcher::narrow_oop_use_complex_address()) {
2927         //
2928         // x86, ARM and friends can handle 2 adds in addressing mode
2929         // and Matcher can fold a DecodeN node into address by using
2930         // a narrow oop directly and do implicit NULL check in address:
2931         //
2932         // [R12 + narrow_oop_reg<<3 + offset]
2933         // NullCheck narrow_oop_reg
2934         //
2935         // On other platforms (Sparc) we have to keep new DecodeN node and
2936         // use it to do implicit NULL check in address:
2937         //
2938         // decode_not_null narrow_oop_reg, base_reg
2939         // [base_reg + offset]
2940         // NullCheck base_reg
2941         //
2942         // Pin the new DecodeN node to non-null path on these platform (Sparc)
2943         // to keep the information to which NULL check the new DecodeN node
2944         // corresponds to use it as value in implicit_null_check().
2945         //
2946         new_in1->set_req(0, n->in(0));
2947       }
2948 
2949       n->subsume_by(new_in1, this);
2950       if (in1->outcnt() == 0) {
2951         in1->disconnect_inputs(NULL, this);
2952       }
2953     } else {
2954       n->subsume_by(n->in(1), this);
2955       if (n->outcnt() == 0) {
2956         n->disconnect_inputs(NULL, this);
2957       }
2958     }
2959     break;
2960   }
2961 #ifdef _LP64
2962   case Op_CmpP:
2963     // Do this transformation here to preserve CmpPNode::sub() and
2964     // other TypePtr related Ideal optimizations (for example, ptr nullness).
2965     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2966       Node* in1 = n->in(1);
2967       Node* in2 = n->in(2);
2968       if (!in1->is_DecodeNarrowPtr()) {
2969         in2 = in1;
2970         in1 = n->in(2);
2971       }
2972       assert(in1->is_DecodeNarrowPtr(), "sanity");
2973 
2974       Node* new_in2 = NULL;
2975       if (in2->is_DecodeNarrowPtr()) {
2976         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2977         new_in2 = in2->in(1);
2978       } else if (in2->Opcode() == Op_ConP) {
2979         const Type* t = in2->bottom_type();
2980         if (t == TypePtr::NULL_PTR) {
2981           assert(in1->is_DecodeN(), "compare klass to null?");
2982           // Don't convert CmpP null check into CmpN if compressed
2983           // oops implicit null check is not generated.
2984           // This will allow to generate normal oop implicit null check.
2985           if (Matcher::gen_narrow_oop_implicit_null_checks())
2986             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
2987           //
2988           // This transformation together with CastPP transformation above
2989           // will generated code for implicit NULL checks for compressed oops.
2990           //
2991           // The original code after Optimize()
2992           //
2993           //    LoadN memory, narrow_oop_reg
2994           //    decode narrow_oop_reg, base_reg
2995           //    CmpP base_reg, NULL
2996           //    CastPP base_reg // NotNull
2997           //    Load [base_reg + offset], val_reg
2998           //
2999           // after these transformations will be
3000           //
3001           //    LoadN memory, narrow_oop_reg
3002           //    CmpN narrow_oop_reg, NULL
3003           //    decode_not_null narrow_oop_reg, base_reg
3004           //    Load [base_reg + offset], val_reg
3005           //
3006           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3007           // since narrow oops can be used in debug info now (see the code in
3008           // final_graph_reshaping_walk()).
3009           //
3010           // At the end the code will be matched to
3011           // on x86:
3012           //
3013           //    Load_narrow_oop memory, narrow_oop_reg
3014           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3015           //    NullCheck narrow_oop_reg
3016           //
3017           // and on sparc:
3018           //
3019           //    Load_narrow_oop memory, narrow_oop_reg
3020           //    decode_not_null narrow_oop_reg, base_reg
3021           //    Load [base_reg + offset], val_reg
3022           //    NullCheck base_reg
3023           //
3024         } else if (t->isa_oopptr()) {
3025           new_in2 = ConNode::make(t->make_narrowoop());
3026         } else if (t->isa_klassptr()) {
3027           new_in2 = ConNode::make(t->make_narrowklass());
3028         }
3029       }
3030       if (new_in2 != NULL) {
3031         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
3032         n->subsume_by(cmpN, this);
3033         if (in1->outcnt() == 0) {
3034           in1->disconnect_inputs(NULL, this);
3035         }
3036         if (in2->outcnt() == 0) {
3037           in2->disconnect_inputs(NULL, this);
3038         }
3039       }
3040     }
3041     break;
3042 
3043   case Op_DecodeN:
3044   case Op_DecodeNKlass:
3045     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3046     // DecodeN could be pinned when it can't be fold into
3047     // an address expression, see the code for Op_CastPP above.
3048     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3049     break;
3050 
3051   case Op_EncodeP:
3052   case Op_EncodePKlass: {
3053     Node* in1 = n->in(1);
3054     if (in1->is_DecodeNarrowPtr()) {
3055       n->subsume_by(in1->in(1), this);
3056     } else if (in1->Opcode() == Op_ConP) {
3057       const Type* t = in1->bottom_type();
3058       if (t == TypePtr::NULL_PTR) {
3059         assert(t->isa_oopptr(), "null klass?");
3060         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
3061       } else if (t->isa_oopptr()) {
3062         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
3063       } else if (t->isa_klassptr()) {
3064         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
3065       }
3066     }
3067     if (in1->outcnt() == 0) {
3068       in1->disconnect_inputs(NULL, this);
3069     }
3070     break;
3071   }
3072 
3073   case Op_Proj: {
3074     if (OptimizeStringConcat) {
3075       ProjNode* p = n->as_Proj();
3076       if (p->_is_io_use) {
3077         // Separate projections were used for the exception path which
3078         // are normally removed by a late inline.  If it wasn't inlined
3079         // then they will hang around and should just be replaced with
3080         // the original one.
3081         Node* proj = NULL;
3082         // Replace with just one
3083         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3084           Node *use = i.get();
3085           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3086             proj = use;
3087             break;
3088           }
3089         }
3090         assert(proj != NULL, "must be found");
3091         p->subsume_by(proj, this);
3092       }
3093     }
3094     break;
3095   }
3096 
3097   case Op_Phi:
3098     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3099       // The EncodeP optimization may create Phi with the same edges
3100       // for all paths. It is not handled well by Register Allocator.
3101       Node* unique_in = n->in(1);
3102       assert(unique_in != NULL, "");
3103       uint cnt = n->req();
3104       for (uint i = 2; i < cnt; i++) {
3105         Node* m = n->in(i);
3106         assert(m != NULL, "");
3107         if (unique_in != m)
3108           unique_in = NULL;
3109       }
3110       if (unique_in != NULL) {
3111         n->subsume_by(unique_in, this);
3112       }
3113     }
3114     break;
3115 
3116 #endif
3117 
3118 #ifdef ASSERT
3119   case Op_CastII:
3120     // Verify that all range check dependent CastII nodes were removed.
3121     if (n->isa_CastII()->has_range_check()) {
3122       n->dump(3);
3123       assert(false, "Range check dependent CastII node was not removed");
3124     }
3125     break;
3126 #endif
3127 
3128   case Op_ModI:
3129     if (UseDivMod) {
3130       // Check if a%b and a/b both exist
3131       Node* d = n->find_similar(Op_DivI);
3132       if (d) {
3133         // Replace them with a fused divmod if supported
3134         if (Matcher::has_match_rule(Op_DivModI)) {
3135           DivModINode* divmod = DivModINode::make(n);
3136           d->subsume_by(divmod->div_proj(), this);
3137           n->subsume_by(divmod->mod_proj(), this);
3138         } else {
3139           // replace a%b with a-((a/b)*b)
3140           Node* mult = new MulINode(d, d->in(2));
3141           Node* sub  = new SubINode(d->in(1), mult);
3142           n->subsume_by(sub, this);
3143         }
3144       }
3145     }
3146     break;
3147 
3148   case Op_ModL:
3149     if (UseDivMod) {
3150       // Check if a%b and a/b both exist
3151       Node* d = n->find_similar(Op_DivL);
3152       if (d) {
3153         // Replace them with a fused divmod if supported
3154         if (Matcher::has_match_rule(Op_DivModL)) {
3155           DivModLNode* divmod = DivModLNode::make(n);
3156           d->subsume_by(divmod->div_proj(), this);
3157           n->subsume_by(divmod->mod_proj(), this);
3158         } else {
3159           // replace a%b with a-((a/b)*b)
3160           Node* mult = new MulLNode(d, d->in(2));
3161           Node* sub  = new SubLNode(d->in(1), mult);
3162           n->subsume_by(sub, this);
3163         }
3164       }
3165     }
3166     break;
3167 
3168   case Op_LoadVector:
3169   case Op_StoreVector:
3170     break;
3171 
3172   case Op_AddReductionVI:
3173   case Op_AddReductionVL:
3174   case Op_AddReductionVF:
3175   case Op_AddReductionVD:
3176   case Op_MulReductionVI:
3177   case Op_MulReductionVL:
3178   case Op_MulReductionVF:
3179   case Op_MulReductionVD:
3180     break;
3181 
3182   case Op_PackB:
3183   case Op_PackS:
3184   case Op_PackI:
3185   case Op_PackF:
3186   case Op_PackL:
3187   case Op_PackD:
3188     if (n->req()-1 > 2) {
3189       // Replace many operand PackNodes with a binary tree for matching
3190       PackNode* p = (PackNode*) n;
3191       Node* btp = p->binary_tree_pack(1, n->req());
3192       n->subsume_by(btp, this);
3193     }
3194     break;
3195   case Op_Loop:
3196   case Op_CountedLoop:
3197     if (n->as_Loop()->is_inner_loop()) {
3198       frc.inc_inner_loop_count();
3199     }
3200     break;
3201   case Op_LShiftI:
3202   case Op_RShiftI:
3203   case Op_URShiftI:
3204   case Op_LShiftL:
3205   case Op_RShiftL:
3206   case Op_URShiftL:
3207     if (Matcher::need_masked_shift_count) {
3208       // The cpu's shift instructions don't restrict the count to the
3209       // lower 5/6 bits. We need to do the masking ourselves.
3210       Node* in2 = n->in(2);
3211       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3212       const TypeInt* t = in2->find_int_type();
3213       if (t != NULL && t->is_con()) {
3214         juint shift = t->get_con();
3215         if (shift > mask) { // Unsigned cmp
3216           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3217         }
3218       } else {
3219         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3220           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3221           n->set_req(2, shift);
3222         }
3223       }
3224       if (in2->outcnt() == 0) { // Remove dead node
3225         in2->disconnect_inputs(NULL, this);
3226       }
3227     }
3228     break;
3229   case Op_MemBarStoreStore:
3230   case Op_MemBarRelease:
3231     // Break the link with AllocateNode: it is no longer useful and
3232     // confuses register allocation.
3233     if (n->req() > MemBarNode::Precedent) {
3234       n->set_req(MemBarNode::Precedent, top());
3235     }
3236     break;
3237   case Op_RangeCheck: {
3238     RangeCheckNode* rc = n->as_RangeCheck();
3239     Node* iff = new IfNode(rc->in(0), rc->in(1), rc->_prob, rc->_fcnt);
3240     n->subsume_by(iff, this);
3241     frc._tests.push(iff);
3242     break;
3243   }
3244   default:
3245     assert( !n->is_Call(), "" );
3246     assert( !n->is_Mem(), "" );
3247     assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3248     break;
3249   }
3250 
3251   // Collect CFG split points
3252   if (n->is_MultiBranch() && !n->is_RangeCheck()) {
3253     frc._tests.push(n);
3254   }
3255 }
3256 
3257 //------------------------------final_graph_reshaping_walk---------------------
3258 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3259 // requires that the walk visits a node's inputs before visiting the node.
3260 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3261   ResourceArea *area = Thread::current()->resource_area();
3262   Unique_Node_List sfpt(area);
3263 
3264   frc._visited.set(root->_idx); // first, mark node as visited
3265   uint cnt = root->req();
3266   Node *n = root;
3267   uint  i = 0;
3268   while (true) {
3269     if (i < cnt) {
3270       // Place all non-visited non-null inputs onto stack
3271       Node* m = n->in(i);
3272       ++i;
3273       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3274         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3275           // compute worst case interpreter size in case of a deoptimization
3276           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3277 
3278           sfpt.push(m);
3279         }
3280         cnt = m->req();
3281         nstack.push(n, i); // put on stack parent and next input's index
3282         n = m;
3283         i = 0;
3284       }
3285     } else {
3286       // Now do post-visit work
3287       final_graph_reshaping_impl( n, frc );
3288       if (nstack.is_empty())
3289         break;             // finished
3290       n = nstack.node();   // Get node from stack
3291       cnt = n->req();
3292       i = nstack.index();
3293       nstack.pop();        // Shift to the next node on stack
3294     }
3295   }
3296 
3297   // Skip next transformation if compressed oops are not used.
3298   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3299       (!UseCompressedOops && !UseCompressedClassPointers))
3300     return;
3301 
3302   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3303   // It could be done for an uncommon traps or any safepoints/calls
3304   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3305   while (sfpt.size() > 0) {
3306     n = sfpt.pop();
3307     JVMState *jvms = n->as_SafePoint()->jvms();
3308     assert(jvms != NULL, "sanity");
3309     int start = jvms->debug_start();
3310     int end   = n->req();
3311     bool is_uncommon = (n->is_CallStaticJava() &&
3312                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3313     for (int j = start; j < end; j++) {
3314       Node* in = n->in(j);
3315       if (in->is_DecodeNarrowPtr()) {
3316         bool safe_to_skip = true;
3317         if (!is_uncommon ) {
3318           // Is it safe to skip?
3319           for (uint i = 0; i < in->outcnt(); i++) {
3320             Node* u = in->raw_out(i);
3321             if (!u->is_SafePoint() ||
3322                  u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3323               safe_to_skip = false;
3324             }
3325           }
3326         }
3327         if (safe_to_skip) {
3328           n->set_req(j, in->in(1));
3329         }
3330         if (in->outcnt() == 0) {
3331           in->disconnect_inputs(NULL, this);
3332         }
3333       }
3334     }
3335   }
3336 }
3337 
3338 //------------------------------final_graph_reshaping--------------------------
3339 // Final Graph Reshaping.
3340 //
3341 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3342 //     and not commoned up and forced early.  Must come after regular
3343 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3344 //     inputs to Loop Phis; these will be split by the allocator anyways.
3345 //     Remove Opaque nodes.
3346 // (2) Move last-uses by commutative operations to the left input to encourage
3347 //     Intel update-in-place two-address operations and better register usage
3348 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3349 //     calls canonicalizing them back.
3350 // (3) Count the number of double-precision FP ops, single-precision FP ops
3351 //     and call sites.  On Intel, we can get correct rounding either by
3352 //     forcing singles to memory (requires extra stores and loads after each
3353 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3354 //     clearing the mode bit around call sites).  The mode bit is only used
3355 //     if the relative frequency of single FP ops to calls is low enough.
3356 //     This is a key transform for SPEC mpeg_audio.
3357 // (4) Detect infinite loops; blobs of code reachable from above but not
3358 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3359 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3360 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3361 //     Detection is by looking for IfNodes where only 1 projection is
3362 //     reachable from below or CatchNodes missing some targets.
3363 // (5) Assert for insane oop offsets in debug mode.
3364 
3365 bool Compile::final_graph_reshaping() {
3366   // an infinite loop may have been eliminated by the optimizer,
3367   // in which case the graph will be empty.
3368   if (root()->req() == 1) {
3369     record_method_not_compilable("trivial infinite loop");
3370     return true;
3371   }
3372 
3373   // Expensive nodes have their control input set to prevent the GVN
3374   // from freely commoning them. There's no GVN beyond this point so
3375   // no need to keep the control input. We want the expensive nodes to
3376   // be freely moved to the least frequent code path by gcm.
3377   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3378   for (int i = 0; i < expensive_count(); i++) {
3379     _expensive_nodes->at(i)->set_req(0, NULL);
3380   }
3381 
3382   Final_Reshape_Counts frc;
3383 
3384   // Visit everybody reachable!
3385   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3386   Node_Stack nstack(live_nodes() >> 1);
3387   final_graph_reshaping_walk(nstack, root(), frc);
3388 
3389   // Check for unreachable (from below) code (i.e., infinite loops).
3390   for( uint i = 0; i < frc._tests.size(); i++ ) {
3391     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3392     // Get number of CFG targets.
3393     // Note that PCTables include exception targets after calls.
3394     uint required_outcnt = n->required_outcnt();
3395     if (n->outcnt() != required_outcnt) {
3396       // Check for a few special cases.  Rethrow Nodes never take the
3397       // 'fall-thru' path, so expected kids is 1 less.
3398       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3399         if (n->in(0)->in(0)->is_Call()) {
3400           CallNode *call = n->in(0)->in(0)->as_Call();
3401           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3402             required_outcnt--;      // Rethrow always has 1 less kid
3403           } else if (call->req() > TypeFunc::Parms &&
3404                      call->is_CallDynamicJava()) {
3405             // Check for null receiver. In such case, the optimizer has
3406             // detected that the virtual call will always result in a null
3407             // pointer exception. The fall-through projection of this CatchNode
3408             // will not be populated.
3409             Node *arg0 = call->in(TypeFunc::Parms);
3410             if (arg0->is_Type() &&
3411                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3412               required_outcnt--;
3413             }
3414           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3415                      call->req() > TypeFunc::Parms+1 &&
3416                      call->is_CallStaticJava()) {
3417             // Check for negative array length. In such case, the optimizer has
3418             // detected that the allocation attempt will always result in an
3419             // exception. There is no fall-through projection of this CatchNode .
3420             Node *arg1 = call->in(TypeFunc::Parms+1);
3421             if (arg1->is_Type() &&
3422                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3423               required_outcnt--;
3424             }
3425           }
3426         }
3427       }
3428       // Recheck with a better notion of 'required_outcnt'
3429       if (n->outcnt() != required_outcnt) {
3430         record_method_not_compilable("malformed control flow");
3431         return true;            // Not all targets reachable!
3432       }
3433     }
3434     // Check that I actually visited all kids.  Unreached kids
3435     // must be infinite loops.
3436     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3437       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3438         record_method_not_compilable("infinite loop");
3439         return true;            // Found unvisited kid; must be unreach
3440       }
3441   }
3442 
3443   // If original bytecodes contained a mixture of floats and doubles
3444   // check if the optimizer has made it homogenous, item (3).
3445   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3446       frc.get_float_count() > 32 &&
3447       frc.get_double_count() == 0 &&
3448       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3449     set_24_bit_selection_and_mode( false,  true );
3450   }
3451 
3452   set_java_calls(frc.get_java_call_count());
3453   set_inner_loops(frc.get_inner_loop_count());
3454 
3455   // No infinite loops, no reason to bail out.
3456   return false;
3457 }
3458 
3459 //-----------------------------too_many_traps----------------------------------
3460 // Report if there are too many traps at the current method and bci.
3461 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3462 bool Compile::too_many_traps(ciMethod* method,
3463                              int bci,
3464                              Deoptimization::DeoptReason reason) {
3465   ciMethodData* md = method->method_data();
3466   if (md->is_empty()) {
3467     // Assume the trap has not occurred, or that it occurred only
3468     // because of a transient condition during start-up in the interpreter.
3469     return false;
3470   }
3471   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3472   if (md->has_trap_at(bci, m, reason) != 0) {
3473     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3474     // Also, if there are multiple reasons, or if there is no per-BCI record,
3475     // assume the worst.
3476     if (log())
3477       log()->elem("observe trap='%s' count='%d'",
3478                   Deoptimization::trap_reason_name(reason),
3479                   md->trap_count(reason));
3480     return true;
3481   } else {
3482     // Ignore method/bci and see if there have been too many globally.
3483     return too_many_traps(reason, md);
3484   }
3485 }
3486 
3487 // Less-accurate variant which does not require a method and bci.
3488 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3489                              ciMethodData* logmd) {
3490   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3491     // Too many traps globally.
3492     // Note that we use cumulative trap_count, not just md->trap_count.
3493     if (log()) {
3494       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3495       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3496                   Deoptimization::trap_reason_name(reason),
3497                   mcount, trap_count(reason));
3498     }
3499     return true;
3500   } else {
3501     // The coast is clear.
3502     return false;
3503   }
3504 }
3505 
3506 //--------------------------too_many_recompiles--------------------------------
3507 // Report if there are too many recompiles at the current method and bci.
3508 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3509 // Is not eager to return true, since this will cause the compiler to use
3510 // Action_none for a trap point, to avoid too many recompilations.
3511 bool Compile::too_many_recompiles(ciMethod* method,
3512                                   int bci,
3513                                   Deoptimization::DeoptReason reason) {
3514   ciMethodData* md = method->method_data();
3515   if (md->is_empty()) {
3516     // Assume the trap has not occurred, or that it occurred only
3517     // because of a transient condition during start-up in the interpreter.
3518     return false;
3519   }
3520   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3521   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3522   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3523   Deoptimization::DeoptReason per_bc_reason
3524     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3525   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3526   if ((per_bc_reason == Deoptimization::Reason_none
3527        || md->has_trap_at(bci, m, reason) != 0)
3528       // The trap frequency measure we care about is the recompile count:
3529       && md->trap_recompiled_at(bci, m)
3530       && md->overflow_recompile_count() >= bc_cutoff) {
3531     // Do not emit a trap here if it has already caused recompilations.
3532     // Also, if there are multiple reasons, or if there is no per-BCI record,
3533     // assume the worst.
3534     if (log())
3535       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3536                   Deoptimization::trap_reason_name(reason),
3537                   md->trap_count(reason),
3538                   md->overflow_recompile_count());
3539     return true;
3540   } else if (trap_count(reason) != 0
3541              && decompile_count() >= m_cutoff) {
3542     // Too many recompiles globally, and we have seen this sort of trap.
3543     // Use cumulative decompile_count, not just md->decompile_count.
3544     if (log())
3545       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3546                   Deoptimization::trap_reason_name(reason),
3547                   md->trap_count(reason), trap_count(reason),
3548                   md->decompile_count(), decompile_count());
3549     return true;
3550   } else {
3551     // The coast is clear.
3552     return false;
3553   }
3554 }
3555 
3556 // Compute when not to trap. Used by matching trap based nodes and
3557 // NullCheck optimization.
3558 void Compile::set_allowed_deopt_reasons() {
3559   _allowed_reasons = 0;
3560   if (is_method_compilation()) {
3561     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3562       assert(rs < BitsPerInt, "recode bit map");
3563       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3564         _allowed_reasons |= nth_bit(rs);
3565       }
3566     }
3567   }
3568 }
3569 
3570 #ifndef PRODUCT
3571 //------------------------------verify_graph_edges---------------------------
3572 // Walk the Graph and verify that there is a one-to-one correspondence
3573 // between Use-Def edges and Def-Use edges in the graph.
3574 void Compile::verify_graph_edges(bool no_dead_code) {
3575   if (VerifyGraphEdges) {
3576     ResourceArea *area = Thread::current()->resource_area();
3577     Unique_Node_List visited(area);
3578     // Call recursive graph walk to check edges
3579     _root->verify_edges(visited);
3580     if (no_dead_code) {
3581       // Now make sure that no visited node is used by an unvisited node.
3582       bool dead_nodes = false;
3583       Unique_Node_List checked(area);
3584       while (visited.size() > 0) {
3585         Node* n = visited.pop();
3586         checked.push(n);
3587         for (uint i = 0; i < n->outcnt(); i++) {
3588           Node* use = n->raw_out(i);
3589           if (checked.member(use))  continue;  // already checked
3590           if (visited.member(use))  continue;  // already in the graph
3591           if (use->is_Con())        continue;  // a dead ConNode is OK
3592           // At this point, we have found a dead node which is DU-reachable.
3593           if (!dead_nodes) {
3594             tty->print_cr("*** Dead nodes reachable via DU edges:");
3595             dead_nodes = true;
3596           }
3597           use->dump(2);
3598           tty->print_cr("---");
3599           checked.push(use);  // No repeats; pretend it is now checked.
3600         }
3601       }
3602       assert(!dead_nodes, "using nodes must be reachable from root");
3603     }
3604   }
3605 }
3606 
3607 // Verify GC barriers consistency
3608 // Currently supported:
3609 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3610 void Compile::verify_barriers() {
3611   if (UseG1GC) {
3612     // Verify G1 pre-barriers
3613     const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + SATBMarkQueue::byte_offset_of_active());
3614 
3615     ResourceArea *area = Thread::current()->resource_area();
3616     Unique_Node_List visited(area);
3617     Node_List worklist(area);
3618     // We're going to walk control flow backwards starting from the Root
3619     worklist.push(_root);
3620     while (worklist.size() > 0) {
3621       Node* x = worklist.pop();
3622       if (x == NULL || x == top()) continue;
3623       if (visited.member(x)) {
3624         continue;
3625       } else {
3626         visited.push(x);
3627       }
3628 
3629       if (x->is_Region()) {
3630         for (uint i = 1; i < x->req(); i++) {
3631           worklist.push(x->in(i));
3632         }
3633       } else {
3634         worklist.push(x->in(0));
3635         // We are looking for the pattern:
3636         //                            /->ThreadLocal
3637         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3638         //              \->ConI(0)
3639         // We want to verify that the If and the LoadB have the same control
3640         // See GraphKit::g1_write_barrier_pre()
3641         if (x->is_If()) {
3642           IfNode *iff = x->as_If();
3643           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3644             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3645             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3646                 && cmp->in(1)->is_Load()) {
3647               LoadNode* load = cmp->in(1)->as_Load();
3648               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3649                   && load->in(2)->in(3)->is_Con()
3650                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3651 
3652                 Node* if_ctrl = iff->in(0);
3653                 Node* load_ctrl = load->in(0);
3654 
3655                 if (if_ctrl != load_ctrl) {
3656                   // Skip possible CProj->NeverBranch in infinite loops
3657                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3658                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3659                     if_ctrl = if_ctrl->in(0)->in(0);
3660                   }
3661                 }
3662                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3663               }
3664             }
3665           }
3666         }
3667       }
3668     }
3669   }
3670 }
3671 
3672 #endif
3673 
3674 // The Compile object keeps track of failure reasons separately from the ciEnv.
3675 // This is required because there is not quite a 1-1 relation between the
3676 // ciEnv and its compilation task and the Compile object.  Note that one
3677 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3678 // to backtrack and retry without subsuming loads.  Other than this backtracking
3679 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3680 // by the logic in C2Compiler.
3681 void Compile::record_failure(const char* reason) {
3682   if (log() != NULL) {
3683     log()->elem("failure reason='%s' phase='compile'", reason);
3684   }
3685   if (_failure_reason == NULL) {
3686     // Record the first failure reason.
3687     _failure_reason = reason;
3688   }
3689 
3690   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3691     C->print_method(PHASE_FAILURE);
3692   }
3693   _root = NULL;  // flush the graph, too
3694 }
3695 
3696 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3697   : TraceTime(name, accumulator, CITime, CITimeVerbose),
3698     _phase_name(name), _dolog(CITimeVerbose)
3699 {
3700   if (_dolog) {
3701     C = Compile::current();
3702     _log = C->log();
3703   } else {
3704     C = NULL;
3705     _log = NULL;
3706   }
3707   if (_log != NULL) {
3708     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3709     _log->stamp();
3710     _log->end_head();
3711   }
3712 }
3713 
3714 Compile::TracePhase::~TracePhase() {
3715 
3716   C = Compile::current();
3717   if (_dolog) {
3718     _log = C->log();
3719   } else {
3720     _log = NULL;
3721   }
3722 
3723 #ifdef ASSERT
3724   if (PrintIdealNodeCount) {
3725     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3726                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3727   }
3728 
3729   if (VerifyIdealNodeCount) {
3730     Compile::current()->print_missing_nodes();
3731   }
3732 #endif
3733 
3734   if (_log != NULL) {
3735     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3736   }
3737 }
3738 
3739 //=============================================================================
3740 // Two Constant's are equal when the type and the value are equal.
3741 bool Compile::Constant::operator==(const Constant& other) {
3742   if (type()          != other.type()         )  return false;
3743   if (can_be_reused() != other.can_be_reused())  return false;
3744   // For floating point values we compare the bit pattern.
3745   switch (type()) {
3746   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3747   case T_LONG:
3748   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
3749   case T_OBJECT:
3750   case T_ADDRESS: return (_v._value.l == other._v._value.l);
3751   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
3752   case T_METADATA: return (_v._metadata == other._v._metadata);
3753   default: ShouldNotReachHere();
3754   }
3755   return false;
3756 }
3757 
3758 static int type_to_size_in_bytes(BasicType t) {
3759   switch (t) {
3760   case T_LONG:    return sizeof(jlong  );
3761   case T_FLOAT:   return sizeof(jfloat );
3762   case T_DOUBLE:  return sizeof(jdouble);
3763   case T_METADATA: return sizeof(Metadata*);
3764     // We use T_VOID as marker for jump-table entries (labels) which
3765     // need an internal word relocation.
3766   case T_VOID:
3767   case T_ADDRESS:
3768   case T_OBJECT:  return sizeof(jobject);
3769   }
3770 
3771   ShouldNotReachHere();
3772   return -1;
3773 }
3774 
3775 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3776   // sort descending
3777   if (a->freq() > b->freq())  return -1;
3778   if (a->freq() < b->freq())  return  1;
3779   return 0;
3780 }
3781 
3782 void Compile::ConstantTable::calculate_offsets_and_size() {
3783   // First, sort the array by frequencies.
3784   _constants.sort(qsort_comparator);
3785 
3786 #ifdef ASSERT
3787   // Make sure all jump-table entries were sorted to the end of the
3788   // array (they have a negative frequency).
3789   bool found_void = false;
3790   for (int i = 0; i < _constants.length(); i++) {
3791     Constant con = _constants.at(i);
3792     if (con.type() == T_VOID)
3793       found_void = true;  // jump-tables
3794     else
3795       assert(!found_void, "wrong sorting");
3796   }
3797 #endif
3798 
3799   int offset = 0;
3800   for (int i = 0; i < _constants.length(); i++) {
3801     Constant* con = _constants.adr_at(i);
3802 
3803     // Align offset for type.
3804     int typesize = type_to_size_in_bytes(con->type());
3805     offset = align_size_up(offset, typesize);
3806     con->set_offset(offset);   // set constant's offset
3807 
3808     if (con->type() == T_VOID) {
3809       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3810       offset = offset + typesize * n->outcnt();  // expand jump-table
3811     } else {
3812       offset = offset + typesize;
3813     }
3814   }
3815 
3816   // Align size up to the next section start (which is insts; see
3817   // CodeBuffer::align_at_start).
3818   assert(_size == -1, "already set?");
3819   _size = align_size_up(offset, CodeEntryAlignment);
3820 }
3821 
3822 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3823   MacroAssembler _masm(&cb);
3824   for (int i = 0; i < _constants.length(); i++) {
3825     Constant con = _constants.at(i);
3826     address constant_addr = NULL;
3827     switch (con.type()) {
3828     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
3829     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3830     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3831     case T_OBJECT: {
3832       jobject obj = con.get_jobject();
3833       int oop_index = _masm.oop_recorder()->find_index(obj);
3834       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3835       break;
3836     }
3837     case T_ADDRESS: {
3838       address addr = (address) con.get_jobject();
3839       constant_addr = _masm.address_constant(addr);
3840       break;
3841     }
3842     // We use T_VOID as marker for jump-table entries (labels) which
3843     // need an internal word relocation.
3844     case T_VOID: {
3845       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3846       // Fill the jump-table with a dummy word.  The real value is
3847       // filled in later in fill_jump_table.
3848       address dummy = (address) n;
3849       constant_addr = _masm.address_constant(dummy);
3850       // Expand jump-table
3851       for (uint i = 1; i < n->outcnt(); i++) {
3852         address temp_addr = _masm.address_constant(dummy + i);
3853         assert(temp_addr, "consts section too small");
3854       }
3855       break;
3856     }
3857     case T_METADATA: {
3858       Metadata* obj = con.get_metadata();
3859       int metadata_index = _masm.oop_recorder()->find_index(obj);
3860       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3861       break;
3862     }
3863     default: ShouldNotReachHere();
3864     }
3865     assert(constant_addr, "consts section too small");
3866     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3867             "must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset()));
3868   }
3869 }
3870 
3871 int Compile::ConstantTable::find_offset(Constant& con) const {
3872   int idx = _constants.find(con);
3873   assert(idx != -1, "constant must be in constant table");
3874   int offset = _constants.at(idx).offset();
3875   assert(offset != -1, "constant table not emitted yet?");
3876   return offset;
3877 }
3878 
3879 void Compile::ConstantTable::add(Constant& con) {
3880   if (con.can_be_reused()) {
3881     int idx = _constants.find(con);
3882     if (idx != -1 && _constants.at(idx).can_be_reused()) {
3883       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
3884       return;
3885     }
3886   }
3887   (void) _constants.append(con);
3888 }
3889 
3890 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3891   Block* b = Compile::current()->cfg()->get_block_for_node(n);
3892   Constant con(type, value, b->_freq);
3893   add(con);
3894   return con;
3895 }
3896 
3897 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3898   Constant con(metadata);
3899   add(con);
3900   return con;
3901 }
3902 
3903 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3904   jvalue value;
3905   BasicType type = oper->type()->basic_type();
3906   switch (type) {
3907   case T_LONG:    value.j = oper->constantL(); break;
3908   case T_FLOAT:   value.f = oper->constantF(); break;
3909   case T_DOUBLE:  value.d = oper->constantD(); break;
3910   case T_OBJECT:
3911   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3912   case T_METADATA: return add((Metadata*)oper->constant()); break;
3913   default: guarantee(false, "unhandled type: %s", type2name(type));
3914   }
3915   return add(n, type, value);
3916 }
3917 
3918 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3919   jvalue value;
3920   // We can use the node pointer here to identify the right jump-table
3921   // as this method is called from Compile::Fill_buffer right before
3922   // the MachNodes are emitted and the jump-table is filled (means the
3923   // MachNode pointers do not change anymore).
3924   value.l = (jobject) n;
3925   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
3926   add(con);
3927   return con;
3928 }
3929 
3930 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3931   // If called from Compile::scratch_emit_size do nothing.
3932   if (Compile::current()->in_scratch_emit_size())  return;
3933 
3934   assert(labels.is_nonempty(), "must be");
3935   assert((uint) labels.length() == n->outcnt(), "must be equal: %d == %d", labels.length(), n->outcnt());
3936 
3937   // Since MachConstantNode::constant_offset() also contains
3938   // table_base_offset() we need to subtract the table_base_offset()
3939   // to get the plain offset into the constant table.
3940   int offset = n->constant_offset() - table_base_offset();
3941 
3942   MacroAssembler _masm(&cb);
3943   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3944 
3945   for (uint i = 0; i < n->outcnt(); i++) {
3946     address* constant_addr = &jump_table_base[i];
3947     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));
3948     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3949     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3950   }
3951 }
3952 
3953 //----------------------------static_subtype_check-----------------------------
3954 // Shortcut important common cases when superklass is exact:
3955 // (0) superklass is java.lang.Object (can occur in reflective code)
3956 // (1) subklass is already limited to a subtype of superklass => always ok
3957 // (2) subklass does not overlap with superklass => always fail
3958 // (3) superklass has NO subtypes and we can check with a simple compare.
3959 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
3960   if (StressReflectiveCode) {
3961     return SSC_full_test;       // Let caller generate the general case.
3962   }
3963 
3964   if (superk == env()->Object_klass()) {
3965     return SSC_always_true;     // (0) this test cannot fail
3966   }
3967 
3968   ciType* superelem = superk;
3969   if (superelem->is_array_klass())
3970     superelem = superelem->as_array_klass()->base_element_type();
3971 
3972   if (!subk->is_interface()) {  // cannot trust static interface types yet
3973     if (subk->is_subtype_of(superk)) {
3974       return SSC_always_true;   // (1) false path dead; no dynamic test needed
3975     }
3976     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
3977         !superk->is_subtype_of(subk)) {
3978       return SSC_always_false;
3979     }
3980   }
3981 
3982   // If casting to an instance klass, it must have no subtypes
3983   if (superk->is_interface()) {
3984     // Cannot trust interfaces yet.
3985     // %%% S.B. superk->nof_implementors() == 1
3986   } else if (superelem->is_instance_klass()) {
3987     ciInstanceKlass* ik = superelem->as_instance_klass();
3988     if (!ik->has_subklass() && !ik->is_interface()) {
3989       if (!ik->is_final()) {
3990         // Add a dependency if there is a chance of a later subclass.
3991         dependencies()->assert_leaf_type(ik);
3992       }
3993       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
3994     }
3995   } else {
3996     // A primitive array type has no subtypes.
3997     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
3998   }
3999 
4000   return SSC_full_test;
4001 }
4002 
4003 Node* Compile::conv_I2X_index(PhaseGVN* phase, Node* idx, const TypeInt* sizetype, Node* ctrl) {
4004 #ifdef _LP64
4005   // The scaled index operand to AddP must be a clean 64-bit value.
4006   // Java allows a 32-bit int to be incremented to a negative
4007   // value, which appears in a 64-bit register as a large
4008   // positive number.  Using that large positive number as an
4009   // operand in pointer arithmetic has bad consequences.
4010   // On the other hand, 32-bit overflow is rare, and the possibility
4011   // can often be excluded, if we annotate the ConvI2L node with
4012   // a type assertion that its value is known to be a small positive
4013   // number.  (The prior range check has ensured this.)
4014   // This assertion is used by ConvI2LNode::Ideal.
4015   int index_max = max_jint - 1;  // array size is max_jint, index is one less
4016   if (sizetype != NULL) index_max = sizetype->_hi - 1;
4017   const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
4018   idx = constrained_convI2L(phase, idx, iidxtype, ctrl);
4019 #endif
4020   return idx;
4021 }
4022 
4023 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4024 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4025   if (ctrl != NULL) {
4026     // Express control dependency by a CastII node with a narrow type.
4027     value = new CastIINode(value, itype, false, true /* range check dependency */);
4028     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4029     // node from floating above the range check during loop optimizations. Otherwise, the
4030     // ConvI2L node may be eliminated independently of the range check, causing the data path
4031     // to become TOP while the control path is still there (although it's unreachable).
4032     value->set_req(0, ctrl);
4033     // Save CastII node to remove it after loop optimizations.
4034     phase->C->add_range_check_cast(value);
4035     value = phase->transform(value);
4036   }
4037   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4038   return phase->transform(new ConvI2LNode(value, ltype));
4039 }
4040 
4041 // The message about the current inlining is accumulated in
4042 // _print_inlining_stream and transfered into the _print_inlining_list
4043 // once we know whether inlining succeeds or not. For regular
4044 // inlining, messages are appended to the buffer pointed by
4045 // _print_inlining_idx in the _print_inlining_list. For late inlining,
4046 // a new buffer is added after _print_inlining_idx in the list. This
4047 // way we can update the inlining message for late inlining call site
4048 // when the inlining is attempted again.
4049 void Compile::print_inlining_init() {
4050   if (print_inlining() || print_intrinsics()) {
4051     _print_inlining_stream = new stringStream();
4052     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
4053   }
4054 }
4055 
4056 void Compile::print_inlining_reinit() {
4057   if (print_inlining() || print_intrinsics()) {
4058     // Re allocate buffer when we change ResourceMark
4059     _print_inlining_stream = new stringStream();
4060   }
4061 }
4062 
4063 void Compile::print_inlining_reset() {
4064   _print_inlining_stream->reset();
4065 }
4066 
4067 void Compile::print_inlining_commit() {
4068   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
4069   // Transfer the message from _print_inlining_stream to the current
4070   // _print_inlining_list buffer and clear _print_inlining_stream.
4071   _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
4072   print_inlining_reset();
4073 }
4074 
4075 void Compile::print_inlining_push() {
4076   // Add new buffer to the _print_inlining_list at current position
4077   _print_inlining_idx++;
4078   _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
4079 }
4080 
4081 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
4082   return _print_inlining_list->at(_print_inlining_idx);
4083 }
4084 
4085 void Compile::print_inlining_update(CallGenerator* cg) {
4086   if (print_inlining() || print_intrinsics()) {
4087     if (!cg->is_late_inline()) {
4088       if (print_inlining_current().cg() != NULL) {
4089         print_inlining_push();
4090       }
4091       print_inlining_commit();
4092     } else {
4093       if (print_inlining_current().cg() != cg &&
4094           (print_inlining_current().cg() != NULL ||
4095            print_inlining_current().ss()->size() != 0)) {
4096         print_inlining_push();
4097       }
4098       print_inlining_commit();
4099       print_inlining_current().set_cg(cg);
4100     }
4101   }
4102 }
4103 
4104 void Compile::print_inlining_move_to(CallGenerator* cg) {
4105   // We resume inlining at a late inlining call site. Locate the
4106   // corresponding inlining buffer so that we can update it.
4107   if (print_inlining()) {
4108     for (int i = 0; i < _print_inlining_list->length(); i++) {
4109       if (_print_inlining_list->adr_at(i)->cg() == cg) {
4110         _print_inlining_idx = i;
4111         return;
4112       }
4113     }
4114     ShouldNotReachHere();
4115   }
4116 }
4117 
4118 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
4119   if (print_inlining()) {
4120     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
4121     assert(print_inlining_current().cg() == cg, "wrong entry");
4122     // replace message with new message
4123     _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
4124     print_inlining_commit();
4125     print_inlining_current().set_cg(cg);
4126   }
4127 }
4128 
4129 void Compile::print_inlining_assert_ready() {
4130   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
4131 }
4132 
4133 void Compile::process_print_inlining() {
4134   bool do_print_inlining = print_inlining() || print_intrinsics();
4135   if (do_print_inlining || log() != NULL) {
4136     // Print inlining message for candidates that we couldn't inline
4137     // for lack of space
4138     for (int i = 0; i < _late_inlines.length(); i++) {
4139       CallGenerator* cg = _late_inlines.at(i);
4140       if (!cg->is_mh_late_inline()) {
4141         const char* msg = "live nodes > LiveNodeCountInliningCutoff";
4142         if (do_print_inlining) {
4143           cg->print_inlining_late(msg);
4144         }
4145         log_late_inline_failure(cg, msg);
4146       }
4147     }
4148   }
4149   if (do_print_inlining) {
4150     ResourceMark rm;
4151     stringStream ss;
4152     for (int i = 0; i < _print_inlining_list->length(); i++) {
4153       ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
4154     }
4155     size_t end = ss.size();
4156     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
4157     strncpy(_print_inlining_output, ss.base(), end+1);
4158     _print_inlining_output[end] = 0;
4159   }
4160 }
4161 
4162 void Compile::dump_print_inlining() {
4163   if (_print_inlining_output != NULL) {
4164     tty->print_raw(_print_inlining_output);
4165   }
4166 }
4167 
4168 void Compile::log_late_inline(CallGenerator* cg) {
4169   if (log() != NULL) {
4170     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
4171                 cg->unique_id());
4172     JVMState* p = cg->call_node()->jvms();
4173     while (p != NULL) {
4174       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
4175       p = p->caller();
4176     }
4177     log()->tail("late_inline");
4178   }
4179 }
4180 
4181 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
4182   log_late_inline(cg);
4183   if (log() != NULL) {
4184     log()->inline_fail(msg);
4185   }
4186 }
4187 
4188 void Compile::log_inline_id(CallGenerator* cg) {
4189   if (log() != NULL) {
4190     // The LogCompilation tool needs a unique way to identify late
4191     // inline call sites. This id must be unique for this call site in
4192     // this compilation. Try to have it unique across compilations as
4193     // well because it can be convenient when grepping through the log
4194     // file.
4195     // Distinguish OSR compilations from others in case CICountOSR is
4196     // on.
4197     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4198     cg->set_unique_id(id);
4199     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4200   }
4201 }
4202 
4203 void Compile::log_inline_failure(const char* msg) {
4204   if (C->log() != NULL) {
4205     C->log()->inline_fail(msg);
4206   }
4207 }
4208 
4209 
4210 // Dump inlining replay data to the stream.
4211 // Don't change thread state and acquire any locks.
4212 void Compile::dump_inline_data(outputStream* out) {
4213   InlineTree* inl_tree = ilt();
4214   if (inl_tree != NULL) {
4215     out->print(" inline %d", inl_tree->count());
4216     inl_tree->dump_replay_data(out);
4217   }
4218 }
4219 
4220 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4221   if (n1->Opcode() < n2->Opcode())      return -1;
4222   else if (n1->Opcode() > n2->Opcode()) return 1;
4223 
4224   assert(n1->req() == n2->req(), "can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req());
4225   for (uint i = 1; i < n1->req(); i++) {
4226     if (n1->in(i) < n2->in(i))      return -1;
4227     else if (n1->in(i) > n2->in(i)) return 1;
4228   }
4229 
4230   return 0;
4231 }
4232 
4233 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4234   Node* n1 = *n1p;
4235   Node* n2 = *n2p;
4236 
4237   return cmp_expensive_nodes(n1, n2);
4238 }
4239 
4240 void Compile::sort_expensive_nodes() {
4241   if (!expensive_nodes_sorted()) {
4242     _expensive_nodes->sort(cmp_expensive_nodes);
4243   }
4244 }
4245 
4246 bool Compile::expensive_nodes_sorted() const {
4247   for (int i = 1; i < _expensive_nodes->length(); i++) {
4248     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4249       return false;
4250     }
4251   }
4252   return true;
4253 }
4254 
4255 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4256   if (_expensive_nodes->length() == 0) {
4257     return false;
4258   }
4259 
4260   assert(OptimizeExpensiveOps, "optimization off?");
4261 
4262   // Take this opportunity to remove dead nodes from the list
4263   int j = 0;
4264   for (int i = 0; i < _expensive_nodes->length(); i++) {
4265     Node* n = _expensive_nodes->at(i);
4266     if (!n->is_unreachable(igvn)) {
4267       assert(n->is_expensive(), "should be expensive");
4268       _expensive_nodes->at_put(j, n);
4269       j++;
4270     }
4271   }
4272   _expensive_nodes->trunc_to(j);
4273 
4274   // Then sort the list so that similar nodes are next to each other
4275   // and check for at least two nodes of identical kind with same data
4276   // inputs.
4277   sort_expensive_nodes();
4278 
4279   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4280     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4281       return true;
4282     }
4283   }
4284 
4285   return false;
4286 }
4287 
4288 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4289   if (_expensive_nodes->length() == 0) {
4290     return;
4291   }
4292 
4293   assert(OptimizeExpensiveOps, "optimization off?");
4294 
4295   // Sort to bring similar nodes next to each other and clear the
4296   // control input of nodes for which there's only a single copy.
4297   sort_expensive_nodes();
4298 
4299   int j = 0;
4300   int identical = 0;
4301   int i = 0;
4302   bool modified = false;
4303   for (; i < _expensive_nodes->length()-1; i++) {
4304     assert(j <= i, "can't write beyond current index");
4305     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4306       identical++;
4307       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4308       continue;
4309     }
4310     if (identical > 0) {
4311       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4312       identical = 0;
4313     } else {
4314       Node* n = _expensive_nodes->at(i);
4315       igvn.replace_input_of(n, 0, NULL);
4316       igvn.hash_insert(n);
4317       modified = true;
4318     }
4319   }
4320   if (identical > 0) {
4321     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4322   } else if (_expensive_nodes->length() >= 1) {
4323     Node* n = _expensive_nodes->at(i);
4324     igvn.replace_input_of(n, 0, NULL);
4325     igvn.hash_insert(n);
4326     modified = true;
4327   }
4328   _expensive_nodes->trunc_to(j);
4329   if (modified) {
4330     igvn.optimize();
4331   }
4332 }
4333 
4334 void Compile::add_expensive_node(Node * n) {
4335   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4336   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4337   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4338   if (OptimizeExpensiveOps) {
4339     _expensive_nodes->append(n);
4340   } else {
4341     // Clear control input and let IGVN optimize expensive nodes if
4342     // OptimizeExpensiveOps is off.
4343     n->set_req(0, NULL);
4344   }
4345 }
4346 
4347 /**
4348  * Remove the speculative part of types and clean up the graph
4349  */
4350 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4351   if (UseTypeSpeculation) {
4352     Unique_Node_List worklist;
4353     worklist.push(root());
4354     int modified = 0;
4355     // Go over all type nodes that carry a speculative type, drop the
4356     // speculative part of the type and enqueue the node for an igvn
4357     // which may optimize it out.
4358     for (uint next = 0; next < worklist.size(); ++next) {
4359       Node *n  = worklist.at(next);
4360       if (n->is_Type()) {
4361         TypeNode* tn = n->as_Type();
4362         const Type* t = tn->type();
4363         const Type* t_no_spec = t->remove_speculative();
4364         if (t_no_spec != t) {
4365           bool in_hash = igvn.hash_delete(n);
4366           assert(in_hash, "node should be in igvn hash table");
4367           tn->set_type(t_no_spec);
4368           igvn.hash_insert(n);
4369           igvn._worklist.push(n); // give it a chance to go away
4370           modified++;
4371         }
4372       }
4373       uint max = n->len();
4374       for( uint i = 0; i < max; ++i ) {
4375         Node *m = n->in(i);
4376         if (not_a_node(m))  continue;
4377         worklist.push(m);
4378       }
4379     }
4380     // Drop the speculative part of all types in the igvn's type table
4381     igvn.remove_speculative_types();
4382     if (modified > 0) {
4383       igvn.optimize();
4384     }
4385 #ifdef ASSERT
4386     // Verify that after the IGVN is over no speculative type has resurfaced
4387     worklist.clear();
4388     worklist.push(root());
4389     for (uint next = 0; next < worklist.size(); ++next) {
4390       Node *n  = worklist.at(next);
4391       const Type* t = igvn.type_or_null(n);
4392       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4393       if (n->is_Type()) {
4394         t = n->as_Type()->type();
4395         assert(t == t->remove_speculative(), "no more speculative types");
4396       }
4397       uint max = n->len();
4398       for( uint i = 0; i < max; ++i ) {
4399         Node *m = n->in(i);
4400         if (not_a_node(m))  continue;
4401         worklist.push(m);
4402       }
4403     }
4404     igvn.check_no_speculative_types();
4405 #endif
4406   }
4407 }
4408 
4409 // Auxiliary method to support randomized stressing/fuzzing.
4410 //
4411 // This method can be called the arbitrary number of times, with current count
4412 // as the argument. The logic allows selecting a single candidate from the
4413 // running list of candidates as follows:
4414 //    int count = 0;
4415 //    Cand* selected = null;
4416 //    while(cand = cand->next()) {
4417 //      if (randomized_select(++count)) {
4418 //        selected = cand;
4419 //      }
4420 //    }
4421 //
4422 // Including count equalizes the chances any candidate is "selected".
4423 // This is useful when we don't have the complete list of candidates to choose
4424 // from uniformly. In this case, we need to adjust the randomicity of the
4425 // selection, or else we will end up biasing the selection towards the latter
4426 // candidates.
4427 //
4428 // Quick back-envelope calculation shows that for the list of n candidates
4429 // the equal probability for the candidate to persist as "best" can be
4430 // achieved by replacing it with "next" k-th candidate with the probability
4431 // of 1/k. It can be easily shown that by the end of the run, the
4432 // probability for any candidate is converged to 1/n, thus giving the
4433 // uniform distribution among all the candidates.
4434 //
4435 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4436 #define RANDOMIZED_DOMAIN_POW 29
4437 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4438 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4439 bool Compile::randomized_select(int count) {
4440   assert(count > 0, "only positive");
4441   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4442 }
4443 
4444 CloneMap&     Compile::clone_map()                 { return _clone_map; }
4445 void          Compile::set_clone_map(Dict* d)      { _clone_map._dict = d; }
4446 
4447 void NodeCloneInfo::dump() const {
4448   tty->print(" {%d:%d} ", idx(), gen());
4449 }
4450 
4451 void CloneMap::clone(Node* old, Node* nnn, int gen) {
4452   uint64_t val = value(old->_idx);
4453   NodeCloneInfo cio(val);
4454   assert(val != 0, "old node should be in the map");
4455   NodeCloneInfo cin(cio.idx(), gen + cio.gen());
4456   insert(nnn->_idx, cin.get());
4457 #ifndef PRODUCT
4458   if (is_debug()) {
4459     tty->print_cr("CloneMap::clone inserted node %d info {%d:%d} into CloneMap", nnn->_idx, cin.idx(), cin.gen());
4460   }
4461 #endif
4462 }
4463 
4464 void CloneMap::verify_insert_and_clone(Node* old, Node* nnn, int gen) {
4465   NodeCloneInfo cio(value(old->_idx));
4466   if (cio.get() == 0) {
4467     cio.set(old->_idx, 0);
4468     insert(old->_idx, cio.get());
4469 #ifndef PRODUCT
4470     if (is_debug()) {
4471       tty->print_cr("CloneMap::verify_insert_and_clone inserted node %d info {%d:%d} into CloneMap", old->_idx, cio.idx(), cio.gen());
4472     }
4473 #endif
4474   }
4475   clone(old, nnn, gen);
4476 }
4477 
4478 int CloneMap::max_gen() const {
4479   int g = 0;
4480   DictI di(_dict);
4481   for(; di.test(); ++di) {
4482     int t = gen(di._key);
4483     if (g < t) {
4484       g = t;
4485 #ifndef PRODUCT
4486       if (is_debug()) {
4487         tty->print_cr("CloneMap::max_gen() update max=%d from %d", g, _2_node_idx_t(di._key));
4488       }
4489 #endif
4490     }
4491   }
4492   return g;
4493 }
4494 
4495 void CloneMap::dump(node_idx_t key) const {
4496   uint64_t val = value(key);
4497   if (val != 0) {
4498     NodeCloneInfo ni(val);
4499     ni.dump();
4500   }
4501 }