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