1 /*
   2  * Copyright (c) 1997, 2014, 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/cfgnode.hpp"
  42 #include "opto/chaitin.hpp"
  43 #include "opto/compile.hpp"
  44 #include "opto/connode.hpp"
  45 #include "opto/divnode.hpp"
  46 #include "opto/escape.hpp"
  47 #include "opto/idealGraphPrinter.hpp"
  48 #include "opto/loopnode.hpp"
  49 #include "opto/machnode.hpp"
  50 #include "opto/macro.hpp"
  51 #include "opto/matcher.hpp"
  52 #include "opto/mathexactnode.hpp"
  53 #include "opto/memnode.hpp"
  54 #include "opto/mulnode.hpp"
  55 #include "opto/narrowptrnode.hpp"
  56 #include "opto/node.hpp"
  57 #include "opto/opcodes.hpp"
  58 #include "opto/output.hpp"
  59 #include "opto/parse.hpp"
  60 #include "opto/phaseX.hpp"
  61 #include "opto/rootnode.hpp"
  62 #include "opto/runtime.hpp"
  63 #include "opto/stringopts.hpp"
  64 #include "opto/type.hpp"
  65 #include "opto/vectornode.hpp"
  66 #include "runtime/arguments.hpp"
  67 #include "runtime/signature.hpp"
  68 #include "runtime/stubRoutines.hpp"
  69 #include "runtime/timer.hpp"
  70 #include "trace/tracing.hpp"
  71 #include "utilities/copy.hpp"
  72 
  73 
  74 // -------------------- Compile::mach_constant_base_node -----------------------
  75 // Constant table base node singleton.
  76 MachConstantBaseNode* Compile::mach_constant_base_node() {
  77   if (_mach_constant_base_node == NULL) {
  78     _mach_constant_base_node = new MachConstantBaseNode();
  79     _mach_constant_base_node->add_req(C->root());
  80   }
  81   return _mach_constant_base_node;
  82 }
  83 
  84 
  85 /// Support for intrinsics.
  86 
  87 // Return the index at which m must be inserted (or already exists).
  88 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
  89 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
  90 #ifdef ASSERT
  91   for (int i = 1; i < _intrinsics->length(); i++) {
  92     CallGenerator* cg1 = _intrinsics->at(i-1);
  93     CallGenerator* cg2 = _intrinsics->at(i);
  94     assert(cg1->method() != cg2->method()
  95            ? cg1->method()     < cg2->method()
  96            : cg1->is_virtual() < cg2->is_virtual(),
  97            "compiler intrinsics list must stay sorted");
  98   }
  99 #endif
 100   // Binary search sorted list, in decreasing intervals [lo, hi].
 101   int lo = 0, hi = _intrinsics->length()-1;
 102   while (lo <= hi) {
 103     int mid = (uint)(hi + lo) / 2;
 104     ciMethod* mid_m = _intrinsics->at(mid)->method();
 105     if (m < mid_m) {
 106       hi = mid-1;
 107     } else if (m > mid_m) {
 108       lo = mid+1;
 109     } else {
 110       // look at minor sort key
 111       bool mid_virt = _intrinsics->at(mid)->is_virtual();
 112       if (is_virtual < mid_virt) {
 113         hi = mid-1;
 114       } else if (is_virtual > mid_virt) {
 115         lo = mid+1;
 116       } else {
 117         return mid;  // exact match
 118       }
 119     }
 120   }
 121   return lo;  // inexact match
 122 }
 123 
 124 void Compile::register_intrinsic(CallGenerator* cg) {
 125   if (_intrinsics == NULL) {
 126     _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 127   }
 128   // This code is stolen from ciObjectFactory::insert.
 129   // Really, GrowableArray should have methods for
 130   // insert_at, remove_at, and binary_search.
 131   int len = _intrinsics->length();
 132   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
 133   if (index == len) {
 134     _intrinsics->append(cg);
 135   } else {
 136 #ifdef ASSERT
 137     CallGenerator* oldcg = _intrinsics->at(index);
 138     assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
 139 #endif
 140     _intrinsics->append(_intrinsics->at(len-1));
 141     int pos;
 142     for (pos = len-2; pos >= index; pos--) {
 143       _intrinsics->at_put(pos+1,_intrinsics->at(pos));
 144     }
 145     _intrinsics->at_put(index, cg);
 146   }
 147   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 148 }
 149 
 150 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 151   assert(m->is_loaded(), "don't try this on unloaded methods");
 152   if (_intrinsics != NULL) {
 153     int index = intrinsic_insertion_index(m, is_virtual);
 154     if (index < _intrinsics->length()
 155         && _intrinsics->at(index)->method() == m
 156         && _intrinsics->at(index)->is_virtual() == is_virtual) {
 157       return _intrinsics->at(index);
 158     }
 159   }
 160   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 161   if (m->intrinsic_id() != vmIntrinsics::_none &&
 162       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 163     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 164     if (cg != NULL) {
 165       // Save it for next time:
 166       register_intrinsic(cg);
 167       return cg;
 168     } else {
 169       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 170     }
 171   }
 172   return NULL;
 173 }
 174 
 175 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 176 // in library_call.cpp.
 177 
 178 
 179 #ifndef PRODUCT
 180 // statistics gathering...
 181 
 182 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 183 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 184 
 185 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 186   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 187   int oflags = _intrinsic_hist_flags[id];
 188   assert(flags != 0, "what happened?");
 189   if (is_virtual) {
 190     flags |= _intrinsic_virtual;
 191   }
 192   bool changed = (flags != oflags);
 193   if ((flags & _intrinsic_worked) != 0) {
 194     juint count = (_intrinsic_hist_count[id] += 1);
 195     if (count == 1) {
 196       changed = true;           // first time
 197     }
 198     // increment the overall count also:
 199     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 200   }
 201   if (changed) {
 202     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 203       // Something changed about the intrinsic's virtuality.
 204       if ((flags & _intrinsic_virtual) != 0) {
 205         // This is the first use of this intrinsic as a virtual call.
 206         if (oflags != 0) {
 207           // We already saw it as a non-virtual, so note both cases.
 208           flags |= _intrinsic_both;
 209         }
 210       } else if ((oflags & _intrinsic_both) == 0) {
 211         // This is the first use of this intrinsic as a non-virtual
 212         flags |= _intrinsic_both;
 213       }
 214     }
 215     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 216   }
 217   // update the overall flags also:
 218   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 219   return changed;
 220 }
 221 
 222 static char* format_flags(int flags, char* buf) {
 223   buf[0] = 0;
 224   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 225   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 226   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 227   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 228   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 229   if (buf[0] == 0)  strcat(buf, ",");
 230   assert(buf[0] == ',', "must be");
 231   return &buf[1];
 232 }
 233 
 234 void Compile::print_intrinsic_statistics() {
 235   char flagsbuf[100];
 236   ttyLocker ttyl;
 237   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 238   tty->print_cr("Compiler intrinsic usage:");
 239   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 240   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 241   #define PRINT_STAT_LINE(name, c, f) \
 242     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 243   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 244     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 245     int   flags = _intrinsic_hist_flags[id];
 246     juint count = _intrinsic_hist_count[id];
 247     if ((flags | count) != 0) {
 248       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 249     }
 250   }
 251   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 252   if (xtty != NULL)  xtty->tail("statistics");
 253 }
 254 
 255 void Compile::print_statistics() {
 256   { ttyLocker ttyl;
 257     if (xtty != NULL)  xtty->head("statistics type='opto'");
 258     Parse::print_statistics();
 259     PhaseCCP::print_statistics();
 260     PhaseRegAlloc::print_statistics();
 261     Scheduling::print_statistics();
 262     PhasePeephole::print_statistics();
 263     PhaseIdealLoop::print_statistics();
 264     if (xtty != NULL)  xtty->tail("statistics");
 265   }
 266   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 267     // put this under its own <statistics> element.
 268     print_intrinsic_statistics();
 269   }
 270 }
 271 #endif //PRODUCT
 272 
 273 // Support for bundling info
 274 Bundle* Compile::node_bundling(const Node *n) {
 275   assert(valid_bundle_info(n), "oob");
 276   return &_node_bundling_base[n->_idx];
 277 }
 278 
 279 bool Compile::valid_bundle_info(const Node *n) {
 280   return (_node_bundling_limit > n->_idx);
 281 }
 282 
 283 
 284 void Compile::gvn_replace_by(Node* n, Node* nn) {
 285   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 286     Node* use = n->last_out(i);
 287     bool is_in_table = initial_gvn()->hash_delete(use);
 288     uint uses_found = 0;
 289     for (uint j = 0; j < use->len(); j++) {
 290       if (use->in(j) == n) {
 291         if (j < use->req())
 292           use->set_req(j, nn);
 293         else
 294           use->set_prec(j, nn);
 295         uses_found++;
 296       }
 297     }
 298     if (is_in_table) {
 299       // reinsert into table
 300       initial_gvn()->hash_find_insert(use);
 301     }
 302     record_for_igvn(use);
 303     i -= uses_found;    // we deleted 1 or more copies of this edge
 304   }
 305 }
 306 
 307 
 308 static inline bool not_a_node(const Node* n) {
 309   if (n == NULL)                   return true;
 310   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 311   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 312   return false;
 313 }
 314 
 315 // Identify all nodes that are reachable from below, useful.
 316 // Use breadth-first pass that records state in a Unique_Node_List,
 317 // recursive traversal is slower.
 318 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 319   int estimated_worklist_size = unique();
 320   useful.map( estimated_worklist_size, NULL );  // preallocate space
 321 
 322   // Initialize worklist
 323   if (root() != NULL)     { useful.push(root()); }
 324   // If 'top' is cached, declare it useful to preserve cached node
 325   if( cached_top_node() ) { useful.push(cached_top_node()); }
 326 
 327   // Push all useful nodes onto the list, breadthfirst
 328   for( uint next = 0; next < useful.size(); ++next ) {
 329     assert( next < unique(), "Unique useful nodes < total nodes");
 330     Node *n  = useful.at(next);
 331     uint max = n->len();
 332     for( uint i = 0; i < max; ++i ) {
 333       Node *m = n->in(i);
 334       if (not_a_node(m))  continue;
 335       useful.push(m);
 336     }
 337   }
 338 }
 339 
 340 // Update dead_node_list with any missing dead nodes using useful
 341 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 342 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 343   uint max_idx = unique();
 344   VectorSet& useful_node_set = useful.member_set();
 345 
 346   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 347     // If node with index node_idx is not in useful set,
 348     // mark it as dead in dead node list.
 349     if (! useful_node_set.test(node_idx) ) {
 350       record_dead_node(node_idx);
 351     }
 352   }
 353 }
 354 
 355 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 356   int shift = 0;
 357   for (int i = 0; i < inlines->length(); i++) {
 358     CallGenerator* cg = inlines->at(i);
 359     CallNode* call = cg->call_node();
 360     if (shift > 0) {
 361       inlines->at_put(i-shift, cg);
 362     }
 363     if (!useful.member(call)) {
 364       shift++;
 365     }
 366   }
 367   inlines->trunc_to(inlines->length()-shift);
 368 }
 369 
 370 // Disconnect all useless nodes by disconnecting those at the boundary.
 371 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 372   uint next = 0;
 373   while (next < useful.size()) {
 374     Node *n = useful.at(next++);
 375     if (n->is_SafePoint()) {
 376       // We're done with a parsing phase. Replaced nodes are not valid
 377       // beyond that point.
 378       n->as_SafePoint()->delete_replaced_nodes();
 379     }
 380     // Use raw traversal of out edges since this code removes out edges
 381     int max = n->outcnt();
 382     for (int j = 0; j < max; ++j) {
 383       Node* child = n->raw_out(j);
 384       if (! useful.member(child)) {
 385         assert(!child->is_top() || child != top(),
 386                "If top is cached in Compile object it is in useful list");
 387         // Only need to remove this out-edge to the useless node
 388         n->raw_del_out(j);
 389         --j;
 390         --max;
 391       }
 392     }
 393     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 394       record_for_igvn(n->unique_out());
 395     }
 396   }
 397   // Remove useless macro and predicate opaq nodes
 398   for (int i = C->macro_count()-1; i >= 0; i--) {
 399     Node* n = C->macro_node(i);
 400     if (!useful.member(n)) {
 401       remove_macro_node(n);
 402     }
 403   }
 404   // Remove useless expensive node
 405   for (int i = C->expensive_count()-1; i >= 0; i--) {
 406     Node* n = C->expensive_node(i);
 407     if (!useful.member(n)) {
 408       remove_expensive_node(n);
 409     }
 410   }
 411   // clean up the late inline lists
 412   remove_useless_late_inlines(&_string_late_inlines, useful);
 413   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 414   remove_useless_late_inlines(&_late_inlines, useful);
 415   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 416 }
 417 
 418 //------------------------------frame_size_in_words-----------------------------
 419 // frame_slots in units of words
 420 int Compile::frame_size_in_words() const {
 421   // shift is 0 in LP32 and 1 in LP64
 422   const int shift = (LogBytesPerWord - LogBytesPerInt);
 423   int words = _frame_slots >> shift;
 424   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 425   return words;
 426 }
 427 
 428 // To bang the stack of this compiled method we use the stack size
 429 // that the interpreter would need in case of a deoptimization. This
 430 // removes the need to bang the stack in the deoptimization blob which
 431 // in turn simplifies stack overflow handling.
 432 int Compile::bang_size_in_bytes() const {
 433   return MAX2(frame_size_in_bytes() + os::extra_bang_size_in_bytes(), _interpreter_frame_size);
 434 }
 435 
 436 // ============================================================================
 437 //------------------------------CompileWrapper---------------------------------
 438 class CompileWrapper : public StackObj {
 439   Compile *const _compile;
 440  public:
 441   CompileWrapper(Compile* compile);
 442 
 443   ~CompileWrapper();
 444 };
 445 
 446 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 447   // the Compile* pointer is stored in the current ciEnv:
 448   ciEnv* env = compile->env();
 449   assert(env == ciEnv::current(), "must already be a ciEnv active");
 450   assert(env->compiler_data() == NULL, "compile already active?");
 451   env->set_compiler_data(compile);
 452   assert(compile == Compile::current(), "sanity");
 453 
 454   compile->set_type_dict(NULL);
 455   compile->set_type_hwm(NULL);
 456   compile->set_type_last_size(0);
 457   compile->set_last_tf(NULL, NULL);
 458   compile->set_indexSet_arena(NULL);
 459   compile->set_indexSet_free_block_list(NULL);
 460   compile->init_type_arena();
 461   Type::Initialize(compile);
 462   _compile->set_scratch_buffer_blob(NULL);
 463   _compile->begin_method();
 464 }
 465 CompileWrapper::~CompileWrapper() {
 466   _compile->end_method();
 467   if (_compile->scratch_buffer_blob() != NULL)
 468     BufferBlob::free(_compile->scratch_buffer_blob());
 469   _compile->env()->set_compiler_data(NULL);
 470 }
 471 
 472 
 473 //----------------------------print_compile_messages---------------------------
 474 void Compile::print_compile_messages() {
 475 #ifndef PRODUCT
 476   // Check if recompiling
 477   if (_subsume_loads == false && PrintOpto) {
 478     // Recompiling without allowing machine instructions to subsume loads
 479     tty->print_cr("*********************************************************");
 480     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 481     tty->print_cr("*********************************************************");
 482   }
 483   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 484     // Recompiling without escape analysis
 485     tty->print_cr("*********************************************************");
 486     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 487     tty->print_cr("*********************************************************");
 488   }
 489   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 490     // Recompiling without boxing elimination
 491     tty->print_cr("*********************************************************");
 492     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 493     tty->print_cr("*********************************************************");
 494   }
 495   if (env()->break_at_compile()) {
 496     // Open the debugger when compiling this method.
 497     tty->print("### Breaking when compiling: ");
 498     method()->print_short_name();
 499     tty->cr();
 500     BREAKPOINT;
 501   }
 502 
 503   if( PrintOpto ) {
 504     if (is_osr_compilation()) {
 505       tty->print("[OSR]%3d", _compile_id);
 506     } else {
 507       tty->print("%3d", _compile_id);
 508     }
 509   }
 510 #endif
 511 }
 512 
 513 
 514 //-----------------------init_scratch_buffer_blob------------------------------
 515 // Construct a temporary BufferBlob and cache it for this compile.
 516 void Compile::init_scratch_buffer_blob(int const_size) {
 517   // If there is already a scratch buffer blob allocated and the
 518   // constant section is big enough, use it.  Otherwise free the
 519   // current and allocate a new one.
 520   BufferBlob* blob = scratch_buffer_blob();
 521   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 522     // Use the current blob.
 523   } else {
 524     if (blob != NULL) {
 525       BufferBlob::free(blob);
 526     }
 527 
 528     ResourceMark rm;
 529     _scratch_const_size = const_size;
 530     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
 531     blob = BufferBlob::create("Compile::scratch_buffer", size);
 532     // Record the buffer blob for next time.
 533     set_scratch_buffer_blob(blob);
 534     // Have we run out of code space?
 535     if (scratch_buffer_blob() == NULL) {
 536       // Let CompilerBroker disable further compilations.
 537       record_failure("Not enough space for scratch buffer in CodeCache");
 538       CompileBroker::handle_full_code_cache();
 539       return;
 540     }
 541   }
 542 
 543   // Initialize the relocation buffers
 544   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 545   set_scratch_locs_memory(locs_buf);
 546 }
 547 
 548 
 549 //-----------------------scratch_emit_size-------------------------------------
 550 // Helper function that computes size by emitting code
 551 uint Compile::scratch_emit_size(const Node* n) {
 552   // Start scratch_emit_size section.
 553   set_in_scratch_emit_size(true);
 554 
 555   // Emit into a trash buffer and count bytes emitted.
 556   // This is a pretty expensive way to compute a size,
 557   // but it works well enough if seldom used.
 558   // All common fixed-size instructions are given a size
 559   // method by the AD file.
 560   // Note that the scratch buffer blob and locs memory are
 561   // allocated at the beginning of the compile task, and
 562   // may be shared by several calls to scratch_emit_size.
 563   // The allocation of the scratch buffer blob is particularly
 564   // expensive, since it has to grab the code cache lock.
 565   BufferBlob* blob = this->scratch_buffer_blob();
 566   assert(blob != NULL, "Initialize BufferBlob at start");
 567   assert(blob->size() > MAX_inst_size, "sanity");
 568   relocInfo* locs_buf = scratch_locs_memory();
 569   address blob_begin = blob->content_begin();
 570   address blob_end   = (address)locs_buf;
 571   assert(blob->content_contains(blob_end), "sanity");
 572   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 573   buf.initialize_consts_size(_scratch_const_size);
 574   buf.initialize_stubs_size(MAX_stubs_size);
 575   assert(locs_buf != NULL, "sanity");
 576   int lsize = MAX_locs_size / 3;
 577   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 578   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 579   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 580 
 581   // Do the emission.
 582 
 583   Label fakeL; // Fake label for branch instructions.
 584   Label*   saveL = NULL;
 585   uint save_bnum = 0;
 586   bool is_branch = n->is_MachBranch();
 587   if (is_branch) {
 588     MacroAssembler masm(&buf);
 589     masm.bind(fakeL);
 590     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 591     n->as_MachBranch()->label_set(&fakeL, 0);
 592   }
 593   n->emit(buf, this->regalloc());
 594   if (is_branch) // Restore label.
 595     n->as_MachBranch()->label_set(saveL, save_bnum);
 596 
 597   // End scratch_emit_size section.
 598   set_in_scratch_emit_size(false);
 599 
 600   return buf.insts_size();
 601 }
 602 
 603 
 604 // ============================================================================
 605 //------------------------------Compile standard-------------------------------
 606 debug_only( int Compile::_debug_idx = 100000; )
 607 
 608 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 609 // the continuation bci for on stack replacement.
 610 
 611 
 612 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 613                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
 614                 : Phase(Compiler),
 615                   _env(ci_env),
 616                   _log(ci_env->log()),
 617                   _compile_id(ci_env->compile_id()),
 618                   _save_argument_registers(false),
 619                   _stub_name(NULL),
 620                   _stub_function(NULL),
 621                   _stub_entry_point(NULL),
 622                   _method(target),
 623                   _entry_bci(osr_bci),
 624                   _initial_gvn(NULL),
 625                   _for_igvn(NULL),
 626                   _warm_calls(NULL),
 627                   _subsume_loads(subsume_loads),
 628                   _do_escape_analysis(do_escape_analysis),
 629                   _eliminate_boxing(eliminate_boxing),
 630                   _failure_reason(NULL),
 631                   _code_buffer("Compile::Fill_buffer"),
 632                   _orig_pc_slot(0),
 633                   _orig_pc_slot_offset_in_bytes(0),
 634                   _has_method_handle_invokes(false),
 635                   _mach_constant_base_node(NULL),
 636                   _node_bundling_limit(0),
 637                   _node_bundling_base(NULL),
 638                   _java_calls(0),
 639                   _inner_loops(0),
 640                   _scratch_const_size(-1),
 641                   _in_scratch_emit_size(false),
 642                   _dead_node_list(comp_arena()),
 643                   _dead_node_count(0),
 644 #ifndef PRODUCT
 645                   _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
 646                   _in_dump_cnt(0),
 647                   _printer(IdealGraphPrinter::printer()),
 648 #endif
 649                   _congraph(NULL),
 650                   _comp_arena(mtCompiler),
 651                   _node_arena(mtCompiler),
 652                   _old_arena(mtCompiler),
 653                   _Compile_types(mtCompiler),
 654                   _replay_inline_data(NULL),
 655                   _late_inlines(comp_arena(), 2, 0, NULL),
 656                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 657                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 658                   _late_inlines_pos(0),
 659                   _number_of_mh_late_inlines(0),
 660                   _inlining_progress(false),
 661                   _inlining_incrementally(false),
 662                   _print_inlining_list(NULL),
 663                   _print_inlining_stream(NULL),
 664                   _print_inlining_idx(0),
 665                   _print_inlining_output(NULL),
 666                   _interpreter_frame_size(0) {
 667   C = this;
 668 
 669   CompileWrapper cw(this);
 670 
 671   if (CITimeVerbose) {
 672     tty->print(" ");
 673     target->holder()->name()->print();
 674     tty->print(".");
 675     target->print_short_name();
 676     tty->print("  ");
 677   }
 678   TraceTime t1("Total compilation time", &_t_totalCompilation, CITime, CITimeVerbose);
 679   TraceTime t2(NULL, &_t_methodCompilation, CITime, false);
 680 
 681 #ifndef PRODUCT
 682   bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
 683   if (!print_opto_assembly) {
 684     bool print_assembly = (PrintAssembly || _method->should_print_assembly());
 685     if (print_assembly && !Disassembler::can_decode()) {
 686       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 687       print_opto_assembly = true;
 688     }
 689   }
 690   set_print_assembly(print_opto_assembly);
 691   set_parsed_irreducible_loop(false);
 692 
 693   if (method()->has_option("ReplayInline")) {
 694     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 695   }
 696 #endif
 697   set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
 698   set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
 699   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 700 
 701   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 702     // Make sure the method being compiled gets its own MDO,
 703     // so we can at least track the decompile_count().
 704     // Need MDO to record RTM code generation state.
 705     method()->ensure_method_data();
 706   }
 707 
 708   Init(::AliasLevel);
 709 
 710 
 711   print_compile_messages();
 712 
 713   _ilt = InlineTree::build_inline_tree_root();
 714 
 715   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 716   assert(num_alias_types() >= AliasIdxRaw, "");
 717 
 718 #define MINIMUM_NODE_HASH  1023
 719   // Node list that Iterative GVN will start with
 720   Unique_Node_List for_igvn(comp_arena());
 721   set_for_igvn(&for_igvn);
 722 
 723   // GVN that will be run immediately on new nodes
 724   uint estimated_size = method()->code_size()*4+64;
 725   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 726   PhaseGVN gvn(node_arena(), estimated_size);
 727   set_initial_gvn(&gvn);
 728 
 729   print_inlining_init();
 730   { // Scope for timing the parser
 731     TracePhase t3("parse", &timers[_t_parser]);
 732 
 733     // Put top into the hash table ASAP.
 734     initial_gvn()->transform_no_reclaim(top());
 735 
 736     // Set up tf(), start(), and find a CallGenerator.
 737     CallGenerator* cg = NULL;
 738     if (is_osr_compilation()) {
 739       const TypeTuple *domain = StartOSRNode::osr_domain();
 740       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 741       init_tf(TypeFunc::make(domain, range));
 742       StartNode* s = new StartOSRNode(root(), domain);
 743       initial_gvn()->set_type_bottom(s);
 744       init_start(s);
 745       cg = CallGenerator::for_osr(method(), entry_bci());
 746     } else {
 747       // Normal case.
 748       init_tf(TypeFunc::make(method()));
 749       StartNode* s = new StartNode(root(), tf()->domain());
 750       initial_gvn()->set_type_bottom(s);
 751       init_start(s);
 752       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
 753         // With java.lang.ref.reference.get() we must go through the
 754         // intrinsic when G1 is enabled - even when get() is the root
 755         // method of the compile - so that, if necessary, the value in
 756         // the referent field of the reference object gets recorded by
 757         // the pre-barrier code.
 758         // Specifically, if G1 is enabled, the value in the referent
 759         // field is recorded by the G1 SATB pre barrier. This will
 760         // result in the referent being marked live and the reference
 761         // object removed from the list of discovered references during
 762         // reference processing.
 763         cg = find_intrinsic(method(), false);
 764       }
 765       if (cg == NULL) {
 766         float past_uses = method()->interpreter_invocation_count();
 767         float expected_uses = past_uses;
 768         cg = CallGenerator::for_inline(method(), expected_uses);
 769       }
 770     }
 771     if (failing())  return;
 772     if (cg == NULL) {
 773       record_method_not_compilable_all_tiers("cannot parse method");
 774       return;
 775     }
 776     JVMState* jvms = build_start_state(start(), tf());
 777     if ((jvms = cg->generate(jvms)) == NULL) {
 778       record_method_not_compilable("method parse failed");
 779       return;
 780     }
 781     GraphKit kit(jvms);
 782 
 783     if (!kit.stopped()) {
 784       // Accept return values, and transfer control we know not where.
 785       // This is done by a special, unique ReturnNode bound to root.
 786       return_values(kit.jvms());
 787     }
 788 
 789     if (kit.has_exceptions()) {
 790       // Any exceptions that escape from this call must be rethrown
 791       // to whatever caller is dynamically above us on the stack.
 792       // This is done by a special, unique RethrowNode bound to root.
 793       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 794     }
 795 
 796     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 797 
 798     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 799       inline_string_calls(true);
 800     }
 801 
 802     if (failing())  return;
 803 
 804     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 805 
 806     // Remove clutter produced by parsing.
 807     if (!failing()) {
 808       ResourceMark rm;
 809       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 810     }
 811   }
 812 
 813   // Note:  Large methods are capped off in do_one_bytecode().
 814   if (failing())  return;
 815 
 816   // After parsing, node notes are no longer automagic.
 817   // They must be propagated by register_new_node_with_optimizer(),
 818   // clone(), or the like.
 819   set_default_node_notes(NULL);
 820 
 821   for (;;) {
 822     int successes = Inline_Warm();
 823     if (failing())  return;
 824     if (successes == 0)  break;
 825   }
 826 
 827   // Drain the list.
 828   Finish_Warm();
 829 #ifndef PRODUCT
 830   if (_printer) {
 831     _printer->print_inlining(this);
 832   }
 833 #endif
 834 
 835   if (failing())  return;
 836   NOT_PRODUCT( verify_graph_edges(); )
 837 
 838   // Now optimize
 839   Optimize();
 840   if (failing())  return;
 841   NOT_PRODUCT( verify_graph_edges(); )
 842 
 843 #ifndef PRODUCT
 844   if (PrintIdeal) {
 845     ttyLocker ttyl;  // keep the following output all in one block
 846     // This output goes directly to the tty, not the compiler log.
 847     // To enable tools to match it up with the compilation activity,
 848     // be sure to tag this tty output with the compile ID.
 849     if (xtty != NULL) {
 850       xtty->head("ideal compile_id='%d'%s", compile_id(),
 851                  is_osr_compilation()    ? " compile_kind='osr'" :
 852                  "");
 853     }
 854     root()->dump(9999);
 855     if (xtty != NULL) {
 856       xtty->tail("ideal");
 857     }
 858   }
 859 #endif
 860 
 861   NOT_PRODUCT( verify_barriers(); )
 862 
 863   // Dump compilation data to replay it.
 864   if (method()->has_option("DumpReplay")) {
 865     env()->dump_replay_data(_compile_id);
 866   }
 867   if (method()->has_option("DumpInline") && (ilt() != NULL)) {
 868     env()->dump_inline_data(_compile_id);
 869   }
 870 
 871   // Now that we know the size of all the monitors we can add a fixed slot
 872   // for the original deopt pc.
 873 
 874   _orig_pc_slot =  fixed_slots();
 875   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 876   set_fixed_slots(next_slot);
 877 
 878   // Compute when to use implicit null checks. Used by matching trap based
 879   // nodes and NullCheck optimization.
 880   set_allowed_deopt_reasons();
 881 
 882   // Now generate code
 883   Code_Gen();
 884   if (failing())  return;
 885 
 886   // Check if we want to skip execution of all compiled code.
 887   {
 888 #ifndef PRODUCT
 889     if (OptoNoExecute) {
 890       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 891       return;
 892     }
 893     TracePhase t2("install_code", &_t_registerMethod);
 894 #endif
 895 
 896     if (is_osr_compilation()) {
 897       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 898       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 899     } else {
 900       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 901       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 902     }
 903 
 904     env()->register_method(_method, _entry_bci,
 905                            &_code_offsets,
 906                            _orig_pc_slot_offset_in_bytes,
 907                            code_buffer(),
 908                            frame_size_in_words(), _oop_map_set,
 909                            &_handler_table, &_inc_table,
 910                            compiler,
 911                            env()->comp_level(),
 912                            has_unsafe_access(),
 913                            SharedRuntime::is_wide_vector(max_vector_size()),
 914                            rtm_state()
 915                            );
 916 
 917     if (log() != NULL) // Print code cache state into compiler log
 918       log()->code_cache_state();
 919   }
 920 }
 921 
 922 //------------------------------Compile----------------------------------------
 923 // Compile a runtime stub
 924 Compile::Compile( ciEnv* ci_env,
 925                   TypeFunc_generator generator,
 926                   address stub_function,
 927                   const char *stub_name,
 928                   int is_fancy_jump,
 929                   bool pass_tls,
 930                   bool save_arg_registers,
 931                   bool return_pc )
 932   : Phase(Compiler),
 933     _env(ci_env),
 934     _log(ci_env->log()),
 935     _compile_id(0),
 936     _save_argument_registers(save_arg_registers),
 937     _method(NULL),
 938     _stub_name(stub_name),
 939     _stub_function(stub_function),
 940     _stub_entry_point(NULL),
 941     _entry_bci(InvocationEntryBci),
 942     _initial_gvn(NULL),
 943     _for_igvn(NULL),
 944     _warm_calls(NULL),
 945     _orig_pc_slot(0),
 946     _orig_pc_slot_offset_in_bytes(0),
 947     _subsume_loads(true),
 948     _do_escape_analysis(false),
 949     _eliminate_boxing(false),
 950     _failure_reason(NULL),
 951     _code_buffer("Compile::Fill_buffer"),
 952     _has_method_handle_invokes(false),
 953     _mach_constant_base_node(NULL),
 954     _node_bundling_limit(0),
 955     _node_bundling_base(NULL),
 956     _java_calls(0),
 957     _inner_loops(0),
 958 #ifndef PRODUCT
 959     _trace_opto_output(TraceOptoOutput),
 960     _in_dump_cnt(0),
 961     _printer(NULL),
 962 #endif
 963     _comp_arena(mtCompiler),
 964     _node_arena(mtCompiler),
 965     _old_arena(mtCompiler),
 966     _Compile_types(mtCompiler),
 967     _dead_node_list(comp_arena()),
 968     _dead_node_count(0),
 969     _congraph(NULL),
 970     _replay_inline_data(NULL),
 971     _number_of_mh_late_inlines(0),
 972     _inlining_progress(false),
 973     _inlining_incrementally(false),
 974     _print_inlining_list(NULL),
 975     _print_inlining_stream(NULL),
 976     _print_inlining_idx(0),
 977     _print_inlining_output(NULL),
 978     _allowed_reasons(0),
 979     _interpreter_frame_size(0) {
 980   C = this;
 981 
 982   TraceTime t1(NULL, &_t_totalCompilation, CITime, false);
 983   TraceTime t2(NULL, &_t_stubCompilation, CITime, false);
 984 
 985 #ifndef PRODUCT
 986   set_print_assembly(PrintFrameConverterAssembly);
 987   set_parsed_irreducible_loop(false);
 988 #endif
 989   set_has_irreducible_loop(false); // no loops
 990 
 991   CompileWrapper cw(this);
 992   Init(/*AliasLevel=*/ 0);
 993   init_tf((*generator)());
 994 
 995   {
 996     // The following is a dummy for the sake of GraphKit::gen_stub
 997     Unique_Node_List for_igvn(comp_arena());
 998     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
 999     PhaseGVN gvn(Thread::current()->resource_area(),255);
1000     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1001     gvn.transform_no_reclaim(top());
1002 
1003     GraphKit kit;
1004     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1005   }
1006 
1007   NOT_PRODUCT( verify_graph_edges(); )
1008   Code_Gen();
1009   if (failing())  return;
1010 
1011 
1012   // Entry point will be accessed using compile->stub_entry_point();
1013   if (code_buffer() == NULL) {
1014     Matcher::soft_match_failure();
1015   } else {
1016     if (PrintAssembly && (WizardMode || Verbose))
1017       tty->print_cr("### Stub::%s", stub_name);
1018 
1019     if (!failing()) {
1020       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1021 
1022       // Make the NMethod
1023       // For now we mark the frame as never safe for profile stackwalking
1024       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1025                                                       code_buffer(),
1026                                                       CodeOffsets::frame_never_safe,
1027                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1028                                                       frame_size_in_words(),
1029                                                       _oop_map_set,
1030                                                       save_arg_registers);
1031       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1032 
1033       _stub_entry_point = rs->entry_point();
1034     }
1035   }
1036 }
1037 
1038 //------------------------------Init-------------------------------------------
1039 // Prepare for a single compilation
1040 void Compile::Init(int aliaslevel) {
1041   _unique  = 0;
1042   _regalloc = NULL;
1043 
1044   _tf      = NULL;  // filled in later
1045   _top     = NULL;  // cached later
1046   _matcher = NULL;  // filled in later
1047   _cfg     = NULL;  // filled in later
1048 
1049   set_24_bit_selection_and_mode(Use24BitFP, false);
1050 
1051   _node_note_array = NULL;
1052   _default_node_notes = NULL;
1053   DEBUG_ONLY( _modified_nodes = NULL; ) // Used in Optimize()
1054 
1055   _immutable_memory = NULL; // filled in at first inquiry
1056 
1057   // Globally visible Nodes
1058   // First set TOP to NULL to give safe behavior during creation of RootNode
1059   set_cached_top_node(NULL);
1060   set_root(new RootNode());
1061   // Now that you have a Root to point to, create the real TOP
1062   set_cached_top_node( new ConNode(Type::TOP) );
1063   set_recent_alloc(NULL, NULL);
1064 
1065   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1066   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1067   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1068   env()->set_dependencies(new Dependencies(env()));
1069 
1070   _fixed_slots = 0;
1071   set_has_split_ifs(false);
1072   set_has_loops(has_method() && method()->has_loops()); // first approximation
1073   set_has_stringbuilder(false);
1074   set_has_boxed_value(false);
1075   _trap_can_recompile = false;  // no traps emitted yet
1076   _major_progress = true; // start out assuming good things will happen
1077   set_has_unsafe_access(false);
1078   set_max_vector_size(0);
1079   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1080   set_decompile_count(0);
1081 
1082   set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1083   set_num_loop_opts(LoopOptsCount);
1084   set_do_inlining(Inline);
1085   set_max_inline_size(MaxInlineSize);
1086   set_freq_inline_size(FreqInlineSize);
1087   set_do_scheduling(OptoScheduling);
1088   set_do_count_invocations(false);
1089   set_do_method_data_update(false);
1090   set_age_code(has_method() && method()->profile_aging());
1091   set_rtm_state(NoRTM); // No RTM lock eliding by default
1092 #if INCLUDE_RTM_OPT
1093   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1094     int rtm_state = method()->method_data()->rtm_state();
1095     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1096       // Don't generate RTM lock eliding code.
1097       set_rtm_state(NoRTM);
1098     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1099       // Generate RTM lock eliding code without abort ratio calculation code.
1100       set_rtm_state(UseRTM);
1101     } else if (UseRTMDeopt) {
1102       // Generate RTM lock eliding code and include abort ratio calculation
1103       // code if UseRTMDeopt is on.
1104       set_rtm_state(ProfileRTM);
1105     }
1106   }
1107 #endif
1108   if (debug_info()->recording_non_safepoints()) {
1109     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1110                         (comp_arena(), 8, 0, NULL));
1111     set_default_node_notes(Node_Notes::make(this));
1112   }
1113 
1114   // // -- Initialize types before each compile --
1115   // // Update cached type information
1116   // if( _method && _method->constants() )
1117   //   Type::update_loaded_types(_method, _method->constants());
1118 
1119   // Init alias_type map.
1120   if (!_do_escape_analysis && aliaslevel == 3)
1121     aliaslevel = 2;  // No unique types without escape analysis
1122   _AliasLevel = aliaslevel;
1123   const int grow_ats = 16;
1124   _max_alias_types = grow_ats;
1125   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1126   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1127   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1128   {
1129     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1130   }
1131   // Initialize the first few types.
1132   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1133   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1134   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1135   _num_alias_types = AliasIdxRaw+1;
1136   // Zero out the alias type cache.
1137   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1138   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1139   probe_alias_cache(NULL)->_index = AliasIdxTop;
1140 
1141   _intrinsics = NULL;
1142   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1143   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1144   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1145   register_library_intrinsics();
1146 }
1147 
1148 //---------------------------init_start----------------------------------------
1149 // Install the StartNode on this compile object.
1150 void Compile::init_start(StartNode* s) {
1151   if (failing())
1152     return; // already failing
1153   assert(s == start(), "");
1154 }
1155 
1156 StartNode* Compile::start() const {
1157   assert(!failing(), "");
1158   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1159     Node* start = root()->fast_out(i);
1160     if( start->is_Start() )
1161       return start->as_Start();
1162   }
1163   fatal("Did not find Start node!");
1164   return NULL;
1165 }
1166 
1167 //-------------------------------immutable_memory-------------------------------------
1168 // Access immutable memory
1169 Node* Compile::immutable_memory() {
1170   if (_immutable_memory != NULL) {
1171     return _immutable_memory;
1172   }
1173   StartNode* s = start();
1174   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1175     Node *p = s->fast_out(i);
1176     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1177       _immutable_memory = p;
1178       return _immutable_memory;
1179     }
1180   }
1181   ShouldNotReachHere();
1182   return NULL;
1183 }
1184 
1185 //----------------------set_cached_top_node------------------------------------
1186 // Install the cached top node, and make sure Node::is_top works correctly.
1187 void Compile::set_cached_top_node(Node* tn) {
1188   if (tn != NULL)  verify_top(tn);
1189   Node* old_top = _top;
1190   _top = tn;
1191   // Calling Node::setup_is_top allows the nodes the chance to adjust
1192   // their _out arrays.
1193   if (_top != NULL)     _top->setup_is_top();
1194   if (old_top != NULL)  old_top->setup_is_top();
1195   assert(_top == NULL || top()->is_top(), "");
1196 }
1197 
1198 #ifdef ASSERT
1199 uint Compile::count_live_nodes_by_graph_walk() {
1200   Unique_Node_List useful(comp_arena());
1201   // Get useful node list by walking the graph.
1202   identify_useful_nodes(useful);
1203   return useful.size();
1204 }
1205 
1206 void Compile::print_missing_nodes() {
1207 
1208   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1209   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1210     return;
1211   }
1212 
1213   // This is an expensive function. It is executed only when the user
1214   // specifies VerifyIdealNodeCount option or otherwise knows the
1215   // additional work that needs to be done to identify reachable nodes
1216   // by walking the flow graph and find the missing ones using
1217   // _dead_node_list.
1218 
1219   Unique_Node_List useful(comp_arena());
1220   // Get useful node list by walking the graph.
1221   identify_useful_nodes(useful);
1222 
1223   uint l_nodes = C->live_nodes();
1224   uint l_nodes_by_walk = useful.size();
1225 
1226   if (l_nodes != l_nodes_by_walk) {
1227     if (_log != NULL) {
1228       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1229       _log->stamp();
1230       _log->end_head();
1231     }
1232     VectorSet& useful_member_set = useful.member_set();
1233     int last_idx = l_nodes_by_walk;
1234     for (int i = 0; i < last_idx; i++) {
1235       if (useful_member_set.test(i)) {
1236         if (_dead_node_list.test(i)) {
1237           if (_log != NULL) {
1238             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1239           }
1240           if (PrintIdealNodeCount) {
1241             // Print the log message to tty
1242               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1243               useful.at(i)->dump();
1244           }
1245         }
1246       }
1247       else if (! _dead_node_list.test(i)) {
1248         if (_log != NULL) {
1249           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1250         }
1251         if (PrintIdealNodeCount) {
1252           // Print the log message to tty
1253           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1254         }
1255       }
1256     }
1257     if (_log != NULL) {
1258       _log->tail("mismatched_nodes");
1259     }
1260   }
1261 }
1262 void Compile::record_modified_node(Node* n) {
1263   if (_modified_nodes != NULL && !_inlining_incrementally &&
1264       n->outcnt() != 0 && !n->is_Con()) {
1265     _modified_nodes->push(n);
1266   }
1267 }
1268 
1269 void Compile::remove_modified_node(Node* n) {
1270   if (_modified_nodes != NULL) {
1271     _modified_nodes->remove(n);
1272   }
1273 }
1274 #endif
1275 
1276 #ifndef PRODUCT
1277 void Compile::verify_top(Node* tn) const {
1278   if (tn != NULL) {
1279     assert(tn->is_Con(), "top node must be a constant");
1280     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1281     assert(tn->in(0) != NULL, "must have live top node");
1282   }
1283 }
1284 #endif
1285 
1286 
1287 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1288 
1289 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1290   guarantee(arr != NULL, "");
1291   int num_blocks = arr->length();
1292   if (grow_by < num_blocks)  grow_by = num_blocks;
1293   int num_notes = grow_by * _node_notes_block_size;
1294   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1295   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1296   while (num_notes > 0) {
1297     arr->append(notes);
1298     notes     += _node_notes_block_size;
1299     num_notes -= _node_notes_block_size;
1300   }
1301   assert(num_notes == 0, "exact multiple, please");
1302 }
1303 
1304 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1305   if (source == NULL || dest == NULL)  return false;
1306 
1307   if (dest->is_Con())
1308     return false;               // Do not push debug info onto constants.
1309 
1310 #ifdef ASSERT
1311   // Leave a bread crumb trail pointing to the original node:
1312   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1313     dest->set_debug_orig(source);
1314   }
1315 #endif
1316 
1317   if (node_note_array() == NULL)
1318     return false;               // Not collecting any notes now.
1319 
1320   // This is a copy onto a pre-existing node, which may already have notes.
1321   // If both nodes have notes, do not overwrite any pre-existing notes.
1322   Node_Notes* source_notes = node_notes_at(source->_idx);
1323   if (source_notes == NULL || source_notes->is_clear())  return false;
1324   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1325   if (dest_notes == NULL || dest_notes->is_clear()) {
1326     return set_node_notes_at(dest->_idx, source_notes);
1327   }
1328 
1329   Node_Notes merged_notes = (*source_notes);
1330   // The order of operations here ensures that dest notes will win...
1331   merged_notes.update_from(dest_notes);
1332   return set_node_notes_at(dest->_idx, &merged_notes);
1333 }
1334 
1335 
1336 //--------------------------allow_range_check_smearing-------------------------
1337 // Gating condition for coalescing similar range checks.
1338 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1339 // single covering check that is at least as strong as any of them.
1340 // If the optimization succeeds, the simplified (strengthened) range check
1341 // will always succeed.  If it fails, we will deopt, and then give up
1342 // on the optimization.
1343 bool Compile::allow_range_check_smearing() const {
1344   // If this method has already thrown a range-check,
1345   // assume it was because we already tried range smearing
1346   // and it failed.
1347   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1348   return !already_trapped;
1349 }
1350 
1351 
1352 //------------------------------flatten_alias_type-----------------------------
1353 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1354   int offset = tj->offset();
1355   TypePtr::PTR ptr = tj->ptr();
1356 
1357   // Known instance (scalarizable allocation) alias only with itself.
1358   bool is_known_inst = tj->isa_oopptr() != NULL &&
1359                        tj->is_oopptr()->is_known_instance();
1360 
1361   // Process weird unsafe references.
1362   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1363     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1364     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1365     tj = TypeOopPtr::BOTTOM;
1366     ptr = tj->ptr();
1367     offset = tj->offset();
1368   }
1369 
1370   // Array pointers need some flattening
1371   const TypeAryPtr *ta = tj->isa_aryptr();
1372   if (ta && ta->is_stable()) {
1373     // Erase stability property for alias analysis.
1374     tj = ta = ta->cast_to_stable(false);
1375   }
1376   if( ta && is_known_inst ) {
1377     if ( offset != Type::OffsetBot &&
1378          offset > arrayOopDesc::length_offset_in_bytes() ) {
1379       offset = Type::OffsetBot; // Flatten constant access into array body only
1380       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1381     }
1382   } else if( ta && _AliasLevel >= 2 ) {
1383     // For arrays indexed by constant indices, we flatten the alias
1384     // space to include all of the array body.  Only the header, klass
1385     // and array length can be accessed un-aliased.
1386     if( offset != Type::OffsetBot ) {
1387       if( ta->const_oop() ) { // MethodData* or Method*
1388         offset = Type::OffsetBot;   // Flatten constant access into array body
1389         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1390       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1391         // range is OK as-is.
1392         tj = ta = TypeAryPtr::RANGE;
1393       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1394         tj = TypeInstPtr::KLASS; // all klass loads look alike
1395         ta = TypeAryPtr::RANGE; // generic ignored junk
1396         ptr = TypePtr::BotPTR;
1397       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1398         tj = TypeInstPtr::MARK;
1399         ta = TypeAryPtr::RANGE; // generic ignored junk
1400         ptr = TypePtr::BotPTR;
1401       } else {                  // Random constant offset into array body
1402         offset = Type::OffsetBot;   // Flatten constant access into array body
1403         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1404       }
1405     }
1406     // Arrays of fixed size alias with arrays of unknown size.
1407     if (ta->size() != TypeInt::POS) {
1408       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1409       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1410     }
1411     // Arrays of known objects become arrays of unknown objects.
1412     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1413       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1414       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1415     }
1416     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1417       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1418       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1419     }
1420     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1421     // cannot be distinguished by bytecode alone.
1422     if (ta->elem() == TypeInt::BOOL) {
1423       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1424       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1425       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1426     }
1427     // During the 2nd round of IterGVN, NotNull castings are removed.
1428     // Make sure the Bottom and NotNull variants alias the same.
1429     // Also, make sure exact and non-exact variants alias the same.
1430     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1431       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1432     }
1433   }
1434 
1435   // Oop pointers need some flattening
1436   const TypeInstPtr *to = tj->isa_instptr();
1437   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1438     ciInstanceKlass *k = to->klass()->as_instance_klass();
1439     if( ptr == TypePtr::Constant ) {
1440       if (to->klass() != ciEnv::current()->Class_klass() ||
1441           offset < k->size_helper() * wordSize) {
1442         // No constant oop pointers (such as Strings); they alias with
1443         // unknown strings.
1444         assert(!is_known_inst, "not scalarizable allocation");
1445         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1446       }
1447     } else if( is_known_inst ) {
1448       tj = to; // Keep NotNull and klass_is_exact for instance type
1449     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1450       // During the 2nd round of IterGVN, NotNull castings are removed.
1451       // Make sure the Bottom and NotNull variants alias the same.
1452       // Also, make sure exact and non-exact variants alias the same.
1453       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1454     }
1455     if (to->speculative() != NULL) {
1456       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1457     }
1458     // Canonicalize the holder of this field
1459     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1460       // First handle header references such as a LoadKlassNode, even if the
1461       // object's klass is unloaded at compile time (4965979).
1462       if (!is_known_inst) { // Do it only for non-instance types
1463         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1464       }
1465     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1466       // Static fields are in the space above the normal instance
1467       // fields in the java.lang.Class instance.
1468       if (to->klass() != ciEnv::current()->Class_klass()) {
1469         to = NULL;
1470         tj = TypeOopPtr::BOTTOM;
1471         offset = tj->offset();
1472       }
1473     } else {
1474       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1475       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1476         if( is_known_inst ) {
1477           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1478         } else {
1479           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1480         }
1481       }
1482     }
1483   }
1484 
1485   // Klass pointers to object array klasses need some flattening
1486   const TypeKlassPtr *tk = tj->isa_klassptr();
1487   if( tk ) {
1488     // If we are referencing a field within a Klass, we need
1489     // to assume the worst case of an Object.  Both exact and
1490     // inexact types must flatten to the same alias class so
1491     // use NotNull as the PTR.
1492     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1493 
1494       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1495                                    TypeKlassPtr::OBJECT->klass(),
1496                                    offset);
1497     }
1498 
1499     ciKlass* klass = tk->klass();
1500     if( klass->is_obj_array_klass() ) {
1501       ciKlass* k = TypeAryPtr::OOPS->klass();
1502       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1503         k = TypeInstPtr::BOTTOM->klass();
1504       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1505     }
1506 
1507     // Check for precise loads from the primary supertype array and force them
1508     // to the supertype cache alias index.  Check for generic array loads from
1509     // the primary supertype array and also force them to the supertype cache
1510     // alias index.  Since the same load can reach both, we need to merge
1511     // these 2 disparate memories into the same alias class.  Since the
1512     // primary supertype array is read-only, there's no chance of confusion
1513     // where we bypass an array load and an array store.
1514     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1515     if (offset == Type::OffsetBot ||
1516         (offset >= primary_supers_offset &&
1517          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1518         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1519       offset = in_bytes(Klass::secondary_super_cache_offset());
1520       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1521     }
1522   }
1523 
1524   // Flatten all Raw pointers together.
1525   if (tj->base() == Type::RawPtr)
1526     tj = TypeRawPtr::BOTTOM;
1527 
1528   if (tj->base() == Type::AnyPtr)
1529     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1530 
1531   // Flatten all to bottom for now
1532   switch( _AliasLevel ) {
1533   case 0:
1534     tj = TypePtr::BOTTOM;
1535     break;
1536   case 1:                       // Flatten to: oop, static, field or array
1537     switch (tj->base()) {
1538     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1539     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1540     case Type::AryPtr:   // do not distinguish arrays at all
1541     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1542     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1543     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1544     default: ShouldNotReachHere();
1545     }
1546     break;
1547   case 2:                       // No collapsing at level 2; keep all splits
1548   case 3:                       // No collapsing at level 3; keep all splits
1549     break;
1550   default:
1551     Unimplemented();
1552   }
1553 
1554   offset = tj->offset();
1555   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1556 
1557   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1558           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1559           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1560           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1561           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1562           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1563           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr)  ,
1564           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1565   assert( tj->ptr() != TypePtr::TopPTR &&
1566           tj->ptr() != TypePtr::AnyNull &&
1567           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1568 //    assert( tj->ptr() != TypePtr::Constant ||
1569 //            tj->base() == Type::RawPtr ||
1570 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1571 
1572   return tj;
1573 }
1574 
1575 void Compile::AliasType::Init(int i, const TypePtr* at) {
1576   _index = i;
1577   _adr_type = at;
1578   _field = NULL;
1579   _element = NULL;
1580   _is_rewritable = true; // default
1581   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1582   if (atoop != NULL && atoop->is_known_instance()) {
1583     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1584     _general_index = Compile::current()->get_alias_index(gt);
1585   } else {
1586     _general_index = 0;
1587   }
1588 }
1589 
1590 //---------------------------------print_on------------------------------------
1591 #ifndef PRODUCT
1592 void Compile::AliasType::print_on(outputStream* st) {
1593   if (index() < 10)
1594         st->print("@ <%d> ", index());
1595   else  st->print("@ <%d>",  index());
1596   st->print(is_rewritable() ? "   " : " RO");
1597   int offset = adr_type()->offset();
1598   if (offset == Type::OffsetBot)
1599         st->print(" +any");
1600   else  st->print(" +%-3d", offset);
1601   st->print(" in ");
1602   adr_type()->dump_on(st);
1603   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1604   if (field() != NULL && tjp) {
1605     if (tjp->klass()  != field()->holder() ||
1606         tjp->offset() != field()->offset_in_bytes()) {
1607       st->print(" != ");
1608       field()->print();
1609       st->print(" ***");
1610     }
1611   }
1612 }
1613 
1614 void print_alias_types() {
1615   Compile* C = Compile::current();
1616   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1617   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1618     C->alias_type(idx)->print_on(tty);
1619     tty->cr();
1620   }
1621 }
1622 #endif
1623 
1624 
1625 //----------------------------probe_alias_cache--------------------------------
1626 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1627   intptr_t key = (intptr_t) adr_type;
1628   key ^= key >> logAliasCacheSize;
1629   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1630 }
1631 
1632 
1633 //-----------------------------grow_alias_types--------------------------------
1634 void Compile::grow_alias_types() {
1635   const int old_ats  = _max_alias_types; // how many before?
1636   const int new_ats  = old_ats;          // how many more?
1637   const int grow_ats = old_ats+new_ats;  // how many now?
1638   _max_alias_types = grow_ats;
1639   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1640   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1641   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1642   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1643 }
1644 
1645 
1646 //--------------------------------find_alias_type------------------------------
1647 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1648   if (_AliasLevel == 0)
1649     return alias_type(AliasIdxBot);
1650 
1651   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1652   if (ace->_adr_type == adr_type) {
1653     return alias_type(ace->_index);
1654   }
1655 
1656   // Handle special cases.
1657   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1658   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1659 
1660   // Do it the slow way.
1661   const TypePtr* flat = flatten_alias_type(adr_type);
1662 
1663 #ifdef ASSERT
1664   assert(flat == flatten_alias_type(flat), "idempotent");
1665   assert(flat != TypePtr::BOTTOM,     "cannot alias-analyze an untyped ptr");
1666   if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1667     const TypeOopPtr* foop = flat->is_oopptr();
1668     // Scalarizable allocations have exact klass always.
1669     bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1670     const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1671     assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1672   }
1673   assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1674 #endif
1675 
1676   int idx = AliasIdxTop;
1677   for (int i = 0; i < num_alias_types(); i++) {
1678     if (alias_type(i)->adr_type() == flat) {
1679       idx = i;
1680       break;
1681     }
1682   }
1683 
1684   if (idx == AliasIdxTop) {
1685     if (no_create)  return NULL;
1686     // Grow the array if necessary.
1687     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1688     // Add a new alias type.
1689     idx = _num_alias_types++;
1690     _alias_types[idx]->Init(idx, flat);
1691     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1692     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1693     if (flat->isa_instptr()) {
1694       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1695           && flat->is_instptr()->klass() == env()->Class_klass())
1696         alias_type(idx)->set_rewritable(false);
1697     }
1698     if (flat->isa_aryptr()) {
1699 #ifdef ASSERT
1700       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1701       // (T_BYTE has the weakest alignment and size restrictions...)
1702       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1703 #endif
1704       if (flat->offset() == TypePtr::OffsetBot) {
1705         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1706       }
1707     }
1708     if (flat->isa_klassptr()) {
1709       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1710         alias_type(idx)->set_rewritable(false);
1711       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1712         alias_type(idx)->set_rewritable(false);
1713       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1714         alias_type(idx)->set_rewritable(false);
1715       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1716         alias_type(idx)->set_rewritable(false);
1717     }
1718     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1719     // but the base pointer type is not distinctive enough to identify
1720     // references into JavaThread.)
1721 
1722     // Check for final fields.
1723     const TypeInstPtr* tinst = flat->isa_instptr();
1724     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1725       ciField* field;
1726       if (tinst->const_oop() != NULL &&
1727           tinst->klass() == ciEnv::current()->Class_klass() &&
1728           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1729         // static field
1730         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1731         field = k->get_field_by_offset(tinst->offset(), true);
1732       } else {
1733         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1734         field = k->get_field_by_offset(tinst->offset(), false);
1735       }
1736       assert(field == NULL ||
1737              original_field == NULL ||
1738              (field->holder() == original_field->holder() &&
1739               field->offset() == original_field->offset() &&
1740               field->is_static() == original_field->is_static()), "wrong field?");
1741       // Set field() and is_rewritable() attributes.
1742       if (field != NULL)  alias_type(idx)->set_field(field);
1743     }
1744   }
1745 
1746   // Fill the cache for next time.
1747   ace->_adr_type = adr_type;
1748   ace->_index    = idx;
1749   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1750 
1751   // Might as well try to fill the cache for the flattened version, too.
1752   AliasCacheEntry* face = probe_alias_cache(flat);
1753   if (face->_adr_type == NULL) {
1754     face->_adr_type = flat;
1755     face->_index    = idx;
1756     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1757   }
1758 
1759   return alias_type(idx);
1760 }
1761 
1762 
1763 Compile::AliasType* Compile::alias_type(ciField* field) {
1764   const TypeOopPtr* t;
1765   if (field->is_static())
1766     t = TypeInstPtr::make(field->holder()->java_mirror());
1767   else
1768     t = TypeOopPtr::make_from_klass_raw(field->holder());
1769   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1770   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1771   return atp;
1772 }
1773 
1774 
1775 //------------------------------have_alias_type--------------------------------
1776 bool Compile::have_alias_type(const TypePtr* adr_type) {
1777   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1778   if (ace->_adr_type == adr_type) {
1779     return true;
1780   }
1781 
1782   // Handle special cases.
1783   if (adr_type == NULL)             return true;
1784   if (adr_type == TypePtr::BOTTOM)  return true;
1785 
1786   return find_alias_type(adr_type, true, NULL) != NULL;
1787 }
1788 
1789 //-----------------------------must_alias--------------------------------------
1790 // True if all values of the given address type are in the given alias category.
1791 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1792   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1793   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1794   if (alias_idx == AliasIdxTop)         return false; // the empty category
1795   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1796 
1797   // the only remaining possible overlap is identity
1798   int adr_idx = get_alias_index(adr_type);
1799   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1800   assert(adr_idx == alias_idx ||
1801          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1802           && adr_type                       != TypeOopPtr::BOTTOM),
1803          "should not be testing for overlap with an unsafe pointer");
1804   return adr_idx == alias_idx;
1805 }
1806 
1807 //------------------------------can_alias--------------------------------------
1808 // True if any values of the given address type are in the given alias category.
1809 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1810   if (alias_idx == AliasIdxTop)         return false; // the empty category
1811   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1812   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1813   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1814 
1815   // the only remaining possible overlap is identity
1816   int adr_idx = get_alias_index(adr_type);
1817   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1818   return adr_idx == alias_idx;
1819 }
1820 
1821 
1822 
1823 //---------------------------pop_warm_call-------------------------------------
1824 WarmCallInfo* Compile::pop_warm_call() {
1825   WarmCallInfo* wci = _warm_calls;
1826   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1827   return wci;
1828 }
1829 
1830 //----------------------------Inline_Warm--------------------------------------
1831 int Compile::Inline_Warm() {
1832   // If there is room, try to inline some more warm call sites.
1833   // %%% Do a graph index compaction pass when we think we're out of space?
1834   if (!InlineWarmCalls)  return 0;
1835 
1836   int calls_made_hot = 0;
1837   int room_to_grow   = NodeCountInliningCutoff - unique();
1838   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1839   int amount_grown   = 0;
1840   WarmCallInfo* call;
1841   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1842     int est_size = (int)call->size();
1843     if (est_size > (room_to_grow - amount_grown)) {
1844       // This one won't fit anyway.  Get rid of it.
1845       call->make_cold();
1846       continue;
1847     }
1848     call->make_hot();
1849     calls_made_hot++;
1850     amount_grown   += est_size;
1851     amount_to_grow -= est_size;
1852   }
1853 
1854   if (calls_made_hot > 0)  set_major_progress();
1855   return calls_made_hot;
1856 }
1857 
1858 
1859 //----------------------------Finish_Warm--------------------------------------
1860 void Compile::Finish_Warm() {
1861   if (!InlineWarmCalls)  return;
1862   if (failing())  return;
1863   if (warm_calls() == NULL)  return;
1864 
1865   // Clean up loose ends, if we are out of space for inlining.
1866   WarmCallInfo* call;
1867   while ((call = pop_warm_call()) != NULL) {
1868     call->make_cold();
1869   }
1870 }
1871 
1872 //---------------------cleanup_loop_predicates-----------------------
1873 // Remove the opaque nodes that protect the predicates so that all unused
1874 // checks and uncommon_traps will be eliminated from the ideal graph
1875 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1876   if (predicate_count()==0) return;
1877   for (int i = predicate_count(); i > 0; i--) {
1878     Node * n = predicate_opaque1_node(i-1);
1879     assert(n->Opcode() == Op_Opaque1, "must be");
1880     igvn.replace_node(n, n->in(1));
1881   }
1882   assert(predicate_count()==0, "should be clean!");
1883 }
1884 
1885 // StringOpts and late inlining of string methods
1886 void Compile::inline_string_calls(bool parse_time) {
1887   {
1888     // remove useless nodes to make the usage analysis simpler
1889     ResourceMark rm;
1890     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1891   }
1892 
1893   {
1894     ResourceMark rm;
1895     print_method(PHASE_BEFORE_STRINGOPTS, 3);
1896     PhaseStringOpts pso(initial_gvn(), for_igvn());
1897     print_method(PHASE_AFTER_STRINGOPTS, 3);
1898   }
1899 
1900   // now inline anything that we skipped the first time around
1901   if (!parse_time) {
1902     _late_inlines_pos = _late_inlines.length();
1903   }
1904 
1905   while (_string_late_inlines.length() > 0) {
1906     CallGenerator* cg = _string_late_inlines.pop();
1907     cg->do_late_inline();
1908     if (failing())  return;
1909   }
1910   _string_late_inlines.trunc_to(0);
1911 }
1912 
1913 // Late inlining of boxing methods
1914 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1915   if (_boxing_late_inlines.length() > 0) {
1916     assert(has_boxed_value(), "inconsistent");
1917 
1918     PhaseGVN* gvn = initial_gvn();
1919     set_inlining_incrementally(true);
1920 
1921     assert( igvn._worklist.size() == 0, "should be done with igvn" );
1922     for_igvn()->clear();
1923     gvn->replace_with(&igvn);
1924 
1925     _late_inlines_pos = _late_inlines.length();
1926 
1927     while (_boxing_late_inlines.length() > 0) {
1928       CallGenerator* cg = _boxing_late_inlines.pop();
1929       cg->do_late_inline();
1930       if (failing())  return;
1931     }
1932     _boxing_late_inlines.trunc_to(0);
1933 
1934     {
1935       ResourceMark rm;
1936       PhaseRemoveUseless pru(gvn, for_igvn());
1937     }
1938 
1939     igvn = PhaseIterGVN(gvn);
1940     igvn.optimize();
1941 
1942     set_inlining_progress(false);
1943     set_inlining_incrementally(false);
1944   }
1945 }
1946 
1947 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1948   assert(IncrementalInline, "incremental inlining should be on");
1949   PhaseGVN* gvn = initial_gvn();
1950 
1951   set_inlining_progress(false);
1952   for_igvn()->clear();
1953   gvn->replace_with(&igvn);
1954 
1955   int i = 0;
1956 
1957   for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1958     CallGenerator* cg = _late_inlines.at(i);
1959     _late_inlines_pos = i+1;
1960     cg->do_late_inline();
1961     if (failing())  return;
1962   }
1963   int j = 0;
1964   for (; i < _late_inlines.length(); i++, j++) {
1965     _late_inlines.at_put(j, _late_inlines.at(i));
1966   }
1967   _late_inlines.trunc_to(j);
1968 
1969   {
1970     ResourceMark rm;
1971     PhaseRemoveUseless pru(gvn, for_igvn());
1972   }
1973 
1974   igvn = PhaseIterGVN(gvn);
1975 }
1976 
1977 // Perform incremental inlining until bound on number of live nodes is reached
1978 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
1979   PhaseGVN* gvn = initial_gvn();
1980 
1981   set_inlining_incrementally(true);
1982   set_inlining_progress(true);
1983   uint low_live_nodes = 0;
1984 
1985   while(inlining_progress() && _late_inlines.length() > 0) {
1986 
1987     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1988       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
1989         // PhaseIdealLoop is expensive so we only try it once we are
1990         // out of live nodes and we only try it again if the previous
1991         // helped got the number of nodes down significantly
1992         PhaseIdealLoop ideal_loop( igvn, false, true );
1993         if (failing())  return;
1994         low_live_nodes = live_nodes();
1995         _major_progress = true;
1996       }
1997 
1998       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
1999         break;
2000       }
2001     }
2002 
2003     inline_incrementally_one(igvn);
2004 
2005     if (failing())  return;
2006 
2007     igvn.optimize();
2008 
2009     if (failing())  return;
2010   }
2011 
2012   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2013 
2014   if (_string_late_inlines.length() > 0) {
2015     assert(has_stringbuilder(), "inconsistent");
2016     for_igvn()->clear();
2017     initial_gvn()->replace_with(&igvn);
2018 
2019     inline_string_calls(false);
2020 
2021     if (failing())  return;
2022 
2023     {
2024       ResourceMark rm;
2025       PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2026     }
2027 
2028     igvn = PhaseIterGVN(gvn);
2029 
2030     igvn.optimize();
2031   }
2032 
2033   set_inlining_incrementally(false);
2034 }
2035 
2036 
2037 //------------------------------Optimize---------------------------------------
2038 // Given a graph, optimize it.
2039 void Compile::Optimize() {
2040   TracePhase t1("optimizer", &timers[_t_optimizer]);
2041 
2042 #ifndef PRODUCT
2043   if (env()->break_at_compile()) {
2044     BREAKPOINT;
2045   }
2046 
2047 #endif
2048 
2049   ResourceMark rm;
2050   int          loop_opts_cnt;
2051 
2052   print_inlining_reinit();
2053 
2054   NOT_PRODUCT( verify_graph_edges(); )
2055 
2056   print_method(PHASE_AFTER_PARSING);
2057 
2058  {
2059   // Iterative Global Value Numbering, including ideal transforms
2060   // Initialize IterGVN with types and values from parse-time GVN
2061   PhaseIterGVN igvn(initial_gvn());
2062 #ifdef ASSERT
2063   _modified_nodes = new (comp_arena()) Unique_Node_List(comp_arena());
2064 #endif
2065   {
2066     TracePhase t2("iterGVN", &timers[_t_iterGVN]);
2067     igvn.optimize();
2068   }
2069 
2070   print_method(PHASE_ITER_GVN1, 2);
2071 
2072   if (failing())  return;
2073 
2074   {
2075     TracePhase t2("incrementalInline", &timers[_t_incrInline]);
2076     inline_incrementally(igvn);
2077   }
2078 
2079   print_method(PHASE_INCREMENTAL_INLINE, 2);
2080 
2081   if (failing())  return;
2082 
2083   if (eliminate_boxing()) {
2084     TracePhase t2("incrementalInline", &timers[_t_incrInline]);
2085     // Inline valueOf() methods now.
2086     inline_boxing_calls(igvn);
2087 
2088     if (AlwaysIncrementalInline) {
2089       inline_incrementally(igvn);
2090     }
2091 
2092     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2093 
2094     if (failing())  return;
2095   }
2096 
2097   // Remove the speculative part of types and clean up the graph from
2098   // the extra CastPP nodes whose only purpose is to carry them. Do
2099   // that early so that optimizations are not disrupted by the extra
2100   // CastPP nodes.
2101   remove_speculative_types(igvn);
2102 
2103   // No more new expensive nodes will be added to the list from here
2104   // so keep only the actual candidates for optimizations.
2105   cleanup_expensive_nodes(igvn);
2106 
2107   // Perform escape analysis
2108   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2109     if (has_loops()) {
2110       // Cleanup graph (remove dead nodes).
2111       TracePhase t2("idealLoop", &timers[_t_idealLoop]);
2112       PhaseIdealLoop ideal_loop( igvn, false, true );
2113       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2114       if (failing())  return;
2115     }
2116     ConnectionGraph::do_analysis(this, &igvn);
2117 
2118     if (failing())  return;
2119 
2120     // Optimize out fields loads from scalar replaceable allocations.
2121     igvn.optimize();
2122     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2123 
2124     if (failing())  return;
2125 
2126     if (congraph() != NULL && macro_count() > 0) {
2127       TracePhase t2("macroEliminate", &timers[_t_macroEliminate]);
2128       PhaseMacroExpand mexp(igvn);
2129       mexp.eliminate_macro_nodes();
2130       igvn.set_delay_transform(false);
2131 
2132       igvn.optimize();
2133       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2134 
2135       if (failing())  return;
2136     }
2137   }
2138 
2139   // Loop transforms on the ideal graph.  Range Check Elimination,
2140   // peeling, unrolling, etc.
2141 
2142   // Set loop opts counter
2143   loop_opts_cnt = num_loop_opts();
2144   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2145     {
2146       TracePhase t2("idealLoop", &timers[_t_idealLoop]);
2147       PhaseIdealLoop ideal_loop( igvn, true );
2148       loop_opts_cnt--;
2149       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2150       if (failing())  return;
2151     }
2152     // Loop opts pass if partial peeling occurred in previous pass
2153     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2154       TracePhase t3("idealLoop", &timers[_t_idealLoop]);
2155       PhaseIdealLoop ideal_loop( igvn, false );
2156       loop_opts_cnt--;
2157       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2158       if (failing())  return;
2159     }
2160     // Loop opts pass for loop-unrolling before CCP
2161     if(major_progress() && (loop_opts_cnt > 0)) {
2162       TracePhase t4("idealLoop", &timers[_t_idealLoop]);
2163       PhaseIdealLoop ideal_loop( igvn, false );
2164       loop_opts_cnt--;
2165       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2166     }
2167     if (!failing()) {
2168       // Verify that last round of loop opts produced a valid graph
2169       TracePhase t2("idealLoopVerify", &timers[_t_idealLoopVerify]);
2170       PhaseIdealLoop::verify(igvn);
2171     }
2172   }
2173   if (failing())  return;
2174 
2175   // Conditional Constant Propagation;
2176   PhaseCCP ccp( &igvn );
2177   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2178   {
2179     TracePhase t2("ccp", &timers[_t_ccp]);
2180     ccp.do_transform();
2181   }
2182   print_method(PHASE_CPP1, 2);
2183 
2184   assert( true, "Break here to ccp.dump_old2new_map()");
2185 
2186   // Iterative Global Value Numbering, including ideal transforms
2187   {
2188     TracePhase t2("iterGVN2", &timers[_t_iterGVN2]);
2189     igvn = ccp;
2190     igvn.optimize();
2191   }
2192 
2193   print_method(PHASE_ITER_GVN2, 2);
2194 
2195   if (failing())  return;
2196 
2197   // Loop transforms on the ideal graph.  Range Check Elimination,
2198   // peeling, unrolling, etc.
2199   if(loop_opts_cnt > 0) {
2200     debug_only( int cnt = 0; );
2201     while(major_progress() && (loop_opts_cnt > 0)) {
2202       TracePhase t2("idealLoop", &timers[_t_idealLoop]);
2203       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2204       PhaseIdealLoop ideal_loop( igvn, true);
2205       loop_opts_cnt--;
2206       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2207       if (failing())  return;
2208     }
2209   }
2210 
2211   {
2212     // Verify that all previous optimizations produced a valid graph
2213     // at least to this point, even if no loop optimizations were done.
2214     TracePhase t2("idealLoopVerify", &timers[_t_idealLoopVerify]);
2215     PhaseIdealLoop::verify(igvn);
2216   }
2217 
2218   {
2219     TracePhase t2("macroExpand", &timers[_t_macroExpand]);
2220     PhaseMacroExpand  mex(igvn);
2221     if (mex.expand_macro_nodes()) {
2222       assert(failing(), "must bail out w/ explicit message");
2223       return;
2224     }
2225   }
2226 
2227   DEBUG_ONLY( _modified_nodes = NULL; )
2228  } // (End scope of igvn; run destructor if necessary for asserts.)
2229 
2230   process_print_inlining();
2231   // A method with only infinite loops has no edges entering loops from root
2232   {
2233     TracePhase t2("graphReshape", &timers[_t_graphReshaping]);
2234     if (final_graph_reshaping()) {
2235       assert(failing(), "must bail out w/ explicit message");
2236       return;
2237     }
2238   }
2239 
2240   print_method(PHASE_OPTIMIZE_FINISHED, 2);
2241 }
2242 
2243 
2244 //------------------------------Code_Gen---------------------------------------
2245 // Given a graph, generate code for it
2246 void Compile::Code_Gen() {
2247   if (failing()) {
2248     return;
2249   }
2250 
2251   // Perform instruction selection.  You might think we could reclaim Matcher
2252   // memory PDQ, but actually the Matcher is used in generating spill code.
2253   // Internals of the Matcher (including some VectorSets) must remain live
2254   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2255   // set a bit in reclaimed memory.
2256 
2257   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2258   // nodes.  Mapping is only valid at the root of each matched subtree.
2259   NOT_PRODUCT( verify_graph_edges(); )
2260 
2261   Matcher matcher;
2262   _matcher = &matcher;
2263   {
2264     TracePhase t2("matcher", &timers[_t_matcher]);
2265     matcher.match();
2266   }
2267   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2268   // nodes.  Mapping is only valid at the root of each matched subtree.
2269   NOT_PRODUCT( verify_graph_edges(); )
2270 
2271   // If you have too many nodes, or if matching has failed, bail out
2272   check_node_count(0, "out of nodes matching instructions");
2273   if (failing()) {
2274     return;
2275   }
2276 
2277   // Build a proper-looking CFG
2278   PhaseCFG cfg(node_arena(), root(), matcher);
2279   _cfg = &cfg;
2280   {
2281     TracePhase t2("scheduler", &timers[_t_scheduler]);
2282     bool success = cfg.do_global_code_motion();
2283     if (!success) {
2284       return;
2285     }
2286 
2287     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2288     NOT_PRODUCT( verify_graph_edges(); )
2289     debug_only( cfg.verify(); )
2290   }
2291 
2292   PhaseChaitin regalloc(unique(), cfg, matcher);
2293   _regalloc = &regalloc;
2294   {
2295     TracePhase t2("regalloc", &timers[_t_registerAllocation]);
2296     // Perform register allocation.  After Chaitin, use-def chains are
2297     // no longer accurate (at spill code) and so must be ignored.
2298     // Node->LRG->reg mappings are still accurate.
2299     _regalloc->Register_Allocate();
2300 
2301     // Bail out if the allocator builds too many nodes
2302     if (failing()) {
2303       return;
2304     }
2305   }
2306 
2307   // Prior to register allocation we kept empty basic blocks in case the
2308   // the allocator needed a place to spill.  After register allocation we
2309   // are not adding any new instructions.  If any basic block is empty, we
2310   // can now safely remove it.
2311   {
2312     TracePhase t2("blockOrdering", &timers[_t_blockOrdering]);
2313     cfg.remove_empty_blocks();
2314     if (do_freq_based_layout()) {
2315       PhaseBlockLayout layout(cfg);
2316     } else {
2317       cfg.set_loop_alignment();
2318     }
2319     cfg.fixup_flow();
2320   }
2321 
2322   // Apply peephole optimizations
2323   if( OptoPeephole ) {
2324     TracePhase t2("peephole", &timers[_t_peephole]);
2325     PhasePeephole peep( _regalloc, cfg);
2326     peep.do_transform();
2327   }
2328 
2329   // Do late expand if CPU requires this.
2330   if (Matcher::require_postalloc_expand) {
2331     TracePhase t2c("postalloc_expand", &timers[_t_postalloc_expand]);
2332     cfg.postalloc_expand(_regalloc);
2333   }
2334 
2335   // Convert Nodes to instruction bits in a buffer
2336   {
2337     TraceTime t2b("output", &timers[_t_output], CITime);
2338     Output();
2339   }
2340 
2341   print_method(PHASE_FINAL_CODE);
2342 
2343   // He's dead, Jim.
2344   _cfg     = (PhaseCFG*)0xdeadbeef;
2345   _regalloc = (PhaseChaitin*)0xdeadbeef;
2346 }
2347 
2348 
2349 //------------------------------dump_asm---------------------------------------
2350 // Dump formatted assembly
2351 #ifndef PRODUCT
2352 void Compile::dump_asm(int *pcs, uint pc_limit) {
2353   bool cut_short = false;
2354   tty->print_cr("#");
2355   tty->print("#  ");  _tf->dump();  tty->cr();
2356   tty->print_cr("#");
2357 
2358   // For all blocks
2359   int pc = 0x0;                 // Program counter
2360   char starts_bundle = ' ';
2361   _regalloc->dump_frame();
2362 
2363   Node *n = NULL;
2364   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2365     if (VMThread::should_terminate()) {
2366       cut_short = true;
2367       break;
2368     }
2369     Block* block = _cfg->get_block(i);
2370     if (block->is_connector() && !Verbose) {
2371       continue;
2372     }
2373     n = block->head();
2374     if (pcs && n->_idx < pc_limit) {
2375       tty->print("%3.3x   ", pcs[n->_idx]);
2376     } else {
2377       tty->print("      ");
2378     }
2379     block->dump_head(_cfg);
2380     if (block->is_connector()) {
2381       tty->print_cr("        # Empty connector block");
2382     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2383       tty->print_cr("        # Block is sole successor of call");
2384     }
2385 
2386     // For all instructions
2387     Node *delay = NULL;
2388     for (uint j = 0; j < block->number_of_nodes(); j++) {
2389       if (VMThread::should_terminate()) {
2390         cut_short = true;
2391         break;
2392       }
2393       n = block->get_node(j);
2394       if (valid_bundle_info(n)) {
2395         Bundle* bundle = node_bundling(n);
2396         if (bundle->used_in_unconditional_delay()) {
2397           delay = n;
2398           continue;
2399         }
2400         if (bundle->starts_bundle()) {
2401           starts_bundle = '+';
2402         }
2403       }
2404 
2405       if (WizardMode) {
2406         n->dump();
2407       }
2408 
2409       if( !n->is_Region() &&    // Dont print in the Assembly
2410           !n->is_Phi() &&       // a few noisely useless nodes
2411           !n->is_Proj() &&
2412           !n->is_MachTemp() &&
2413           !n->is_SafePointScalarObject() &&
2414           !n->is_Catch() &&     // Would be nice to print exception table targets
2415           !n->is_MergeMem() &&  // Not very interesting
2416           !n->is_top() &&       // Debug info table constants
2417           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2418           ) {
2419         if (pcs && n->_idx < pc_limit)
2420           tty->print("%3.3x", pcs[n->_idx]);
2421         else
2422           tty->print("   ");
2423         tty->print(" %c ", starts_bundle);
2424         starts_bundle = ' ';
2425         tty->print("\t");
2426         n->format(_regalloc, tty);
2427         tty->cr();
2428       }
2429 
2430       // If we have an instruction with a delay slot, and have seen a delay,
2431       // then back up and print it
2432       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2433         assert(delay != NULL, "no unconditional delay instruction");
2434         if (WizardMode) delay->dump();
2435 
2436         if (node_bundling(delay)->starts_bundle())
2437           starts_bundle = '+';
2438         if (pcs && n->_idx < pc_limit)
2439           tty->print("%3.3x", pcs[n->_idx]);
2440         else
2441           tty->print("   ");
2442         tty->print(" %c ", starts_bundle);
2443         starts_bundle = ' ';
2444         tty->print("\t");
2445         delay->format(_regalloc, tty);
2446         tty->cr();
2447         delay = NULL;
2448       }
2449 
2450       // Dump the exception table as well
2451       if( n->is_Catch() && (Verbose || WizardMode) ) {
2452         // Print the exception table for this offset
2453         _handler_table.print_subtable_for(pc);
2454       }
2455     }
2456 
2457     if (pcs && n->_idx < pc_limit)
2458       tty->print_cr("%3.3x", pcs[n->_idx]);
2459     else
2460       tty->cr();
2461 
2462     assert(cut_short || delay == NULL, "no unconditional delay branch");
2463 
2464   } // End of per-block dump
2465   tty->cr();
2466 
2467   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2468 }
2469 #endif
2470 
2471 //------------------------------Final_Reshape_Counts---------------------------
2472 // This class defines counters to help identify when a method
2473 // may/must be executed using hardware with only 24-bit precision.
2474 struct Final_Reshape_Counts : public StackObj {
2475   int  _call_count;             // count non-inlined 'common' calls
2476   int  _float_count;            // count float ops requiring 24-bit precision
2477   int  _double_count;           // count double ops requiring more precision
2478   int  _java_call_count;        // count non-inlined 'java' calls
2479   int  _inner_loop_count;       // count loops which need alignment
2480   VectorSet _visited;           // Visitation flags
2481   Node_List _tests;             // Set of IfNodes & PCTableNodes
2482 
2483   Final_Reshape_Counts() :
2484     _call_count(0), _float_count(0), _double_count(0),
2485     _java_call_count(0), _inner_loop_count(0),
2486     _visited( Thread::current()->resource_area() ) { }
2487 
2488   void inc_call_count  () { _call_count  ++; }
2489   void inc_float_count () { _float_count ++; }
2490   void inc_double_count() { _double_count++; }
2491   void inc_java_call_count() { _java_call_count++; }
2492   void inc_inner_loop_count() { _inner_loop_count++; }
2493 
2494   int  get_call_count  () const { return _call_count  ; }
2495   int  get_float_count () const { return _float_count ; }
2496   int  get_double_count() const { return _double_count; }
2497   int  get_java_call_count() const { return _java_call_count; }
2498   int  get_inner_loop_count() const { return _inner_loop_count; }
2499 };
2500 
2501 #ifdef ASSERT
2502 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2503   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2504   // Make sure the offset goes inside the instance layout.
2505   return k->contains_field_offset(tp->offset());
2506   // Note that OffsetBot and OffsetTop are very negative.
2507 }
2508 #endif
2509 
2510 // Eliminate trivially redundant StoreCMs and accumulate their
2511 // precedence edges.
2512 void Compile::eliminate_redundant_card_marks(Node* n) {
2513   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2514   if (n->in(MemNode::Address)->outcnt() > 1) {
2515     // There are multiple users of the same address so it might be
2516     // possible to eliminate some of the StoreCMs
2517     Node* mem = n->in(MemNode::Memory);
2518     Node* adr = n->in(MemNode::Address);
2519     Node* val = n->in(MemNode::ValueIn);
2520     Node* prev = n;
2521     bool done = false;
2522     // Walk the chain of StoreCMs eliminating ones that match.  As
2523     // long as it's a chain of single users then the optimization is
2524     // safe.  Eliminating partially redundant StoreCMs would require
2525     // cloning copies down the other paths.
2526     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2527       if (adr == mem->in(MemNode::Address) &&
2528           val == mem->in(MemNode::ValueIn)) {
2529         // redundant StoreCM
2530         if (mem->req() > MemNode::OopStore) {
2531           // Hasn't been processed by this code yet.
2532           n->add_prec(mem->in(MemNode::OopStore));
2533         } else {
2534           // Already converted to precedence edge
2535           for (uint i = mem->req(); i < mem->len(); i++) {
2536             // Accumulate any precedence edges
2537             if (mem->in(i) != NULL) {
2538               n->add_prec(mem->in(i));
2539             }
2540           }
2541           // Everything above this point has been processed.
2542           done = true;
2543         }
2544         // Eliminate the previous StoreCM
2545         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2546         assert(mem->outcnt() == 0, "should be dead");
2547         mem->disconnect_inputs(NULL, this);
2548       } else {
2549         prev = mem;
2550       }
2551       mem = prev->in(MemNode::Memory);
2552     }
2553   }
2554 }
2555 
2556 //------------------------------final_graph_reshaping_impl----------------------
2557 // Implement items 1-5 from final_graph_reshaping below.
2558 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2559 
2560   if ( n->outcnt() == 0 ) return; // dead node
2561   uint nop = n->Opcode();
2562 
2563   // Check for 2-input instruction with "last use" on right input.
2564   // Swap to left input.  Implements item (2).
2565   if( n->req() == 3 &&          // two-input instruction
2566       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2567       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2568       n->in(2)->outcnt() == 1 &&// right use IS a last use
2569       !n->in(2)->is_Con() ) {   // right use is not a constant
2570     // Check for commutative opcode
2571     switch( nop ) {
2572     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2573     case Op_MaxI:  case Op_MinI:
2574     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2575     case Op_AndL:  case Op_XorL:  case Op_OrL:
2576     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2577       // Move "last use" input to left by swapping inputs
2578       n->swap_edges(1, 2);
2579       break;
2580     }
2581     default:
2582       break;
2583     }
2584   }
2585 
2586 #ifdef ASSERT
2587   if( n->is_Mem() ) {
2588     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2589     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2590             // oop will be recorded in oop map if load crosses safepoint
2591             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2592                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2593             "raw memory operations should have control edge");
2594   }
2595 #endif
2596   // Count FPU ops and common calls, implements item (3)
2597   switch( nop ) {
2598   // Count all float operations that may use FPU
2599   case Op_AddF:
2600   case Op_SubF:
2601   case Op_MulF:
2602   case Op_DivF:
2603   case Op_NegF:
2604   case Op_ModF:
2605   case Op_ConvI2F:
2606   case Op_ConF:
2607   case Op_CmpF:
2608   case Op_CmpF3:
2609   // case Op_ConvL2F: // longs are split into 32-bit halves
2610     frc.inc_float_count();
2611     break;
2612 
2613   case Op_ConvF2D:
2614   case Op_ConvD2F:
2615     frc.inc_float_count();
2616     frc.inc_double_count();
2617     break;
2618 
2619   // Count all double operations that may use FPU
2620   case Op_AddD:
2621   case Op_SubD:
2622   case Op_MulD:
2623   case Op_DivD:
2624   case Op_NegD:
2625   case Op_ModD:
2626   case Op_ConvI2D:
2627   case Op_ConvD2I:
2628   // case Op_ConvL2D: // handled by leaf call
2629   // case Op_ConvD2L: // handled by leaf call
2630   case Op_ConD:
2631   case Op_CmpD:
2632   case Op_CmpD3:
2633     frc.inc_double_count();
2634     break;
2635   case Op_Opaque1:              // Remove Opaque Nodes before matching
2636   case Op_Opaque2:              // Remove Opaque Nodes before matching
2637   case Op_Opaque3:
2638     n->subsume_by(n->in(1), this);
2639     break;
2640   case Op_CallStaticJava:
2641   case Op_CallJava:
2642   case Op_CallDynamicJava:
2643     frc.inc_java_call_count(); // Count java call site;
2644   case Op_CallRuntime:
2645   case Op_CallLeaf:
2646   case Op_CallLeafNoFP: {
2647     assert( n->is_Call(), "" );
2648     CallNode *call = n->as_Call();
2649     // Count call sites where the FP mode bit would have to be flipped.
2650     // Do not count uncommon runtime calls:
2651     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2652     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2653     if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2654       frc.inc_call_count();   // Count the call site
2655     } else {                  // See if uncommon argument is shared
2656       Node *n = call->in(TypeFunc::Parms);
2657       int nop = n->Opcode();
2658       // Clone shared simple arguments to uncommon calls, item (1).
2659       if( n->outcnt() > 1 &&
2660           !n->is_Proj() &&
2661           nop != Op_CreateEx &&
2662           nop != Op_CheckCastPP &&
2663           nop != Op_DecodeN &&
2664           nop != Op_DecodeNKlass &&
2665           !n->is_Mem() ) {
2666         Node *x = n->clone();
2667         call->set_req( TypeFunc::Parms, x );
2668       }
2669     }
2670     break;
2671   }
2672 
2673   case Op_StoreD:
2674   case Op_LoadD:
2675   case Op_LoadD_unaligned:
2676     frc.inc_double_count();
2677     goto handle_mem;
2678   case Op_StoreF:
2679   case Op_LoadF:
2680     frc.inc_float_count();
2681     goto handle_mem;
2682 
2683   case Op_StoreCM:
2684     {
2685       // Convert OopStore dependence into precedence edge
2686       Node* prec = n->in(MemNode::OopStore);
2687       n->del_req(MemNode::OopStore);
2688       n->add_prec(prec);
2689       eliminate_redundant_card_marks(n);
2690     }
2691 
2692     // fall through
2693 
2694   case Op_StoreB:
2695   case Op_StoreC:
2696   case Op_StorePConditional:
2697   case Op_StoreI:
2698   case Op_StoreL:
2699   case Op_StoreIConditional:
2700   case Op_StoreLConditional:
2701   case Op_CompareAndSwapI:
2702   case Op_CompareAndSwapL:
2703   case Op_CompareAndSwapP:
2704   case Op_CompareAndSwapN:
2705   case Op_GetAndAddI:
2706   case Op_GetAndAddL:
2707   case Op_GetAndSetI:
2708   case Op_GetAndSetL:
2709   case Op_GetAndSetP:
2710   case Op_GetAndSetN:
2711   case Op_StoreP:
2712   case Op_StoreN:
2713   case Op_StoreNKlass:
2714   case Op_LoadB:
2715   case Op_LoadUB:
2716   case Op_LoadUS:
2717   case Op_LoadI:
2718   case Op_LoadKlass:
2719   case Op_LoadNKlass:
2720   case Op_LoadL:
2721   case Op_LoadL_unaligned:
2722   case Op_LoadPLocked:
2723   case Op_LoadP:
2724   case Op_LoadN:
2725   case Op_LoadRange:
2726   case Op_LoadS: {
2727   handle_mem:
2728 #ifdef ASSERT
2729     if( VerifyOptoOopOffsets ) {
2730       assert( n->is_Mem(), "" );
2731       MemNode *mem  = (MemNode*)n;
2732       // Check to see if address types have grounded out somehow.
2733       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2734       assert( !tp || oop_offset_is_sane(tp), "" );
2735     }
2736 #endif
2737     break;
2738   }
2739 
2740   case Op_AddP: {               // Assert sane base pointers
2741     Node *addp = n->in(AddPNode::Address);
2742     assert( !addp->is_AddP() ||
2743             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2744             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2745             "Base pointers must match" );
2746 #ifdef _LP64
2747     if ((UseCompressedOops || UseCompressedClassPointers) &&
2748         addp->Opcode() == Op_ConP &&
2749         addp == n->in(AddPNode::Base) &&
2750         n->in(AddPNode::Offset)->is_Con()) {
2751       // Use addressing with narrow klass to load with offset on x86.
2752       // On sparc loading 32-bits constant and decoding it have less
2753       // instructions (4) then load 64-bits constant (7).
2754       // Do this transformation here since IGVN will convert ConN back to ConP.
2755       const Type* t = addp->bottom_type();
2756       if (t->isa_oopptr() || t->isa_klassptr()) {
2757         Node* nn = NULL;
2758 
2759         int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2760 
2761         // Look for existing ConN node of the same exact type.
2762         Node* r  = root();
2763         uint cnt = r->outcnt();
2764         for (uint i = 0; i < cnt; i++) {
2765           Node* m = r->raw_out(i);
2766           if (m!= NULL && m->Opcode() == op &&
2767               m->bottom_type()->make_ptr() == t) {
2768             nn = m;
2769             break;
2770           }
2771         }
2772         if (nn != NULL) {
2773           // Decode a narrow oop to match address
2774           // [R12 + narrow_oop_reg<<3 + offset]
2775           if (t->isa_oopptr()) {
2776             nn = new DecodeNNode(nn, t);
2777           } else {
2778             nn = new DecodeNKlassNode(nn, t);
2779           }
2780           n->set_req(AddPNode::Base, nn);
2781           n->set_req(AddPNode::Address, nn);
2782           if (addp->outcnt() == 0) {
2783             addp->disconnect_inputs(NULL, this);
2784           }
2785         }
2786       }
2787     }
2788 #endif
2789     break;
2790   }
2791 
2792 #ifdef _LP64
2793   case Op_CastPP:
2794     if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2795       Node* in1 = n->in(1);
2796       const Type* t = n->bottom_type();
2797       Node* new_in1 = in1->clone();
2798       new_in1->as_DecodeN()->set_type(t);
2799 
2800       if (!Matcher::narrow_oop_use_complex_address()) {
2801         //
2802         // x86, ARM and friends can handle 2 adds in addressing mode
2803         // and Matcher can fold a DecodeN node into address by using
2804         // a narrow oop directly and do implicit NULL check in address:
2805         //
2806         // [R12 + narrow_oop_reg<<3 + offset]
2807         // NullCheck narrow_oop_reg
2808         //
2809         // On other platforms (Sparc) we have to keep new DecodeN node and
2810         // use it to do implicit NULL check in address:
2811         //
2812         // decode_not_null narrow_oop_reg, base_reg
2813         // [base_reg + offset]
2814         // NullCheck base_reg
2815         //
2816         // Pin the new DecodeN node to non-null path on these platform (Sparc)
2817         // to keep the information to which NULL check the new DecodeN node
2818         // corresponds to use it as value in implicit_null_check().
2819         //
2820         new_in1->set_req(0, n->in(0));
2821       }
2822 
2823       n->subsume_by(new_in1, this);
2824       if (in1->outcnt() == 0) {
2825         in1->disconnect_inputs(NULL, this);
2826       }
2827     }
2828     break;
2829 
2830   case Op_CmpP:
2831     // Do this transformation here to preserve CmpPNode::sub() and
2832     // other TypePtr related Ideal optimizations (for example, ptr nullness).
2833     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2834       Node* in1 = n->in(1);
2835       Node* in2 = n->in(2);
2836       if (!in1->is_DecodeNarrowPtr()) {
2837         in2 = in1;
2838         in1 = n->in(2);
2839       }
2840       assert(in1->is_DecodeNarrowPtr(), "sanity");
2841 
2842       Node* new_in2 = NULL;
2843       if (in2->is_DecodeNarrowPtr()) {
2844         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2845         new_in2 = in2->in(1);
2846       } else if (in2->Opcode() == Op_ConP) {
2847         const Type* t = in2->bottom_type();
2848         if (t == TypePtr::NULL_PTR) {
2849           assert(in1->is_DecodeN(), "compare klass to null?");
2850           // Don't convert CmpP null check into CmpN if compressed
2851           // oops implicit null check is not generated.
2852           // This will allow to generate normal oop implicit null check.
2853           if (Matcher::gen_narrow_oop_implicit_null_checks())
2854             new_in2 = ConNode::make(TypeNarrowOop::NULL_PTR);
2855           //
2856           // This transformation together with CastPP transformation above
2857           // will generated code for implicit NULL checks for compressed oops.
2858           //
2859           // The original code after Optimize()
2860           //
2861           //    LoadN memory, narrow_oop_reg
2862           //    decode narrow_oop_reg, base_reg
2863           //    CmpP base_reg, NULL
2864           //    CastPP base_reg // NotNull
2865           //    Load [base_reg + offset], val_reg
2866           //
2867           // after these transformations will be
2868           //
2869           //    LoadN memory, narrow_oop_reg
2870           //    CmpN narrow_oop_reg, NULL
2871           //    decode_not_null narrow_oop_reg, base_reg
2872           //    Load [base_reg + offset], val_reg
2873           //
2874           // and the uncommon path (== NULL) will use narrow_oop_reg directly
2875           // since narrow oops can be used in debug info now (see the code in
2876           // final_graph_reshaping_walk()).
2877           //
2878           // At the end the code will be matched to
2879           // on x86:
2880           //
2881           //    Load_narrow_oop memory, narrow_oop_reg
2882           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2883           //    NullCheck narrow_oop_reg
2884           //
2885           // and on sparc:
2886           //
2887           //    Load_narrow_oop memory, narrow_oop_reg
2888           //    decode_not_null narrow_oop_reg, base_reg
2889           //    Load [base_reg + offset], val_reg
2890           //    NullCheck base_reg
2891           //
2892         } else if (t->isa_oopptr()) {
2893           new_in2 = ConNode::make(t->make_narrowoop());
2894         } else if (t->isa_klassptr()) {
2895           new_in2 = ConNode::make(t->make_narrowklass());
2896         }
2897       }
2898       if (new_in2 != NULL) {
2899         Node* cmpN = new CmpNNode(in1->in(1), new_in2);
2900         n->subsume_by(cmpN, this);
2901         if (in1->outcnt() == 0) {
2902           in1->disconnect_inputs(NULL, this);
2903         }
2904         if (in2->outcnt() == 0) {
2905           in2->disconnect_inputs(NULL, this);
2906         }
2907       }
2908     }
2909     break;
2910 
2911   case Op_DecodeN:
2912   case Op_DecodeNKlass:
2913     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2914     // DecodeN could be pinned when it can't be fold into
2915     // an address expression, see the code for Op_CastPP above.
2916     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2917     break;
2918 
2919   case Op_EncodeP:
2920   case Op_EncodePKlass: {
2921     Node* in1 = n->in(1);
2922     if (in1->is_DecodeNarrowPtr()) {
2923       n->subsume_by(in1->in(1), this);
2924     } else if (in1->Opcode() == Op_ConP) {
2925       const Type* t = in1->bottom_type();
2926       if (t == TypePtr::NULL_PTR) {
2927         assert(t->isa_oopptr(), "null klass?");
2928         n->subsume_by(ConNode::make(TypeNarrowOop::NULL_PTR), this);
2929       } else if (t->isa_oopptr()) {
2930         n->subsume_by(ConNode::make(t->make_narrowoop()), this);
2931       } else if (t->isa_klassptr()) {
2932         n->subsume_by(ConNode::make(t->make_narrowklass()), this);
2933       }
2934     }
2935     if (in1->outcnt() == 0) {
2936       in1->disconnect_inputs(NULL, this);
2937     }
2938     break;
2939   }
2940 
2941   case Op_Proj: {
2942     if (OptimizeStringConcat) {
2943       ProjNode* p = n->as_Proj();
2944       if (p->_is_io_use) {
2945         // Separate projections were used for the exception path which
2946         // are normally removed by a late inline.  If it wasn't inlined
2947         // then they will hang around and should just be replaced with
2948         // the original one.
2949         Node* proj = NULL;
2950         // Replace with just one
2951         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2952           Node *use = i.get();
2953           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2954             proj = use;
2955             break;
2956           }
2957         }
2958         assert(proj != NULL, "must be found");
2959         p->subsume_by(proj, this);
2960       }
2961     }
2962     break;
2963   }
2964 
2965   case Op_Phi:
2966     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
2967       // The EncodeP optimization may create Phi with the same edges
2968       // for all paths. It is not handled well by Register Allocator.
2969       Node* unique_in = n->in(1);
2970       assert(unique_in != NULL, "");
2971       uint cnt = n->req();
2972       for (uint i = 2; i < cnt; i++) {
2973         Node* m = n->in(i);
2974         assert(m != NULL, "");
2975         if (unique_in != m)
2976           unique_in = NULL;
2977       }
2978       if (unique_in != NULL) {
2979         n->subsume_by(unique_in, this);
2980       }
2981     }
2982     break;
2983 
2984 #endif
2985 
2986   case Op_ModI:
2987     if (UseDivMod) {
2988       // Check if a%b and a/b both exist
2989       Node* d = n->find_similar(Op_DivI);
2990       if (d) {
2991         // Replace them with a fused divmod if supported
2992         if (Matcher::has_match_rule(Op_DivModI)) {
2993           DivModINode* divmod = DivModINode::make(n);
2994           d->subsume_by(divmod->div_proj(), this);
2995           n->subsume_by(divmod->mod_proj(), this);
2996         } else {
2997           // replace a%b with a-((a/b)*b)
2998           Node* mult = new MulINode(d, d->in(2));
2999           Node* sub  = new SubINode(d->in(1), mult);
3000           n->subsume_by(sub, this);
3001         }
3002       }
3003     }
3004     break;
3005 
3006   case Op_ModL:
3007     if (UseDivMod) {
3008       // Check if a%b and a/b both exist
3009       Node* d = n->find_similar(Op_DivL);
3010       if (d) {
3011         // Replace them with a fused divmod if supported
3012         if (Matcher::has_match_rule(Op_DivModL)) {
3013           DivModLNode* divmod = DivModLNode::make(n);
3014           d->subsume_by(divmod->div_proj(), this);
3015           n->subsume_by(divmod->mod_proj(), this);
3016         } else {
3017           // replace a%b with a-((a/b)*b)
3018           Node* mult = new MulLNode(d, d->in(2));
3019           Node* sub  = new SubLNode(d->in(1), mult);
3020           n->subsume_by(sub, this);
3021         }
3022       }
3023     }
3024     break;
3025 
3026   case Op_LoadVector:
3027   case Op_StoreVector:
3028     break;
3029 
3030   case Op_PackB:
3031   case Op_PackS:
3032   case Op_PackI:
3033   case Op_PackF:
3034   case Op_PackL:
3035   case Op_PackD:
3036     if (n->req()-1 > 2) {
3037       // Replace many operand PackNodes with a binary tree for matching
3038       PackNode* p = (PackNode*) n;
3039       Node* btp = p->binary_tree_pack(1, n->req());
3040       n->subsume_by(btp, this);
3041     }
3042     break;
3043   case Op_Loop:
3044   case Op_CountedLoop:
3045     if (n->as_Loop()->is_inner_loop()) {
3046       frc.inc_inner_loop_count();
3047     }
3048     break;
3049   case Op_LShiftI:
3050   case Op_RShiftI:
3051   case Op_URShiftI:
3052   case Op_LShiftL:
3053   case Op_RShiftL:
3054   case Op_URShiftL:
3055     if (Matcher::need_masked_shift_count) {
3056       // The cpu's shift instructions don't restrict the count to the
3057       // lower 5/6 bits. We need to do the masking ourselves.
3058       Node* in2 = n->in(2);
3059       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3060       const TypeInt* t = in2->find_int_type();
3061       if (t != NULL && t->is_con()) {
3062         juint shift = t->get_con();
3063         if (shift > mask) { // Unsigned cmp
3064           n->set_req(2, ConNode::make(TypeInt::make(shift & mask)));
3065         }
3066       } else {
3067         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3068           Node* shift = new AndINode(in2, ConNode::make(TypeInt::make(mask)));
3069           n->set_req(2, shift);
3070         }
3071       }
3072       if (in2->outcnt() == 0) { // Remove dead node
3073         in2->disconnect_inputs(NULL, this);
3074       }
3075     }
3076     break;
3077   case Op_MemBarStoreStore:
3078   case Op_MemBarRelease:
3079     // Break the link with AllocateNode: it is no longer useful and
3080     // confuses register allocation.
3081     if (n->req() > MemBarNode::Precedent) {
3082       n->set_req(MemBarNode::Precedent, top());
3083     }
3084     break;
3085   default:
3086     assert( !n->is_Call(), "" );
3087     assert( !n->is_Mem(), "" );
3088     break;
3089   }
3090 
3091   // Collect CFG split points
3092   if (n->is_MultiBranch())
3093     frc._tests.push(n);
3094 }
3095 
3096 //------------------------------final_graph_reshaping_walk---------------------
3097 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3098 // requires that the walk visits a node's inputs before visiting the node.
3099 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3100   ResourceArea *area = Thread::current()->resource_area();
3101   Unique_Node_List sfpt(area);
3102 
3103   frc._visited.set(root->_idx); // first, mark node as visited
3104   uint cnt = root->req();
3105   Node *n = root;
3106   uint  i = 0;
3107   while (true) {
3108     if (i < cnt) {
3109       // Place all non-visited non-null inputs onto stack
3110       Node* m = n->in(i);
3111       ++i;
3112       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3113         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3114           // compute worst case interpreter size in case of a deoptimization
3115           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3116 
3117           sfpt.push(m);
3118         }
3119         cnt = m->req();
3120         nstack.push(n, i); // put on stack parent and next input's index
3121         n = m;
3122         i = 0;
3123       }
3124     } else {
3125       // Now do post-visit work
3126       final_graph_reshaping_impl( n, frc );
3127       if (nstack.is_empty())
3128         break;             // finished
3129       n = nstack.node();   // Get node from stack
3130       cnt = n->req();
3131       i = nstack.index();
3132       nstack.pop();        // Shift to the next node on stack
3133     }
3134   }
3135 
3136   // Skip next transformation if compressed oops are not used.
3137   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3138       (!UseCompressedOops && !UseCompressedClassPointers))
3139     return;
3140 
3141   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3142   // It could be done for an uncommon traps or any safepoints/calls
3143   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3144   while (sfpt.size() > 0) {
3145     n = sfpt.pop();
3146     JVMState *jvms = n->as_SafePoint()->jvms();
3147     assert(jvms != NULL, "sanity");
3148     int start = jvms->debug_start();
3149     int end   = n->req();
3150     bool is_uncommon = (n->is_CallStaticJava() &&
3151                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3152     for (int j = start; j < end; j++) {
3153       Node* in = n->in(j);
3154       if (in->is_DecodeNarrowPtr()) {
3155         bool safe_to_skip = true;
3156         if (!is_uncommon ) {
3157           // Is it safe to skip?
3158           for (uint i = 0; i < in->outcnt(); i++) {
3159             Node* u = in->raw_out(i);
3160             if (!u->is_SafePoint() ||
3161                  u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3162               safe_to_skip = false;
3163             }
3164           }
3165         }
3166         if (safe_to_skip) {
3167           n->set_req(j, in->in(1));
3168         }
3169         if (in->outcnt() == 0) {
3170           in->disconnect_inputs(NULL, this);
3171         }
3172       }
3173     }
3174   }
3175 }
3176 
3177 //------------------------------final_graph_reshaping--------------------------
3178 // Final Graph Reshaping.
3179 //
3180 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3181 //     and not commoned up and forced early.  Must come after regular
3182 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3183 //     inputs to Loop Phis; these will be split by the allocator anyways.
3184 //     Remove Opaque nodes.
3185 // (2) Move last-uses by commutative operations to the left input to encourage
3186 //     Intel update-in-place two-address operations and better register usage
3187 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3188 //     calls canonicalizing them back.
3189 // (3) Count the number of double-precision FP ops, single-precision FP ops
3190 //     and call sites.  On Intel, we can get correct rounding either by
3191 //     forcing singles to memory (requires extra stores and loads after each
3192 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3193 //     clearing the mode bit around call sites).  The mode bit is only used
3194 //     if the relative frequency of single FP ops to calls is low enough.
3195 //     This is a key transform for SPEC mpeg_audio.
3196 // (4) Detect infinite loops; blobs of code reachable from above but not
3197 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3198 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3199 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3200 //     Detection is by looking for IfNodes where only 1 projection is
3201 //     reachable from below or CatchNodes missing some targets.
3202 // (5) Assert for insane oop offsets in debug mode.
3203 
3204 bool Compile::final_graph_reshaping() {
3205   // an infinite loop may have been eliminated by the optimizer,
3206   // in which case the graph will be empty.
3207   if (root()->req() == 1) {
3208     record_method_not_compilable("trivial infinite loop");
3209     return true;
3210   }
3211 
3212   // Expensive nodes have their control input set to prevent the GVN
3213   // from freely commoning them. There's no GVN beyond this point so
3214   // no need to keep the control input. We want the expensive nodes to
3215   // be freely moved to the least frequent code path by gcm.
3216   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3217   for (int i = 0; i < expensive_count(); i++) {
3218     _expensive_nodes->at(i)->set_req(0, NULL);
3219   }
3220 
3221   Final_Reshape_Counts frc;
3222 
3223   // Visit everybody reachable!
3224   // Allocate stack of size C->unique()/2 to avoid frequent realloc
3225   Node_Stack nstack(unique() >> 1);
3226   final_graph_reshaping_walk(nstack, root(), frc);
3227 
3228   // Check for unreachable (from below) code (i.e., infinite loops).
3229   for( uint i = 0; i < frc._tests.size(); i++ ) {
3230     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3231     // Get number of CFG targets.
3232     // Note that PCTables include exception targets after calls.
3233     uint required_outcnt = n->required_outcnt();
3234     if (n->outcnt() != required_outcnt) {
3235       // Check for a few special cases.  Rethrow Nodes never take the
3236       // 'fall-thru' path, so expected kids is 1 less.
3237       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3238         if (n->in(0)->in(0)->is_Call()) {
3239           CallNode *call = n->in(0)->in(0)->as_Call();
3240           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3241             required_outcnt--;      // Rethrow always has 1 less kid
3242           } else if (call->req() > TypeFunc::Parms &&
3243                      call->is_CallDynamicJava()) {
3244             // Check for null receiver. In such case, the optimizer has
3245             // detected that the virtual call will always result in a null
3246             // pointer exception. The fall-through projection of this CatchNode
3247             // will not be populated.
3248             Node *arg0 = call->in(TypeFunc::Parms);
3249             if (arg0->is_Type() &&
3250                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3251               required_outcnt--;
3252             }
3253           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3254                      call->req() > TypeFunc::Parms+1 &&
3255                      call->is_CallStaticJava()) {
3256             // Check for negative array length. In such case, the optimizer has
3257             // detected that the allocation attempt will always result in an
3258             // exception. There is no fall-through projection of this CatchNode .
3259             Node *arg1 = call->in(TypeFunc::Parms+1);
3260             if (arg1->is_Type() &&
3261                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3262               required_outcnt--;
3263             }
3264           }
3265         }
3266       }
3267       // Recheck with a better notion of 'required_outcnt'
3268       if (n->outcnt() != required_outcnt) {
3269         record_method_not_compilable("malformed control flow");
3270         return true;            // Not all targets reachable!
3271       }
3272     }
3273     // Check that I actually visited all kids.  Unreached kids
3274     // must be infinite loops.
3275     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3276       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3277         record_method_not_compilable("infinite loop");
3278         return true;            // Found unvisited kid; must be unreach
3279       }
3280   }
3281 
3282   // If original bytecodes contained a mixture of floats and doubles
3283   // check if the optimizer has made it homogenous, item (3).
3284   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3285       frc.get_float_count() > 32 &&
3286       frc.get_double_count() == 0 &&
3287       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3288     set_24_bit_selection_and_mode( false,  true );
3289   }
3290 
3291   set_java_calls(frc.get_java_call_count());
3292   set_inner_loops(frc.get_inner_loop_count());
3293 
3294   // No infinite loops, no reason to bail out.
3295   return false;
3296 }
3297 
3298 //-----------------------------too_many_traps----------------------------------
3299 // Report if there are too many traps at the current method and bci.
3300 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3301 bool Compile::too_many_traps(ciMethod* method,
3302                              int bci,
3303                              Deoptimization::DeoptReason reason) {
3304   ciMethodData* md = method->method_data();
3305   if (md->is_empty()) {
3306     // Assume the trap has not occurred, or that it occurred only
3307     // because of a transient condition during start-up in the interpreter.
3308     return false;
3309   }
3310   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3311   if (md->has_trap_at(bci, m, reason) != 0) {
3312     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3313     // Also, if there are multiple reasons, or if there is no per-BCI record,
3314     // assume the worst.
3315     if (log())
3316       log()->elem("observe trap='%s' count='%d'",
3317                   Deoptimization::trap_reason_name(reason),
3318                   md->trap_count(reason));
3319     return true;
3320   } else {
3321     // Ignore method/bci and see if there have been too many globally.
3322     return too_many_traps(reason, md);
3323   }
3324 }
3325 
3326 // Less-accurate variant which does not require a method and bci.
3327 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3328                              ciMethodData* logmd) {
3329   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3330     // Too many traps globally.
3331     // Note that we use cumulative trap_count, not just md->trap_count.
3332     if (log()) {
3333       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3334       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3335                   Deoptimization::trap_reason_name(reason),
3336                   mcount, trap_count(reason));
3337     }
3338     return true;
3339   } else {
3340     // The coast is clear.
3341     return false;
3342   }
3343 }
3344 
3345 //--------------------------too_many_recompiles--------------------------------
3346 // Report if there are too many recompiles at the current method and bci.
3347 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3348 // Is not eager to return true, since this will cause the compiler to use
3349 // Action_none for a trap point, to avoid too many recompilations.
3350 bool Compile::too_many_recompiles(ciMethod* method,
3351                                   int bci,
3352                                   Deoptimization::DeoptReason reason) {
3353   ciMethodData* md = method->method_data();
3354   if (md->is_empty()) {
3355     // Assume the trap has not occurred, or that it occurred only
3356     // because of a transient condition during start-up in the interpreter.
3357     return false;
3358   }
3359   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3360   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3361   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3362   Deoptimization::DeoptReason per_bc_reason
3363     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3364   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3365   if ((per_bc_reason == Deoptimization::Reason_none
3366        || md->has_trap_at(bci, m, reason) != 0)
3367       // The trap frequency measure we care about is the recompile count:
3368       && md->trap_recompiled_at(bci, m)
3369       && md->overflow_recompile_count() >= bc_cutoff) {
3370     // Do not emit a trap here if it has already caused recompilations.
3371     // Also, if there are multiple reasons, or if there is no per-BCI record,
3372     // assume the worst.
3373     if (log())
3374       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3375                   Deoptimization::trap_reason_name(reason),
3376                   md->trap_count(reason),
3377                   md->overflow_recompile_count());
3378     return true;
3379   } else if (trap_count(reason) != 0
3380              && decompile_count() >= m_cutoff) {
3381     // Too many recompiles globally, and we have seen this sort of trap.
3382     // Use cumulative decompile_count, not just md->decompile_count.
3383     if (log())
3384       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3385                   Deoptimization::trap_reason_name(reason),
3386                   md->trap_count(reason), trap_count(reason),
3387                   md->decompile_count(), decompile_count());
3388     return true;
3389   } else {
3390     // The coast is clear.
3391     return false;
3392   }
3393 }
3394 
3395 // Compute when not to trap. Used by matching trap based nodes and
3396 // NullCheck optimization.
3397 void Compile::set_allowed_deopt_reasons() {
3398   _allowed_reasons = 0;
3399   if (is_method_compilation()) {
3400     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3401       assert(rs < BitsPerInt, "recode bit map");
3402       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3403         _allowed_reasons |= nth_bit(rs);
3404       }
3405     }
3406   }
3407 }
3408 
3409 #ifndef PRODUCT
3410 //------------------------------verify_graph_edges---------------------------
3411 // Walk the Graph and verify that there is a one-to-one correspondence
3412 // between Use-Def edges and Def-Use edges in the graph.
3413 void Compile::verify_graph_edges(bool no_dead_code) {
3414   if (VerifyGraphEdges) {
3415     ResourceArea *area = Thread::current()->resource_area();
3416     Unique_Node_List visited(area);
3417     // Call recursive graph walk to check edges
3418     _root->verify_edges(visited);
3419     if (no_dead_code) {
3420       // Now make sure that no visited node is used by an unvisited node.
3421       bool dead_nodes = false;
3422       Unique_Node_List checked(area);
3423       while (visited.size() > 0) {
3424         Node* n = visited.pop();
3425         checked.push(n);
3426         for (uint i = 0; i < n->outcnt(); i++) {
3427           Node* use = n->raw_out(i);
3428           if (checked.member(use))  continue;  // already checked
3429           if (visited.member(use))  continue;  // already in the graph
3430           if (use->is_Con())        continue;  // a dead ConNode is OK
3431           // At this point, we have found a dead node which is DU-reachable.
3432           if (!dead_nodes) {
3433             tty->print_cr("*** Dead nodes reachable via DU edges:");
3434             dead_nodes = true;
3435           }
3436           use->dump(2);
3437           tty->print_cr("---");
3438           checked.push(use);  // No repeats; pretend it is now checked.
3439         }
3440       }
3441       assert(!dead_nodes, "using nodes must be reachable from root");
3442     }
3443   }
3444 }
3445 
3446 // Verify GC barriers consistency
3447 // Currently supported:
3448 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3449 void Compile::verify_barriers() {
3450   if (UseG1GC) {
3451     // Verify G1 pre-barriers
3452     const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3453 
3454     ResourceArea *area = Thread::current()->resource_area();
3455     Unique_Node_List visited(area);
3456     Node_List worklist(area);
3457     // We're going to walk control flow backwards starting from the Root
3458     worklist.push(_root);
3459     while (worklist.size() > 0) {
3460       Node* x = worklist.pop();
3461       if (x == NULL || x == top()) continue;
3462       if (visited.member(x)) {
3463         continue;
3464       } else {
3465         visited.push(x);
3466       }
3467 
3468       if (x->is_Region()) {
3469         for (uint i = 1; i < x->req(); i++) {
3470           worklist.push(x->in(i));
3471         }
3472       } else {
3473         worklist.push(x->in(0));
3474         // We are looking for the pattern:
3475         //                            /->ThreadLocal
3476         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3477         //              \->ConI(0)
3478         // We want to verify that the If and the LoadB have the same control
3479         // See GraphKit::g1_write_barrier_pre()
3480         if (x->is_If()) {
3481           IfNode *iff = x->as_If();
3482           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3483             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3484             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3485                 && cmp->in(1)->is_Load()) {
3486               LoadNode* load = cmp->in(1)->as_Load();
3487               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3488                   && load->in(2)->in(3)->is_Con()
3489                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3490 
3491                 Node* if_ctrl = iff->in(0);
3492                 Node* load_ctrl = load->in(0);
3493 
3494                 if (if_ctrl != load_ctrl) {
3495                   // Skip possible CProj->NeverBranch in infinite loops
3496                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3497                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3498                     if_ctrl = if_ctrl->in(0)->in(0);
3499                   }
3500                 }
3501                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3502               }
3503             }
3504           }
3505         }
3506       }
3507     }
3508   }
3509 }
3510 
3511 #endif
3512 
3513 // The Compile object keeps track of failure reasons separately from the ciEnv.
3514 // This is required because there is not quite a 1-1 relation between the
3515 // ciEnv and its compilation task and the Compile object.  Note that one
3516 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3517 // to backtrack and retry without subsuming loads.  Other than this backtracking
3518 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3519 // by the logic in C2Compiler.
3520 void Compile::record_failure(const char* reason) {
3521   if (log() != NULL) {
3522     log()->elem("failure reason='%s' phase='compile'", reason);
3523   }
3524   if (_failure_reason == NULL) {
3525     // Record the first failure reason.
3526     _failure_reason = reason;
3527   }
3528 
3529   EventCompilerFailure event;
3530   if (event.should_commit()) {
3531     event.set_compileID(Compile::compile_id());
3532     event.set_failure(reason);
3533     event.commit();
3534   }
3535 
3536   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3537     C->print_method(PHASE_FAILURE);
3538   }
3539   _root = NULL;  // flush the graph, too
3540 }
3541 
3542 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator)
3543   : TraceTime(name, accumulator, CITime, CITimeVerbose),
3544     _phase_name(name), _dolog(CITimeVerbose)
3545 {
3546   if (_dolog) {
3547     C = Compile::current();
3548     _log = C->log();
3549   } else {
3550     C = NULL;
3551     _log = NULL;
3552   }
3553   if (_log != NULL) {
3554     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3555     _log->stamp();
3556     _log->end_head();
3557   }
3558 }
3559 
3560 Compile::TracePhase::~TracePhase() {
3561 
3562   C = Compile::current();
3563   if (_dolog) {
3564     _log = C->log();
3565   } else {
3566     _log = NULL;
3567   }
3568 
3569 #ifdef ASSERT
3570   if (PrintIdealNodeCount) {
3571     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3572                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3573   }
3574 
3575   if (VerifyIdealNodeCount) {
3576     Compile::current()->print_missing_nodes();
3577   }
3578 #endif
3579 
3580   if (_log != NULL) {
3581     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3582   }
3583 }
3584 
3585 //=============================================================================
3586 // Two Constant's are equal when the type and the value are equal.
3587 bool Compile::Constant::operator==(const Constant& other) {
3588   if (type()          != other.type()         )  return false;
3589   if (can_be_reused() != other.can_be_reused())  return false;
3590   // For floating point values we compare the bit pattern.
3591   switch (type()) {
3592   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3593   case T_LONG:
3594   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
3595   case T_OBJECT:
3596   case T_ADDRESS: return (_v._value.l == other._v._value.l);
3597   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
3598   case T_METADATA: return (_v._metadata == other._v._metadata);
3599   default: ShouldNotReachHere();
3600   }
3601   return false;
3602 }
3603 
3604 static int type_to_size_in_bytes(BasicType t) {
3605   switch (t) {
3606   case T_LONG:    return sizeof(jlong  );
3607   case T_FLOAT:   return sizeof(jfloat );
3608   case T_DOUBLE:  return sizeof(jdouble);
3609   case T_METADATA: return sizeof(Metadata*);
3610     // We use T_VOID as marker for jump-table entries (labels) which
3611     // need an internal word relocation.
3612   case T_VOID:
3613   case T_ADDRESS:
3614   case T_OBJECT:  return sizeof(jobject);
3615   }
3616 
3617   ShouldNotReachHere();
3618   return -1;
3619 }
3620 
3621 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3622   // sort descending
3623   if (a->freq() > b->freq())  return -1;
3624   if (a->freq() < b->freq())  return  1;
3625   return 0;
3626 }
3627 
3628 void Compile::ConstantTable::calculate_offsets_and_size() {
3629   // First, sort the array by frequencies.
3630   _constants.sort(qsort_comparator);
3631 
3632 #ifdef ASSERT
3633   // Make sure all jump-table entries were sorted to the end of the
3634   // array (they have a negative frequency).
3635   bool found_void = false;
3636   for (int i = 0; i < _constants.length(); i++) {
3637     Constant con = _constants.at(i);
3638     if (con.type() == T_VOID)
3639       found_void = true;  // jump-tables
3640     else
3641       assert(!found_void, "wrong sorting");
3642   }
3643 #endif
3644 
3645   int offset = 0;
3646   for (int i = 0; i < _constants.length(); i++) {
3647     Constant* con = _constants.adr_at(i);
3648 
3649     // Align offset for type.
3650     int typesize = type_to_size_in_bytes(con->type());
3651     offset = align_size_up(offset, typesize);
3652     con->set_offset(offset);   // set constant's offset
3653 
3654     if (con->type() == T_VOID) {
3655       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3656       offset = offset + typesize * n->outcnt();  // expand jump-table
3657     } else {
3658       offset = offset + typesize;
3659     }
3660   }
3661 
3662   // Align size up to the next section start (which is insts; see
3663   // CodeBuffer::align_at_start).
3664   assert(_size == -1, "already set?");
3665   _size = align_size_up(offset, CodeEntryAlignment);
3666 }
3667 
3668 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3669   MacroAssembler _masm(&cb);
3670   for (int i = 0; i < _constants.length(); i++) {
3671     Constant con = _constants.at(i);
3672     address constant_addr;
3673     switch (con.type()) {
3674     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
3675     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3676     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3677     case T_OBJECT: {
3678       jobject obj = con.get_jobject();
3679       int oop_index = _masm.oop_recorder()->find_index(obj);
3680       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3681       break;
3682     }
3683     case T_ADDRESS: {
3684       address addr = (address) con.get_jobject();
3685       constant_addr = _masm.address_constant(addr);
3686       break;
3687     }
3688     // We use T_VOID as marker for jump-table entries (labels) which
3689     // need an internal word relocation.
3690     case T_VOID: {
3691       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3692       // Fill the jump-table with a dummy word.  The real value is
3693       // filled in later in fill_jump_table.
3694       address dummy = (address) n;
3695       constant_addr = _masm.address_constant(dummy);
3696       // Expand jump-table
3697       for (uint i = 1; i < n->outcnt(); i++) {
3698         address temp_addr = _masm.address_constant(dummy + i);
3699         assert(temp_addr, "consts section too small");
3700       }
3701       break;
3702     }
3703     case T_METADATA: {
3704       Metadata* obj = con.get_metadata();
3705       int metadata_index = _masm.oop_recorder()->find_index(obj);
3706       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3707       break;
3708     }
3709     default: ShouldNotReachHere();
3710     }
3711     assert(constant_addr, "consts section too small");
3712     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3713             err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3714   }
3715 }
3716 
3717 int Compile::ConstantTable::find_offset(Constant& con) const {
3718   int idx = _constants.find(con);
3719   assert(idx != -1, "constant must be in constant table");
3720   int offset = _constants.at(idx).offset();
3721   assert(offset != -1, "constant table not emitted yet?");
3722   return offset;
3723 }
3724 
3725 void Compile::ConstantTable::add(Constant& con) {
3726   if (con.can_be_reused()) {
3727     int idx = _constants.find(con);
3728     if (idx != -1 && _constants.at(idx).can_be_reused()) {
3729       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
3730       return;
3731     }
3732   }
3733   (void) _constants.append(con);
3734 }
3735 
3736 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3737   Block* b = Compile::current()->cfg()->get_block_for_node(n);
3738   Constant con(type, value, b->_freq);
3739   add(con);
3740   return con;
3741 }
3742 
3743 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3744   Constant con(metadata);
3745   add(con);
3746   return con;
3747 }
3748 
3749 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3750   jvalue value;
3751   BasicType type = oper->type()->basic_type();
3752   switch (type) {
3753   case T_LONG:    value.j = oper->constantL(); break;
3754   case T_FLOAT:   value.f = oper->constantF(); break;
3755   case T_DOUBLE:  value.d = oper->constantD(); break;
3756   case T_OBJECT:
3757   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3758   case T_METADATA: return add((Metadata*)oper->constant()); break;
3759   default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3760   }
3761   return add(n, type, value);
3762 }
3763 
3764 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3765   jvalue value;
3766   // We can use the node pointer here to identify the right jump-table
3767   // as this method is called from Compile::Fill_buffer right before
3768   // the MachNodes are emitted and the jump-table is filled (means the
3769   // MachNode pointers do not change anymore).
3770   value.l = (jobject) n;
3771   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
3772   add(con);
3773   return con;
3774 }
3775 
3776 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3777   // If called from Compile::scratch_emit_size do nothing.
3778   if (Compile::current()->in_scratch_emit_size())  return;
3779 
3780   assert(labels.is_nonempty(), "must be");
3781   assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3782 
3783   // Since MachConstantNode::constant_offset() also contains
3784   // table_base_offset() we need to subtract the table_base_offset()
3785   // to get the plain offset into the constant table.
3786   int offset = n->constant_offset() - table_base_offset();
3787 
3788   MacroAssembler _masm(&cb);
3789   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3790 
3791   for (uint i = 0; i < n->outcnt(); i++) {
3792     address* constant_addr = &jump_table_base[i];
3793     assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
3794     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3795     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3796   }
3797 }
3798 
3799 //----------------------------static_subtype_check-----------------------------
3800 // Shortcut important common cases when superklass is exact:
3801 // (0) superklass is java.lang.Object (can occur in reflective code)
3802 // (1) subklass is already limited to a subtype of superklass => always ok
3803 // (2) subklass does not overlap with superklass => always fail
3804 // (3) superklass has NO subtypes and we can check with a simple compare.
3805 int Compile::static_subtype_check(ciKlass* superk, ciKlass* subk) {
3806   if (StressReflectiveCode) {
3807     return SSC_full_test;       // Let caller generate the general case.
3808   }
3809 
3810   if (superk == env()->Object_klass()) {
3811     return SSC_always_true;     // (0) this test cannot fail
3812   }
3813 
3814   ciType* superelem = superk;
3815   if (superelem->is_array_klass())
3816     superelem = superelem->as_array_klass()->base_element_type();
3817 
3818   if (!subk->is_interface()) {  // cannot trust static interface types yet
3819     if (subk->is_subtype_of(superk)) {
3820       return SSC_always_true;   // (1) false path dead; no dynamic test needed
3821     }
3822     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
3823         !superk->is_subtype_of(subk)) {
3824       return SSC_always_false;
3825     }
3826   }
3827 
3828   // If casting to an instance klass, it must have no subtypes
3829   if (superk->is_interface()) {
3830     // Cannot trust interfaces yet.
3831     // %%% S.B. superk->nof_implementors() == 1
3832   } else if (superelem->is_instance_klass()) {
3833     ciInstanceKlass* ik = superelem->as_instance_klass();
3834     if (!ik->has_subklass() && !ik->is_interface()) {
3835       if (!ik->is_final()) {
3836         // Add a dependency if there is a chance of a later subclass.
3837         dependencies()->assert_leaf_type(ik);
3838       }
3839       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
3840     }
3841   } else {
3842     // A primitive array type has no subtypes.
3843     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
3844   }
3845 
3846   return SSC_full_test;
3847 }
3848 
3849 // The message about the current inlining is accumulated in
3850 // _print_inlining_stream and transfered into the _print_inlining_list
3851 // once we know whether inlining succeeds or not. For regular
3852 // inlining, messages are appended to the buffer pointed by
3853 // _print_inlining_idx in the _print_inlining_list. For late inlining,
3854 // a new buffer is added after _print_inlining_idx in the list. This
3855 // way we can update the inlining message for late inlining call site
3856 // when the inlining is attempted again.
3857 void Compile::print_inlining_init() {
3858   if (print_inlining() || print_intrinsics()) {
3859     _print_inlining_stream = new stringStream();
3860     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
3861   }
3862 }
3863 
3864 void Compile::print_inlining_reinit() {
3865   if (print_inlining() || print_intrinsics()) {
3866     // Re allocate buffer when we change ResourceMark
3867     _print_inlining_stream = new stringStream();
3868   }
3869 }
3870 
3871 void Compile::print_inlining_reset() {
3872   _print_inlining_stream->reset();
3873 }
3874 
3875 void Compile::print_inlining_commit() {
3876   assert(print_inlining() || print_intrinsics(), "PrintInlining off?");
3877   // Transfer the message from _print_inlining_stream to the current
3878   // _print_inlining_list buffer and clear _print_inlining_stream.
3879   _print_inlining_list->at(_print_inlining_idx).ss()->write(_print_inlining_stream->as_string(), _print_inlining_stream->size());
3880   print_inlining_reset();
3881 }
3882 
3883 void Compile::print_inlining_push() {
3884   // Add new buffer to the _print_inlining_list at current position
3885   _print_inlining_idx++;
3886   _print_inlining_list->insert_before(_print_inlining_idx, PrintInliningBuffer());
3887 }
3888 
3889 Compile::PrintInliningBuffer& Compile::print_inlining_current() {
3890   return _print_inlining_list->at(_print_inlining_idx);
3891 }
3892 
3893 void Compile::print_inlining_update(CallGenerator* cg) {
3894   if (print_inlining() || print_intrinsics()) {
3895     if (!cg->is_late_inline()) {
3896       if (print_inlining_current().cg() != NULL) {
3897         print_inlining_push();
3898       }
3899       print_inlining_commit();
3900     } else {
3901       if (print_inlining_current().cg() != cg &&
3902           (print_inlining_current().cg() != NULL ||
3903            print_inlining_current().ss()->size() != 0)) {
3904         print_inlining_push();
3905       }
3906       print_inlining_commit();
3907       print_inlining_current().set_cg(cg);
3908     }
3909   }
3910 }
3911 
3912 void Compile::print_inlining_move_to(CallGenerator* cg) {
3913   // We resume inlining at a late inlining call site. Locate the
3914   // corresponding inlining buffer so that we can update it.
3915   if (print_inlining()) {
3916     for (int i = 0; i < _print_inlining_list->length(); i++) {
3917       if (_print_inlining_list->adr_at(i)->cg() == cg) {
3918         _print_inlining_idx = i;
3919         return;
3920       }
3921     }
3922     ShouldNotReachHere();
3923   }
3924 }
3925 
3926 void Compile::print_inlining_update_delayed(CallGenerator* cg) {
3927   if (print_inlining()) {
3928     assert(_print_inlining_stream->size() > 0, "missing inlining msg");
3929     assert(print_inlining_current().cg() == cg, "wrong entry");
3930     // replace message with new message
3931     _print_inlining_list->at_put(_print_inlining_idx, PrintInliningBuffer());
3932     print_inlining_commit();
3933     print_inlining_current().set_cg(cg);
3934   }
3935 }
3936 
3937 void Compile::print_inlining_assert_ready() {
3938   assert(!_print_inlining || _print_inlining_stream->size() == 0, "loosing data");
3939 }
3940 
3941 void Compile::process_print_inlining() {
3942   bool do_print_inlining = print_inlining() || print_intrinsics();
3943   if (do_print_inlining || log() != NULL) {
3944     // Print inlining message for candidates that we couldn't inline
3945     // for lack of space
3946     for (int i = 0; i < _late_inlines.length(); i++) {
3947       CallGenerator* cg = _late_inlines.at(i);
3948       if (!cg->is_mh_late_inline()) {
3949         const char* msg = "live nodes > LiveNodeCountInliningCutoff";
3950         if (do_print_inlining) {
3951           cg->print_inlining_late(msg);
3952         }
3953         log_late_inline_failure(cg, msg);
3954       }
3955     }
3956   }
3957   if (do_print_inlining) {
3958     ResourceMark rm;
3959     stringStream ss;
3960     for (int i = 0; i < _print_inlining_list->length(); i++) {
3961       ss.print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3962     }
3963     size_t end = ss.size();
3964     _print_inlining_output = NEW_ARENA_ARRAY(comp_arena(), char, end+1);
3965     strncpy(_print_inlining_output, ss.base(), end+1);
3966     _print_inlining_output[end] = 0;
3967   }
3968 }
3969 
3970 void Compile::dump_print_inlining() {
3971   if (_print_inlining_output != NULL) {
3972     tty->print_raw(_print_inlining_output);
3973   }
3974 }
3975 
3976 void Compile::log_late_inline(CallGenerator* cg) {
3977   if (log() != NULL) {
3978     log()->head("late_inline method='%d'  inline_id='" JLONG_FORMAT "'", log()->identify(cg->method()),
3979                 cg->unique_id());
3980     JVMState* p = cg->call_node()->jvms();
3981     while (p != NULL) {
3982       log()->elem("jvms bci='%d' method='%d'", p->bci(), log()->identify(p->method()));
3983       p = p->caller();
3984     }
3985     log()->tail("late_inline");
3986   }
3987 }
3988 
3989 void Compile::log_late_inline_failure(CallGenerator* cg, const char* msg) {
3990   log_late_inline(cg);
3991   if (log() != NULL) {
3992     log()->inline_fail(msg);
3993   }
3994 }
3995 
3996 void Compile::log_inline_id(CallGenerator* cg) {
3997   if (log() != NULL) {
3998     // The LogCompilation tool needs a unique way to identify late
3999     // inline call sites. This id must be unique for this call site in
4000     // this compilation. Try to have it unique across compilations as
4001     // well because it can be convenient when grepping through the log
4002     // file.
4003     // Distinguish OSR compilations from others in case CICountOSR is
4004     // on.
4005     jlong id = ((jlong)unique()) + (((jlong)compile_id()) << 33) + (CICountOSR && is_osr_compilation() ? ((jlong)1) << 32 : 0);
4006     cg->set_unique_id(id);
4007     log()->elem("inline_id id='" JLONG_FORMAT "'", id);
4008   }
4009 }
4010 
4011 void Compile::log_inline_failure(const char* msg) {
4012   if (C->log() != NULL) {
4013     C->log()->inline_fail(msg);
4014   }
4015 }
4016 
4017 
4018 // Dump inlining replay data to the stream.
4019 // Don't change thread state and acquire any locks.
4020 void Compile::dump_inline_data(outputStream* out) {
4021   InlineTree* inl_tree = ilt();
4022   if (inl_tree != NULL) {
4023     out->print(" inline %d", inl_tree->count());
4024     inl_tree->dump_replay_data(out);
4025   }
4026 }
4027 
4028 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
4029   if (n1->Opcode() < n2->Opcode())      return -1;
4030   else if (n1->Opcode() > n2->Opcode()) return 1;
4031 
4032   assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
4033   for (uint i = 1; i < n1->req(); i++) {
4034     if (n1->in(i) < n2->in(i))      return -1;
4035     else if (n1->in(i) > n2->in(i)) return 1;
4036   }
4037 
4038   return 0;
4039 }
4040 
4041 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4042   Node* n1 = *n1p;
4043   Node* n2 = *n2p;
4044 
4045   return cmp_expensive_nodes(n1, n2);
4046 }
4047 
4048 void Compile::sort_expensive_nodes() {
4049   if (!expensive_nodes_sorted()) {
4050     _expensive_nodes->sort(cmp_expensive_nodes);
4051   }
4052 }
4053 
4054 bool Compile::expensive_nodes_sorted() const {
4055   for (int i = 1; i < _expensive_nodes->length(); i++) {
4056     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4057       return false;
4058     }
4059   }
4060   return true;
4061 }
4062 
4063 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4064   if (_expensive_nodes->length() == 0) {
4065     return false;
4066   }
4067 
4068   assert(OptimizeExpensiveOps, "optimization off?");
4069 
4070   // Take this opportunity to remove dead nodes from the list
4071   int j = 0;
4072   for (int i = 0; i < _expensive_nodes->length(); i++) {
4073     Node* n = _expensive_nodes->at(i);
4074     if (!n->is_unreachable(igvn)) {
4075       assert(n->is_expensive(), "should be expensive");
4076       _expensive_nodes->at_put(j, n);
4077       j++;
4078     }
4079   }
4080   _expensive_nodes->trunc_to(j);
4081 
4082   // Then sort the list so that similar nodes are next to each other
4083   // and check for at least two nodes of identical kind with same data
4084   // inputs.
4085   sort_expensive_nodes();
4086 
4087   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4088     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4089       return true;
4090     }
4091   }
4092 
4093   return false;
4094 }
4095 
4096 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4097   if (_expensive_nodes->length() == 0) {
4098     return;
4099   }
4100 
4101   assert(OptimizeExpensiveOps, "optimization off?");
4102 
4103   // Sort to bring similar nodes next to each other and clear the
4104   // control input of nodes for which there's only a single copy.
4105   sort_expensive_nodes();
4106 
4107   int j = 0;
4108   int identical = 0;
4109   int i = 0;
4110   bool modified = false;
4111   for (; i < _expensive_nodes->length()-1; i++) {
4112     assert(j <= i, "can't write beyond current index");
4113     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4114       identical++;
4115       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4116       continue;
4117     }
4118     if (identical > 0) {
4119       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4120       identical = 0;
4121     } else {
4122       Node* n = _expensive_nodes->at(i);
4123       igvn.replace_input_of(n, 0, NULL);
4124       igvn.hash_insert(n);
4125       modified = true;
4126     }
4127   }
4128   if (identical > 0) {
4129     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4130   } else if (_expensive_nodes->length() >= 1) {
4131     Node* n = _expensive_nodes->at(i);
4132     igvn.replace_input_of(n, 0, NULL);
4133     igvn.hash_insert(n);
4134     modified = true;
4135   }
4136   _expensive_nodes->trunc_to(j);
4137   if (modified) {
4138     igvn.optimize();
4139   }
4140 }
4141 
4142 void Compile::add_expensive_node(Node * n) {
4143   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4144   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4145   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4146   if (OptimizeExpensiveOps) {
4147     _expensive_nodes->append(n);
4148   } else {
4149     // Clear control input and let IGVN optimize expensive nodes if
4150     // OptimizeExpensiveOps is off.
4151     n->set_req(0, NULL);
4152   }
4153 }
4154 
4155 /**
4156  * Remove the speculative part of types and clean up the graph
4157  */
4158 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4159   if (UseTypeSpeculation) {
4160     Unique_Node_List worklist;
4161     worklist.push(root());
4162     int modified = 0;
4163     // Go over all type nodes that carry a speculative type, drop the
4164     // speculative part of the type and enqueue the node for an igvn
4165     // which may optimize it out.
4166     for (uint next = 0; next < worklist.size(); ++next) {
4167       Node *n  = worklist.at(next);
4168       if (n->is_Type()) {
4169         TypeNode* tn = n->as_Type();
4170         const Type* t = tn->type();
4171         const Type* t_no_spec = t->remove_speculative();
4172         if (t_no_spec != t) {
4173           bool in_hash = igvn.hash_delete(n);
4174           assert(in_hash, "node should be in igvn hash table");
4175           tn->set_type(t_no_spec);
4176           igvn.hash_insert(n);
4177           igvn._worklist.push(n); // give it a chance to go away
4178           modified++;
4179         }
4180       }
4181       uint max = n->len();
4182       for( uint i = 0; i < max; ++i ) {
4183         Node *m = n->in(i);
4184         if (not_a_node(m))  continue;
4185         worklist.push(m);
4186       }
4187     }
4188     // Drop the speculative part of all types in the igvn's type table
4189     igvn.remove_speculative_types();
4190     if (modified > 0) {
4191       igvn.optimize();
4192     }
4193 #ifdef ASSERT
4194     // Verify that after the IGVN is over no speculative type has resurfaced
4195     worklist.clear();
4196     worklist.push(root());
4197     for (uint next = 0; next < worklist.size(); ++next) {
4198       Node *n  = worklist.at(next);
4199       const Type* t = igvn.type_or_null(n);
4200       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4201       if (n->is_Type()) {
4202         t = n->as_Type()->type();
4203         assert(t == t->remove_speculative(), "no more speculative types");
4204       }
4205       uint max = n->len();
4206       for( uint i = 0; i < max; ++i ) {
4207         Node *m = n->in(i);
4208         if (not_a_node(m))  continue;
4209         worklist.push(m);
4210       }
4211     }
4212     igvn.check_no_speculative_types();
4213 #endif
4214   }
4215 }
4216 
4217 // Auxiliary method to support randomized stressing/fuzzing.
4218 //
4219 // This method can be called the arbitrary number of times, with current count
4220 // as the argument. The logic allows selecting a single candidate from the
4221 // running list of candidates as follows:
4222 //    int count = 0;
4223 //    Cand* selected = null;
4224 //    while(cand = cand->next()) {
4225 //      if (randomized_select(++count)) {
4226 //        selected = cand;
4227 //      }
4228 //    }
4229 //
4230 // Including count equalizes the chances any candidate is "selected".
4231 // This is useful when we don't have the complete list of candidates to choose
4232 // from uniformly. In this case, we need to adjust the randomicity of the
4233 // selection, or else we will end up biasing the selection towards the latter
4234 // candidates.
4235 //
4236 // Quick back-envelope calculation shows that for the list of n candidates
4237 // the equal probability for the candidate to persist as "best" can be
4238 // achieved by replacing it with "next" k-th candidate with the probability
4239 // of 1/k. It can be easily shown that by the end of the run, the
4240 // probability for any candidate is converged to 1/n, thus giving the
4241 // uniform distribution among all the candidates.
4242 //
4243 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4244 #define RANDOMIZED_DOMAIN_POW 29
4245 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4246 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4247 bool Compile::randomized_select(int count) {
4248   assert(count > 0, "only positive");
4249   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4250 }