1 /*
   2  * Copyright (c) 1997, 2011, 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/assembler.hpp"
  27 #include "classfile/systemDictionary.hpp"
  28 #include "code/exceptionHandlerTable.hpp"
  29 #include "code/nmethod.hpp"
  30 #include "compiler/compileLog.hpp"
  31 #include "compiler/oopMap.hpp"
  32 #include "opto/addnode.hpp"
  33 #include "opto/block.hpp"
  34 #include "opto/c2compiler.hpp"
  35 #include "opto/callGenerator.hpp"
  36 #include "opto/callnode.hpp"
  37 #include "opto/cfgnode.hpp"
  38 #include "opto/chaitin.hpp"
  39 #include "opto/compile.hpp"
  40 #include "opto/connode.hpp"
  41 #include "opto/divnode.hpp"
  42 #include "opto/escape.hpp"
  43 #include "opto/idealGraphPrinter.hpp"
  44 #include "opto/loopnode.hpp"
  45 #include "opto/machnode.hpp"
  46 #include "opto/macro.hpp"
  47 #include "opto/matcher.hpp"
  48 #include "opto/memnode.hpp"
  49 #include "opto/mulnode.hpp"
  50 #include "opto/node.hpp"
  51 #include "opto/opcodes.hpp"
  52 #include "opto/output.hpp"
  53 #include "opto/parse.hpp"
  54 #include "opto/phaseX.hpp"
  55 #include "opto/rootnode.hpp"
  56 #include "opto/runtime.hpp"
  57 #include "opto/stringopts.hpp"
  58 #include "opto/type.hpp"
  59 #include "opto/vectornode.hpp"
  60 #include "runtime/arguments.hpp"
  61 #include "runtime/signature.hpp"
  62 #include "runtime/stubRoutines.hpp"
  63 #include "runtime/timer.hpp"
  64 #include "utilities/copy.hpp"
  65 #ifdef TARGET_ARCH_MODEL_x86_32
  66 # include "adfiles/ad_x86_32.hpp"
  67 #endif
  68 #ifdef TARGET_ARCH_MODEL_x86_64
  69 # include "adfiles/ad_x86_64.hpp"
  70 #endif
  71 #ifdef TARGET_ARCH_MODEL_sparc
  72 # include "adfiles/ad_sparc.hpp"
  73 #endif
  74 #ifdef TARGET_ARCH_MODEL_zero
  75 # include "adfiles/ad_zero.hpp"
  76 #endif
  77 #ifdef TARGET_ARCH_MODEL_arm
  78 # include "adfiles/ad_arm.hpp"
  79 #endif
  80 #ifdef TARGET_ARCH_MODEL_ppc
  81 # include "adfiles/ad_ppc.hpp"
  82 #endif
  83 
  84 
  85 // -------------------- Compile::mach_constant_base_node -----------------------
  86 // Constant table base node singleton.
  87 MachConstantBaseNode* Compile::mach_constant_base_node() {
  88   if (_mach_constant_base_node == NULL) {
  89     _mach_constant_base_node = new (C) MachConstantBaseNode();
  90     _mach_constant_base_node->add_req(C->root());
  91   }
  92   return _mach_constant_base_node;
  93 }
  94 
  95 
  96 /// Support for intrinsics.
  97 
  98 // Return the index at which m must be inserted (or already exists).
  99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
 100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
 101 #ifdef ASSERT
 102   for (int i = 1; i < _intrinsics->length(); i++) {
 103     CallGenerator* cg1 = _intrinsics->at(i-1);
 104     CallGenerator* cg2 = _intrinsics->at(i);
 105     assert(cg1->method() != cg2->method()
 106            ? cg1->method()     < cg2->method()
 107            : cg1->is_virtual() < cg2->is_virtual(),
 108            "compiler intrinsics list must stay sorted");
 109   }
 110 #endif
 111   // Binary search sorted list, in decreasing intervals [lo, hi].
 112   int lo = 0, hi = _intrinsics->length()-1;
 113   while (lo <= hi) {
 114     int mid = (uint)(hi + lo) / 2;
 115     ciMethod* mid_m = _intrinsics->at(mid)->method();
 116     if (m < mid_m) {
 117       hi = mid-1;
 118     } else if (m > mid_m) {
 119       lo = mid+1;
 120     } else {
 121       // look at minor sort key
 122       bool mid_virt = _intrinsics->at(mid)->is_virtual();
 123       if (is_virtual < mid_virt) {
 124         hi = mid-1;
 125       } else if (is_virtual > mid_virt) {
 126         lo = mid+1;
 127       } else {
 128         return mid;  // exact match
 129       }
 130     }
 131   }
 132   return lo;  // inexact match
 133 }
 134 
 135 void Compile::register_intrinsic(CallGenerator* cg) {
 136   if (_intrinsics == NULL) {
 137     _intrinsics = new GrowableArray<CallGenerator*>(60);
 138   }
 139   // This code is stolen from ciObjectFactory::insert.
 140   // Really, GrowableArray should have methods for
 141   // insert_at, remove_at, and binary_search.
 142   int len = _intrinsics->length();
 143   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
 144   if (index == len) {
 145     _intrinsics->append(cg);
 146   } else {
 147 #ifdef ASSERT
 148     CallGenerator* oldcg = _intrinsics->at(index);
 149     assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
 150 #endif
 151     _intrinsics->append(_intrinsics->at(len-1));
 152     int pos;
 153     for (pos = len-2; pos >= index; pos--) {
 154       _intrinsics->at_put(pos+1,_intrinsics->at(pos));
 155     }
 156     _intrinsics->at_put(index, cg);
 157   }
 158   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 159 }
 160 
 161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 162   assert(m->is_loaded(), "don't try this on unloaded methods");
 163   if (_intrinsics != NULL) {
 164     int index = intrinsic_insertion_index(m, is_virtual);
 165     if (index < _intrinsics->length()
 166         && _intrinsics->at(index)->method() == m
 167         && _intrinsics->at(index)->is_virtual() == is_virtual) {
 168       return _intrinsics->at(index);
 169     }
 170   }
 171   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 172   if (m->intrinsic_id() != vmIntrinsics::_none &&
 173       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 174     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 175     if (cg != NULL) {
 176       // Save it for next time:
 177       register_intrinsic(cg);
 178       return cg;
 179     } else {
 180       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 181     }
 182   }
 183   return NULL;
 184 }
 185 
 186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 187 // in library_call.cpp.
 188 
 189 
 190 #ifndef PRODUCT
 191 // statistics gathering...
 192 
 193 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 195 
 196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 197   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 198   int oflags = _intrinsic_hist_flags[id];
 199   assert(flags != 0, "what happened?");
 200   if (is_virtual) {
 201     flags |= _intrinsic_virtual;
 202   }
 203   bool changed = (flags != oflags);
 204   if ((flags & _intrinsic_worked) != 0) {
 205     juint count = (_intrinsic_hist_count[id] += 1);
 206     if (count == 1) {
 207       changed = true;           // first time
 208     }
 209     // increment the overall count also:
 210     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 211   }
 212   if (changed) {
 213     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 214       // Something changed about the intrinsic's virtuality.
 215       if ((flags & _intrinsic_virtual) != 0) {
 216         // This is the first use of this intrinsic as a virtual call.
 217         if (oflags != 0) {
 218           // We already saw it as a non-virtual, so note both cases.
 219           flags |= _intrinsic_both;
 220         }
 221       } else if ((oflags & _intrinsic_both) == 0) {
 222         // This is the first use of this intrinsic as a non-virtual
 223         flags |= _intrinsic_both;
 224       }
 225     }
 226     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 227   }
 228   // update the overall flags also:
 229   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 230   return changed;
 231 }
 232 
 233 static char* format_flags(int flags, char* buf) {
 234   buf[0] = 0;
 235   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 236   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 237   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 238   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 239   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 240   if (buf[0] == 0)  strcat(buf, ",");
 241   assert(buf[0] == ',', "must be");
 242   return &buf[1];
 243 }
 244 
 245 void Compile::print_intrinsic_statistics() {
 246   char flagsbuf[100];
 247   ttyLocker ttyl;
 248   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 249   tty->print_cr("Compiler intrinsic usage:");
 250   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 251   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 252   #define PRINT_STAT_LINE(name, c, f) \
 253     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 254   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 255     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 256     int   flags = _intrinsic_hist_flags[id];
 257     juint count = _intrinsic_hist_count[id];
 258     if ((flags | count) != 0) {
 259       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 260     }
 261   }
 262   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 263   if (xtty != NULL)  xtty->tail("statistics");
 264 }
 265 
 266 void Compile::print_statistics() {
 267   { ttyLocker ttyl;
 268     if (xtty != NULL)  xtty->head("statistics type='opto'");
 269     Parse::print_statistics();
 270     PhaseCCP::print_statistics();
 271     PhaseRegAlloc::print_statistics();
 272     Scheduling::print_statistics();
 273     PhasePeephole::print_statistics();
 274     PhaseIdealLoop::print_statistics();
 275     if (xtty != NULL)  xtty->tail("statistics");
 276   }
 277   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 278     // put this under its own <statistics> element.
 279     print_intrinsic_statistics();
 280   }
 281 }
 282 #endif //PRODUCT
 283 
 284 // Support for bundling info
 285 Bundle* Compile::node_bundling(const Node *n) {
 286   assert(valid_bundle_info(n), "oob");
 287   return &_node_bundling_base[n->_idx];
 288 }
 289 
 290 bool Compile::valid_bundle_info(const Node *n) {
 291   return (_node_bundling_limit > n->_idx);
 292 }
 293 
 294 
 295 void Compile::gvn_replace_by(Node* n, Node* nn) {
 296   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 297     Node* use = n->last_out(i);
 298     bool is_in_table = initial_gvn()->hash_delete(use);
 299     uint uses_found = 0;
 300     for (uint j = 0; j < use->len(); j++) {
 301       if (use->in(j) == n) {
 302         if (j < use->req())
 303           use->set_req(j, nn);
 304         else
 305           use->set_prec(j, nn);
 306         uses_found++;
 307       }
 308     }
 309     if (is_in_table) {
 310       // reinsert into table
 311       initial_gvn()->hash_find_insert(use);
 312     }
 313     record_for_igvn(use);
 314     i -= uses_found;    // we deleted 1 or more copies of this edge
 315   }
 316 }
 317 
 318 
 319 
 320 
 321 // Identify all nodes that are reachable from below, useful.
 322 // Use breadth-first pass that records state in a Unique_Node_List,
 323 // recursive traversal is slower.
 324 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 325   int estimated_worklist_size = unique();
 326   useful.map( estimated_worklist_size, NULL );  // preallocate space
 327 
 328   // Initialize worklist
 329   if (root() != NULL)     { useful.push(root()); }
 330   // If 'top' is cached, declare it useful to preserve cached node
 331   if( cached_top_node() ) { useful.push(cached_top_node()); }
 332 
 333   // Push all useful nodes onto the list, breadthfirst
 334   for( uint next = 0; next < useful.size(); ++next ) {
 335     assert( next < unique(), "Unique useful nodes < total nodes");
 336     Node *n  = useful.at(next);
 337     uint max = n->len();
 338     for( uint i = 0; i < max; ++i ) {
 339       Node *m = n->in(i);
 340       if( m == NULL ) continue;
 341       useful.push(m);
 342     }
 343   }
 344 }
 345 
 346 // Disconnect all useless nodes by disconnecting those at the boundary.
 347 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 348   uint next = 0;
 349   while( next < useful.size() ) {
 350     Node *n = useful.at(next++);
 351     // Use raw traversal of out edges since this code removes out edges
 352     int max = n->outcnt();
 353     for (int j = 0; j < max; ++j ) {
 354       Node* child = n->raw_out(j);
 355       if( ! useful.member(child) ) {
 356         assert( !child->is_top() || child != top(),
 357                 "If top is cached in Compile object it is in useful list");
 358         // Only need to remove this out-edge to the useless node
 359         n->raw_del_out(j);
 360         --j;
 361         --max;
 362       }
 363     }
 364     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 365       record_for_igvn( n->unique_out() );
 366     }
 367   }
 368   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 369 }
 370 
 371 //------------------------------frame_size_in_words-----------------------------
 372 // frame_slots in units of words
 373 int Compile::frame_size_in_words() const {
 374   // shift is 0 in LP32 and 1 in LP64
 375   const int shift = (LogBytesPerWord - LogBytesPerInt);
 376   int words = _frame_slots >> shift;
 377   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 378   return words;
 379 }
 380 
 381 // ============================================================================
 382 //------------------------------CompileWrapper---------------------------------
 383 class CompileWrapper : public StackObj {
 384   Compile *const _compile;
 385  public:
 386   CompileWrapper(Compile* compile);
 387 
 388   ~CompileWrapper();
 389 };
 390 
 391 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 392   // the Compile* pointer is stored in the current ciEnv:
 393   ciEnv* env = compile->env();
 394   assert(env == ciEnv::current(), "must already be a ciEnv active");
 395   assert(env->compiler_data() == NULL, "compile already active?");
 396   env->set_compiler_data(compile);
 397   assert(compile == Compile::current(), "sanity");
 398 
 399   compile->set_type_dict(NULL);
 400   compile->set_type_hwm(NULL);
 401   compile->set_type_last_size(0);
 402   compile->set_last_tf(NULL, NULL);
 403   compile->set_indexSet_arena(NULL);
 404   compile->set_indexSet_free_block_list(NULL);
 405   compile->init_type_arena();
 406   Type::Initialize(compile);
 407   _compile->set_scratch_buffer_blob(NULL);
 408   _compile->begin_method();
 409 }
 410 CompileWrapper::~CompileWrapper() {
 411   _compile->end_method();
 412   if (_compile->scratch_buffer_blob() != NULL)
 413     BufferBlob::free(_compile->scratch_buffer_blob());
 414   _compile->env()->set_compiler_data(NULL);
 415 }
 416 
 417 
 418 //----------------------------print_compile_messages---------------------------
 419 void Compile::print_compile_messages() {
 420 #ifndef PRODUCT
 421   // Check if recompiling
 422   if (_subsume_loads == false && PrintOpto) {
 423     // Recompiling without allowing machine instructions to subsume loads
 424     tty->print_cr("*********************************************************");
 425     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 426     tty->print_cr("*********************************************************");
 427   }
 428   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 429     // Recompiling without escape analysis
 430     tty->print_cr("*********************************************************");
 431     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 432     tty->print_cr("*********************************************************");
 433   }
 434   if (env()->break_at_compile()) {
 435     // Open the debugger when compiling this method.
 436     tty->print("### Breaking when compiling: ");
 437     method()->print_short_name();
 438     tty->cr();
 439     BREAKPOINT;
 440   }
 441 
 442   if( PrintOpto ) {
 443     if (is_osr_compilation()) {
 444       tty->print("[OSR]%3d", _compile_id);
 445     } else {
 446       tty->print("%3d", _compile_id);
 447     }
 448   }
 449 #endif
 450 }
 451 
 452 
 453 //-----------------------init_scratch_buffer_blob------------------------------
 454 // Construct a temporary BufferBlob and cache it for this compile.
 455 void Compile::init_scratch_buffer_blob(int const_size) {
 456   // If there is already a scratch buffer blob allocated and the
 457   // constant section is big enough, use it.  Otherwise free the
 458   // current and allocate a new one.
 459   BufferBlob* blob = scratch_buffer_blob();
 460   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 461     // Use the current blob.
 462   } else {
 463     if (blob != NULL) {
 464       BufferBlob::free(blob);
 465     }
 466 
 467     ResourceMark rm;
 468     _scratch_const_size = const_size;
 469     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
 470     blob = BufferBlob::create("Compile::scratch_buffer", size);
 471     // Record the buffer blob for next time.
 472     set_scratch_buffer_blob(blob);
 473     // Have we run out of code space?
 474     if (scratch_buffer_blob() == NULL) {
 475       // Let CompilerBroker disable further compilations.
 476       record_failure("Not enough space for scratch buffer in CodeCache");
 477       return;
 478     }
 479   }
 480 
 481   // Initialize the relocation buffers
 482   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 483   set_scratch_locs_memory(locs_buf);
 484 }
 485 
 486 
 487 //-----------------------scratch_emit_size-------------------------------------
 488 // Helper function that computes size by emitting code
 489 uint Compile::scratch_emit_size(const Node* n) {
 490   // Start scratch_emit_size section.
 491   set_in_scratch_emit_size(true);
 492 
 493   // Emit into a trash buffer and count bytes emitted.
 494   // This is a pretty expensive way to compute a size,
 495   // but it works well enough if seldom used.
 496   // All common fixed-size instructions are given a size
 497   // method by the AD file.
 498   // Note that the scratch buffer blob and locs memory are
 499   // allocated at the beginning of the compile task, and
 500   // may be shared by several calls to scratch_emit_size.
 501   // The allocation of the scratch buffer blob is particularly
 502   // expensive, since it has to grab the code cache lock.
 503   BufferBlob* blob = this->scratch_buffer_blob();
 504   assert(blob != NULL, "Initialize BufferBlob at start");
 505   assert(blob->size() > MAX_inst_size, "sanity");
 506   relocInfo* locs_buf = scratch_locs_memory();
 507   address blob_begin = blob->content_begin();
 508   address blob_end   = (address)locs_buf;
 509   assert(blob->content_contains(blob_end), "sanity");
 510   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 511   buf.initialize_consts_size(_scratch_const_size);
 512   buf.initialize_stubs_size(MAX_stubs_size);
 513   assert(locs_buf != NULL, "sanity");
 514   int lsize = MAX_locs_size / 3;
 515   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 516   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 517   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 518 
 519   // Do the emission.
 520 
 521   Label fakeL; // Fake label for branch instructions.
 522   Label* saveL = NULL;
 523   uint save_bnum = 0;
 524   bool is_branch = n->is_MachBranch();
 525   if (is_branch) {
 526     MacroAssembler masm(&buf);
 527     masm.bind(fakeL);
 528     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 529     n->as_MachBranch()->label_set(&fakeL, 0);
 530   }
 531   n->emit(buf, this->regalloc());
 532   if (is_branch) // Restore label.
 533     n->as_MachBranch()->label_set(saveL, save_bnum);
 534 
 535   // End scratch_emit_size section.
 536   set_in_scratch_emit_size(false);
 537 
 538   return buf.insts_size();
 539 }
 540 
 541 
 542 // ============================================================================
 543 //------------------------------Compile standard-------------------------------
 544 debug_only( int Compile::_debug_idx = 100000; )
 545 
 546 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 547 // the continuation bci for on stack replacement.
 548 
 549 
 550 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
 551                 : Phase(Compiler),
 552                   _env(ci_env),
 553                   _log(ci_env->log()),
 554                   _compile_id(ci_env->compile_id()),
 555                   _save_argument_registers(false),
 556                   _stub_name(NULL),
 557                   _stub_function(NULL),
 558                   _stub_entry_point(NULL),
 559                   _method(target),
 560                   _entry_bci(osr_bci),
 561                   _initial_gvn(NULL),
 562                   _for_igvn(NULL),
 563                   _warm_calls(NULL),
 564                   _subsume_loads(subsume_loads),
 565                   _do_escape_analysis(do_escape_analysis),
 566                   _failure_reason(NULL),
 567                   _code_buffer("Compile::Fill_buffer"),
 568                   _orig_pc_slot(0),
 569                   _orig_pc_slot_offset_in_bytes(0),
 570                   _has_method_handle_invokes(false),
 571                   _mach_constant_base_node(NULL),
 572                   _node_bundling_limit(0),
 573                   _node_bundling_base(NULL),
 574                   _java_calls(0),
 575                   _inner_loops(0),
 576                   _scratch_const_size(-1),
 577                   _in_scratch_emit_size(false),
 578 #ifndef PRODUCT
 579                   _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
 580                   _printer(IdealGraphPrinter::printer()),
 581 #endif
 582                   _congraph(NULL) {
 583   C = this;
 584 
 585   CompileWrapper cw(this);
 586 #ifndef PRODUCT
 587   if (TimeCompiler2) {
 588     tty->print(" ");
 589     target->holder()->name()->print();
 590     tty->print(".");
 591     target->print_short_name();
 592     tty->print("  ");
 593   }
 594   TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
 595   TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
 596   bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
 597   if (!print_opto_assembly) {
 598     bool print_assembly = (PrintAssembly || _method->should_print_assembly());
 599     if (print_assembly && !Disassembler::can_decode()) {
 600       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 601       print_opto_assembly = true;
 602     }
 603   }
 604   set_print_assembly(print_opto_assembly);
 605   set_parsed_irreducible_loop(false);
 606 #endif
 607 
 608   if (ProfileTraps) {
 609     // Make sure the method being compiled gets its own MDO,
 610     // so we can at least track the decompile_count().
 611     method()->ensure_method_data();
 612   }
 613 
 614   Init(::AliasLevel);
 615 
 616 
 617   print_compile_messages();
 618 
 619   if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
 620     _ilt = InlineTree::build_inline_tree_root();
 621   else
 622     _ilt = NULL;
 623 
 624   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 625   assert(num_alias_types() >= AliasIdxRaw, "");
 626 
 627 #define MINIMUM_NODE_HASH  1023
 628   // Node list that Iterative GVN will start with
 629   Unique_Node_List for_igvn(comp_arena());
 630   set_for_igvn(&for_igvn);
 631 
 632   // GVN that will be run immediately on new nodes
 633   uint estimated_size = method()->code_size()*4+64;
 634   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 635   PhaseGVN gvn(node_arena(), estimated_size);
 636   set_initial_gvn(&gvn);
 637 
 638   { // Scope for timing the parser
 639     TracePhase t3("parse", &_t_parser, true);
 640 
 641     // Put top into the hash table ASAP.
 642     initial_gvn()->transform_no_reclaim(top());
 643 
 644     // Set up tf(), start(), and find a CallGenerator.
 645     CallGenerator* cg = NULL;
 646     if (is_osr_compilation()) {
 647       const TypeTuple *domain = StartOSRNode::osr_domain();
 648       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 649       init_tf(TypeFunc::make(domain, range));
 650       StartNode* s = new (this, 2) StartOSRNode(root(), domain);
 651       initial_gvn()->set_type_bottom(s);
 652       init_start(s);
 653       cg = CallGenerator::for_osr(method(), entry_bci());
 654     } else {
 655       // Normal case.
 656       init_tf(TypeFunc::make(method()));
 657       StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
 658       initial_gvn()->set_type_bottom(s);
 659       init_start(s);
 660       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
 661         // With java.lang.ref.reference.get() we must go through the
 662         // intrinsic when G1 is enabled - even when get() is the root
 663         // method of the compile - so that, if necessary, the value in
 664         // the referent field of the reference object gets recorded by
 665         // the pre-barrier code.
 666         // Specifically, if G1 is enabled, the value in the referent
 667         // field is recorded by the G1 SATB pre barrier. This will
 668         // result in the referent being marked live and the reference
 669         // object removed from the list of discovered references during
 670         // reference processing.
 671         cg = find_intrinsic(method(), false);
 672       }
 673       if (cg == NULL) {
 674         float past_uses = method()->interpreter_invocation_count();
 675         float expected_uses = past_uses;
 676         cg = CallGenerator::for_inline(method(), expected_uses);
 677       }
 678     }
 679     if (failing())  return;
 680     if (cg == NULL) {
 681       record_method_not_compilable_all_tiers("cannot parse method");
 682       return;
 683     }
 684     JVMState* jvms = build_start_state(start(), tf());
 685     if ((jvms = cg->generate(jvms)) == NULL) {
 686       record_method_not_compilable("method parse failed");
 687       return;
 688     }
 689     GraphKit kit(jvms);
 690 
 691     if (!kit.stopped()) {
 692       // Accept return values, and transfer control we know not where.
 693       // This is done by a special, unique ReturnNode bound to root.
 694       return_values(kit.jvms());
 695     }
 696 
 697     if (kit.has_exceptions()) {
 698       // Any exceptions that escape from this call must be rethrown
 699       // to whatever caller is dynamically above us on the stack.
 700       // This is done by a special, unique RethrowNode bound to root.
 701       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 702     }
 703 
 704     if (!failing() && has_stringbuilder()) {
 705       {
 706         // remove useless nodes to make the usage analysis simpler
 707         ResourceMark rm;
 708         PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 709       }
 710 
 711       {
 712         ResourceMark rm;
 713         print_method("Before StringOpts", 3);
 714         PhaseStringOpts pso(initial_gvn(), &for_igvn);
 715         print_method("After StringOpts", 3);
 716       }
 717 
 718       // now inline anything that we skipped the first time around
 719       while (_late_inlines.length() > 0) {
 720         CallGenerator* cg = _late_inlines.pop();
 721         cg->do_late_inline();
 722       }
 723     }
 724     assert(_late_inlines.length() == 0, "should have been processed");
 725 
 726     print_method("Before RemoveUseless", 3);
 727 
 728     // Remove clutter produced by parsing.
 729     if (!failing()) {
 730       ResourceMark rm;
 731       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 732     }
 733   }
 734 
 735   // Note:  Large methods are capped off in do_one_bytecode().
 736   if (failing())  return;
 737 
 738   // After parsing, node notes are no longer automagic.
 739   // They must be propagated by register_new_node_with_optimizer(),
 740   // clone(), or the like.
 741   set_default_node_notes(NULL);
 742 
 743   for (;;) {
 744     int successes = Inline_Warm();
 745     if (failing())  return;
 746     if (successes == 0)  break;
 747   }
 748 
 749   // Drain the list.
 750   Finish_Warm();
 751 #ifndef PRODUCT
 752   if (_printer) {
 753     _printer->print_inlining(this);
 754   }
 755 #endif
 756 
 757   if (failing())  return;
 758   NOT_PRODUCT( verify_graph_edges(); )
 759 
 760   // Now optimize
 761   Optimize();
 762   if (failing())  return;
 763   NOT_PRODUCT( verify_graph_edges(); )
 764 
 765 #ifndef PRODUCT
 766   if (PrintIdeal) {
 767     ttyLocker ttyl;  // keep the following output all in one block
 768     // This output goes directly to the tty, not the compiler log.
 769     // To enable tools to match it up with the compilation activity,
 770     // be sure to tag this tty output with the compile ID.
 771     if (xtty != NULL) {
 772       xtty->head("ideal compile_id='%d'%s", compile_id(),
 773                  is_osr_compilation()    ? " compile_kind='osr'" :
 774                  "");
 775     }
 776     root()->dump(9999);
 777     if (xtty != NULL) {
 778       xtty->tail("ideal");
 779     }
 780   }
 781 #endif
 782 
 783   // Now that we know the size of all the monitors we can add a fixed slot
 784   // for the original deopt pc.
 785 
 786   _orig_pc_slot =  fixed_slots();
 787   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 788   set_fixed_slots(next_slot);
 789 
 790   // Now generate code
 791   Code_Gen();
 792   if (failing())  return;
 793 
 794   // Check if we want to skip execution of all compiled code.
 795   {
 796 #ifndef PRODUCT
 797     if (OptoNoExecute) {
 798       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 799       return;
 800     }
 801     TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
 802 #endif
 803 
 804     if (is_osr_compilation()) {
 805       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 806       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 807     } else {
 808       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 809       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 810     }
 811 
 812     env()->register_method(_method, _entry_bci,
 813                            &_code_offsets,
 814                            _orig_pc_slot_offset_in_bytes,
 815                            code_buffer(),
 816                            frame_size_in_words(), _oop_map_set,
 817                            &_handler_table, &_inc_table,
 818                            compiler,
 819                            env()->comp_level(),
 820                            true, /*has_debug_info*/
 821                            has_unsafe_access()
 822                            );
 823   }
 824 }
 825 
 826 //------------------------------Compile----------------------------------------
 827 // Compile a runtime stub
 828 Compile::Compile( ciEnv* ci_env,
 829                   TypeFunc_generator generator,
 830                   address stub_function,
 831                   const char *stub_name,
 832                   int is_fancy_jump,
 833                   bool pass_tls,
 834                   bool save_arg_registers,
 835                   bool return_pc )
 836   : Phase(Compiler),
 837     _env(ci_env),
 838     _log(ci_env->log()),
 839     _compile_id(-1),
 840     _save_argument_registers(save_arg_registers),
 841     _method(NULL),
 842     _stub_name(stub_name),
 843     _stub_function(stub_function),
 844     _stub_entry_point(NULL),
 845     _entry_bci(InvocationEntryBci),
 846     _initial_gvn(NULL),
 847     _for_igvn(NULL),
 848     _warm_calls(NULL),
 849     _orig_pc_slot(0),
 850     _orig_pc_slot_offset_in_bytes(0),
 851     _subsume_loads(true),
 852     _do_escape_analysis(false),
 853     _failure_reason(NULL),
 854     _code_buffer("Compile::Fill_buffer"),
 855     _has_method_handle_invokes(false),
 856     _mach_constant_base_node(NULL),
 857     _node_bundling_limit(0),
 858     _node_bundling_base(NULL),
 859     _java_calls(0),
 860     _inner_loops(0),
 861 #ifndef PRODUCT
 862     _trace_opto_output(TraceOptoOutput),
 863     _printer(NULL),
 864 #endif
 865     _congraph(NULL) {
 866   C = this;
 867 
 868 #ifndef PRODUCT
 869   TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
 870   TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
 871   set_print_assembly(PrintFrameConverterAssembly);
 872   set_parsed_irreducible_loop(false);
 873 #endif
 874   CompileWrapper cw(this);
 875   Init(/*AliasLevel=*/ 0);
 876   init_tf((*generator)());
 877 
 878   {
 879     // The following is a dummy for the sake of GraphKit::gen_stub
 880     Unique_Node_List for_igvn(comp_arena());
 881     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
 882     PhaseGVN gvn(Thread::current()->resource_area(),255);
 883     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
 884     gvn.transform_no_reclaim(top());
 885 
 886     GraphKit kit;
 887     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
 888   }
 889 
 890   NOT_PRODUCT( verify_graph_edges(); )
 891   Code_Gen();
 892   if (failing())  return;
 893 
 894 
 895   // Entry point will be accessed using compile->stub_entry_point();
 896   if (code_buffer() == NULL) {
 897     Matcher::soft_match_failure();
 898   } else {
 899     if (PrintAssembly && (WizardMode || Verbose))
 900       tty->print_cr("### Stub::%s", stub_name);
 901 
 902     if (!failing()) {
 903       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
 904 
 905       // Make the NMethod
 906       // For now we mark the frame as never safe for profile stackwalking
 907       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
 908                                                       code_buffer(),
 909                                                       CodeOffsets::frame_never_safe,
 910                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
 911                                                       frame_size_in_words(),
 912                                                       _oop_map_set,
 913                                                       save_arg_registers);
 914       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
 915 
 916       _stub_entry_point = rs->entry_point();
 917     }
 918   }
 919 }
 920 
 921 #ifndef PRODUCT
 922 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
 923   if(PrintOpto && Verbose) {
 924     tty->print("%s   ", stub_name); j_sig->print_flattened(); tty->cr();
 925   }
 926 }
 927 #endif
 928 
 929 void Compile::print_codes() {
 930 }
 931 
 932 //------------------------------Init-------------------------------------------
 933 // Prepare for a single compilation
 934 void Compile::Init(int aliaslevel) {
 935   _unique  = 0;
 936   _regalloc = NULL;
 937 
 938   _tf      = NULL;  // filled in later
 939   _top     = NULL;  // cached later
 940   _matcher = NULL;  // filled in later
 941   _cfg     = NULL;  // filled in later
 942 
 943   set_24_bit_selection_and_mode(Use24BitFP, false);
 944 
 945   _node_note_array = NULL;
 946   _default_node_notes = NULL;
 947 
 948   _immutable_memory = NULL; // filled in at first inquiry
 949 
 950   // Globally visible Nodes
 951   // First set TOP to NULL to give safe behavior during creation of RootNode
 952   set_cached_top_node(NULL);
 953   set_root(new (this, 3) RootNode());
 954   // Now that you have a Root to point to, create the real TOP
 955   set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
 956   set_recent_alloc(NULL, NULL);
 957 
 958   // Create Debug Information Recorder to record scopes, oopmaps, etc.
 959   env()->set_oop_recorder(new OopRecorder(comp_arena()));
 960   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
 961   env()->set_dependencies(new Dependencies(env()));
 962 
 963   _fixed_slots = 0;
 964   set_has_split_ifs(false);
 965   set_has_loops(has_method() && method()->has_loops()); // first approximation
 966   set_has_stringbuilder(false);
 967   _trap_can_recompile = false;  // no traps emitted yet
 968   _major_progress = true; // start out assuming good things will happen
 969   set_has_unsafe_access(false);
 970   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
 971   set_decompile_count(0);
 972 
 973   set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
 974   set_num_loop_opts(LoopOptsCount);
 975   set_do_inlining(Inline);
 976   set_max_inline_size(MaxInlineSize);
 977   set_freq_inline_size(FreqInlineSize);
 978   set_do_scheduling(OptoScheduling);
 979   set_do_count_invocations(false);
 980   set_do_method_data_update(false);
 981 
 982   if (debug_info()->recording_non_safepoints()) {
 983     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
 984                         (comp_arena(), 8, 0, NULL));
 985     set_default_node_notes(Node_Notes::make(this));
 986   }
 987 
 988   // // -- Initialize types before each compile --
 989   // // Update cached type information
 990   // if( _method && _method->constants() )
 991   //   Type::update_loaded_types(_method, _method->constants());
 992 
 993   // Init alias_type map.
 994   if (!_do_escape_analysis && aliaslevel == 3)
 995     aliaslevel = 2;  // No unique types without escape analysis
 996   _AliasLevel = aliaslevel;
 997   const int grow_ats = 16;
 998   _max_alias_types = grow_ats;
 999   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1000   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1001   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1002   {
1003     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1004   }
1005   // Initialize the first few types.
1006   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1007   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1008   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1009   _num_alias_types = AliasIdxRaw+1;
1010   // Zero out the alias type cache.
1011   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1012   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1013   probe_alias_cache(NULL)->_index = AliasIdxTop;
1014 
1015   _intrinsics = NULL;
1016   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1017   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1018   register_library_intrinsics();
1019 }
1020 
1021 //---------------------------init_start----------------------------------------
1022 // Install the StartNode on this compile object.
1023 void Compile::init_start(StartNode* s) {
1024   if (failing())
1025     return; // already failing
1026   assert(s == start(), "");
1027 }
1028 
1029 StartNode* Compile::start() const {
1030   assert(!failing(), "");
1031   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1032     Node* start = root()->fast_out(i);
1033     if( start->is_Start() )
1034       return start->as_Start();
1035   }
1036   ShouldNotReachHere();
1037   return NULL;
1038 }
1039 
1040 //-------------------------------immutable_memory-------------------------------------
1041 // Access immutable memory
1042 Node* Compile::immutable_memory() {
1043   if (_immutable_memory != NULL) {
1044     return _immutable_memory;
1045   }
1046   StartNode* s = start();
1047   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1048     Node *p = s->fast_out(i);
1049     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1050       _immutable_memory = p;
1051       return _immutable_memory;
1052     }
1053   }
1054   ShouldNotReachHere();
1055   return NULL;
1056 }
1057 
1058 //----------------------set_cached_top_node------------------------------------
1059 // Install the cached top node, and make sure Node::is_top works correctly.
1060 void Compile::set_cached_top_node(Node* tn) {
1061   if (tn != NULL)  verify_top(tn);
1062   Node* old_top = _top;
1063   _top = tn;
1064   // Calling Node::setup_is_top allows the nodes the chance to adjust
1065   // their _out arrays.
1066   if (_top != NULL)     _top->setup_is_top();
1067   if (old_top != NULL)  old_top->setup_is_top();
1068   assert(_top == NULL || top()->is_top(), "");
1069 }
1070 
1071 #ifndef PRODUCT
1072 void Compile::verify_top(Node* tn) const {
1073   if (tn != NULL) {
1074     assert(tn->is_Con(), "top node must be a constant");
1075     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1076     assert(tn->in(0) != NULL, "must have live top node");
1077   }
1078 }
1079 #endif
1080 
1081 
1082 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1083 
1084 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1085   guarantee(arr != NULL, "");
1086   int num_blocks = arr->length();
1087   if (grow_by < num_blocks)  grow_by = num_blocks;
1088   int num_notes = grow_by * _node_notes_block_size;
1089   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1090   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1091   while (num_notes > 0) {
1092     arr->append(notes);
1093     notes     += _node_notes_block_size;
1094     num_notes -= _node_notes_block_size;
1095   }
1096   assert(num_notes == 0, "exact multiple, please");
1097 }
1098 
1099 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1100   if (source == NULL || dest == NULL)  return false;
1101 
1102   if (dest->is_Con())
1103     return false;               // Do not push debug info onto constants.
1104 
1105 #ifdef ASSERT
1106   // Leave a bread crumb trail pointing to the original node:
1107   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1108     dest->set_debug_orig(source);
1109   }
1110 #endif
1111 
1112   if (node_note_array() == NULL)
1113     return false;               // Not collecting any notes now.
1114 
1115   // This is a copy onto a pre-existing node, which may already have notes.
1116   // If both nodes have notes, do not overwrite any pre-existing notes.
1117   Node_Notes* source_notes = node_notes_at(source->_idx);
1118   if (source_notes == NULL || source_notes->is_clear())  return false;
1119   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1120   if (dest_notes == NULL || dest_notes->is_clear()) {
1121     return set_node_notes_at(dest->_idx, source_notes);
1122   }
1123 
1124   Node_Notes merged_notes = (*source_notes);
1125   // The order of operations here ensures that dest notes will win...
1126   merged_notes.update_from(dest_notes);
1127   return set_node_notes_at(dest->_idx, &merged_notes);
1128 }
1129 
1130 
1131 //--------------------------allow_range_check_smearing-------------------------
1132 // Gating condition for coalescing similar range checks.
1133 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1134 // single covering check that is at least as strong as any of them.
1135 // If the optimization succeeds, the simplified (strengthened) range check
1136 // will always succeed.  If it fails, we will deopt, and then give up
1137 // on the optimization.
1138 bool Compile::allow_range_check_smearing() const {
1139   // If this method has already thrown a range-check,
1140   // assume it was because we already tried range smearing
1141   // and it failed.
1142   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1143   return !already_trapped;
1144 }
1145 
1146 
1147 //------------------------------flatten_alias_type-----------------------------
1148 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1149   int offset = tj->offset();
1150   TypePtr::PTR ptr = tj->ptr();
1151 
1152   // Known instance (scalarizable allocation) alias only with itself.
1153   bool is_known_inst = tj->isa_oopptr() != NULL &&
1154                        tj->is_oopptr()->is_known_instance();
1155 
1156   // Process weird unsafe references.
1157   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1158     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1159     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1160     tj = TypeOopPtr::BOTTOM;
1161     ptr = tj->ptr();
1162     offset = tj->offset();
1163   }
1164 
1165   // Array pointers need some flattening
1166   const TypeAryPtr *ta = tj->isa_aryptr();
1167   if( ta && is_known_inst ) {
1168     if ( offset != Type::OffsetBot &&
1169          offset > arrayOopDesc::length_offset_in_bytes() ) {
1170       offset = Type::OffsetBot; // Flatten constant access into array body only
1171       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1172     }
1173   } else if( ta && _AliasLevel >= 2 ) {
1174     // For arrays indexed by constant indices, we flatten the alias
1175     // space to include all of the array body.  Only the header, klass
1176     // and array length can be accessed un-aliased.
1177     if( offset != Type::OffsetBot ) {
1178       if( ta->const_oop() ) { // methodDataOop or methodOop
1179         offset = Type::OffsetBot;   // Flatten constant access into array body
1180         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1181       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1182         // range is OK as-is.
1183         tj = ta = TypeAryPtr::RANGE;
1184       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1185         tj = TypeInstPtr::KLASS; // all klass loads look alike
1186         ta = TypeAryPtr::RANGE; // generic ignored junk
1187         ptr = TypePtr::BotPTR;
1188       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1189         tj = TypeInstPtr::MARK;
1190         ta = TypeAryPtr::RANGE; // generic ignored junk
1191         ptr = TypePtr::BotPTR;
1192       } else {                  // Random constant offset into array body
1193         offset = Type::OffsetBot;   // Flatten constant access into array body
1194         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1195       }
1196     }
1197     // Arrays of fixed size alias with arrays of unknown size.
1198     if (ta->size() != TypeInt::POS) {
1199       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1200       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1201     }
1202     // Arrays of known objects become arrays of unknown objects.
1203     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1204       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1205       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1206     }
1207     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1208       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1209       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1210     }
1211     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1212     // cannot be distinguished by bytecode alone.
1213     if (ta->elem() == TypeInt::BOOL) {
1214       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1215       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1216       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1217     }
1218     // During the 2nd round of IterGVN, NotNull castings are removed.
1219     // Make sure the Bottom and NotNull variants alias the same.
1220     // Also, make sure exact and non-exact variants alias the same.
1221     if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
1222       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1223     }
1224   }
1225 
1226   // Oop pointers need some flattening
1227   const TypeInstPtr *to = tj->isa_instptr();
1228   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1229     ciInstanceKlass *k = to->klass()->as_instance_klass();
1230     if( ptr == TypePtr::Constant ) {
1231       if (to->klass() != ciEnv::current()->Class_klass() ||
1232           offset < k->size_helper() * wordSize) {
1233         // No constant oop pointers (such as Strings); they alias with
1234         // unknown strings.
1235         assert(!is_known_inst, "not scalarizable allocation");
1236         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1237       }
1238     } else if( is_known_inst ) {
1239       tj = to; // Keep NotNull and klass_is_exact for instance type
1240     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1241       // During the 2nd round of IterGVN, NotNull castings are removed.
1242       // Make sure the Bottom and NotNull variants alias the same.
1243       // Also, make sure exact and non-exact variants alias the same.
1244       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1245     }
1246     // Canonicalize the holder of this field
1247     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1248       // First handle header references such as a LoadKlassNode, even if the
1249       // object's klass is unloaded at compile time (4965979).
1250       if (!is_known_inst) { // Do it only for non-instance types
1251         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1252       }
1253     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1254       // Static fields are in the space above the normal instance
1255       // fields in the java.lang.Class instance.
1256       if (to->klass() != ciEnv::current()->Class_klass()) {
1257         to = NULL;
1258         tj = TypeOopPtr::BOTTOM;
1259         offset = tj->offset();
1260       }
1261     } else {
1262       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1263       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1264         if( is_known_inst ) {
1265           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1266         } else {
1267           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1268         }
1269       }
1270     }
1271   }
1272 
1273   // Klass pointers to object array klasses need some flattening
1274   const TypeKlassPtr *tk = tj->isa_klassptr();
1275   if( tk ) {
1276     // If we are referencing a field within a Klass, we need
1277     // to assume the worst case of an Object.  Both exact and
1278     // inexact types must flatten to the same alias class.
1279     // Since the flattened result for a klass is defined to be
1280     // precisely java.lang.Object, use a constant ptr.
1281     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1282 
1283       tj = tk = TypeKlassPtr::make(TypePtr::Constant,
1284                                    TypeKlassPtr::OBJECT->klass(),
1285                                    offset);
1286     }
1287 
1288     ciKlass* klass = tk->klass();
1289     if( klass->is_obj_array_klass() ) {
1290       ciKlass* k = TypeAryPtr::OOPS->klass();
1291       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1292         k = TypeInstPtr::BOTTOM->klass();
1293       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1294     }
1295 
1296     // Check for precise loads from the primary supertype array and force them
1297     // to the supertype cache alias index.  Check for generic array loads from
1298     // the primary supertype array and also force them to the supertype cache
1299     // alias index.  Since the same load can reach both, we need to merge
1300     // these 2 disparate memories into the same alias class.  Since the
1301     // primary supertype array is read-only, there's no chance of confusion
1302     // where we bypass an array load and an array store.
1303     uint off2 = offset - Klass::primary_supers_offset_in_bytes();
1304     if( offset == Type::OffsetBot ||
1305         off2 < Klass::primary_super_limit()*wordSize ) {
1306       offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
1307       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1308     }
1309   }
1310 
1311   // Flatten all Raw pointers together.
1312   if (tj->base() == Type::RawPtr)
1313     tj = TypeRawPtr::BOTTOM;
1314 
1315   if (tj->base() == Type::AnyPtr)
1316     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1317 
1318   // Flatten all to bottom for now
1319   switch( _AliasLevel ) {
1320   case 0:
1321     tj = TypePtr::BOTTOM;
1322     break;
1323   case 1:                       // Flatten to: oop, static, field or array
1324     switch (tj->base()) {
1325     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1326     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1327     case Type::AryPtr:   // do not distinguish arrays at all
1328     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1329     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1330     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1331     default: ShouldNotReachHere();
1332     }
1333     break;
1334   case 2:                       // No collapsing at level 2; keep all splits
1335   case 3:                       // No collapsing at level 3; keep all splits
1336     break;
1337   default:
1338     Unimplemented();
1339   }
1340 
1341   offset = tj->offset();
1342   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1343 
1344   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1345           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1346           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1347           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1348           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1349           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1350           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr)  ,
1351           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1352   assert( tj->ptr() != TypePtr::TopPTR &&
1353           tj->ptr() != TypePtr::AnyNull &&
1354           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1355 //    assert( tj->ptr() != TypePtr::Constant ||
1356 //            tj->base() == Type::RawPtr ||
1357 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1358 
1359   return tj;
1360 }
1361 
1362 void Compile::AliasType::Init(int i, const TypePtr* at) {
1363   _index = i;
1364   _adr_type = at;
1365   _field = NULL;
1366   _is_rewritable = true; // default
1367   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1368   if (atoop != NULL && atoop->is_known_instance()) {
1369     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1370     _general_index = Compile::current()->get_alias_index(gt);
1371   } else {
1372     _general_index = 0;
1373   }
1374 }
1375 
1376 //---------------------------------print_on------------------------------------
1377 #ifndef PRODUCT
1378 void Compile::AliasType::print_on(outputStream* st) {
1379   if (index() < 10)
1380         st->print("@ <%d> ", index());
1381   else  st->print("@ <%d>",  index());
1382   st->print(is_rewritable() ? "   " : " RO");
1383   int offset = adr_type()->offset();
1384   if (offset == Type::OffsetBot)
1385         st->print(" +any");
1386   else  st->print(" +%-3d", offset);
1387   st->print(" in ");
1388   adr_type()->dump_on(st);
1389   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1390   if (field() != NULL && tjp) {
1391     if (tjp->klass()  != field()->holder() ||
1392         tjp->offset() != field()->offset_in_bytes()) {
1393       st->print(" != ");
1394       field()->print();
1395       st->print(" ***");
1396     }
1397   }
1398 }
1399 
1400 void print_alias_types() {
1401   Compile* C = Compile::current();
1402   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1403   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1404     C->alias_type(idx)->print_on(tty);
1405     tty->cr();
1406   }
1407 }
1408 #endif
1409 
1410 
1411 //----------------------------probe_alias_cache--------------------------------
1412 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1413   intptr_t key = (intptr_t) adr_type;
1414   key ^= key >> logAliasCacheSize;
1415   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1416 }
1417 
1418 
1419 //-----------------------------grow_alias_types--------------------------------
1420 void Compile::grow_alias_types() {
1421   const int old_ats  = _max_alias_types; // how many before?
1422   const int new_ats  = old_ats;          // how many more?
1423   const int grow_ats = old_ats+new_ats;  // how many now?
1424   _max_alias_types = grow_ats;
1425   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1426   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1427   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1428   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1429 }
1430 
1431 
1432 //--------------------------------find_alias_type------------------------------
1433 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1434   if (_AliasLevel == 0)
1435     return alias_type(AliasIdxBot);
1436 
1437   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1438   if (ace->_adr_type == adr_type) {
1439     return alias_type(ace->_index);
1440   }
1441 
1442   // Handle special cases.
1443   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1444   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1445 
1446   // Do it the slow way.
1447   const TypePtr* flat = flatten_alias_type(adr_type);
1448 
1449 #ifdef ASSERT
1450   assert(flat == flatten_alias_type(flat), "idempotent");
1451   assert(flat != TypePtr::BOTTOM,     "cannot alias-analyze an untyped ptr");
1452   if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1453     const TypeOopPtr* foop = flat->is_oopptr();
1454     // Scalarizable allocations have exact klass always.
1455     bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1456     const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1457     assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1458   }
1459   assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1460 #endif
1461 
1462   int idx = AliasIdxTop;
1463   for (int i = 0; i < num_alias_types(); i++) {
1464     if (alias_type(i)->adr_type() == flat) {
1465       idx = i;
1466       break;
1467     }
1468   }
1469 
1470   if (idx == AliasIdxTop) {
1471     if (no_create)  return NULL;
1472     // Grow the array if necessary.
1473     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1474     // Add a new alias type.
1475     idx = _num_alias_types++;
1476     _alias_types[idx]->Init(idx, flat);
1477     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1478     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1479     if (flat->isa_instptr()) {
1480       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1481           && flat->is_instptr()->klass() == env()->Class_klass())
1482         alias_type(idx)->set_rewritable(false);
1483     }
1484     if (flat->isa_klassptr()) {
1485       if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
1486         alias_type(idx)->set_rewritable(false);
1487       if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1488         alias_type(idx)->set_rewritable(false);
1489       if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
1490         alias_type(idx)->set_rewritable(false);
1491       if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
1492         alias_type(idx)->set_rewritable(false);
1493     }
1494     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1495     // but the base pointer type is not distinctive enough to identify
1496     // references into JavaThread.)
1497 
1498     // Check for final fields.
1499     const TypeInstPtr* tinst = flat->isa_instptr();
1500     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1501       ciField* field;
1502       if (tinst->const_oop() != NULL &&
1503           tinst->klass() == ciEnv::current()->Class_klass() &&
1504           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1505         // static field
1506         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1507         field = k->get_field_by_offset(tinst->offset(), true);
1508       } else {
1509         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1510         field = k->get_field_by_offset(tinst->offset(), false);
1511       }
1512       assert(field == NULL ||
1513              original_field == NULL ||
1514              (field->holder() == original_field->holder() &&
1515               field->offset() == original_field->offset() &&
1516               field->is_static() == original_field->is_static()), "wrong field?");
1517       // Set field() and is_rewritable() attributes.
1518       if (field != NULL)  alias_type(idx)->set_field(field);
1519     }
1520   }
1521 
1522   // Fill the cache for next time.
1523   ace->_adr_type = adr_type;
1524   ace->_index    = idx;
1525   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1526 
1527   // Might as well try to fill the cache for the flattened version, too.
1528   AliasCacheEntry* face = probe_alias_cache(flat);
1529   if (face->_adr_type == NULL) {
1530     face->_adr_type = flat;
1531     face->_index    = idx;
1532     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1533   }
1534 
1535   return alias_type(idx);
1536 }
1537 
1538 
1539 Compile::AliasType* Compile::alias_type(ciField* field) {
1540   const TypeOopPtr* t;
1541   if (field->is_static())
1542     t = TypeInstPtr::make(field->holder()->java_mirror());
1543   else
1544     t = TypeOopPtr::make_from_klass_raw(field->holder());
1545   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1546   assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
1547   return atp;
1548 }
1549 
1550 
1551 //------------------------------have_alias_type--------------------------------
1552 bool Compile::have_alias_type(const TypePtr* adr_type) {
1553   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1554   if (ace->_adr_type == adr_type) {
1555     return true;
1556   }
1557 
1558   // Handle special cases.
1559   if (adr_type == NULL)             return true;
1560   if (adr_type == TypePtr::BOTTOM)  return true;
1561 
1562   return find_alias_type(adr_type, true, NULL) != NULL;
1563 }
1564 
1565 //-----------------------------must_alias--------------------------------------
1566 // True if all values of the given address type are in the given alias category.
1567 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1568   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1569   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1570   if (alias_idx == AliasIdxTop)         return false; // the empty category
1571   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1572 
1573   // the only remaining possible overlap is identity
1574   int adr_idx = get_alias_index(adr_type);
1575   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1576   assert(adr_idx == alias_idx ||
1577          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1578           && adr_type                       != TypeOopPtr::BOTTOM),
1579          "should not be testing for overlap with an unsafe pointer");
1580   return adr_idx == alias_idx;
1581 }
1582 
1583 //------------------------------can_alias--------------------------------------
1584 // True if any values of the given address type are in the given alias category.
1585 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1586   if (alias_idx == AliasIdxTop)         return false; // the empty category
1587   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1588   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1589   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1590 
1591   // the only remaining possible overlap is identity
1592   int adr_idx = get_alias_index(adr_type);
1593   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1594   return adr_idx == alias_idx;
1595 }
1596 
1597 
1598 
1599 //---------------------------pop_warm_call-------------------------------------
1600 WarmCallInfo* Compile::pop_warm_call() {
1601   WarmCallInfo* wci = _warm_calls;
1602   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1603   return wci;
1604 }
1605 
1606 //----------------------------Inline_Warm--------------------------------------
1607 int Compile::Inline_Warm() {
1608   // If there is room, try to inline some more warm call sites.
1609   // %%% Do a graph index compaction pass when we think we're out of space?
1610   if (!InlineWarmCalls)  return 0;
1611 
1612   int calls_made_hot = 0;
1613   int room_to_grow   = NodeCountInliningCutoff - unique();
1614   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1615   int amount_grown   = 0;
1616   WarmCallInfo* call;
1617   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1618     int est_size = (int)call->size();
1619     if (est_size > (room_to_grow - amount_grown)) {
1620       // This one won't fit anyway.  Get rid of it.
1621       call->make_cold();
1622       continue;
1623     }
1624     call->make_hot();
1625     calls_made_hot++;
1626     amount_grown   += est_size;
1627     amount_to_grow -= est_size;
1628   }
1629 
1630   if (calls_made_hot > 0)  set_major_progress();
1631   return calls_made_hot;
1632 }
1633 
1634 
1635 //----------------------------Finish_Warm--------------------------------------
1636 void Compile::Finish_Warm() {
1637   if (!InlineWarmCalls)  return;
1638   if (failing())  return;
1639   if (warm_calls() == NULL)  return;
1640 
1641   // Clean up loose ends, if we are out of space for inlining.
1642   WarmCallInfo* call;
1643   while ((call = pop_warm_call()) != NULL) {
1644     call->make_cold();
1645   }
1646 }
1647 
1648 //---------------------cleanup_loop_predicates-----------------------
1649 // Remove the opaque nodes that protect the predicates so that all unused
1650 // checks and uncommon_traps will be eliminated from the ideal graph
1651 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1652   if (predicate_count()==0) return;
1653   for (int i = predicate_count(); i > 0; i--) {
1654     Node * n = predicate_opaque1_node(i-1);
1655     assert(n->Opcode() == Op_Opaque1, "must be");
1656     igvn.replace_node(n, n->in(1));
1657   }
1658   assert(predicate_count()==0, "should be clean!");
1659 }
1660 
1661 //------------------------------Optimize---------------------------------------
1662 // Given a graph, optimize it.
1663 void Compile::Optimize() {
1664   TracePhase t1("optimizer", &_t_optimizer, true);
1665 
1666 #ifndef PRODUCT
1667   if (env()->break_at_compile()) {
1668     BREAKPOINT;
1669   }
1670 
1671 #endif
1672 
1673   ResourceMark rm;
1674   int          loop_opts_cnt;
1675 
1676   NOT_PRODUCT( verify_graph_edges(); )
1677 
1678   print_method("After Parsing");
1679 
1680  {
1681   // Iterative Global Value Numbering, including ideal transforms
1682   // Initialize IterGVN with types and values from parse-time GVN
1683   PhaseIterGVN igvn(initial_gvn());
1684   {
1685     NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
1686     igvn.optimize();
1687   }
1688 
1689   print_method("Iter GVN 1", 2);
1690 
1691   if (failing())  return;
1692 
1693   // Perform escape analysis
1694   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
1695     TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
1696     ConnectionGraph::do_analysis(this, &igvn);
1697 
1698     if (failing())  return;
1699 
1700     igvn.optimize();
1701     print_method("Iter GVN 3", 2);
1702 
1703     if (failing())  return;
1704 
1705   }
1706 
1707   // Loop transforms on the ideal graph.  Range Check Elimination,
1708   // peeling, unrolling, etc.
1709 
1710   // Set loop opts counter
1711   loop_opts_cnt = num_loop_opts();
1712   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
1713     {
1714       TracePhase t2("idealLoop", &_t_idealLoop, true);
1715       PhaseIdealLoop ideal_loop( igvn, true );
1716       loop_opts_cnt--;
1717       if (major_progress()) print_method("PhaseIdealLoop 1", 2);
1718       if (failing())  return;
1719     }
1720     // Loop opts pass if partial peeling occurred in previous pass
1721     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
1722       TracePhase t3("idealLoop", &_t_idealLoop, true);
1723       PhaseIdealLoop ideal_loop( igvn, false );
1724       loop_opts_cnt--;
1725       if (major_progress()) print_method("PhaseIdealLoop 2", 2);
1726       if (failing())  return;
1727     }
1728     // Loop opts pass for loop-unrolling before CCP
1729     if(major_progress() && (loop_opts_cnt > 0)) {
1730       TracePhase t4("idealLoop", &_t_idealLoop, true);
1731       PhaseIdealLoop ideal_loop( igvn, false );
1732       loop_opts_cnt--;
1733       if (major_progress()) print_method("PhaseIdealLoop 3", 2);
1734     }
1735     if (!failing()) {
1736       // Verify that last round of loop opts produced a valid graph
1737       NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1738       PhaseIdealLoop::verify(igvn);
1739     }
1740   }
1741   if (failing())  return;
1742 
1743   // Conditional Constant Propagation;
1744   PhaseCCP ccp( &igvn );
1745   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
1746   {
1747     TracePhase t2("ccp", &_t_ccp, true);
1748     ccp.do_transform();
1749   }
1750   print_method("PhaseCPP 1", 2);
1751 
1752   assert( true, "Break here to ccp.dump_old2new_map()");
1753 
1754   // Iterative Global Value Numbering, including ideal transforms
1755   {
1756     NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
1757     igvn = ccp;
1758     igvn.optimize();
1759   }
1760 
1761   print_method("Iter GVN 2", 2);
1762 
1763   if (failing())  return;
1764 
1765   // Loop transforms on the ideal graph.  Range Check Elimination,
1766   // peeling, unrolling, etc.
1767   if(loop_opts_cnt > 0) {
1768     debug_only( int cnt = 0; );
1769     while(major_progress() && (loop_opts_cnt > 0)) {
1770       TracePhase t2("idealLoop", &_t_idealLoop, true);
1771       assert( cnt++ < 40, "infinite cycle in loop optimization" );
1772       PhaseIdealLoop ideal_loop( igvn, true);
1773       loop_opts_cnt--;
1774       if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
1775       if (failing())  return;
1776     }
1777   }
1778 
1779   {
1780     // Verify that all previous optimizations produced a valid graph
1781     // at least to this point, even if no loop optimizations were done.
1782     NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
1783     PhaseIdealLoop::verify(igvn);
1784   }
1785 
1786   {
1787     NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
1788     PhaseMacroExpand  mex(igvn);
1789     if (mex.expand_macro_nodes()) {
1790       assert(failing(), "must bail out w/ explicit message");
1791       return;
1792     }
1793   }
1794 
1795  } // (End scope of igvn; run destructor if necessary for asserts.)
1796 
1797   // A method with only infinite loops has no edges entering loops from root
1798   {
1799     NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
1800     if (final_graph_reshaping()) {
1801       assert(failing(), "must bail out w/ explicit message");
1802       return;
1803     }
1804   }
1805 
1806   print_method("Optimize finished", 2);
1807 }
1808 
1809 
1810 //------------------------------Code_Gen---------------------------------------
1811 // Given a graph, generate code for it
1812 void Compile::Code_Gen() {
1813   if (failing())  return;
1814 
1815   // Perform instruction selection.  You might think we could reclaim Matcher
1816   // memory PDQ, but actually the Matcher is used in generating spill code.
1817   // Internals of the Matcher (including some VectorSets) must remain live
1818   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
1819   // set a bit in reclaimed memory.
1820 
1821   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1822   // nodes.  Mapping is only valid at the root of each matched subtree.
1823   NOT_PRODUCT( verify_graph_edges(); )
1824 
1825   Node_List proj_list;
1826   Matcher m(proj_list);
1827   _matcher = &m;
1828   {
1829     TracePhase t2("matcher", &_t_matcher, true);
1830     m.match();
1831   }
1832   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
1833   // nodes.  Mapping is only valid at the root of each matched subtree.
1834   NOT_PRODUCT( verify_graph_edges(); )
1835 
1836   // If you have too many nodes, or if matching has failed, bail out
1837   check_node_count(0, "out of nodes matching instructions");
1838   if (failing())  return;
1839 
1840   // Build a proper-looking CFG
1841   PhaseCFG cfg(node_arena(), root(), m);
1842   _cfg = &cfg;
1843   {
1844     NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
1845     cfg.Dominators();
1846     if (failing())  return;
1847 
1848     NOT_PRODUCT( verify_graph_edges(); )
1849 
1850     cfg.Estimate_Block_Frequency();
1851     cfg.GlobalCodeMotion(m,unique(),proj_list);
1852 
1853     print_method("Global code motion", 2);
1854 
1855     if (failing())  return;
1856     NOT_PRODUCT( verify_graph_edges(); )
1857 
1858     debug_only( cfg.verify(); )
1859   }
1860   NOT_PRODUCT( verify_graph_edges(); )
1861 
1862   PhaseChaitin regalloc(unique(),cfg,m);
1863   _regalloc = &regalloc;
1864   {
1865     TracePhase t2("regalloc", &_t_registerAllocation, true);
1866     // Perform any platform dependent preallocation actions.  This is used,
1867     // for example, to avoid taking an implicit null pointer exception
1868     // using the frame pointer on win95.
1869     _regalloc->pd_preallocate_hook();
1870 
1871     // Perform register allocation.  After Chaitin, use-def chains are
1872     // no longer accurate (at spill code) and so must be ignored.
1873     // Node->LRG->reg mappings are still accurate.
1874     _regalloc->Register_Allocate();
1875 
1876     // Bail out if the allocator builds too many nodes
1877     if (failing())  return;
1878   }
1879 
1880   // Prior to register allocation we kept empty basic blocks in case the
1881   // the allocator needed a place to spill.  After register allocation we
1882   // are not adding any new instructions.  If any basic block is empty, we
1883   // can now safely remove it.
1884   {
1885     NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
1886     cfg.remove_empty();
1887     if (do_freq_based_layout()) {
1888       PhaseBlockLayout layout(cfg);
1889     } else {
1890       cfg.set_loop_alignment();
1891     }
1892     cfg.fixup_flow();
1893   }
1894 
1895   // Perform any platform dependent postallocation verifications.
1896   debug_only( _regalloc->pd_postallocate_verify_hook(); )
1897 
1898   // Apply peephole optimizations
1899   if( OptoPeephole ) {
1900     NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
1901     PhasePeephole peep( _regalloc, cfg);
1902     peep.do_transform();
1903   }
1904 
1905   // Convert Nodes to instruction bits in a buffer
1906   {
1907     // %%%% workspace merge brought two timers together for one job
1908     TracePhase t2a("output", &_t_output, true);
1909     NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
1910     Output();
1911   }
1912 
1913   print_method("Final Code");
1914 
1915   // He's dead, Jim.
1916   _cfg     = (PhaseCFG*)0xdeadbeef;
1917   _regalloc = (PhaseChaitin*)0xdeadbeef;
1918 }
1919 
1920 
1921 //------------------------------dump_asm---------------------------------------
1922 // Dump formatted assembly
1923 #ifndef PRODUCT
1924 void Compile::dump_asm(int *pcs, uint pc_limit) {
1925   bool cut_short = false;
1926   tty->print_cr("#");
1927   tty->print("#  ");  _tf->dump();  tty->cr();
1928   tty->print_cr("#");
1929 
1930   // For all blocks
1931   int pc = 0x0;                 // Program counter
1932   char starts_bundle = ' ';
1933   _regalloc->dump_frame();
1934 
1935   Node *n = NULL;
1936   for( uint i=0; i<_cfg->_num_blocks; i++ ) {
1937     if (VMThread::should_terminate()) { cut_short = true; break; }
1938     Block *b = _cfg->_blocks[i];
1939     if (b->is_connector() && !Verbose) continue;
1940     n = b->_nodes[0];
1941     if (pcs && n->_idx < pc_limit)
1942       tty->print("%3.3x   ", pcs[n->_idx]);
1943     else
1944       tty->print("      ");
1945     b->dump_head( &_cfg->_bbs );
1946     if (b->is_connector()) {
1947       tty->print_cr("        # Empty connector block");
1948     } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
1949       tty->print_cr("        # Block is sole successor of call");
1950     }
1951 
1952     // For all instructions
1953     Node *delay = NULL;
1954     for( uint j = 0; j<b->_nodes.size(); j++ ) {
1955       if (VMThread::should_terminate()) { cut_short = true; break; }
1956       n = b->_nodes[j];
1957       if (valid_bundle_info(n)) {
1958         Bundle *bundle = node_bundling(n);
1959         if (bundle->used_in_unconditional_delay()) {
1960           delay = n;
1961           continue;
1962         }
1963         if (bundle->starts_bundle())
1964           starts_bundle = '+';
1965       }
1966 
1967       if (WizardMode) n->dump();
1968 
1969       if( !n->is_Region() &&    // Dont print in the Assembly
1970           !n->is_Phi() &&       // a few noisely useless nodes
1971           !n->is_Proj() &&
1972           !n->is_MachTemp() &&
1973           !n->is_SafePointScalarObject() &&
1974           !n->is_Catch() &&     // Would be nice to print exception table targets
1975           !n->is_MergeMem() &&  // Not very interesting
1976           !n->is_top() &&       // Debug info table constants
1977           !(n->is_Con() && !n->is_Mach())// Debug info table constants
1978           ) {
1979         if (pcs && n->_idx < pc_limit)
1980           tty->print("%3.3x", pcs[n->_idx]);
1981         else
1982           tty->print("   ");
1983         tty->print(" %c ", starts_bundle);
1984         starts_bundle = ' ';
1985         tty->print("\t");
1986         n->format(_regalloc, tty);
1987         tty->cr();
1988       }
1989 
1990       // If we have an instruction with a delay slot, and have seen a delay,
1991       // then back up and print it
1992       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
1993         assert(delay != NULL, "no unconditional delay instruction");
1994         if (WizardMode) delay->dump();
1995 
1996         if (node_bundling(delay)->starts_bundle())
1997           starts_bundle = '+';
1998         if (pcs && n->_idx < pc_limit)
1999           tty->print("%3.3x", pcs[n->_idx]);
2000         else
2001           tty->print("   ");
2002         tty->print(" %c ", starts_bundle);
2003         starts_bundle = ' ';
2004         tty->print("\t");
2005         delay->format(_regalloc, tty);
2006         tty->print_cr("");
2007         delay = NULL;
2008       }
2009 
2010       // Dump the exception table as well
2011       if( n->is_Catch() && (Verbose || WizardMode) ) {
2012         // Print the exception table for this offset
2013         _handler_table.print_subtable_for(pc);
2014       }
2015     }
2016 
2017     if (pcs && n->_idx < pc_limit)
2018       tty->print_cr("%3.3x", pcs[n->_idx]);
2019     else
2020       tty->print_cr("");
2021 
2022     assert(cut_short || delay == NULL, "no unconditional delay branch");
2023 
2024   } // End of per-block dump
2025   tty->print_cr("");
2026 
2027   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2028 }
2029 #endif
2030 
2031 //------------------------------Final_Reshape_Counts---------------------------
2032 // This class defines counters to help identify when a method
2033 // may/must be executed using hardware with only 24-bit precision.
2034 struct Final_Reshape_Counts : public StackObj {
2035   int  _call_count;             // count non-inlined 'common' calls
2036   int  _float_count;            // count float ops requiring 24-bit precision
2037   int  _double_count;           // count double ops requiring more precision
2038   int  _java_call_count;        // count non-inlined 'java' calls
2039   int  _inner_loop_count;       // count loops which need alignment
2040   VectorSet _visited;           // Visitation flags
2041   Node_List _tests;             // Set of IfNodes & PCTableNodes
2042 
2043   Final_Reshape_Counts() :
2044     _call_count(0), _float_count(0), _double_count(0),
2045     _java_call_count(0), _inner_loop_count(0),
2046     _visited( Thread::current()->resource_area() ) { }
2047 
2048   void inc_call_count  () { _call_count  ++; }
2049   void inc_float_count () { _float_count ++; }
2050   void inc_double_count() { _double_count++; }
2051   void inc_java_call_count() { _java_call_count++; }
2052   void inc_inner_loop_count() { _inner_loop_count++; }
2053 
2054   int  get_call_count  () const { return _call_count  ; }
2055   int  get_float_count () const { return _float_count ; }
2056   int  get_double_count() const { return _double_count; }
2057   int  get_java_call_count() const { return _java_call_count; }
2058   int  get_inner_loop_count() const { return _inner_loop_count; }
2059 };
2060 
2061 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2062   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2063   // Make sure the offset goes inside the instance layout.
2064   return k->contains_field_offset(tp->offset());
2065   // Note that OffsetBot and OffsetTop are very negative.
2066 }
2067 
2068 // Eliminate trivially redundant StoreCMs and accumulate their
2069 // precedence edges.
2070 static void eliminate_redundant_card_marks(Node* n) {
2071   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2072   if (n->in(MemNode::Address)->outcnt() > 1) {
2073     // There are multiple users of the same address so it might be
2074     // possible to eliminate some of the StoreCMs
2075     Node* mem = n->in(MemNode::Memory);
2076     Node* adr = n->in(MemNode::Address);
2077     Node* val = n->in(MemNode::ValueIn);
2078     Node* prev = n;
2079     bool done = false;
2080     // Walk the chain of StoreCMs eliminating ones that match.  As
2081     // long as it's a chain of single users then the optimization is
2082     // safe.  Eliminating partially redundant StoreCMs would require
2083     // cloning copies down the other paths.
2084     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2085       if (adr == mem->in(MemNode::Address) &&
2086           val == mem->in(MemNode::ValueIn)) {
2087         // redundant StoreCM
2088         if (mem->req() > MemNode::OopStore) {
2089           // Hasn't been processed by this code yet.
2090           n->add_prec(mem->in(MemNode::OopStore));
2091         } else {
2092           // Already converted to precedence edge
2093           for (uint i = mem->req(); i < mem->len(); i++) {
2094             // Accumulate any precedence edges
2095             if (mem->in(i) != NULL) {
2096               n->add_prec(mem->in(i));
2097             }
2098           }
2099           // Everything above this point has been processed.
2100           done = true;
2101         }
2102         // Eliminate the previous StoreCM
2103         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2104         assert(mem->outcnt() == 0, "should be dead");
2105         mem->disconnect_inputs(NULL);
2106       } else {
2107         prev = mem;
2108       }
2109       mem = prev->in(MemNode::Memory);
2110     }
2111   }
2112 }
2113 
2114 //------------------------------final_graph_reshaping_impl----------------------
2115 // Implement items 1-5 from final_graph_reshaping below.
2116 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
2117 
2118   if ( n->outcnt() == 0 ) return; // dead node
2119   uint nop = n->Opcode();
2120 
2121   // Check for 2-input instruction with "last use" on right input.
2122   // Swap to left input.  Implements item (2).
2123   if( n->req() == 3 &&          // two-input instruction
2124       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2125       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2126       n->in(2)->outcnt() == 1 &&// right use IS a last use
2127       !n->in(2)->is_Con() ) {   // right use is not a constant
2128     // Check for commutative opcode
2129     switch( nop ) {
2130     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2131     case Op_MaxI:  case Op_MinI:
2132     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2133     case Op_AndL:  case Op_XorL:  case Op_OrL:
2134     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2135       // Move "last use" input to left by swapping inputs
2136       n->swap_edges(1, 2);
2137       break;
2138     }
2139     default:
2140       break;
2141     }
2142   }
2143 
2144 #ifdef ASSERT
2145   if( n->is_Mem() ) {
2146     Compile* C = Compile::current();
2147     int alias_idx = C->get_alias_index(n->as_Mem()->adr_type());
2148     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2149             // oop will be recorded in oop map if load crosses safepoint
2150             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2151                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2152             "raw memory operations should have control edge");
2153   }
2154 #endif
2155   // Count FPU ops and common calls, implements item (3)
2156   switch( nop ) {
2157   // Count all float operations that may use FPU
2158   case Op_AddF:
2159   case Op_SubF:
2160   case Op_MulF:
2161   case Op_DivF:
2162   case Op_NegF:
2163   case Op_ModF:
2164   case Op_ConvI2F:
2165   case Op_ConF:
2166   case Op_CmpF:
2167   case Op_CmpF3:
2168   // case Op_ConvL2F: // longs are split into 32-bit halves
2169     frc.inc_float_count();
2170     break;
2171 
2172   case Op_ConvF2D:
2173   case Op_ConvD2F:
2174     frc.inc_float_count();
2175     frc.inc_double_count();
2176     break;
2177 
2178   // Count all double operations that may use FPU
2179   case Op_AddD:
2180   case Op_SubD:
2181   case Op_MulD:
2182   case Op_DivD:
2183   case Op_NegD:
2184   case Op_ModD:
2185   case Op_ConvI2D:
2186   case Op_ConvD2I:
2187   // case Op_ConvL2D: // handled by leaf call
2188   // case Op_ConvD2L: // handled by leaf call
2189   case Op_ConD:
2190   case Op_CmpD:
2191   case Op_CmpD3:
2192     frc.inc_double_count();
2193     break;
2194   case Op_Opaque1:              // Remove Opaque Nodes before matching
2195   case Op_Opaque2:              // Remove Opaque Nodes before matching
2196     n->subsume_by(n->in(1));
2197     break;
2198   case Op_CallStaticJava:
2199   case Op_CallJava:
2200   case Op_CallDynamicJava:
2201     frc.inc_java_call_count(); // Count java call site;
2202   case Op_CallRuntime:
2203   case Op_CallLeaf:
2204   case Op_CallLeafNoFP: {
2205     assert( n->is_Call(), "" );
2206     CallNode *call = n->as_Call();
2207     // Count call sites where the FP mode bit would have to be flipped.
2208     // Do not count uncommon runtime calls:
2209     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2210     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2211     if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2212       frc.inc_call_count();   // Count the call site
2213     } else {                  // See if uncommon argument is shared
2214       Node *n = call->in(TypeFunc::Parms);
2215       int nop = n->Opcode();
2216       // Clone shared simple arguments to uncommon calls, item (1).
2217       if( n->outcnt() > 1 &&
2218           !n->is_Proj() &&
2219           nop != Op_CreateEx &&
2220           nop != Op_CheckCastPP &&
2221           nop != Op_DecodeN &&
2222           !n->is_Mem() ) {
2223         Node *x = n->clone();
2224         call->set_req( TypeFunc::Parms, x );
2225       }
2226     }
2227     break;
2228   }
2229 
2230   case Op_StoreD:
2231   case Op_LoadD:
2232   case Op_LoadD_unaligned:
2233     frc.inc_double_count();
2234     goto handle_mem;
2235   case Op_StoreF:
2236   case Op_LoadF:
2237     frc.inc_float_count();
2238     goto handle_mem;
2239 
2240   case Op_StoreCM:
2241     {
2242       // Convert OopStore dependence into precedence edge
2243       Node* prec = n->in(MemNode::OopStore);
2244       n->del_req(MemNode::OopStore);
2245       n->add_prec(prec);
2246       eliminate_redundant_card_marks(n);
2247     }
2248 
2249     // fall through
2250 
2251   case Op_StoreB:
2252   case Op_StoreC:
2253   case Op_StorePConditional:
2254   case Op_StoreI:
2255   case Op_StoreL:
2256   case Op_StoreIConditional:
2257   case Op_StoreLConditional:
2258   case Op_CompareAndSwapI:
2259   case Op_CompareAndSwapL:
2260   case Op_CompareAndSwapP:
2261   case Op_CompareAndSwapN:
2262   case Op_StoreP:
2263   case Op_StoreN:
2264   case Op_LoadB:
2265   case Op_LoadUB:
2266   case Op_LoadUS:
2267   case Op_LoadI:
2268   case Op_LoadUI2L:
2269   case Op_LoadKlass:
2270   case Op_LoadNKlass:
2271   case Op_LoadL:
2272   case Op_LoadL_unaligned:
2273   case Op_LoadPLocked:
2274   case Op_LoadLLocked:
2275   case Op_LoadP:
2276   case Op_LoadN:
2277   case Op_LoadRange:
2278   case Op_LoadS: {
2279   handle_mem:
2280 #ifdef ASSERT
2281     if( VerifyOptoOopOffsets ) {
2282       assert( n->is_Mem(), "" );
2283       MemNode *mem  = (MemNode*)n;
2284       // Check to see if address types have grounded out somehow.
2285       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2286       assert( !tp || oop_offset_is_sane(tp), "" );
2287     }
2288 #endif
2289     break;
2290   }
2291 
2292   case Op_AddP: {               // Assert sane base pointers
2293     Node *addp = n->in(AddPNode::Address);
2294     assert( !addp->is_AddP() ||
2295             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2296             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2297             "Base pointers must match" );
2298 #ifdef _LP64
2299     if (UseCompressedOops &&
2300         addp->Opcode() == Op_ConP &&
2301         addp == n->in(AddPNode::Base) &&
2302         n->in(AddPNode::Offset)->is_Con()) {
2303       // Use addressing with narrow klass to load with offset on x86.
2304       // On sparc loading 32-bits constant and decoding it have less
2305       // instructions (4) then load 64-bits constant (7).
2306       // Do this transformation here since IGVN will convert ConN back to ConP.
2307       const Type* t = addp->bottom_type();
2308       if (t->isa_oopptr()) {
2309         Node* nn = NULL;
2310 
2311         // Look for existing ConN node of the same exact type.
2312         Compile* C = Compile::current();
2313         Node* r  = C->root();
2314         uint cnt = r->outcnt();
2315         for (uint i = 0; i < cnt; i++) {
2316           Node* m = r->raw_out(i);
2317           if (m!= NULL && m->Opcode() == Op_ConN &&
2318               m->bottom_type()->make_ptr() == t) {
2319             nn = m;
2320             break;
2321           }
2322         }
2323         if (nn != NULL) {
2324           // Decode a narrow oop to match address
2325           // [R12 + narrow_oop_reg<<3 + offset]
2326           nn = new (C,  2) DecodeNNode(nn, t);
2327           n->set_req(AddPNode::Base, nn);
2328           n->set_req(AddPNode::Address, nn);
2329           if (addp->outcnt() == 0) {
2330             addp->disconnect_inputs(NULL);
2331           }
2332         }
2333       }
2334     }
2335 #endif
2336     break;
2337   }
2338 
2339 #ifdef _LP64
2340   case Op_CastPP:
2341     if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2342       Compile* C = Compile::current();
2343       Node* in1 = n->in(1);
2344       const Type* t = n->bottom_type();
2345       Node* new_in1 = in1->clone();
2346       new_in1->as_DecodeN()->set_type(t);
2347 
2348       if (!Matcher::narrow_oop_use_complex_address()) {
2349         //
2350         // x86, ARM and friends can handle 2 adds in addressing mode
2351         // and Matcher can fold a DecodeN node into address by using
2352         // a narrow oop directly and do implicit NULL check in address:
2353         //
2354         // [R12 + narrow_oop_reg<<3 + offset]
2355         // NullCheck narrow_oop_reg
2356         //
2357         // On other platforms (Sparc) we have to keep new DecodeN node and
2358         // use it to do implicit NULL check in address:
2359         //
2360         // decode_not_null narrow_oop_reg, base_reg
2361         // [base_reg + offset]
2362         // NullCheck base_reg
2363         //
2364         // Pin the new DecodeN node to non-null path on these platform (Sparc)
2365         // to keep the information to which NULL check the new DecodeN node
2366         // corresponds to use it as value in implicit_null_check().
2367         //
2368         new_in1->set_req(0, n->in(0));
2369       }
2370 
2371       n->subsume_by(new_in1);
2372       if (in1->outcnt() == 0) {
2373         in1->disconnect_inputs(NULL);
2374       }
2375     }
2376     break;
2377 
2378   case Op_CmpP:
2379     // Do this transformation here to preserve CmpPNode::sub() and
2380     // other TypePtr related Ideal optimizations (for example, ptr nullness).
2381     if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
2382       Node* in1 = n->in(1);
2383       Node* in2 = n->in(2);
2384       if (!in1->is_DecodeN()) {
2385         in2 = in1;
2386         in1 = n->in(2);
2387       }
2388       assert(in1->is_DecodeN(), "sanity");
2389 
2390       Compile* C = Compile::current();
2391       Node* new_in2 = NULL;
2392       if (in2->is_DecodeN()) {
2393         new_in2 = in2->in(1);
2394       } else if (in2->Opcode() == Op_ConP) {
2395         const Type* t = in2->bottom_type();
2396         if (t == TypePtr::NULL_PTR) {
2397           // Don't convert CmpP null check into CmpN if compressed
2398           // oops implicit null check is not generated.
2399           // This will allow to generate normal oop implicit null check.
2400           if (Matcher::gen_narrow_oop_implicit_null_checks())
2401             new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
2402           //
2403           // This transformation together with CastPP transformation above
2404           // will generated code for implicit NULL checks for compressed oops.
2405           //
2406           // The original code after Optimize()
2407           //
2408           //    LoadN memory, narrow_oop_reg
2409           //    decode narrow_oop_reg, base_reg
2410           //    CmpP base_reg, NULL
2411           //    CastPP base_reg // NotNull
2412           //    Load [base_reg + offset], val_reg
2413           //
2414           // after these transformations will be
2415           //
2416           //    LoadN memory, narrow_oop_reg
2417           //    CmpN narrow_oop_reg, NULL
2418           //    decode_not_null narrow_oop_reg, base_reg
2419           //    Load [base_reg + offset], val_reg
2420           //
2421           // and the uncommon path (== NULL) will use narrow_oop_reg directly
2422           // since narrow oops can be used in debug info now (see the code in
2423           // final_graph_reshaping_walk()).
2424           //
2425           // At the end the code will be matched to
2426           // on x86:
2427           //
2428           //    Load_narrow_oop memory, narrow_oop_reg
2429           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2430           //    NullCheck narrow_oop_reg
2431           //
2432           // and on sparc:
2433           //
2434           //    Load_narrow_oop memory, narrow_oop_reg
2435           //    decode_not_null narrow_oop_reg, base_reg
2436           //    Load [base_reg + offset], val_reg
2437           //    NullCheck base_reg
2438           //
2439         } else if (t->isa_oopptr()) {
2440           new_in2 = ConNode::make(C, t->make_narrowoop());
2441         }
2442       }
2443       if (new_in2 != NULL) {
2444         Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
2445         n->subsume_by( cmpN );
2446         if (in1->outcnt() == 0) {
2447           in1->disconnect_inputs(NULL);
2448         }
2449         if (in2->outcnt() == 0) {
2450           in2->disconnect_inputs(NULL);
2451         }
2452       }
2453     }
2454     break;
2455 
2456   case Op_DecodeN:
2457     assert(!n->in(1)->is_EncodeP(), "should be optimized out");
2458     // DecodeN could be pinned when it can't be fold into
2459     // an address expression, see the code for Op_CastPP above.
2460     assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control");
2461     break;
2462 
2463   case Op_EncodeP: {
2464     Node* in1 = n->in(1);
2465     if (in1->is_DecodeN()) {
2466       n->subsume_by(in1->in(1));
2467     } else if (in1->Opcode() == Op_ConP) {
2468       Compile* C = Compile::current();
2469       const Type* t = in1->bottom_type();
2470       if (t == TypePtr::NULL_PTR) {
2471         n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
2472       } else if (t->isa_oopptr()) {
2473         n->subsume_by(ConNode::make(C, t->make_narrowoop()));
2474       }
2475     }
2476     if (in1->outcnt() == 0) {
2477       in1->disconnect_inputs(NULL);
2478     }
2479     break;
2480   }
2481 
2482   case Op_Proj: {
2483     if (OptimizeStringConcat) {
2484       ProjNode* p = n->as_Proj();
2485       if (p->_is_io_use) {
2486         // Separate projections were used for the exception path which
2487         // are normally removed by a late inline.  If it wasn't inlined
2488         // then they will hang around and should just be replaced with
2489         // the original one.
2490         Node* proj = NULL;
2491         // Replace with just one
2492         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2493           Node *use = i.get();
2494           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2495             proj = use;
2496             break;
2497           }
2498         }
2499         assert(p != NULL, "must be found");
2500         p->subsume_by(proj);
2501       }
2502     }
2503     break;
2504   }
2505 
2506   case Op_Phi:
2507     if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
2508       // The EncodeP optimization may create Phi with the same edges
2509       // for all paths. It is not handled well by Register Allocator.
2510       Node* unique_in = n->in(1);
2511       assert(unique_in != NULL, "");
2512       uint cnt = n->req();
2513       for (uint i = 2; i < cnt; i++) {
2514         Node* m = n->in(i);
2515         assert(m != NULL, "");
2516         if (unique_in != m)
2517           unique_in = NULL;
2518       }
2519       if (unique_in != NULL) {
2520         n->subsume_by(unique_in);
2521       }
2522     }
2523     break;
2524 
2525 #endif
2526 
2527   case Op_ModI:
2528     if (UseDivMod) {
2529       // Check if a%b and a/b both exist
2530       Node* d = n->find_similar(Op_DivI);
2531       if (d) {
2532         // Replace them with a fused divmod if supported
2533         Compile* C = Compile::current();
2534         if (Matcher::has_match_rule(Op_DivModI)) {
2535           DivModINode* divmod = DivModINode::make(C, n);
2536           d->subsume_by(divmod->div_proj());
2537           n->subsume_by(divmod->mod_proj());
2538         } else {
2539           // replace a%b with a-((a/b)*b)
2540           Node* mult = new (C, 3) MulINode(d, d->in(2));
2541           Node* sub  = new (C, 3) SubINode(d->in(1), mult);
2542           n->subsume_by( sub );
2543         }
2544       }
2545     }
2546     break;
2547 
2548   case Op_ModL:
2549     if (UseDivMod) {
2550       // Check if a%b and a/b both exist
2551       Node* d = n->find_similar(Op_DivL);
2552       if (d) {
2553         // Replace them with a fused divmod if supported
2554         Compile* C = Compile::current();
2555         if (Matcher::has_match_rule(Op_DivModL)) {
2556           DivModLNode* divmod = DivModLNode::make(C, n);
2557           d->subsume_by(divmod->div_proj());
2558           n->subsume_by(divmod->mod_proj());
2559         } else {
2560           // replace a%b with a-((a/b)*b)
2561           Node* mult = new (C, 3) MulLNode(d, d->in(2));
2562           Node* sub  = new (C, 3) SubLNode(d->in(1), mult);
2563           n->subsume_by( sub );
2564         }
2565       }
2566     }
2567     break;
2568 
2569   case Op_Load16B:
2570   case Op_Load8B:
2571   case Op_Load4B:
2572   case Op_Load8S:
2573   case Op_Load4S:
2574   case Op_Load2S:
2575   case Op_Load8C:
2576   case Op_Load4C:
2577   case Op_Load2C:
2578   case Op_Load4I:
2579   case Op_Load2I:
2580   case Op_Load2L:
2581   case Op_Load4F:
2582   case Op_Load2F:
2583   case Op_Load2D:
2584   case Op_Store16B:
2585   case Op_Store8B:
2586   case Op_Store4B:
2587   case Op_Store8C:
2588   case Op_Store4C:
2589   case Op_Store2C:
2590   case Op_Store4I:
2591   case Op_Store2I:
2592   case Op_Store2L:
2593   case Op_Store4F:
2594   case Op_Store2F:
2595   case Op_Store2D:
2596     break;
2597 
2598   case Op_PackB:
2599   case Op_PackS:
2600   case Op_PackC:
2601   case Op_PackI:
2602   case Op_PackF:
2603   case Op_PackL:
2604   case Op_PackD:
2605     if (n->req()-1 > 2) {
2606       // Replace many operand PackNodes with a binary tree for matching
2607       PackNode* p = (PackNode*) n;
2608       Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
2609       n->subsume_by(btp);
2610     }
2611     break;
2612   case Op_Loop:
2613   case Op_CountedLoop:
2614     if (n->as_Loop()->is_inner_loop()) {
2615       frc.inc_inner_loop_count();
2616     }
2617     break;
2618   case Op_LShiftI:
2619   case Op_RShiftI:
2620   case Op_URShiftI:
2621   case Op_LShiftL:
2622   case Op_RShiftL:
2623   case Op_URShiftL:
2624     if (Matcher::need_masked_shift_count) {
2625       // The cpu's shift instructions don't restrict the count to the
2626       // lower 5/6 bits. We need to do the masking ourselves.
2627       Node* in2 = n->in(2);
2628       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
2629       const TypeInt* t = in2->find_int_type();
2630       if (t != NULL && t->is_con()) {
2631         juint shift = t->get_con();
2632         if (shift > mask) { // Unsigned cmp
2633           Compile* C = Compile::current();
2634           n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask)));
2635         }
2636       } else {
2637         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
2638           Compile* C = Compile::current();
2639           Node* shift = new (C, 3) AndINode(in2, ConNode::make(C, TypeInt::make(mask)));
2640           n->set_req(2, shift);
2641         }
2642       }
2643       if (in2->outcnt() == 0) { // Remove dead node
2644         in2->disconnect_inputs(NULL);
2645       }
2646     }
2647     break;
2648   default:
2649     assert( !n->is_Call(), "" );
2650     assert( !n->is_Mem(), "" );
2651     break;
2652   }
2653 
2654   // Collect CFG split points
2655   if (n->is_MultiBranch())
2656     frc._tests.push(n);
2657 }
2658 
2659 //------------------------------final_graph_reshaping_walk---------------------
2660 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
2661 // requires that the walk visits a node's inputs before visiting the node.
2662 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
2663   ResourceArea *area = Thread::current()->resource_area();
2664   Unique_Node_List sfpt(area);
2665 
2666   frc._visited.set(root->_idx); // first, mark node as visited
2667   uint cnt = root->req();
2668   Node *n = root;
2669   uint  i = 0;
2670   while (true) {
2671     if (i < cnt) {
2672       // Place all non-visited non-null inputs onto stack
2673       Node* m = n->in(i);
2674       ++i;
2675       if (m != NULL && !frc._visited.test_set(m->_idx)) {
2676         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
2677           sfpt.push(m);
2678         cnt = m->req();
2679         nstack.push(n, i); // put on stack parent and next input's index
2680         n = m;
2681         i = 0;
2682       }
2683     } else {
2684       // Now do post-visit work
2685       final_graph_reshaping_impl( n, frc );
2686       if (nstack.is_empty())
2687         break;             // finished
2688       n = nstack.node();   // Get node from stack
2689       cnt = n->req();
2690       i = nstack.index();
2691       nstack.pop();        // Shift to the next node on stack
2692     }
2693   }
2694 
2695   // Skip next transformation if compressed oops are not used.
2696   if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks())
2697     return;
2698 
2699   // Go over safepoints nodes to skip DecodeN nodes for debug edges.
2700   // It could be done for an uncommon traps or any safepoints/calls
2701   // if the DecodeN node is referenced only in a debug info.
2702   while (sfpt.size() > 0) {
2703     n = sfpt.pop();
2704     JVMState *jvms = n->as_SafePoint()->jvms();
2705     assert(jvms != NULL, "sanity");
2706     int start = jvms->debug_start();
2707     int end   = n->req();
2708     bool is_uncommon = (n->is_CallStaticJava() &&
2709                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
2710     for (int j = start; j < end; j++) {
2711       Node* in = n->in(j);
2712       if (in->is_DecodeN()) {
2713         bool safe_to_skip = true;
2714         if (!is_uncommon ) {
2715           // Is it safe to skip?
2716           for (uint i = 0; i < in->outcnt(); i++) {
2717             Node* u = in->raw_out(i);
2718             if (!u->is_SafePoint() ||
2719                  u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
2720               safe_to_skip = false;
2721             }
2722           }
2723         }
2724         if (safe_to_skip) {
2725           n->set_req(j, in->in(1));
2726         }
2727         if (in->outcnt() == 0) {
2728           in->disconnect_inputs(NULL);
2729         }
2730       }
2731     }
2732   }
2733 }
2734 
2735 //------------------------------final_graph_reshaping--------------------------
2736 // Final Graph Reshaping.
2737 //
2738 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
2739 //     and not commoned up and forced early.  Must come after regular
2740 //     optimizations to avoid GVN undoing the cloning.  Clone constant
2741 //     inputs to Loop Phis; these will be split by the allocator anyways.
2742 //     Remove Opaque nodes.
2743 // (2) Move last-uses by commutative operations to the left input to encourage
2744 //     Intel update-in-place two-address operations and better register usage
2745 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
2746 //     calls canonicalizing them back.
2747 // (3) Count the number of double-precision FP ops, single-precision FP ops
2748 //     and call sites.  On Intel, we can get correct rounding either by
2749 //     forcing singles to memory (requires extra stores and loads after each
2750 //     FP bytecode) or we can set a rounding mode bit (requires setting and
2751 //     clearing the mode bit around call sites).  The mode bit is only used
2752 //     if the relative frequency of single FP ops to calls is low enough.
2753 //     This is a key transform for SPEC mpeg_audio.
2754 // (4) Detect infinite loops; blobs of code reachable from above but not
2755 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
2756 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
2757 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
2758 //     Detection is by looking for IfNodes where only 1 projection is
2759 //     reachable from below or CatchNodes missing some targets.
2760 // (5) Assert for insane oop offsets in debug mode.
2761 
2762 bool Compile::final_graph_reshaping() {
2763   // an infinite loop may have been eliminated by the optimizer,
2764   // in which case the graph will be empty.
2765   if (root()->req() == 1) {
2766     record_method_not_compilable("trivial infinite loop");
2767     return true;
2768   }
2769 
2770   Final_Reshape_Counts frc;
2771 
2772   // Visit everybody reachable!
2773   // Allocate stack of size C->unique()/2 to avoid frequent realloc
2774   Node_Stack nstack(unique() >> 1);
2775   final_graph_reshaping_walk(nstack, root(), frc);
2776 
2777   // Check for unreachable (from below) code (i.e., infinite loops).
2778   for( uint i = 0; i < frc._tests.size(); i++ ) {
2779     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
2780     // Get number of CFG targets.
2781     // Note that PCTables include exception targets after calls.
2782     uint required_outcnt = n->required_outcnt();
2783     if (n->outcnt() != required_outcnt) {
2784       // Check for a few special cases.  Rethrow Nodes never take the
2785       // 'fall-thru' path, so expected kids is 1 less.
2786       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
2787         if (n->in(0)->in(0)->is_Call()) {
2788           CallNode *call = n->in(0)->in(0)->as_Call();
2789           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
2790             required_outcnt--;      // Rethrow always has 1 less kid
2791           } else if (call->req() > TypeFunc::Parms &&
2792                      call->is_CallDynamicJava()) {
2793             // Check for null receiver. In such case, the optimizer has
2794             // detected that the virtual call will always result in a null
2795             // pointer exception. The fall-through projection of this CatchNode
2796             // will not be populated.
2797             Node *arg0 = call->in(TypeFunc::Parms);
2798             if (arg0->is_Type() &&
2799                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
2800               required_outcnt--;
2801             }
2802           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
2803                      call->req() > TypeFunc::Parms+1 &&
2804                      call->is_CallStaticJava()) {
2805             // Check for negative array length. In such case, the optimizer has
2806             // detected that the allocation attempt will always result in an
2807             // exception. There is no fall-through projection of this CatchNode .
2808             Node *arg1 = call->in(TypeFunc::Parms+1);
2809             if (arg1->is_Type() &&
2810                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
2811               required_outcnt--;
2812             }
2813           }
2814         }
2815       }
2816       // Recheck with a better notion of 'required_outcnt'
2817       if (n->outcnt() != required_outcnt) {
2818         record_method_not_compilable("malformed control flow");
2819         return true;            // Not all targets reachable!
2820       }
2821     }
2822     // Check that I actually visited all kids.  Unreached kids
2823     // must be infinite loops.
2824     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
2825       if (!frc._visited.test(n->fast_out(j)->_idx)) {
2826         record_method_not_compilable("infinite loop");
2827         return true;            // Found unvisited kid; must be unreach
2828       }
2829   }
2830 
2831   // If original bytecodes contained a mixture of floats and doubles
2832   // check if the optimizer has made it homogenous, item (3).
2833   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
2834       frc.get_float_count() > 32 &&
2835       frc.get_double_count() == 0 &&
2836       (10 * frc.get_call_count() < frc.get_float_count()) ) {
2837     set_24_bit_selection_and_mode( false,  true );
2838   }
2839 
2840   set_java_calls(frc.get_java_call_count());
2841   set_inner_loops(frc.get_inner_loop_count());
2842 
2843   // No infinite loops, no reason to bail out.
2844   return false;
2845 }
2846 
2847 //-----------------------------too_many_traps----------------------------------
2848 // Report if there are too many traps at the current method and bci.
2849 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
2850 bool Compile::too_many_traps(ciMethod* method,
2851                              int bci,
2852                              Deoptimization::DeoptReason reason) {
2853   ciMethodData* md = method->method_data();
2854   if (md->is_empty()) {
2855     // Assume the trap has not occurred, or that it occurred only
2856     // because of a transient condition during start-up in the interpreter.
2857     return false;
2858   }
2859   if (md->has_trap_at(bci, reason) != 0) {
2860     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
2861     // Also, if there are multiple reasons, or if there is no per-BCI record,
2862     // assume the worst.
2863     if (log())
2864       log()->elem("observe trap='%s' count='%d'",
2865                   Deoptimization::trap_reason_name(reason),
2866                   md->trap_count(reason));
2867     return true;
2868   } else {
2869     // Ignore method/bci and see if there have been too many globally.
2870     return too_many_traps(reason, md);
2871   }
2872 }
2873 
2874 // Less-accurate variant which does not require a method and bci.
2875 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
2876                              ciMethodData* logmd) {
2877  if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
2878     // Too many traps globally.
2879     // Note that we use cumulative trap_count, not just md->trap_count.
2880     if (log()) {
2881       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
2882       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
2883                   Deoptimization::trap_reason_name(reason),
2884                   mcount, trap_count(reason));
2885     }
2886     return true;
2887   } else {
2888     // The coast is clear.
2889     return false;
2890   }
2891 }
2892 
2893 //--------------------------too_many_recompiles--------------------------------
2894 // Report if there are too many recompiles at the current method and bci.
2895 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
2896 // Is not eager to return true, since this will cause the compiler to use
2897 // Action_none for a trap point, to avoid too many recompilations.
2898 bool Compile::too_many_recompiles(ciMethod* method,
2899                                   int bci,
2900                                   Deoptimization::DeoptReason reason) {
2901   ciMethodData* md = method->method_data();
2902   if (md->is_empty()) {
2903     // Assume the trap has not occurred, or that it occurred only
2904     // because of a transient condition during start-up in the interpreter.
2905     return false;
2906   }
2907   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
2908   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
2909   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
2910   Deoptimization::DeoptReason per_bc_reason
2911     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
2912   if ((per_bc_reason == Deoptimization::Reason_none
2913        || md->has_trap_at(bci, reason) != 0)
2914       // The trap frequency measure we care about is the recompile count:
2915       && md->trap_recompiled_at(bci)
2916       && md->overflow_recompile_count() >= bc_cutoff) {
2917     // Do not emit a trap here if it has already caused recompilations.
2918     // Also, if there are multiple reasons, or if there is no per-BCI record,
2919     // assume the worst.
2920     if (log())
2921       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
2922                   Deoptimization::trap_reason_name(reason),
2923                   md->trap_count(reason),
2924                   md->overflow_recompile_count());
2925     return true;
2926   } else if (trap_count(reason) != 0
2927              && decompile_count() >= m_cutoff) {
2928     // Too many recompiles globally, and we have seen this sort of trap.
2929     // Use cumulative decompile_count, not just md->decompile_count.
2930     if (log())
2931       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
2932                   Deoptimization::trap_reason_name(reason),
2933                   md->trap_count(reason), trap_count(reason),
2934                   md->decompile_count(), decompile_count());
2935     return true;
2936   } else {
2937     // The coast is clear.
2938     return false;
2939   }
2940 }
2941 
2942 
2943 #ifndef PRODUCT
2944 //------------------------------verify_graph_edges---------------------------
2945 // Walk the Graph and verify that there is a one-to-one correspondence
2946 // between Use-Def edges and Def-Use edges in the graph.
2947 void Compile::verify_graph_edges(bool no_dead_code) {
2948   if (VerifyGraphEdges) {
2949     ResourceArea *area = Thread::current()->resource_area();
2950     Unique_Node_List visited(area);
2951     // Call recursive graph walk to check edges
2952     _root->verify_edges(visited);
2953     if (no_dead_code) {
2954       // Now make sure that no visited node is used by an unvisited node.
2955       bool dead_nodes = 0;
2956       Unique_Node_List checked(area);
2957       while (visited.size() > 0) {
2958         Node* n = visited.pop();
2959         checked.push(n);
2960         for (uint i = 0; i < n->outcnt(); i++) {
2961           Node* use = n->raw_out(i);
2962           if (checked.member(use))  continue;  // already checked
2963           if (visited.member(use))  continue;  // already in the graph
2964           if (use->is_Con())        continue;  // a dead ConNode is OK
2965           // At this point, we have found a dead node which is DU-reachable.
2966           if (dead_nodes++ == 0)
2967             tty->print_cr("*** Dead nodes reachable via DU edges:");
2968           use->dump(2);
2969           tty->print_cr("---");
2970           checked.push(use);  // No repeats; pretend it is now checked.
2971         }
2972       }
2973       assert(dead_nodes == 0, "using nodes must be reachable from root");
2974     }
2975   }
2976 }
2977 #endif
2978 
2979 // The Compile object keeps track of failure reasons separately from the ciEnv.
2980 // This is required because there is not quite a 1-1 relation between the
2981 // ciEnv and its compilation task and the Compile object.  Note that one
2982 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
2983 // to backtrack and retry without subsuming loads.  Other than this backtracking
2984 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
2985 // by the logic in C2Compiler.
2986 void Compile::record_failure(const char* reason) {
2987   if (log() != NULL) {
2988     log()->elem("failure reason='%s' phase='compile'", reason);
2989   }
2990   if (_failure_reason == NULL) {
2991     // Record the first failure reason.
2992     _failure_reason = reason;
2993   }
2994   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
2995     C->print_method(_failure_reason);
2996   }
2997   _root = NULL;  // flush the graph, too
2998 }
2999 
3000 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3001   : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
3002 {
3003   if (dolog) {
3004     C = Compile::current();
3005     _log = C->log();
3006   } else {
3007     C = NULL;
3008     _log = NULL;
3009   }
3010   if (_log != NULL) {
3011     _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
3012     _log->stamp();
3013     _log->end_head();
3014   }
3015 }
3016 
3017 Compile::TracePhase::~TracePhase() {
3018   if (_log != NULL) {
3019     _log->done("phase nodes='%d'", C->unique());
3020   }
3021 }
3022 
3023 //=============================================================================
3024 // Two Constant's are equal when the type and the value are equal.
3025 bool Compile::Constant::operator==(const Constant& other) {
3026   if (type()          != other.type()         )  return false;
3027   if (can_be_reused() != other.can_be_reused())  return false;
3028   // For floating point values we compare the bit pattern.
3029   switch (type()) {
3030   case T_FLOAT:   return (_value.i == other._value.i);
3031   case T_LONG:
3032   case T_DOUBLE:  return (_value.j == other._value.j);
3033   case T_OBJECT:
3034   case T_ADDRESS: return (_value.l == other._value.l);
3035   case T_VOID:    return (_value.l == other._value.l);  // jump-table entries
3036   default: ShouldNotReachHere();
3037   }
3038   return false;
3039 }
3040 
3041 // Emit constants grouped in the following order:
3042 static BasicType type_order[] = {
3043   T_FLOAT,    // 32-bit
3044   T_OBJECT,   // 32 or 64-bit
3045   T_ADDRESS,  // 32 or 64-bit
3046   T_DOUBLE,   // 64-bit
3047   T_LONG,     // 64-bit
3048   T_VOID,     // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons)
3049   T_ILLEGAL
3050 };
3051 
3052 static int type_to_size_in_bytes(BasicType t) {
3053   switch (t) {
3054   case T_LONG:    return sizeof(jlong  );
3055   case T_FLOAT:   return sizeof(jfloat );
3056   case T_DOUBLE:  return sizeof(jdouble);
3057     // We use T_VOID as marker for jump-table entries (labels) which
3058     // need an interal word relocation.
3059   case T_VOID:
3060   case T_ADDRESS:
3061   case T_OBJECT:  return sizeof(jobject);
3062   }
3063 
3064   ShouldNotReachHere();
3065   return -1;
3066 }
3067 
3068 void Compile::ConstantTable::calculate_offsets_and_size() {
3069   int size = 0;
3070   for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
3071     BasicType type = type_order[t];
3072 
3073     for (int i = 0; i < _constants.length(); i++) {
3074       Constant con = _constants.at(i);
3075       if (con.type() != type)  continue;  // Skip other types.
3076 
3077       // Align size for type.
3078       int typesize = type_to_size_in_bytes(con.type());
3079       size = align_size_up(size, typesize);
3080 
3081       // Set offset.
3082       con.set_offset(size);
3083       _constants.at_put(i, con);
3084 
3085       // Add type size.
3086       size = size + typesize;
3087     }
3088   }
3089 
3090   // Align size up to the next section start (which is insts; see
3091   // CodeBuffer::align_at_start).
3092   assert(_size == -1, "already set?");
3093   _size = align_size_up(size, CodeEntryAlignment);
3094 
3095   if (Matcher::constant_table_absolute_addressing) {
3096     set_table_base_offset(0);  // No table base offset required
3097   } else {
3098     if (UseRDPCForConstantTableBase) {
3099       // table base offset is set in MachConstantBaseNode::emit
3100     } else {
3101       // When RDPC is not used, the table base is set into the middle of
3102       // the constant table.
3103       int half_size = _size / 2;
3104       assert(half_size * 2 == _size, "sanity");
3105       set_table_base_offset(-half_size);
3106     }
3107   }
3108 }
3109 
3110 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3111   MacroAssembler _masm(&cb);
3112   for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
3113     BasicType type = type_order[t];
3114 
3115     for (int i = 0; i < _constants.length(); i++) {
3116       Constant con = _constants.at(i);
3117       if (con.type() != type)  continue;  // Skip other types.
3118 
3119       address constant_addr;
3120       switch (con.type()) {
3121       case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
3122       case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3123       case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3124       case T_OBJECT: {
3125         jobject obj = con.get_jobject();
3126         int oop_index = _masm.oop_recorder()->find_index(obj);
3127         constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3128         break;
3129       }
3130       case T_ADDRESS: {
3131         address addr = (address) con.get_jobject();
3132         constant_addr = _masm.address_constant(addr);
3133         break;
3134       }
3135       // We use T_VOID as marker for jump-table entries (labels) which
3136       // need an interal word relocation.
3137       case T_VOID: {
3138         // Write a dummy word.  The real value is filled in later
3139         // in fill_jump_table_in_constant_table.
3140         address addr = (address) con.get_jobject();
3141         constant_addr = _masm.address_constant(addr);
3142         break;
3143       }
3144       default: ShouldNotReachHere();
3145       }
3146       assert(constant_addr != NULL, "consts section too small");
3147       assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
3148     }
3149   }
3150 }
3151 
3152 int Compile::ConstantTable::find_offset(Constant& con) const {
3153   int idx = _constants.find(con);
3154   assert(idx != -1, "constant must be in constant table");
3155   int offset = _constants.at(idx).offset();
3156   assert(offset != -1, "constant table not emitted yet?");
3157   return offset;
3158 }
3159 
3160 void Compile::ConstantTable::add(Constant& con) {
3161   if (con.can_be_reused()) {
3162     int idx = _constants.find(con);
3163     if (idx != -1 && _constants.at(idx).can_be_reused()) {
3164       return;
3165     }
3166   }
3167   (void) _constants.append(con);
3168 }
3169 
3170 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) {
3171   Constant con(type, value);
3172   add(con);
3173   return con;
3174 }
3175 
3176 Compile::Constant Compile::ConstantTable::add(MachOper* oper) {
3177   jvalue value;
3178   BasicType type = oper->type()->basic_type();
3179   switch (type) {
3180   case T_LONG:    value.j = oper->constantL(); break;
3181   case T_FLOAT:   value.f = oper->constantF(); break;
3182   case T_DOUBLE:  value.d = oper->constantD(); break;
3183   case T_OBJECT:
3184   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3185   default: ShouldNotReachHere();
3186   }
3187   return add(type, value);
3188 }
3189 
3190 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) {
3191   jvalue value;
3192   // We can use the node pointer here to identify the right jump-table
3193   // as this method is called from Compile::Fill_buffer right before
3194   // the MachNodes are emitted and the jump-table is filled (means the
3195   // MachNode pointers do not change anymore).
3196   value.l = (jobject) n;
3197   Constant con(T_VOID, value, false);  // Labels of a jump-table cannot be reused.
3198   for (uint i = 0; i < n->outcnt(); i++) {
3199     add(con);
3200   }
3201   return con;
3202 }
3203 
3204 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3205   // If called from Compile::scratch_emit_size do nothing.
3206   if (Compile::current()->in_scratch_emit_size())  return;
3207 
3208   assert(labels.is_nonempty(), "must be");
3209   assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt()));
3210 
3211   // Since MachConstantNode::constant_offset() also contains
3212   // table_base_offset() we need to subtract the table_base_offset()
3213   // to get the plain offset into the constant table.
3214   int offset = n->constant_offset() - table_base_offset();
3215 
3216   MacroAssembler _masm(&cb);
3217   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3218 
3219   for (int i = 0; i < labels.length(); i++) {
3220     address* constant_addr = &jump_table_base[i];
3221     assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer");
3222     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3223     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3224   }
3225 }