1 /*
   2  * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "asm/macroAssembler.hpp"
  27 #include "asm/macroAssembler.inline.hpp"
  28 #include "ci/ciReplay.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "code/exceptionHandlerTable.hpp"
  31 #include "code/nmethod.hpp"
  32 #include "compiler/compileLog.hpp"
  33 #include "compiler/disassembler.hpp"
  34 #include "compiler/oopMap.hpp"
  35 #include "opto/addnode.hpp"
  36 #include "opto/block.hpp"
  37 #include "opto/c2compiler.hpp"
  38 #include "opto/callGenerator.hpp"
  39 #include "opto/callnode.hpp"
  40 #include "opto/cfgnode.hpp"
  41 #include "opto/chaitin.hpp"
  42 #include "opto/compile.hpp"
  43 #include "opto/connode.hpp"
  44 #include "opto/divnode.hpp"
  45 #include "opto/escape.hpp"
  46 #include "opto/idealGraphPrinter.hpp"
  47 #include "opto/loopnode.hpp"
  48 #include "opto/machnode.hpp"
  49 #include "opto/macro.hpp"
  50 #include "opto/matcher.hpp"
  51 #include "opto/mathexactnode.hpp"
  52 #include "opto/memnode.hpp"
  53 #include "opto/mulnode.hpp"
  54 #include "opto/node.hpp"
  55 #include "opto/opcodes.hpp"
  56 #include "opto/output.hpp"
  57 #include "opto/parse.hpp"
  58 #include "opto/phaseX.hpp"
  59 #include "opto/rootnode.hpp"
  60 #include "opto/runtime.hpp"
  61 #include "opto/stringopts.hpp"
  62 #include "opto/type.hpp"
  63 #include "opto/vectornode.hpp"
  64 #include "runtime/arguments.hpp"
  65 #include "runtime/signature.hpp"
  66 #include "runtime/stubRoutines.hpp"
  67 #include "runtime/timer.hpp"
  68 #include "trace/tracing.hpp"
  69 #include "utilities/copy.hpp"
  70 #if defined AD_MD_HPP
  71 # include AD_MD_HPP
  72 #elif defined TARGET_ARCH_MODEL_x86_32
  73 # include "adfiles/ad_x86_32.hpp"
  74 #elif defined TARGET_ARCH_MODEL_x86_64
  75 # include "adfiles/ad_x86_64.hpp"
  76 #elif defined TARGET_ARCH_MODEL_sparc
  77 # include "adfiles/ad_sparc.hpp"
  78 #elif defined TARGET_ARCH_MODEL_zero
  79 # include "adfiles/ad_zero.hpp"
  80 #elif defined TARGET_ARCH_MODEL_ppc_64
  81 # include "adfiles/ad_ppc_64.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 (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 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 static inline bool not_a_node(const Node* n) {
 320   if (n == NULL)                   return true;
 321   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 322   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 323   return false;
 324 }
 325 
 326 // Identify all nodes that are reachable from below, useful.
 327 // Use breadth-first pass that records state in a Unique_Node_List,
 328 // recursive traversal is slower.
 329 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 330   int estimated_worklist_size = live_nodes();
 331   useful.map( estimated_worklist_size, NULL );  // preallocate space
 332 
 333   // Initialize worklist
 334   if (root() != NULL)     { useful.push(root()); }
 335   // If 'top' is cached, declare it useful to preserve cached node
 336   if( cached_top_node() ) { useful.push(cached_top_node()); }
 337 
 338   // Push all useful nodes onto the list, breadthfirst
 339   for( uint next = 0; next < useful.size(); ++next ) {
 340     assert( next < unique(), "Unique useful nodes < total nodes");
 341     Node *n  = useful.at(next);
 342     uint max = n->len();
 343     for( uint i = 0; i < max; ++i ) {
 344       Node *m = n->in(i);
 345       if (not_a_node(m))  continue;
 346       useful.push(m);
 347     }
 348   }
 349 }
 350 
 351 // Update dead_node_list with any missing dead nodes using useful
 352 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 353 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 354   uint max_idx = unique();
 355   VectorSet& useful_node_set = useful.member_set();
 356 
 357   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 358     // If node with index node_idx is not in useful set,
 359     // mark it as dead in dead node list.
 360     if (! useful_node_set.test(node_idx) ) {
 361       record_dead_node(node_idx);
 362     }
 363   }
 364 }
 365 
 366 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 367   int shift = 0;
 368   for (int i = 0; i < inlines->length(); i++) {
 369     CallGenerator* cg = inlines->at(i);
 370     CallNode* call = cg->call_node();
 371     if (shift > 0) {
 372       inlines->at_put(i-shift, cg);
 373     }
 374     if (!useful.member(call)) {
 375       shift++;
 376     }
 377   }
 378   inlines->trunc_to(inlines->length()-shift);
 379 }
 380 
 381 // Disconnect all useless nodes by disconnecting those at the boundary.
 382 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 383   uint next = 0;
 384   while (next < useful.size()) {
 385     Node *n = useful.at(next++);
 386     if (n->is_SafePoint()) {
 387       // We're done with a parsing phase. Replaced nodes are not valid
 388       // beyond that point.
 389       n->as_SafePoint()->delete_replaced_nodes();
 390     }
 391     // Use raw traversal of out edges since this code removes out edges
 392     int max = n->outcnt();
 393     for (int j = 0; j < max; ++j) {
 394       Node* child = n->raw_out(j);
 395       if (! useful.member(child)) {
 396         assert(!child->is_top() || child != top(),
 397                "If top is cached in Compile object it is in useful list");
 398         // Only need to remove this out-edge to the useless node
 399         n->raw_del_out(j);
 400         --j;
 401         --max;
 402       }
 403     }
 404     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 405       record_for_igvn(n->unique_out());
 406     }
 407   }
 408   // Remove useless macro and predicate opaq nodes
 409   for (int i = C->macro_count()-1; i >= 0; i--) {
 410     Node* n = C->macro_node(i);
 411     if (!useful.member(n)) {
 412       remove_macro_node(n);
 413     }
 414   }
 415   // Remove useless CastII nodes with range check dependency
 416   for (int i = range_check_cast_count() - 1; i >= 0; i--) {
 417     Node* cast = range_check_cast_node(i);
 418     if (!useful.member(cast)) {
 419       remove_range_check_cast(cast);
 420     }
 421   }
 422   // Remove useless expensive node
 423   for (int i = C->expensive_count()-1; i >= 0; i--) {
 424     Node* n = C->expensive_node(i);
 425     if (!useful.member(n)) {
 426       remove_expensive_node(n);
 427     }
 428   }
 429   // clean up the late inline lists
 430   remove_useless_late_inlines(&_string_late_inlines, useful);
 431   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 432   remove_useless_late_inlines(&_late_inlines, useful);
 433   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 434 }
 435 
 436 //------------------------------frame_size_in_words-----------------------------
 437 // frame_slots in units of words
 438 int Compile::frame_size_in_words() const {
 439   // shift is 0 in LP32 and 1 in LP64
 440   const int shift = (LogBytesPerWord - LogBytesPerInt);
 441   int words = _frame_slots >> shift;
 442   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 443   return words;
 444 }
 445 
 446 // To bang the stack of this compiled method we use the stack size
 447 // that the interpreter would need in case of a deoptimization. This
 448 // removes the need to bang the stack in the deoptimization blob which
 449 // in turn simplifies stack overflow handling.
 450 int Compile::bang_size_in_bytes() const {
 451   return MAX2(_interpreter_frame_size, frame_size_in_bytes());
 452 }
 453 
 454 // ============================================================================
 455 //------------------------------CompileWrapper---------------------------------
 456 class CompileWrapper : public StackObj {
 457   Compile *const _compile;
 458  public:
 459   CompileWrapper(Compile* compile);
 460 
 461   ~CompileWrapper();
 462 };
 463 
 464 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 465   // the Compile* pointer is stored in the current ciEnv:
 466   ciEnv* env = compile->env();
 467   assert(env == ciEnv::current(), "must already be a ciEnv active");
 468   assert(env->compiler_data() == NULL, "compile already active?");
 469   env->set_compiler_data(compile);
 470   assert(compile == Compile::current(), "sanity");
 471 
 472   compile->set_type_dict(NULL);
 473   compile->set_type_hwm(NULL);
 474   compile->set_type_last_size(0);
 475   compile->set_last_tf(NULL, NULL);
 476   compile->set_indexSet_arena(NULL);
 477   compile->set_indexSet_free_block_list(NULL);
 478   compile->init_type_arena();
 479   Type::Initialize(compile);
 480   _compile->set_scratch_buffer_blob(NULL);
 481   _compile->begin_method();
 482 }
 483 CompileWrapper::~CompileWrapper() {
 484   _compile->end_method();
 485   if (_compile->scratch_buffer_blob() != NULL)
 486     BufferBlob::free(_compile->scratch_buffer_blob());
 487   _compile->env()->set_compiler_data(NULL);
 488 }
 489 
 490 
 491 //----------------------------print_compile_messages---------------------------
 492 void Compile::print_compile_messages() {
 493 #ifndef PRODUCT
 494   // Check if recompiling
 495   if (_subsume_loads == false && PrintOpto) {
 496     // Recompiling without allowing machine instructions to subsume loads
 497     tty->print_cr("*********************************************************");
 498     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 499     tty->print_cr("*********************************************************");
 500   }
 501   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 502     // Recompiling without escape analysis
 503     tty->print_cr("*********************************************************");
 504     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 505     tty->print_cr("*********************************************************");
 506   }
 507   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 508     // Recompiling without boxing elimination
 509     tty->print_cr("*********************************************************");
 510     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 511     tty->print_cr("*********************************************************");
 512   }
 513   if (env()->break_at_compile()) {
 514     // Open the debugger when compiling this method.
 515     tty->print("### Breaking when compiling: ");
 516     method()->print_short_name();
 517     tty->cr();
 518     BREAKPOINT;
 519   }
 520 
 521   if( PrintOpto ) {
 522     if (is_osr_compilation()) {
 523       tty->print("[OSR]%3d", _compile_id);
 524     } else {
 525       tty->print("%3d", _compile_id);
 526     }
 527   }
 528 #endif
 529 }
 530 
 531 
 532 //-----------------------init_scratch_buffer_blob------------------------------
 533 // Construct a temporary BufferBlob and cache it for this compile.
 534 void Compile::init_scratch_buffer_blob(int const_size) {
 535   // If there is already a scratch buffer blob allocated and the
 536   // constant section is big enough, use it.  Otherwise free the
 537   // current and allocate a new one.
 538   BufferBlob* blob = scratch_buffer_blob();
 539   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 540     // Use the current blob.
 541   } else {
 542     if (blob != NULL) {
 543       BufferBlob::free(blob);
 544     }
 545 
 546     ResourceMark rm;
 547     _scratch_const_size = const_size;
 548     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
 549     blob = BufferBlob::create("Compile::scratch_buffer", size);
 550     // Record the buffer blob for next time.
 551     set_scratch_buffer_blob(blob);
 552     // Have we run out of code space?
 553     if (scratch_buffer_blob() == NULL) {
 554       // Let CompilerBroker disable further compilations.
 555       record_failure("Not enough space for scratch buffer in CodeCache");
 556       return;
 557     }
 558   }
 559 
 560   // Initialize the relocation buffers
 561   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 562   set_scratch_locs_memory(locs_buf);
 563 }
 564 
 565 
 566 //-----------------------scratch_emit_size-------------------------------------
 567 // Helper function that computes size by emitting code
 568 uint Compile::scratch_emit_size(const Node* n) {
 569   // Start scratch_emit_size section.
 570   set_in_scratch_emit_size(true);
 571 
 572   // Emit into a trash buffer and count bytes emitted.
 573   // This is a pretty expensive way to compute a size,
 574   // but it works well enough if seldom used.
 575   // All common fixed-size instructions are given a size
 576   // method by the AD file.
 577   // Note that the scratch buffer blob and locs memory are
 578   // allocated at the beginning of the compile task, and
 579   // may be shared by several calls to scratch_emit_size.
 580   // The allocation of the scratch buffer blob is particularly
 581   // expensive, since it has to grab the code cache lock.
 582   BufferBlob* blob = this->scratch_buffer_blob();
 583   assert(blob != NULL, "Initialize BufferBlob at start");
 584   assert(blob->size() > MAX_inst_size, "sanity");
 585   relocInfo* locs_buf = scratch_locs_memory();
 586   address blob_begin = blob->content_begin();
 587   address blob_end   = (address)locs_buf;
 588   assert(blob->content_contains(blob_end), "sanity");
 589   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 590   buf.initialize_consts_size(_scratch_const_size);
 591   buf.initialize_stubs_size(MAX_stubs_size);
 592   assert(locs_buf != NULL, "sanity");
 593   int lsize = MAX_locs_size / 3;
 594   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 595   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 596   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 597 
 598   // Do the emission.
 599 
 600   Label fakeL; // Fake label for branch instructions.
 601   Label*   saveL = NULL;
 602   uint save_bnum = 0;
 603   bool is_branch = n->is_MachBranch();
 604   if (is_branch) {
 605     MacroAssembler masm(&buf);
 606     masm.bind(fakeL);
 607     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 608     n->as_MachBranch()->label_set(&fakeL, 0);
 609   }
 610   n->emit(buf, this->regalloc());
 611 
 612   // Emitting into the scratch buffer should not fail
 613   assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason()));
 614 
 615   if (is_branch) // Restore label.
 616     n->as_MachBranch()->label_set(saveL, save_bnum);
 617 
 618   // End scratch_emit_size section.
 619   set_in_scratch_emit_size(false);
 620 
 621   return buf.insts_size();
 622 }
 623 
 624 
 625 // ============================================================================
 626 //------------------------------Compile standard-------------------------------
 627 debug_only( int Compile::_debug_idx = 100000; )
 628 
 629 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 630 // the continuation bci for on stack replacement.
 631 
 632 
 633 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 634                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
 635                 : Phase(Compiler),
 636                   _env(ci_env),
 637                   _log(ci_env->log()),
 638                   _compile_id(ci_env->compile_id()),
 639                   _save_argument_registers(false),
 640                   _stub_name(NULL),
 641                   _stub_function(NULL),
 642                   _stub_entry_point(NULL),
 643                   _method(target),
 644                   _entry_bci(osr_bci),
 645                   _initial_gvn(NULL),
 646                   _for_igvn(NULL),
 647                   _warm_calls(NULL),
 648                   _subsume_loads(subsume_loads),
 649                   _do_escape_analysis(do_escape_analysis),
 650                   _eliminate_boxing(eliminate_boxing),
 651                   _failure_reason(NULL),
 652                   _code_buffer("Compile::Fill_buffer"),
 653                   _orig_pc_slot(0),
 654                   _orig_pc_slot_offset_in_bytes(0),
 655                   _has_method_handle_invokes(false),
 656                   _mach_constant_base_node(NULL),
 657                   _node_bundling_limit(0),
 658                   _node_bundling_base(NULL),
 659                   _java_calls(0),
 660                   _inner_loops(0),
 661                   _scratch_const_size(-1),
 662                   _in_scratch_emit_size(false),
 663                   _dead_node_list(comp_arena()),
 664                   _dead_node_count(0),
 665 #ifndef PRODUCT
 666                   _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
 667                   _in_dump_cnt(0),
 668                   _printer(IdealGraphPrinter::printer()),
 669 #endif
 670                   _congraph(NULL),
 671                   _comp_arena(mtCompiler),
 672                   _node_arena(mtCompiler),
 673                   _old_arena(mtCompiler),
 674                   _Compile_types(mtCompiler),
 675                   _replay_inline_data(NULL),
 676                   _late_inlines(comp_arena(), 2, 0, NULL),
 677                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 678                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 679                   _late_inlines_pos(0),
 680                   _number_of_mh_late_inlines(0),
 681                   _inlining_progress(false),
 682                   _inlining_incrementally(false),
 683                   _print_inlining_list(NULL),
 684                   _print_inlining_idx(0),
 685                   _interpreter_frame_size(0),
 686                   _max_node_limit(MaxNodeLimit) {
 687   C = this;
 688 
 689   CompileWrapper cw(this);
 690 #ifndef PRODUCT
 691   if (TimeCompiler2) {
 692     tty->print(" ");
 693     target->holder()->name()->print();
 694     tty->print(".");
 695     target->print_short_name();
 696     tty->print("  ");
 697   }
 698   TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
 699   TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
 700   bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
 701   if (!print_opto_assembly) {
 702     bool print_assembly = (PrintAssembly || _method->should_print_assembly());
 703     if (print_assembly && !Disassembler::can_decode()) {
 704       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 705       print_opto_assembly = true;
 706     }
 707   }
 708   set_print_assembly(print_opto_assembly);
 709   set_parsed_irreducible_loop(false);
 710 
 711   if (method()->has_option("ReplayInline")) {
 712     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 713   }
 714 #endif
 715   set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
 716   set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
 717   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 718 
 719   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 720     // Make sure the method being compiled gets its own MDO,
 721     // so we can at least track the decompile_count().
 722     // Need MDO to record RTM code generation state.
 723     method()->ensure_method_data();
 724   }
 725 
 726   Init(::AliasLevel);
 727 
 728 
 729   print_compile_messages();
 730 
 731   _ilt = InlineTree::build_inline_tree_root();
 732 
 733   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 734   assert(num_alias_types() >= AliasIdxRaw, "");
 735 
 736 #define MINIMUM_NODE_HASH  1023
 737   // Node list that Iterative GVN will start with
 738   Unique_Node_List for_igvn(comp_arena());
 739   set_for_igvn(&for_igvn);
 740 
 741   // GVN that will be run immediately on new nodes
 742   uint estimated_size = method()->code_size()*4+64;
 743   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 744   PhaseGVN gvn(node_arena(), estimated_size);
 745   set_initial_gvn(&gvn);
 746 
 747   if (print_inlining() || print_intrinsics()) {
 748     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
 749   }
 750   { // Scope for timing the parser
 751     TracePhase t3("parse", &_t_parser, true);
 752 
 753     // Put top into the hash table ASAP.
 754     initial_gvn()->transform_no_reclaim(top());
 755 
 756     // Set up tf(), start(), and find a CallGenerator.
 757     CallGenerator* cg = NULL;
 758     if (is_osr_compilation()) {
 759       const TypeTuple *domain = StartOSRNode::osr_domain();
 760       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 761       init_tf(TypeFunc::make(domain, range));
 762       StartNode* s = new (this) StartOSRNode(root(), domain);
 763       initial_gvn()->set_type_bottom(s);
 764       init_start(s);
 765       cg = CallGenerator::for_osr(method(), entry_bci());
 766     } else {
 767       // Normal case.
 768       init_tf(TypeFunc::make(method()));
 769       StartNode* s = new (this) StartNode(root(), tf()->domain());
 770       initial_gvn()->set_type_bottom(s);
 771       init_start(s);
 772       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
 773         // With java.lang.ref.reference.get() we must go through the
 774         // intrinsic when G1 is enabled - even when get() is the root
 775         // method of the compile - so that, if necessary, the value in
 776         // the referent field of the reference object gets recorded by
 777         // the pre-barrier code.
 778         // Specifically, if G1 is enabled, the value in the referent
 779         // field is recorded by the G1 SATB pre barrier. This will
 780         // result in the referent being marked live and the reference
 781         // object removed from the list of discovered references during
 782         // reference processing.
 783         cg = find_intrinsic(method(), false);
 784       }
 785       if (cg == NULL) {
 786         float past_uses = method()->interpreter_invocation_count();
 787         float expected_uses = past_uses;
 788         cg = CallGenerator::for_inline(method(), expected_uses);
 789       }
 790     }
 791     if (failing())  return;
 792     if (cg == NULL) {
 793       record_method_not_compilable_all_tiers("cannot parse method");
 794       return;
 795     }
 796     JVMState* jvms = build_start_state(start(), tf());
 797     if ((jvms = cg->generate(jvms)) == NULL) {
 798       record_method_not_compilable("method parse failed");
 799       return;
 800     }
 801     GraphKit kit(jvms);
 802 
 803     if (!kit.stopped()) {
 804       // Accept return values, and transfer control we know not where.
 805       // This is done by a special, unique ReturnNode bound to root.
 806       return_values(kit.jvms());
 807     }
 808 
 809     if (kit.has_exceptions()) {
 810       // Any exceptions that escape from this call must be rethrown
 811       // to whatever caller is dynamically above us on the stack.
 812       // This is done by a special, unique RethrowNode bound to root.
 813       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 814     }
 815 
 816     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 817 
 818     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 819       inline_string_calls(true);
 820     }
 821 
 822     if (failing())  return;
 823 
 824     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 825 
 826     // Remove clutter produced by parsing.
 827     if (!failing()) {
 828       ResourceMark rm;
 829       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 830     }
 831   }
 832 
 833   // Note:  Large methods are capped off in do_one_bytecode().
 834   if (failing())  return;
 835 
 836   // After parsing, node notes are no longer automagic.
 837   // They must be propagated by register_new_node_with_optimizer(),
 838   // clone(), or the like.
 839   set_default_node_notes(NULL);
 840 
 841   for (;;) {
 842     int successes = Inline_Warm();
 843     if (failing())  return;
 844     if (successes == 0)  break;
 845   }
 846 
 847   // Drain the list.
 848   Finish_Warm();
 849 #ifndef PRODUCT
 850   if (_printer) {
 851     _printer->print_inlining(this);
 852   }
 853 #endif
 854 
 855   if (failing())  return;
 856   NOT_PRODUCT( verify_graph_edges(); )
 857 
 858   // Now optimize
 859   Optimize();
 860   if (failing())  return;
 861   NOT_PRODUCT( verify_graph_edges(); )
 862 
 863 #ifndef PRODUCT
 864   if (PrintIdeal) {
 865     ttyLocker ttyl;  // keep the following output all in one block
 866     // This output goes directly to the tty, not the compiler log.
 867     // To enable tools to match it up with the compilation activity,
 868     // be sure to tag this tty output with the compile ID.
 869     if (xtty != NULL) {
 870       xtty->head("ideal compile_id='%d'%s", compile_id(),
 871                  is_osr_compilation()    ? " compile_kind='osr'" :
 872                  "");
 873     }
 874     root()->dump(9999);
 875     if (xtty != NULL) {
 876       xtty->tail("ideal");
 877     }
 878   }
 879 #endif
 880 
 881   NOT_PRODUCT( verify_barriers(); )
 882 
 883   // Dump compilation data to replay it.
 884   if (method()->has_option("DumpReplay")) {
 885     env()->dump_replay_data(_compile_id);
 886   }
 887   if (method()->has_option("DumpInline") && (ilt() != NULL)) {
 888     env()->dump_inline_data(_compile_id);
 889   }
 890 
 891   // Now that we know the size of all the monitors we can add a fixed slot
 892   // for the original deopt pc.
 893 
 894   _orig_pc_slot =  fixed_slots();
 895   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 896   set_fixed_slots(next_slot);
 897 
 898   // Compute when to use implicit null checks. Used by matching trap based
 899   // nodes and NullCheck optimization.
 900   set_allowed_deopt_reasons();
 901 
 902   // Now generate code
 903   Code_Gen();
 904   if (failing())  return;
 905 
 906   // Check if we want to skip execution of all compiled code.
 907   {
 908 #ifndef PRODUCT
 909     if (OptoNoExecute) {
 910       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 911       return;
 912     }
 913     TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
 914 #endif
 915 
 916     if (is_osr_compilation()) {
 917       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 918       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 919     } else {
 920       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 921       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 922     }
 923 
 924     env()->register_method(_method, _entry_bci,
 925                            &_code_offsets,
 926                            _orig_pc_slot_offset_in_bytes,
 927                            code_buffer(),
 928                            frame_size_in_words(), _oop_map_set,
 929                            &_handler_table, &_inc_table,
 930                            compiler,
 931                            env()->comp_level(),
 932                            has_unsafe_access(),
 933                            SharedRuntime::is_wide_vector(max_vector_size()),
 934                            rtm_state()
 935                            );
 936 
 937     if (log() != NULL) // Print code cache state into compiler log
 938       log()->code_cache_state();
 939   }
 940 }
 941 
 942 //------------------------------Compile----------------------------------------
 943 // Compile a runtime stub
 944 Compile::Compile( ciEnv* ci_env,
 945                   TypeFunc_generator generator,
 946                   address stub_function,
 947                   const char *stub_name,
 948                   int is_fancy_jump,
 949                   bool pass_tls,
 950                   bool save_arg_registers,
 951                   bool return_pc )
 952   : Phase(Compiler),
 953     _env(ci_env),
 954     _log(ci_env->log()),
 955     _compile_id(0),
 956     _save_argument_registers(save_arg_registers),
 957     _method(NULL),
 958     _stub_name(stub_name),
 959     _stub_function(stub_function),
 960     _stub_entry_point(NULL),
 961     _entry_bci(InvocationEntryBci),
 962     _initial_gvn(NULL),
 963     _for_igvn(NULL),
 964     _warm_calls(NULL),
 965     _orig_pc_slot(0),
 966     _orig_pc_slot_offset_in_bytes(0),
 967     _subsume_loads(true),
 968     _do_escape_analysis(false),
 969     _eliminate_boxing(false),
 970     _failure_reason(NULL),
 971     _code_buffer("Compile::Fill_buffer"),
 972     _has_method_handle_invokes(false),
 973     _mach_constant_base_node(NULL),
 974     _node_bundling_limit(0),
 975     _node_bundling_base(NULL),
 976     _java_calls(0),
 977     _inner_loops(0),
 978 #ifndef PRODUCT
 979     _trace_opto_output(TraceOptoOutput),
 980     _in_dump_cnt(0),
 981     _printer(NULL),
 982 #endif
 983     _comp_arena(mtCompiler),
 984     _node_arena(mtCompiler),
 985     _old_arena(mtCompiler),
 986     _Compile_types(mtCompiler),
 987     _dead_node_list(comp_arena()),
 988     _dead_node_count(0),
 989     _congraph(NULL),
 990     _replay_inline_data(NULL),
 991     _number_of_mh_late_inlines(0),
 992     _inlining_progress(false),
 993     _inlining_incrementally(false),
 994     _print_inlining_list(NULL),
 995     _print_inlining_idx(0),
 996     _allowed_reasons(0),
 997     _interpreter_frame_size(0),
 998     _max_node_limit(MaxNodeLimit) {
 999   C = this;
1000 
1001 #ifndef PRODUCT
1002   TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
1003   TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
1004   set_print_assembly(PrintFrameConverterAssembly);
1005   set_parsed_irreducible_loop(false);
1006 #endif
1007   set_has_irreducible_loop(false); // no loops
1008 
1009   CompileWrapper cw(this);
1010   Init(/*AliasLevel=*/ 0);
1011   init_tf((*generator)());
1012 
1013   {
1014     // The following is a dummy for the sake of GraphKit::gen_stub
1015     Unique_Node_List for_igvn(comp_arena());
1016     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1017     PhaseGVN gvn(Thread::current()->resource_area(),255);
1018     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1019     gvn.transform_no_reclaim(top());
1020 
1021     GraphKit kit;
1022     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1023   }
1024 
1025   NOT_PRODUCT( verify_graph_edges(); )
1026   Code_Gen();
1027   if (failing())  return;
1028 
1029 
1030   // Entry point will be accessed using compile->stub_entry_point();
1031   if (code_buffer() == NULL) {
1032     Matcher::soft_match_failure();
1033   } else {
1034     if (PrintAssembly && (WizardMode || Verbose))
1035       tty->print_cr("### Stub::%s", stub_name);
1036 
1037     if (!failing()) {
1038       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1039 
1040       // Make the NMethod
1041       // For now we mark the frame as never safe for profile stackwalking
1042       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1043                                                       code_buffer(),
1044                                                       CodeOffsets::frame_never_safe,
1045                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1046                                                       frame_size_in_words(),
1047                                                       _oop_map_set,
1048                                                       save_arg_registers);
1049       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1050 
1051       _stub_entry_point = rs->entry_point();
1052     }
1053   }
1054 }
1055 
1056 //------------------------------Init-------------------------------------------
1057 // Prepare for a single compilation
1058 void Compile::Init(int aliaslevel) {
1059   _unique  = 0;
1060   _regalloc = NULL;
1061 
1062   _tf      = NULL;  // filled in later
1063   _top     = NULL;  // cached later
1064   _matcher = NULL;  // filled in later
1065   _cfg     = NULL;  // filled in later
1066 
1067   set_24_bit_selection_and_mode(Use24BitFP, false);
1068 
1069   _node_note_array = NULL;
1070   _default_node_notes = NULL;
1071 
1072   _immutable_memory = NULL; // filled in at first inquiry
1073 
1074   // Globally visible Nodes
1075   // First set TOP to NULL to give safe behavior during creation of RootNode
1076   set_cached_top_node(NULL);
1077   set_root(new (this) RootNode());
1078   // Now that you have a Root to point to, create the real TOP
1079   set_cached_top_node( new (this) ConNode(Type::TOP) );
1080   set_recent_alloc(NULL, NULL);
1081 
1082   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1083   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1084   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1085   env()->set_dependencies(new Dependencies(env()));
1086 
1087   _fixed_slots = 0;
1088   set_has_split_ifs(false);
1089   set_has_loops(has_method() && method()->has_loops()); // first approximation
1090   set_has_stringbuilder(false);
1091   set_has_boxed_value(false);
1092   _trap_can_recompile = false;  // no traps emitted yet
1093   _major_progress = true; // start out assuming good things will happen
1094   set_has_unsafe_access(false);
1095   set_max_vector_size(0);
1096   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1097   set_decompile_count(0);
1098 
1099   set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1100   set_num_loop_opts(LoopOptsCount);
1101   set_do_inlining(Inline);
1102   set_max_inline_size(MaxInlineSize);
1103   set_freq_inline_size(FreqInlineSize);
1104   set_do_scheduling(OptoScheduling);
1105   set_do_count_invocations(false);
1106   set_do_method_data_update(false);
1107   set_rtm_state(NoRTM); // No RTM lock eliding by default
1108   method_has_option_value("MaxNodeLimit", _max_node_limit);
1109 #if INCLUDE_RTM_OPT
1110   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1111     int rtm_state = method()->method_data()->rtm_state();
1112     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1113       // Don't generate RTM lock eliding code.
1114       set_rtm_state(NoRTM);
1115     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1116       // Generate RTM lock eliding code without abort ratio calculation code.
1117       set_rtm_state(UseRTM);
1118     } else if (UseRTMDeopt) {
1119       // Generate RTM lock eliding code and include abort ratio calculation
1120       // code if UseRTMDeopt is on.
1121       set_rtm_state(ProfileRTM);
1122     }
1123   }
1124 #endif
1125   if (debug_info()->recording_non_safepoints()) {
1126     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1127                         (comp_arena(), 8, 0, NULL));
1128     set_default_node_notes(Node_Notes::make(this));
1129   }
1130 
1131   // // -- Initialize types before each compile --
1132   // // Update cached type information
1133   // if( _method && _method->constants() )
1134   //   Type::update_loaded_types(_method, _method->constants());
1135 
1136   // Init alias_type map.
1137   if (!_do_escape_analysis && aliaslevel == 3)
1138     aliaslevel = 2;  // No unique types without escape analysis
1139   _AliasLevel = aliaslevel;
1140   const int grow_ats = 16;
1141   _max_alias_types = grow_ats;
1142   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1143   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1144   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1145   {
1146     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1147   }
1148   // Initialize the first few types.
1149   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1150   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1151   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1152   _num_alias_types = AliasIdxRaw+1;
1153   // Zero out the alias type cache.
1154   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1155   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1156   probe_alias_cache(NULL)->_index = AliasIdxTop;
1157 
1158   _intrinsics = NULL;
1159   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1160   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1161   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1162   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1163   register_library_intrinsics();
1164 }
1165 
1166 //---------------------------init_start----------------------------------------
1167 // Install the StartNode on this compile object.
1168 void Compile::init_start(StartNode* s) {
1169   if (failing())
1170     return; // already failing
1171   assert(s == start(), "");
1172 }
1173 
1174 StartNode* Compile::start() const {
1175   assert(!failing(), "");
1176   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1177     Node* start = root()->fast_out(i);
1178     if( start->is_Start() )
1179       return start->as_Start();
1180   }
1181   fatal("Did not find Start node!");
1182   return NULL;
1183 }
1184 
1185 //-------------------------------immutable_memory-------------------------------------
1186 // Access immutable memory
1187 Node* Compile::immutable_memory() {
1188   if (_immutable_memory != NULL) {
1189     return _immutable_memory;
1190   }
1191   StartNode* s = start();
1192   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1193     Node *p = s->fast_out(i);
1194     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1195       _immutable_memory = p;
1196       return _immutable_memory;
1197     }
1198   }
1199   ShouldNotReachHere();
1200   return NULL;
1201 }
1202 
1203 //----------------------set_cached_top_node------------------------------------
1204 // Install the cached top node, and make sure Node::is_top works correctly.
1205 void Compile::set_cached_top_node(Node* tn) {
1206   if (tn != NULL)  verify_top(tn);
1207   Node* old_top = _top;
1208   _top = tn;
1209   // Calling Node::setup_is_top allows the nodes the chance to adjust
1210   // their _out arrays.
1211   if (_top != NULL)     _top->setup_is_top();
1212   if (old_top != NULL)  old_top->setup_is_top();
1213   assert(_top == NULL || top()->is_top(), "");
1214 }
1215 
1216 #ifdef ASSERT
1217 uint Compile::count_live_nodes_by_graph_walk() {
1218   Unique_Node_List useful(comp_arena());
1219   // Get useful node list by walking the graph.
1220   identify_useful_nodes(useful);
1221   return useful.size();
1222 }
1223 
1224 void Compile::print_missing_nodes() {
1225 
1226   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1227   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1228     return;
1229   }
1230 
1231   // This is an expensive function. It is executed only when the user
1232   // specifies VerifyIdealNodeCount option or otherwise knows the
1233   // additional work that needs to be done to identify reachable nodes
1234   // by walking the flow graph and find the missing ones using
1235   // _dead_node_list.
1236 
1237   Unique_Node_List useful(comp_arena());
1238   // Get useful node list by walking the graph.
1239   identify_useful_nodes(useful);
1240 
1241   uint l_nodes = C->live_nodes();
1242   uint l_nodes_by_walk = useful.size();
1243 
1244   if (l_nodes != l_nodes_by_walk) {
1245     if (_log != NULL) {
1246       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1247       _log->stamp();
1248       _log->end_head();
1249     }
1250     VectorSet& useful_member_set = useful.member_set();
1251     int last_idx = l_nodes_by_walk;
1252     for (int i = 0; i < last_idx; i++) {
1253       if (useful_member_set.test(i)) {
1254         if (_dead_node_list.test(i)) {
1255           if (_log != NULL) {
1256             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1257           }
1258           if (PrintIdealNodeCount) {
1259             // Print the log message to tty
1260               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1261               useful.at(i)->dump();
1262           }
1263         }
1264       }
1265       else if (! _dead_node_list.test(i)) {
1266         if (_log != NULL) {
1267           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1268         }
1269         if (PrintIdealNodeCount) {
1270           // Print the log message to tty
1271           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1272         }
1273       }
1274     }
1275     if (_log != NULL) {
1276       _log->tail("mismatched_nodes");
1277     }
1278   }
1279 }
1280 #endif
1281 
1282 #ifndef PRODUCT
1283 void Compile::verify_top(Node* tn) const {
1284   if (tn != NULL) {
1285     assert(tn->is_Con(), "top node must be a constant");
1286     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1287     assert(tn->in(0) != NULL, "must have live top node");
1288   }
1289 }
1290 #endif
1291 
1292 
1293 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1294 
1295 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1296   guarantee(arr != NULL, "");
1297   int num_blocks = arr->length();
1298   if (grow_by < num_blocks)  grow_by = num_blocks;
1299   int num_notes = grow_by * _node_notes_block_size;
1300   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1301   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1302   while (num_notes > 0) {
1303     arr->append(notes);
1304     notes     += _node_notes_block_size;
1305     num_notes -= _node_notes_block_size;
1306   }
1307   assert(num_notes == 0, "exact multiple, please");
1308 }
1309 
1310 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1311   if (source == NULL || dest == NULL)  return false;
1312 
1313   if (dest->is_Con())
1314     return false;               // Do not push debug info onto constants.
1315 
1316 #ifdef ASSERT
1317   // Leave a bread crumb trail pointing to the original node:
1318   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1319     dest->set_debug_orig(source);
1320   }
1321 #endif
1322 
1323   if (node_note_array() == NULL)
1324     return false;               // Not collecting any notes now.
1325 
1326   // This is a copy onto a pre-existing node, which may already have notes.
1327   // If both nodes have notes, do not overwrite any pre-existing notes.
1328   Node_Notes* source_notes = node_notes_at(source->_idx);
1329   if (source_notes == NULL || source_notes->is_clear())  return false;
1330   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1331   if (dest_notes == NULL || dest_notes->is_clear()) {
1332     return set_node_notes_at(dest->_idx, source_notes);
1333   }
1334 
1335   Node_Notes merged_notes = (*source_notes);
1336   // The order of operations here ensures that dest notes will win...
1337   merged_notes.update_from(dest_notes);
1338   return set_node_notes_at(dest->_idx, &merged_notes);
1339 }
1340 
1341 
1342 //--------------------------allow_range_check_smearing-------------------------
1343 // Gating condition for coalescing similar range checks.
1344 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1345 // single covering check that is at least as strong as any of them.
1346 // If the optimization succeeds, the simplified (strengthened) range check
1347 // will always succeed.  If it fails, we will deopt, and then give up
1348 // on the optimization.
1349 bool Compile::allow_range_check_smearing() const {
1350   // If this method has already thrown a range-check,
1351   // assume it was because we already tried range smearing
1352   // and it failed.
1353   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1354   return !already_trapped;
1355 }
1356 
1357 
1358 //------------------------------flatten_alias_type-----------------------------
1359 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1360   int offset = tj->offset();
1361   TypePtr::PTR ptr = tj->ptr();
1362 
1363   // Known instance (scalarizable allocation) alias only with itself.
1364   bool is_known_inst = tj->isa_oopptr() != NULL &&
1365                        tj->is_oopptr()->is_known_instance();
1366 
1367   // Process weird unsafe references.
1368   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1369     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1370     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1371     tj = TypeOopPtr::BOTTOM;
1372     ptr = tj->ptr();
1373     offset = tj->offset();
1374   }
1375 
1376   // Array pointers need some flattening
1377   const TypeAryPtr *ta = tj->isa_aryptr();
1378   if (ta && ta->is_stable()) {
1379     // Erase stability property for alias analysis.
1380     tj = ta = ta->cast_to_stable(false);
1381   }
1382   if( ta && is_known_inst ) {
1383     if ( offset != Type::OffsetBot &&
1384          offset > arrayOopDesc::length_offset_in_bytes() ) {
1385       offset = Type::OffsetBot; // Flatten constant access into array body only
1386       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1387     }
1388   } else if( ta && _AliasLevel >= 2 ) {
1389     // For arrays indexed by constant indices, we flatten the alias
1390     // space to include all of the array body.  Only the header, klass
1391     // and array length can be accessed un-aliased.
1392     if( offset != Type::OffsetBot ) {
1393       if( ta->const_oop() ) { // MethodData* or Method*
1394         offset = Type::OffsetBot;   // Flatten constant access into array body
1395         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1396       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1397         // range is OK as-is.
1398         tj = ta = TypeAryPtr::RANGE;
1399       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1400         tj = TypeInstPtr::KLASS; // all klass loads look alike
1401         ta = TypeAryPtr::RANGE; // generic ignored junk
1402         ptr = TypePtr::BotPTR;
1403       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1404         tj = TypeInstPtr::MARK;
1405         ta = TypeAryPtr::RANGE; // generic ignored junk
1406         ptr = TypePtr::BotPTR;
1407       } else {                  // Random constant offset into array body
1408         offset = Type::OffsetBot;   // Flatten constant access into array body
1409         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1410       }
1411     }
1412     // Arrays of fixed size alias with arrays of unknown size.
1413     if (ta->size() != TypeInt::POS) {
1414       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1415       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1416     }
1417     // Arrays of known objects become arrays of unknown objects.
1418     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1419       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1420       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1421     }
1422     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1423       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1424       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1425     }
1426     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1427     // cannot be distinguished by bytecode alone.
1428     if (ta->elem() == TypeInt::BOOL) {
1429       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1430       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1431       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1432     }
1433     // During the 2nd round of IterGVN, NotNull castings are removed.
1434     // Make sure the Bottom and NotNull variants alias the same.
1435     // Also, make sure exact and non-exact variants alias the same.
1436     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1437       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1438     }
1439   }
1440 
1441   // Oop pointers need some flattening
1442   const TypeInstPtr *to = tj->isa_instptr();
1443   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1444     ciInstanceKlass *k = to->klass()->as_instance_klass();
1445     if( ptr == TypePtr::Constant ) {
1446       if (to->klass() != ciEnv::current()->Class_klass() ||
1447           offset < k->size_helper() * wordSize) {
1448         // No constant oop pointers (such as Strings); they alias with
1449         // unknown strings.
1450         assert(!is_known_inst, "not scalarizable allocation");
1451         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1452       }
1453     } else if( is_known_inst ) {
1454       tj = to; // Keep NotNull and klass_is_exact for instance type
1455     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1456       // During the 2nd round of IterGVN, NotNull castings are removed.
1457       // Make sure the Bottom and NotNull variants alias the same.
1458       // Also, make sure exact and non-exact variants alias the same.
1459       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1460     }
1461     if (to->speculative() != NULL) {
1462       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1463     }
1464     // Canonicalize the holder of this field
1465     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1466       // First handle header references such as a LoadKlassNode, even if the
1467       // object's klass is unloaded at compile time (4965979).
1468       if (!is_known_inst) { // Do it only for non-instance types
1469         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1470       }
1471     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1472       // Static fields are in the space above the normal instance
1473       // fields in the java.lang.Class instance.
1474       if (to->klass() != ciEnv::current()->Class_klass()) {
1475         to = NULL;
1476         tj = TypeOopPtr::BOTTOM;
1477         offset = tj->offset();
1478       }
1479     } else {
1480       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1481       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1482         if( is_known_inst ) {
1483           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1484         } else {
1485           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1486         }
1487       }
1488     }
1489   }
1490 
1491   // Klass pointers to object array klasses need some flattening
1492   const TypeKlassPtr *tk = tj->isa_klassptr();
1493   if( tk ) {
1494     // If we are referencing a field within a Klass, we need
1495     // to assume the worst case of an Object.  Both exact and
1496     // inexact types must flatten to the same alias class so
1497     // use NotNull as the PTR.
1498     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1499 
1500       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1501                                    TypeKlassPtr::OBJECT->klass(),
1502                                    offset);
1503     }
1504 
1505     ciKlass* klass = tk->klass();
1506     if( klass->is_obj_array_klass() ) {
1507       ciKlass* k = TypeAryPtr::OOPS->klass();
1508       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1509         k = TypeInstPtr::BOTTOM->klass();
1510       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1511     }
1512 
1513     // Check for precise loads from the primary supertype array and force them
1514     // to the supertype cache alias index.  Check for generic array loads from
1515     // the primary supertype array and also force them to the supertype cache
1516     // alias index.  Since the same load can reach both, we need to merge
1517     // these 2 disparate memories into the same alias class.  Since the
1518     // primary supertype array is read-only, there's no chance of confusion
1519     // where we bypass an array load and an array store.
1520     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1521     if (offset == Type::OffsetBot ||
1522         (offset >= primary_supers_offset &&
1523          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1524         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1525       offset = in_bytes(Klass::secondary_super_cache_offset());
1526       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1527     }
1528   }
1529 
1530   // Flatten all Raw pointers together.
1531   if (tj->base() == Type::RawPtr)
1532     tj = TypeRawPtr::BOTTOM;
1533 
1534   if (tj->base() == Type::AnyPtr)
1535     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1536 
1537   // Flatten all to bottom for now
1538   switch( _AliasLevel ) {
1539   case 0:
1540     tj = TypePtr::BOTTOM;
1541     break;
1542   case 1:                       // Flatten to: oop, static, field or array
1543     switch (tj->base()) {
1544     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1545     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1546     case Type::AryPtr:   // do not distinguish arrays at all
1547     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1548     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1549     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1550     default: ShouldNotReachHere();
1551     }
1552     break;
1553   case 2:                       // No collapsing at level 2; keep all splits
1554   case 3:                       // No collapsing at level 3; keep all splits
1555     break;
1556   default:
1557     Unimplemented();
1558   }
1559 
1560   offset = tj->offset();
1561   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1562 
1563   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1564           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1565           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1566           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1567           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1568           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1569           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr)  ,
1570           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1571   assert( tj->ptr() != TypePtr::TopPTR &&
1572           tj->ptr() != TypePtr::AnyNull &&
1573           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1574 //    assert( tj->ptr() != TypePtr::Constant ||
1575 //            tj->base() == Type::RawPtr ||
1576 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1577 
1578   return tj;
1579 }
1580 
1581 void Compile::AliasType::Init(int i, const TypePtr* at) {
1582   _index = i;
1583   _adr_type = at;
1584   _field = NULL;
1585   _element = NULL;
1586   _is_rewritable = true; // default
1587   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1588   if (atoop != NULL && atoop->is_known_instance()) {
1589     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1590     _general_index = Compile::current()->get_alias_index(gt);
1591   } else {
1592     _general_index = 0;
1593   }
1594 }
1595 
1596 //---------------------------------print_on------------------------------------
1597 #ifndef PRODUCT
1598 void Compile::AliasType::print_on(outputStream* st) {
1599   if (index() < 10)
1600         st->print("@ <%d> ", index());
1601   else  st->print("@ <%d>",  index());
1602   st->print(is_rewritable() ? "   " : " RO");
1603   int offset = adr_type()->offset();
1604   if (offset == Type::OffsetBot)
1605         st->print(" +any");
1606   else  st->print(" +%-3d", offset);
1607   st->print(" in ");
1608   adr_type()->dump_on(st);
1609   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1610   if (field() != NULL && tjp) {
1611     if (tjp->klass()  != field()->holder() ||
1612         tjp->offset() != field()->offset_in_bytes()) {
1613       st->print(" != ");
1614       field()->print();
1615       st->print(" ***");
1616     }
1617   }
1618 }
1619 
1620 void print_alias_types() {
1621   Compile* C = Compile::current();
1622   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1623   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1624     C->alias_type(idx)->print_on(tty);
1625     tty->cr();
1626   }
1627 }
1628 #endif
1629 
1630 
1631 //----------------------------probe_alias_cache--------------------------------
1632 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1633   intptr_t key = (intptr_t) adr_type;
1634   key ^= key >> logAliasCacheSize;
1635   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1636 }
1637 
1638 
1639 //-----------------------------grow_alias_types--------------------------------
1640 void Compile::grow_alias_types() {
1641   const int old_ats  = _max_alias_types; // how many before?
1642   const int new_ats  = old_ats;          // how many more?
1643   const int grow_ats = old_ats+new_ats;  // how many now?
1644   _max_alias_types = grow_ats;
1645   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1646   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1647   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1648   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1649 }
1650 
1651 
1652 //--------------------------------find_alias_type------------------------------
1653 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1654   if (_AliasLevel == 0)
1655     return alias_type(AliasIdxBot);
1656 
1657   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1658   if (ace->_adr_type == adr_type) {
1659     return alias_type(ace->_index);
1660   }
1661 
1662   // Handle special cases.
1663   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1664   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1665 
1666   // Do it the slow way.
1667   const TypePtr* flat = flatten_alias_type(adr_type);
1668 
1669 #ifdef ASSERT
1670   assert(flat == flatten_alias_type(flat), "idempotent");
1671   assert(flat != TypePtr::BOTTOM,     "cannot alias-analyze an untyped ptr");
1672   if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1673     const TypeOopPtr* foop = flat->is_oopptr();
1674     // Scalarizable allocations have exact klass always.
1675     bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1676     const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1677     assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
1678   }
1679   assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
1680 #endif
1681 
1682   int idx = AliasIdxTop;
1683   for (int i = 0; i < num_alias_types(); i++) {
1684     if (alias_type(i)->adr_type() == flat) {
1685       idx = i;
1686       break;
1687     }
1688   }
1689 
1690   if (idx == AliasIdxTop) {
1691     if (no_create)  return NULL;
1692     // Grow the array if necessary.
1693     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1694     // Add a new alias type.
1695     idx = _num_alias_types++;
1696     _alias_types[idx]->Init(idx, flat);
1697     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1698     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1699     if (flat->isa_instptr()) {
1700       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1701           && flat->is_instptr()->klass() == env()->Class_klass())
1702         alias_type(idx)->set_rewritable(false);
1703     }
1704     if (flat->isa_aryptr()) {
1705 #ifdef ASSERT
1706       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1707       // (T_BYTE has the weakest alignment and size restrictions...)
1708       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1709 #endif
1710       if (flat->offset() == TypePtr::OffsetBot) {
1711         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1712       }
1713     }
1714     if (flat->isa_klassptr()) {
1715       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1716         alias_type(idx)->set_rewritable(false);
1717       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1718         alias_type(idx)->set_rewritable(false);
1719       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1720         alias_type(idx)->set_rewritable(false);
1721       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1722         alias_type(idx)->set_rewritable(false);
1723     }
1724     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1725     // but the base pointer type is not distinctive enough to identify
1726     // references into JavaThread.)
1727 
1728     // Check for final fields.
1729     const TypeInstPtr* tinst = flat->isa_instptr();
1730     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1731       ciField* field;
1732       if (tinst->const_oop() != NULL &&
1733           tinst->klass() == ciEnv::current()->Class_klass() &&
1734           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1735         // static field
1736         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1737         field = k->get_field_by_offset(tinst->offset(), true);
1738       } else {
1739         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1740         field = k->get_field_by_offset(tinst->offset(), false);
1741       }
1742       assert(field == NULL ||
1743              original_field == NULL ||
1744              (field->holder() == original_field->holder() &&
1745               field->offset() == original_field->offset() &&
1746               field->is_static() == original_field->is_static()), "wrong field?");
1747       // Set field() and is_rewritable() attributes.
1748       if (field != NULL)  alias_type(idx)->set_field(field);
1749     }
1750   }
1751 
1752   // Fill the cache for next time.
1753   ace->_adr_type = adr_type;
1754   ace->_index    = idx;
1755   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1756 
1757   // Might as well try to fill the cache for the flattened version, too.
1758   AliasCacheEntry* face = probe_alias_cache(flat);
1759   if (face->_adr_type == NULL) {
1760     face->_adr_type = flat;
1761     face->_index    = idx;
1762     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1763   }
1764 
1765   return alias_type(idx);
1766 }
1767 
1768 
1769 Compile::AliasType* Compile::alias_type(ciField* field) {
1770   const TypeOopPtr* t;
1771   if (field->is_static())
1772     t = TypeInstPtr::make(field->holder()->java_mirror());
1773   else
1774     t = TypeOopPtr::make_from_klass_raw(field->holder());
1775   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1776   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1777   return atp;
1778 }
1779 
1780 
1781 //------------------------------have_alias_type--------------------------------
1782 bool Compile::have_alias_type(const TypePtr* adr_type) {
1783   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1784   if (ace->_adr_type == adr_type) {
1785     return true;
1786   }
1787 
1788   // Handle special cases.
1789   if (adr_type == NULL)             return true;
1790   if (adr_type == TypePtr::BOTTOM)  return true;
1791 
1792   return find_alias_type(adr_type, true, NULL) != NULL;
1793 }
1794 
1795 //-----------------------------must_alias--------------------------------------
1796 // True if all values of the given address type are in the given alias category.
1797 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1798   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1799   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1800   if (alias_idx == AliasIdxTop)         return false; // the empty category
1801   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1802 
1803   // the only remaining possible overlap is identity
1804   int adr_idx = get_alias_index(adr_type);
1805   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1806   assert(adr_idx == alias_idx ||
1807          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1808           && adr_type                       != TypeOopPtr::BOTTOM),
1809          "should not be testing for overlap with an unsafe pointer");
1810   return adr_idx == alias_idx;
1811 }
1812 
1813 //------------------------------can_alias--------------------------------------
1814 // True if any values of the given address type are in the given alias category.
1815 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1816   if (alias_idx == AliasIdxTop)         return false; // the empty category
1817   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1818   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1819   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1820 
1821   // the only remaining possible overlap is identity
1822   int adr_idx = get_alias_index(adr_type);
1823   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1824   return adr_idx == alias_idx;
1825 }
1826 
1827 
1828 
1829 //---------------------------pop_warm_call-------------------------------------
1830 WarmCallInfo* Compile::pop_warm_call() {
1831   WarmCallInfo* wci = _warm_calls;
1832   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1833   return wci;
1834 }
1835 
1836 //----------------------------Inline_Warm--------------------------------------
1837 int Compile::Inline_Warm() {
1838   // If there is room, try to inline some more warm call sites.
1839   // %%% Do a graph index compaction pass when we think we're out of space?
1840   if (!InlineWarmCalls)  return 0;
1841 
1842   int calls_made_hot = 0;
1843   int room_to_grow   = NodeCountInliningCutoff - unique();
1844   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1845   int amount_grown   = 0;
1846   WarmCallInfo* call;
1847   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1848     int est_size = (int)call->size();
1849     if (est_size > (room_to_grow - amount_grown)) {
1850       // This one won't fit anyway.  Get rid of it.
1851       call->make_cold();
1852       continue;
1853     }
1854     call->make_hot();
1855     calls_made_hot++;
1856     amount_grown   += est_size;
1857     amount_to_grow -= est_size;
1858   }
1859 
1860   if (calls_made_hot > 0)  set_major_progress();
1861   return calls_made_hot;
1862 }
1863 
1864 
1865 //----------------------------Finish_Warm--------------------------------------
1866 void Compile::Finish_Warm() {
1867   if (!InlineWarmCalls)  return;
1868   if (failing())  return;
1869   if (warm_calls() == NULL)  return;
1870 
1871   // Clean up loose ends, if we are out of space for inlining.
1872   WarmCallInfo* call;
1873   while ((call = pop_warm_call()) != NULL) {
1874     call->make_cold();
1875   }
1876 }
1877 
1878 //---------------------cleanup_loop_predicates-----------------------
1879 // Remove the opaque nodes that protect the predicates so that all unused
1880 // checks and uncommon_traps will be eliminated from the ideal graph
1881 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1882   if (predicate_count()==0) return;
1883   for (int i = predicate_count(); i > 0; i--) {
1884     Node * n = predicate_opaque1_node(i-1);
1885     assert(n->Opcode() == Op_Opaque1, "must be");
1886     igvn.replace_node(n, n->in(1));
1887   }
1888   assert(predicate_count()==0, "should be clean!");
1889 }
1890 
1891 void Compile::add_range_check_cast(Node* n) {
1892   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1893   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1894   _range_check_casts->append(n);
1895 }
1896 
1897 // Remove all range check dependent CastIINodes.
1898 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1899   for (int i = range_check_cast_count(); i > 0; i--) {
1900     Node* cast = range_check_cast_node(i-1);
1901     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1902     igvn.replace_node(cast, cast->in(1));
1903   }
1904   assert(range_check_cast_count() == 0, "should be empty");
1905 }
1906 
1907 // StringOpts and late inlining of string methods
1908 void Compile::inline_string_calls(bool parse_time) {
1909   {
1910     // remove useless nodes to make the usage analysis simpler
1911     ResourceMark rm;
1912     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1913   }
1914 
1915   {
1916     ResourceMark rm;
1917     print_method(PHASE_BEFORE_STRINGOPTS, 3);
1918     PhaseStringOpts pso(initial_gvn(), for_igvn());
1919     print_method(PHASE_AFTER_STRINGOPTS, 3);
1920   }
1921 
1922   // now inline anything that we skipped the first time around
1923   if (!parse_time) {
1924     _late_inlines_pos = _late_inlines.length();
1925   }
1926 
1927   while (_string_late_inlines.length() > 0) {
1928     CallGenerator* cg = _string_late_inlines.pop();
1929     cg->do_late_inline();
1930     if (failing())  return;
1931   }
1932   _string_late_inlines.trunc_to(0);
1933 }
1934 
1935 // Late inlining of boxing methods
1936 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1937   if (_boxing_late_inlines.length() > 0) {
1938     assert(has_boxed_value(), "inconsistent");
1939 
1940     PhaseGVN* gvn = initial_gvn();
1941     set_inlining_incrementally(true);
1942 
1943     assert( igvn._worklist.size() == 0, "should be done with igvn" );
1944     for_igvn()->clear();
1945     gvn->replace_with(&igvn);
1946 
1947     _late_inlines_pos = _late_inlines.length();
1948 
1949     while (_boxing_late_inlines.length() > 0) {
1950       CallGenerator* cg = _boxing_late_inlines.pop();
1951       cg->do_late_inline();
1952       if (failing())  return;
1953     }
1954     _boxing_late_inlines.trunc_to(0);
1955 
1956     {
1957       ResourceMark rm;
1958       PhaseRemoveUseless pru(gvn, for_igvn());
1959     }
1960 
1961     igvn = PhaseIterGVN(gvn);
1962     igvn.optimize();
1963 
1964     set_inlining_progress(false);
1965     set_inlining_incrementally(false);
1966   }
1967 }
1968 
1969 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
1970   assert(IncrementalInline, "incremental inlining should be on");
1971   PhaseGVN* gvn = initial_gvn();
1972 
1973   set_inlining_progress(false);
1974   for_igvn()->clear();
1975   gvn->replace_with(&igvn);
1976 
1977   int i = 0;
1978 
1979   for (; i <_late_inlines.length() && !inlining_progress(); i++) {
1980     CallGenerator* cg = _late_inlines.at(i);
1981     _late_inlines_pos = i+1;
1982     cg->do_late_inline();
1983     if (failing())  return;
1984   }
1985   int j = 0;
1986   for (; i < _late_inlines.length(); i++, j++) {
1987     _late_inlines.at_put(j, _late_inlines.at(i));
1988   }
1989   _late_inlines.trunc_to(j);
1990 
1991   {
1992     ResourceMark rm;
1993     PhaseRemoveUseless pru(gvn, for_igvn());
1994   }
1995 
1996   igvn = PhaseIterGVN(gvn);
1997 }
1998 
1999 // Perform incremental inlining until bound on number of live nodes is reached
2000 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2001   PhaseGVN* gvn = initial_gvn();
2002 
2003   set_inlining_incrementally(true);
2004   set_inlining_progress(true);
2005   uint low_live_nodes = 0;
2006 
2007   while(inlining_progress() && _late_inlines.length() > 0) {
2008 
2009     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2010       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2011         // PhaseIdealLoop is expensive so we only try it once we are
2012         // out of live nodes and we only try it again if the previous
2013         // helped got the number of nodes down significantly
2014         PhaseIdealLoop ideal_loop( igvn, false, true );
2015         if (failing())  return;
2016         low_live_nodes = live_nodes();
2017         _major_progress = true;
2018       }
2019 
2020       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2021         break;
2022       }
2023     }
2024 
2025     inline_incrementally_one(igvn);
2026 
2027     if (failing())  return;
2028 
2029     igvn.optimize();
2030 
2031     if (failing())  return;
2032   }
2033 
2034   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2035 
2036   if (_string_late_inlines.length() > 0) {
2037     assert(has_stringbuilder(), "inconsistent");
2038     for_igvn()->clear();
2039     initial_gvn()->replace_with(&igvn);
2040 
2041     inline_string_calls(false);
2042 
2043     if (failing())  return;
2044 
2045     {
2046       ResourceMark rm;
2047       PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2048     }
2049 
2050     igvn = PhaseIterGVN(gvn);
2051 
2052     igvn.optimize();
2053   }
2054 
2055   set_inlining_incrementally(false);
2056 }
2057 
2058 
2059 //------------------------------Optimize---------------------------------------
2060 // Given a graph, optimize it.
2061 void Compile::Optimize() {
2062   TracePhase t1("optimizer", &_t_optimizer, true);
2063 
2064 #ifndef PRODUCT
2065   if (env()->break_at_compile()) {
2066     BREAKPOINT;
2067   }
2068 
2069 #endif
2070 
2071   ResourceMark rm;
2072   int          loop_opts_cnt;
2073 
2074   NOT_PRODUCT( verify_graph_edges(); )
2075 
2076   print_method(PHASE_AFTER_PARSING);
2077 
2078  {
2079   // Iterative Global Value Numbering, including ideal transforms
2080   // Initialize IterGVN with types and values from parse-time GVN
2081   PhaseIterGVN igvn(initial_gvn());
2082   {
2083     NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2084     igvn.optimize();
2085   }
2086 
2087   print_method(PHASE_ITER_GVN1, 2);
2088 
2089   if (failing())  return;
2090 
2091   {
2092     NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2093     inline_incrementally(igvn);
2094   }
2095 
2096   print_method(PHASE_INCREMENTAL_INLINE, 2);
2097 
2098   if (failing())  return;
2099 
2100   if (eliminate_boxing()) {
2101     NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2102     // Inline valueOf() methods now.
2103     inline_boxing_calls(igvn);
2104 
2105     if (AlwaysIncrementalInline) {
2106       inline_incrementally(igvn);
2107     }
2108 
2109     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2110 
2111     if (failing())  return;
2112   }
2113 
2114   // Remove the speculative part of types and clean up the graph from
2115   // the extra CastPP nodes whose only purpose is to carry them. Do
2116   // that early so that optimizations are not disrupted by the extra
2117   // CastPP nodes.
2118   remove_speculative_types(igvn);
2119 
2120   // No more new expensive nodes will be added to the list from here
2121   // so keep only the actual candidates for optimizations.
2122   cleanup_expensive_nodes(igvn);
2123 
2124   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2125     NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);)
2126     initial_gvn()->replace_with(&igvn);
2127     for_igvn()->clear();
2128     Unique_Node_List new_worklist(C->comp_arena());
2129     {
2130       ResourceMark rm;
2131       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2132     }
2133     set_for_igvn(&new_worklist);
2134     igvn = PhaseIterGVN(initial_gvn());
2135     igvn.optimize();
2136   }
2137 
2138   // Perform escape analysis
2139   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2140     if (has_loops()) {
2141       // Cleanup graph (remove dead nodes).
2142       TracePhase t2("idealLoop", &_t_idealLoop, true);
2143       PhaseIdealLoop ideal_loop( igvn, false, true );
2144       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2145       if (failing())  return;
2146     }
2147     ConnectionGraph::do_analysis(this, &igvn);
2148 
2149     if (failing())  return;
2150 
2151     // Optimize out fields loads from scalar replaceable allocations.
2152     igvn.optimize();
2153     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2154 
2155     if (failing())  return;
2156 
2157     if (congraph() != NULL && macro_count() > 0) {
2158       NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2159       PhaseMacroExpand mexp(igvn);
2160       mexp.eliminate_macro_nodes();
2161       igvn.set_delay_transform(false);
2162 
2163       igvn.optimize();
2164       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2165 
2166       if (failing())  return;
2167     }
2168   }
2169 
2170   // Loop transforms on the ideal graph.  Range Check Elimination,
2171   // peeling, unrolling, etc.
2172 
2173   // Set loop opts counter
2174   loop_opts_cnt = num_loop_opts();
2175   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2176     {
2177       TracePhase t2("idealLoop", &_t_idealLoop, true);
2178       PhaseIdealLoop ideal_loop( igvn, true );
2179       loop_opts_cnt--;
2180       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2181       if (failing())  return;
2182     }
2183     // Loop opts pass if partial peeling occurred in previous pass
2184     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2185       TracePhase t3("idealLoop", &_t_idealLoop, true);
2186       PhaseIdealLoop ideal_loop( igvn, false );
2187       loop_opts_cnt--;
2188       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2189       if (failing())  return;
2190     }
2191     // Loop opts pass for loop-unrolling before CCP
2192     if(major_progress() && (loop_opts_cnt > 0)) {
2193       TracePhase t4("idealLoop", &_t_idealLoop, true);
2194       PhaseIdealLoop ideal_loop( igvn, false );
2195       loop_opts_cnt--;
2196       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2197     }
2198     if (!failing()) {
2199       // Verify that last round of loop opts produced a valid graph
2200       NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2201       PhaseIdealLoop::verify(igvn);
2202     }
2203   }
2204   if (failing())  return;
2205 
2206   // Conditional Constant Propagation;
2207   PhaseCCP ccp( &igvn );
2208   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2209   {
2210     TracePhase t2("ccp", &_t_ccp, true);
2211     ccp.do_transform();
2212   }
2213   print_method(PHASE_CPP1, 2);
2214 
2215   assert( true, "Break here to ccp.dump_old2new_map()");
2216 
2217   // Iterative Global Value Numbering, including ideal transforms
2218   {
2219     NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2220     igvn = ccp;
2221     igvn.optimize();
2222   }
2223 
2224   print_method(PHASE_ITER_GVN2, 2);
2225 
2226   if (failing())  return;
2227 
2228   // Loop transforms on the ideal graph.  Range Check Elimination,
2229   // peeling, unrolling, etc.
2230   if(loop_opts_cnt > 0) {
2231     debug_only( int cnt = 0; );
2232     while(major_progress() && (loop_opts_cnt > 0)) {
2233       TracePhase t2("idealLoop", &_t_idealLoop, true);
2234       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2235       PhaseIdealLoop ideal_loop( igvn, true);
2236       loop_opts_cnt--;
2237       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2238       if (failing())  return;
2239     }
2240   }
2241 
2242   {
2243     // Verify that all previous optimizations produced a valid graph
2244     // at least to this point, even if no loop optimizations were done.
2245     NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2246     PhaseIdealLoop::verify(igvn);
2247   }
2248 
2249   if (range_check_cast_count() > 0) {
2250     // No more loop optimizations. Remove all range check dependent CastIINodes.
2251     C->remove_range_check_casts(igvn);
2252     igvn.optimize();
2253   }
2254 
2255   {
2256     NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2257     PhaseMacroExpand  mex(igvn);
2258     if (mex.expand_macro_nodes()) {
2259       assert(failing(), "must bail out w/ explicit message");
2260       return;
2261     }
2262   }
2263 
2264  } // (End scope of igvn; run destructor if necessary for asserts.)
2265 
2266   dump_inlining();
2267   // A method with only infinite loops has no edges entering loops from root
2268   {
2269     NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2270     if (final_graph_reshaping()) {
2271       assert(failing(), "must bail out w/ explicit message");
2272       return;
2273     }
2274   }
2275 
2276   print_method(PHASE_OPTIMIZE_FINISHED, 2);
2277 }
2278 
2279 
2280 //------------------------------Code_Gen---------------------------------------
2281 // Given a graph, generate code for it
2282 void Compile::Code_Gen() {
2283   if (failing()) {
2284     return;
2285   }
2286 
2287   // Perform instruction selection.  You might think we could reclaim Matcher
2288   // memory PDQ, but actually the Matcher is used in generating spill code.
2289   // Internals of the Matcher (including some VectorSets) must remain live
2290   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2291   // set a bit in reclaimed memory.
2292 
2293   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2294   // nodes.  Mapping is only valid at the root of each matched subtree.
2295   NOT_PRODUCT( verify_graph_edges(); )
2296 
2297   Matcher matcher;
2298   _matcher = &matcher;
2299   {
2300     TracePhase t2("matcher", &_t_matcher, true);
2301     matcher.match();
2302   }
2303   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2304   // nodes.  Mapping is only valid at the root of each matched subtree.
2305   NOT_PRODUCT( verify_graph_edges(); )
2306 
2307   // If you have too many nodes, or if matching has failed, bail out
2308   check_node_count(0, "out of nodes matching instructions");
2309   if (failing()) {
2310     return;
2311   }
2312 
2313   // Build a proper-looking CFG
2314   PhaseCFG cfg(node_arena(), root(), matcher);
2315   _cfg = &cfg;
2316   {
2317     NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2318     bool success = cfg.do_global_code_motion();
2319     if (!success) {
2320       return;
2321     }
2322 
2323     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2324     NOT_PRODUCT( verify_graph_edges(); )
2325     debug_only( cfg.verify(); )
2326   }
2327 
2328   PhaseChaitin regalloc(unique(), cfg, matcher);
2329   _regalloc = &regalloc;
2330   {
2331     TracePhase t2("regalloc", &_t_registerAllocation, true);
2332     // Perform register allocation.  After Chaitin, use-def chains are
2333     // no longer accurate (at spill code) and so must be ignored.
2334     // Node->LRG->reg mappings are still accurate.
2335     _regalloc->Register_Allocate();
2336 
2337     // Bail out if the allocator builds too many nodes
2338     if (failing()) {
2339       return;
2340     }
2341   }
2342 
2343   // Prior to register allocation we kept empty basic blocks in case the
2344   // the allocator needed a place to spill.  After register allocation we
2345   // are not adding any new instructions.  If any basic block is empty, we
2346   // can now safely remove it.
2347   {
2348     NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2349     cfg.remove_empty_blocks();
2350     if (do_freq_based_layout()) {
2351       PhaseBlockLayout layout(cfg);
2352     } else {
2353       cfg.set_loop_alignment();
2354     }
2355     cfg.fixup_flow();
2356   }
2357 
2358   // Apply peephole optimizations
2359   if( OptoPeephole ) {
2360     NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2361     PhasePeephole peep( _regalloc, cfg);
2362     peep.do_transform();
2363   }
2364 
2365   // Do late expand if CPU requires this.
2366   if (Matcher::require_postalloc_expand) {
2367     NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2368     cfg.postalloc_expand(_regalloc);
2369   }
2370 
2371   // Convert Nodes to instruction bits in a buffer
2372   {
2373     // %%%% workspace merge brought two timers together for one job
2374     TracePhase t2a("output", &_t_output, true);
2375     NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2376     Output();
2377   }
2378 
2379   print_method(PHASE_FINAL_CODE);
2380 
2381   // He's dead, Jim.
2382   _cfg     = (PhaseCFG*)0xdeadbeef;
2383   _regalloc = (PhaseChaitin*)0xdeadbeef;
2384 }
2385 
2386 
2387 //------------------------------dump_asm---------------------------------------
2388 // Dump formatted assembly
2389 #ifndef PRODUCT
2390 void Compile::dump_asm(int *pcs, uint pc_limit) {
2391   bool cut_short = false;
2392   tty->print_cr("#");
2393   tty->print("#  ");  _tf->dump();  tty->cr();
2394   tty->print_cr("#");
2395 
2396   // For all blocks
2397   int pc = 0x0;                 // Program counter
2398   char starts_bundle = ' ';
2399   _regalloc->dump_frame();
2400 
2401   Node *n = NULL;
2402   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2403     if (VMThread::should_terminate()) {
2404       cut_short = true;
2405       break;
2406     }
2407     Block* block = _cfg->get_block(i);
2408     if (block->is_connector() && !Verbose) {
2409       continue;
2410     }
2411     n = block->head();
2412     if (pcs && n->_idx < pc_limit) {
2413       tty->print("%3.3x   ", pcs[n->_idx]);
2414     } else {
2415       tty->print("      ");
2416     }
2417     block->dump_head(_cfg);
2418     if (block->is_connector()) {
2419       tty->print_cr("        # Empty connector block");
2420     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2421       tty->print_cr("        # Block is sole successor of call");
2422     }
2423 
2424     // For all instructions
2425     Node *delay = NULL;
2426     for (uint j = 0; j < block->number_of_nodes(); j++) {
2427       if (VMThread::should_terminate()) {
2428         cut_short = true;
2429         break;
2430       }
2431       n = block->get_node(j);
2432       if (valid_bundle_info(n)) {
2433         Bundle* bundle = node_bundling(n);
2434         if (bundle->used_in_unconditional_delay()) {
2435           delay = n;
2436           continue;
2437         }
2438         if (bundle->starts_bundle()) {
2439           starts_bundle = '+';
2440         }
2441       }
2442 
2443       if (WizardMode) {
2444         n->dump();
2445       }
2446 
2447       if( !n->is_Region() &&    // Dont print in the Assembly
2448           !n->is_Phi() &&       // a few noisely useless nodes
2449           !n->is_Proj() &&
2450           !n->is_MachTemp() &&
2451           !n->is_SafePointScalarObject() &&
2452           !n->is_Catch() &&     // Would be nice to print exception table targets
2453           !n->is_MergeMem() &&  // Not very interesting
2454           !n->is_top() &&       // Debug info table constants
2455           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2456           ) {
2457         if (pcs && n->_idx < pc_limit)
2458           tty->print("%3.3x", pcs[n->_idx]);
2459         else
2460           tty->print("   ");
2461         tty->print(" %c ", starts_bundle);
2462         starts_bundle = ' ';
2463         tty->print("\t");
2464         n->format(_regalloc, tty);
2465         tty->cr();
2466       }
2467 
2468       // If we have an instruction with a delay slot, and have seen a delay,
2469       // then back up and print it
2470       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2471         assert(delay != NULL, "no unconditional delay instruction");
2472         if (WizardMode) delay->dump();
2473 
2474         if (node_bundling(delay)->starts_bundle())
2475           starts_bundle = '+';
2476         if (pcs && n->_idx < pc_limit)
2477           tty->print("%3.3x", pcs[n->_idx]);
2478         else
2479           tty->print("   ");
2480         tty->print(" %c ", starts_bundle);
2481         starts_bundle = ' ';
2482         tty->print("\t");
2483         delay->format(_regalloc, tty);
2484         tty->cr();
2485         delay = NULL;
2486       }
2487 
2488       // Dump the exception table as well
2489       if( n->is_Catch() && (Verbose || WizardMode) ) {
2490         // Print the exception table for this offset
2491         _handler_table.print_subtable_for(pc);
2492       }
2493     }
2494 
2495     if (pcs && n->_idx < pc_limit)
2496       tty->print_cr("%3.3x", pcs[n->_idx]);
2497     else
2498       tty->cr();
2499 
2500     assert(cut_short || delay == NULL, "no unconditional delay branch");
2501 
2502   } // End of per-block dump
2503   tty->cr();
2504 
2505   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2506 }
2507 #endif
2508 
2509 //------------------------------Final_Reshape_Counts---------------------------
2510 // This class defines counters to help identify when a method
2511 // may/must be executed using hardware with only 24-bit precision.
2512 struct Final_Reshape_Counts : public StackObj {
2513   int  _call_count;             // count non-inlined 'common' calls
2514   int  _float_count;            // count float ops requiring 24-bit precision
2515   int  _double_count;           // count double ops requiring more precision
2516   int  _java_call_count;        // count non-inlined 'java' calls
2517   int  _inner_loop_count;       // count loops which need alignment
2518   VectorSet _visited;           // Visitation flags
2519   Node_List _tests;             // Set of IfNodes & PCTableNodes
2520 
2521   Final_Reshape_Counts() :
2522     _call_count(0), _float_count(0), _double_count(0),
2523     _java_call_count(0), _inner_loop_count(0),
2524     _visited( Thread::current()->resource_area() ) { }
2525 
2526   void inc_call_count  () { _call_count  ++; }
2527   void inc_float_count () { _float_count ++; }
2528   void inc_double_count() { _double_count++; }
2529   void inc_java_call_count() { _java_call_count++; }
2530   void inc_inner_loop_count() { _inner_loop_count++; }
2531 
2532   int  get_call_count  () const { return _call_count  ; }
2533   int  get_float_count () const { return _float_count ; }
2534   int  get_double_count() const { return _double_count; }
2535   int  get_java_call_count() const { return _java_call_count; }
2536   int  get_inner_loop_count() const { return _inner_loop_count; }
2537 };
2538 
2539 #ifdef ASSERT
2540 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2541   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2542   // Make sure the offset goes inside the instance layout.
2543   return k->contains_field_offset(tp->offset());
2544   // Note that OffsetBot and OffsetTop are very negative.
2545 }
2546 #endif
2547 
2548 // Eliminate trivially redundant StoreCMs and accumulate their
2549 // precedence edges.
2550 void Compile::eliminate_redundant_card_marks(Node* n) {
2551   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2552   if (n->in(MemNode::Address)->outcnt() > 1) {
2553     // There are multiple users of the same address so it might be
2554     // possible to eliminate some of the StoreCMs
2555     Node* mem = n->in(MemNode::Memory);
2556     Node* adr = n->in(MemNode::Address);
2557     Node* val = n->in(MemNode::ValueIn);
2558     Node* prev = n;
2559     bool done = false;
2560     // Walk the chain of StoreCMs eliminating ones that match.  As
2561     // long as it's a chain of single users then the optimization is
2562     // safe.  Eliminating partially redundant StoreCMs would require
2563     // cloning copies down the other paths.
2564     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2565       if (adr == mem->in(MemNode::Address) &&
2566           val == mem->in(MemNode::ValueIn)) {
2567         // redundant StoreCM
2568         if (mem->req() > MemNode::OopStore) {
2569           // Hasn't been processed by this code yet.
2570           n->add_prec(mem->in(MemNode::OopStore));
2571         } else {
2572           // Already converted to precedence edge
2573           for (uint i = mem->req(); i < mem->len(); i++) {
2574             // Accumulate any precedence edges
2575             if (mem->in(i) != NULL) {
2576               n->add_prec(mem->in(i));
2577             }
2578           }
2579           // Everything above this point has been processed.
2580           done = true;
2581         }
2582         // Eliminate the previous StoreCM
2583         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2584         assert(mem->outcnt() == 0, "should be dead");
2585         mem->disconnect_inputs(NULL, this);
2586       } else {
2587         prev = mem;
2588       }
2589       mem = prev->in(MemNode::Memory);
2590     }
2591   }
2592 }
2593 
2594 //------------------------------final_graph_reshaping_impl----------------------
2595 // Implement items 1-5 from final_graph_reshaping below.
2596 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2597 
2598   if ( n->outcnt() == 0 ) return; // dead node
2599   uint nop = n->Opcode();
2600 
2601   // Check for 2-input instruction with "last use" on right input.
2602   // Swap to left input.  Implements item (2).
2603   if( n->req() == 3 &&          // two-input instruction
2604       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2605       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2606       n->in(2)->outcnt() == 1 &&// right use IS a last use
2607       !n->in(2)->is_Con() ) {   // right use is not a constant
2608     // Check for commutative opcode
2609     switch( nop ) {
2610     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2611     case Op_MaxI:  case Op_MinI:
2612     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2613     case Op_AndL:  case Op_XorL:  case Op_OrL:
2614     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2615       // Move "last use" input to left by swapping inputs
2616       n->swap_edges(1, 2);
2617       break;
2618     }
2619     default:
2620       break;
2621     }
2622   }
2623 
2624 #ifdef ASSERT
2625   if( n->is_Mem() ) {
2626     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2627     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2628             // oop will be recorded in oop map if load crosses safepoint
2629             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2630                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2631             "raw memory operations should have control edge");
2632   }
2633 #endif
2634   // Count FPU ops and common calls, implements item (3)
2635   switch( nop ) {
2636   // Count all float operations that may use FPU
2637   case Op_AddF:
2638   case Op_SubF:
2639   case Op_MulF:
2640   case Op_DivF:
2641   case Op_NegF:
2642   case Op_ModF:
2643   case Op_ConvI2F:
2644   case Op_ConF:
2645   case Op_CmpF:
2646   case Op_CmpF3:
2647   // case Op_ConvL2F: // longs are split into 32-bit halves
2648     frc.inc_float_count();
2649     break;
2650 
2651   case Op_ConvF2D:
2652   case Op_ConvD2F:
2653     frc.inc_float_count();
2654     frc.inc_double_count();
2655     break;
2656 
2657   // Count all double operations that may use FPU
2658   case Op_AddD:
2659   case Op_SubD:
2660   case Op_MulD:
2661   case Op_DivD:
2662   case Op_NegD:
2663   case Op_ModD:
2664   case Op_ConvI2D:
2665   case Op_ConvD2I:
2666   // case Op_ConvL2D: // handled by leaf call
2667   // case Op_ConvD2L: // handled by leaf call
2668   case Op_ConD:
2669   case Op_CmpD:
2670   case Op_CmpD3:
2671     frc.inc_double_count();
2672     break;
2673   case Op_Opaque1:              // Remove Opaque Nodes before matching
2674   case Op_Opaque2:              // Remove Opaque Nodes before matching
2675   case Op_Opaque3:
2676     n->subsume_by(n->in(1), this);
2677     break;
2678   case Op_CallStaticJava:
2679   case Op_CallJava:
2680   case Op_CallDynamicJava:
2681     frc.inc_java_call_count(); // Count java call site;
2682   case Op_CallRuntime:
2683   case Op_CallLeaf:
2684   case Op_CallLeafNoFP: {
2685     assert( n->is_Call(), "" );
2686     CallNode *call = n->as_Call();
2687     // Count call sites where the FP mode bit would have to be flipped.
2688     // Do not count uncommon runtime calls:
2689     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2690     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2691     if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2692       frc.inc_call_count();   // Count the call site
2693     } else {                  // See if uncommon argument is shared
2694       Node *n = call->in(TypeFunc::Parms);
2695       int nop = n->Opcode();
2696       // Clone shared simple arguments to uncommon calls, item (1).
2697       if( n->outcnt() > 1 &&
2698           !n->is_Proj() &&
2699           nop != Op_CreateEx &&
2700           nop != Op_CheckCastPP &&
2701           nop != Op_DecodeN &&
2702           nop != Op_DecodeNKlass &&
2703           !n->is_Mem() ) {
2704         Node *x = n->clone();
2705         call->set_req( TypeFunc::Parms, x );
2706       }
2707     }
2708     break;
2709   }
2710 
2711   case Op_StoreD:
2712   case Op_LoadD:
2713   case Op_LoadD_unaligned:
2714     frc.inc_double_count();
2715     goto handle_mem;
2716   case Op_StoreF:
2717   case Op_LoadF:
2718     frc.inc_float_count();
2719     goto handle_mem;
2720 
2721   case Op_StoreCM:
2722     {
2723       // Convert OopStore dependence into precedence edge
2724       Node* prec = n->in(MemNode::OopStore);
2725       n->del_req(MemNode::OopStore);
2726       n->add_prec(prec);
2727       eliminate_redundant_card_marks(n);
2728     }
2729 
2730     // fall through
2731 
2732   case Op_StoreB:
2733   case Op_StoreC:
2734   case Op_StorePConditional:
2735   case Op_StoreI:
2736   case Op_StoreL:
2737   case Op_StoreIConditional:
2738   case Op_StoreLConditional:
2739   case Op_CompareAndSwapI:
2740   case Op_CompareAndSwapL:
2741   case Op_CompareAndSwapP:
2742   case Op_CompareAndSwapN:
2743   case Op_GetAndAddI:
2744   case Op_GetAndAddL:
2745   case Op_GetAndSetI:
2746   case Op_GetAndSetL:
2747   case Op_GetAndSetP:
2748   case Op_GetAndSetN:
2749   case Op_StoreP:
2750   case Op_StoreN:
2751   case Op_StoreNKlass:
2752   case Op_LoadB:
2753   case Op_LoadUB:
2754   case Op_LoadUS:
2755   case Op_LoadI:
2756   case Op_LoadKlass:
2757   case Op_LoadNKlass:
2758   case Op_LoadL:
2759   case Op_LoadL_unaligned:
2760   case Op_LoadPLocked:
2761   case Op_LoadP:
2762   case Op_LoadN:
2763   case Op_LoadRange:
2764   case Op_LoadS: {
2765   handle_mem:
2766 #ifdef ASSERT
2767     if( VerifyOptoOopOffsets ) {
2768       assert( n->is_Mem(), "" );
2769       MemNode *mem  = (MemNode*)n;
2770       // Check to see if address types have grounded out somehow.
2771       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2772       assert( !tp || oop_offset_is_sane(tp), "" );
2773     }
2774 #endif
2775     break;
2776   }
2777 
2778   case Op_AddP: {               // Assert sane base pointers
2779     Node *addp = n->in(AddPNode::Address);
2780     assert( !addp->is_AddP() ||
2781             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2782             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2783             "Base pointers must match" );
2784 #ifdef _LP64
2785     if ((UseCompressedOops || UseCompressedClassPointers) &&
2786         addp->Opcode() == Op_ConP &&
2787         addp == n->in(AddPNode::Base) &&
2788         n->in(AddPNode::Offset)->is_Con()) {
2789       // Use addressing with narrow klass to load with offset on x86.
2790       // On sparc loading 32-bits constant and decoding it have less
2791       // instructions (4) then load 64-bits constant (7).
2792       // Do this transformation here since IGVN will convert ConN back to ConP.
2793       const Type* t = addp->bottom_type();
2794       if (t->isa_oopptr() || t->isa_klassptr()) {
2795         Node* nn = NULL;
2796 
2797         int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2798 
2799         // Look for existing ConN node of the same exact type.
2800         Node* r  = root();
2801         uint cnt = r->outcnt();
2802         for (uint i = 0; i < cnt; i++) {
2803           Node* m = r->raw_out(i);
2804           if (m!= NULL && m->Opcode() == op &&
2805               m->bottom_type()->make_ptr() == t) {
2806             nn = m;
2807             break;
2808           }
2809         }
2810         if (nn != NULL) {
2811           // Decode a narrow oop to match address
2812           // [R12 + narrow_oop_reg<<3 + offset]
2813           if (t->isa_oopptr()) {
2814             nn = new (this) DecodeNNode(nn, t);
2815           } else {
2816             nn = new (this) DecodeNKlassNode(nn, t);
2817           }
2818           n->set_req(AddPNode::Base, nn);
2819           n->set_req(AddPNode::Address, nn);
2820           if (addp->outcnt() == 0) {
2821             addp->disconnect_inputs(NULL, this);
2822           }
2823         }
2824       }
2825     }
2826 #endif
2827     break;
2828   }
2829 
2830 #ifdef _LP64
2831   case Op_CastPP:
2832     if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2833       Node* in1 = n->in(1);
2834       const Type* t = n->bottom_type();
2835       Node* new_in1 = in1->clone();
2836       new_in1->as_DecodeN()->set_type(t);
2837 
2838       if (!Matcher::narrow_oop_use_complex_address()) {
2839         //
2840         // x86, ARM and friends can handle 2 adds in addressing mode
2841         // and Matcher can fold a DecodeN node into address by using
2842         // a narrow oop directly and do implicit NULL check in address:
2843         //
2844         // [R12 + narrow_oop_reg<<3 + offset]
2845         // NullCheck narrow_oop_reg
2846         //
2847         // On other platforms (Sparc) we have to keep new DecodeN node and
2848         // use it to do implicit NULL check in address:
2849         //
2850         // decode_not_null narrow_oop_reg, base_reg
2851         // [base_reg + offset]
2852         // NullCheck base_reg
2853         //
2854         // Pin the new DecodeN node to non-null path on these platform (Sparc)
2855         // to keep the information to which NULL check the new DecodeN node
2856         // corresponds to use it as value in implicit_null_check().
2857         //
2858         new_in1->set_req(0, n->in(0));
2859       }
2860 
2861       n->subsume_by(new_in1, this);
2862       if (in1->outcnt() == 0) {
2863         in1->disconnect_inputs(NULL, this);
2864       }
2865     }
2866     break;
2867 
2868   case Op_CmpP:
2869     // Do this transformation here to preserve CmpPNode::sub() and
2870     // other TypePtr related Ideal optimizations (for example, ptr nullness).
2871     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2872       Node* in1 = n->in(1);
2873       Node* in2 = n->in(2);
2874       if (!in1->is_DecodeNarrowPtr()) {
2875         in2 = in1;
2876         in1 = n->in(2);
2877       }
2878       assert(in1->is_DecodeNarrowPtr(), "sanity");
2879 
2880       Node* new_in2 = NULL;
2881       if (in2->is_DecodeNarrowPtr()) {
2882         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2883         new_in2 = in2->in(1);
2884       } else if (in2->Opcode() == Op_ConP) {
2885         const Type* t = in2->bottom_type();
2886         if (t == TypePtr::NULL_PTR) {
2887           assert(in1->is_DecodeN(), "compare klass to null?");
2888           // Don't convert CmpP null check into CmpN if compressed
2889           // oops implicit null check is not generated.
2890           // This will allow to generate normal oop implicit null check.
2891           if (Matcher::gen_narrow_oop_implicit_null_checks())
2892             new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2893           //
2894           // This transformation together with CastPP transformation above
2895           // will generated code for implicit NULL checks for compressed oops.
2896           //
2897           // The original code after Optimize()
2898           //
2899           //    LoadN memory, narrow_oop_reg
2900           //    decode narrow_oop_reg, base_reg
2901           //    CmpP base_reg, NULL
2902           //    CastPP base_reg // NotNull
2903           //    Load [base_reg + offset], val_reg
2904           //
2905           // after these transformations will be
2906           //
2907           //    LoadN memory, narrow_oop_reg
2908           //    CmpN narrow_oop_reg, NULL
2909           //    decode_not_null narrow_oop_reg, base_reg
2910           //    Load [base_reg + offset], val_reg
2911           //
2912           // and the uncommon path (== NULL) will use narrow_oop_reg directly
2913           // since narrow oops can be used in debug info now (see the code in
2914           // final_graph_reshaping_walk()).
2915           //
2916           // At the end the code will be matched to
2917           // on x86:
2918           //
2919           //    Load_narrow_oop memory, narrow_oop_reg
2920           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
2921           //    NullCheck narrow_oop_reg
2922           //
2923           // and on sparc:
2924           //
2925           //    Load_narrow_oop memory, narrow_oop_reg
2926           //    decode_not_null narrow_oop_reg, base_reg
2927           //    Load [base_reg + offset], val_reg
2928           //    NullCheck base_reg
2929           //
2930         } else if (t->isa_oopptr()) {
2931           new_in2 = ConNode::make(this, t->make_narrowoop());
2932         } else if (t->isa_klassptr()) {
2933           new_in2 = ConNode::make(this, t->make_narrowklass());
2934         }
2935       }
2936       if (new_in2 != NULL) {
2937         Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
2938         n->subsume_by(cmpN, this);
2939         if (in1->outcnt() == 0) {
2940           in1->disconnect_inputs(NULL, this);
2941         }
2942         if (in2->outcnt() == 0) {
2943           in2->disconnect_inputs(NULL, this);
2944         }
2945       }
2946     }
2947     break;
2948 
2949   case Op_DecodeN:
2950   case Op_DecodeNKlass:
2951     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
2952     // DecodeN could be pinned when it can't be fold into
2953     // an address expression, see the code for Op_CastPP above.
2954     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
2955     break;
2956 
2957   case Op_EncodeP:
2958   case Op_EncodePKlass: {
2959     Node* in1 = n->in(1);
2960     if (in1->is_DecodeNarrowPtr()) {
2961       n->subsume_by(in1->in(1), this);
2962     } else if (in1->Opcode() == Op_ConP) {
2963       const Type* t = in1->bottom_type();
2964       if (t == TypePtr::NULL_PTR) {
2965         assert(t->isa_oopptr(), "null klass?");
2966         n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
2967       } else if (t->isa_oopptr()) {
2968         n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
2969       } else if (t->isa_klassptr()) {
2970         n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
2971       }
2972     }
2973     if (in1->outcnt() == 0) {
2974       in1->disconnect_inputs(NULL, this);
2975     }
2976     break;
2977   }
2978 
2979   case Op_Proj: {
2980     if (OptimizeStringConcat) {
2981       ProjNode* p = n->as_Proj();
2982       if (p->_is_io_use) {
2983         // Separate projections were used for the exception path which
2984         // are normally removed by a late inline.  If it wasn't inlined
2985         // then they will hang around and should just be replaced with
2986         // the original one.
2987         Node* proj = NULL;
2988         // Replace with just one
2989         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
2990           Node *use = i.get();
2991           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
2992             proj = use;
2993             break;
2994           }
2995         }
2996         assert(proj != NULL, "must be found");
2997         p->subsume_by(proj, this);
2998       }
2999     }
3000     break;
3001   }
3002 
3003   case Op_Phi:
3004     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3005       // The EncodeP optimization may create Phi with the same edges
3006       // for all paths. It is not handled well by Register Allocator.
3007       Node* unique_in = n->in(1);
3008       assert(unique_in != NULL, "");
3009       uint cnt = n->req();
3010       for (uint i = 2; i < cnt; i++) {
3011         Node* m = n->in(i);
3012         assert(m != NULL, "");
3013         if (unique_in != m)
3014           unique_in = NULL;
3015       }
3016       if (unique_in != NULL) {
3017         n->subsume_by(unique_in, this);
3018       }
3019     }
3020     break;
3021 
3022 #endif
3023 
3024 #ifdef ASSERT
3025   case Op_CastII:
3026     // Verify that all range check dependent CastII nodes were removed.
3027     if (n->isa_CastII()->has_range_check()) {
3028       n->dump(3);
3029       assert(false, "Range check dependent CastII node was not removed");
3030     }
3031     break;
3032 #endif
3033 
3034   case Op_ModI:
3035     if (UseDivMod) {
3036       // Check if a%b and a/b both exist
3037       Node* d = n->find_similar(Op_DivI);
3038       if (d) {
3039         // Replace them with a fused divmod if supported
3040         if (Matcher::has_match_rule(Op_DivModI)) {
3041           DivModINode* divmod = DivModINode::make(this, n);
3042           d->subsume_by(divmod->div_proj(), this);
3043           n->subsume_by(divmod->mod_proj(), this);
3044         } else {
3045           // replace a%b with a-((a/b)*b)
3046           Node* mult = new (this) MulINode(d, d->in(2));
3047           Node* sub  = new (this) SubINode(d->in(1), mult);
3048           n->subsume_by(sub, this);
3049         }
3050       }
3051     }
3052     break;
3053 
3054   case Op_ModL:
3055     if (UseDivMod) {
3056       // Check if a%b and a/b both exist
3057       Node* d = n->find_similar(Op_DivL);
3058       if (d) {
3059         // Replace them with a fused divmod if supported
3060         if (Matcher::has_match_rule(Op_DivModL)) {
3061           DivModLNode* divmod = DivModLNode::make(this, n);
3062           d->subsume_by(divmod->div_proj(), this);
3063           n->subsume_by(divmod->mod_proj(), this);
3064         } else {
3065           // replace a%b with a-((a/b)*b)
3066           Node* mult = new (this) MulLNode(d, d->in(2));
3067           Node* sub  = new (this) SubLNode(d->in(1), mult);
3068           n->subsume_by(sub, this);
3069         }
3070       }
3071     }
3072     break;
3073 
3074   case Op_LoadVector:
3075   case Op_StoreVector:
3076     break;
3077 
3078   case Op_PackB:
3079   case Op_PackS:
3080   case Op_PackI:
3081   case Op_PackF:
3082   case Op_PackL:
3083   case Op_PackD:
3084     if (n->req()-1 > 2) {
3085       // Replace many operand PackNodes with a binary tree for matching
3086       PackNode* p = (PackNode*) n;
3087       Node* btp = p->binary_tree_pack(this, 1, n->req());
3088       n->subsume_by(btp, this);
3089     }
3090     break;
3091   case Op_Loop:
3092   case Op_CountedLoop:
3093     if (n->as_Loop()->is_inner_loop()) {
3094       frc.inc_inner_loop_count();
3095     }
3096     break;
3097   case Op_LShiftI:
3098   case Op_RShiftI:
3099   case Op_URShiftI:
3100   case Op_LShiftL:
3101   case Op_RShiftL:
3102   case Op_URShiftL:
3103     if (Matcher::need_masked_shift_count) {
3104       // The cpu's shift instructions don't restrict the count to the
3105       // lower 5/6 bits. We need to do the masking ourselves.
3106       Node* in2 = n->in(2);
3107       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3108       const TypeInt* t = in2->find_int_type();
3109       if (t != NULL && t->is_con()) {
3110         juint shift = t->get_con();
3111         if (shift > mask) { // Unsigned cmp
3112           n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3113         }
3114       } else {
3115         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3116           Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3117           n->set_req(2, shift);
3118         }
3119       }
3120       if (in2->outcnt() == 0) { // Remove dead node
3121         in2->disconnect_inputs(NULL, this);
3122       }
3123     }
3124     break;
3125   case Op_MemBarStoreStore:
3126   case Op_MemBarRelease:
3127     // Break the link with AllocateNode: it is no longer useful and
3128     // confuses register allocation.
3129     if (n->req() > MemBarNode::Precedent) {
3130       n->set_req(MemBarNode::Precedent, top());
3131     }
3132     break;
3133   default:
3134     assert( !n->is_Call(), "" );
3135     assert( !n->is_Mem(), "" );
3136     assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3137     break;
3138   }
3139 
3140   // Collect CFG split points
3141   if (n->is_MultiBranch())
3142     frc._tests.push(n);
3143 }
3144 
3145 //------------------------------final_graph_reshaping_walk---------------------
3146 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3147 // requires that the walk visits a node's inputs before visiting the node.
3148 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3149   ResourceArea *area = Thread::current()->resource_area();
3150   Unique_Node_List sfpt(area);
3151 
3152   frc._visited.set(root->_idx); // first, mark node as visited
3153   uint cnt = root->req();
3154   Node *n = root;
3155   uint  i = 0;
3156   while (true) {
3157     if (i < cnt) {
3158       // Place all non-visited non-null inputs onto stack
3159       Node* m = n->in(i);
3160       ++i;
3161       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3162         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3163           // compute worst case interpreter size in case of a deoptimization
3164           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3165 
3166           sfpt.push(m);
3167         }
3168         cnt = m->req();
3169         nstack.push(n, i); // put on stack parent and next input's index
3170         n = m;
3171         i = 0;
3172       }
3173     } else {
3174       // Now do post-visit work
3175       final_graph_reshaping_impl( n, frc );
3176       if (nstack.is_empty())
3177         break;             // finished
3178       n = nstack.node();   // Get node from stack
3179       cnt = n->req();
3180       i = nstack.index();
3181       nstack.pop();        // Shift to the next node on stack
3182     }
3183   }
3184 
3185   // Skip next transformation if compressed oops are not used.
3186   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3187       (!UseCompressedOops && !UseCompressedClassPointers))
3188     return;
3189 
3190   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3191   // It could be done for an uncommon traps or any safepoints/calls
3192   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3193   while (sfpt.size() > 0) {
3194     n = sfpt.pop();
3195     JVMState *jvms = n->as_SafePoint()->jvms();
3196     assert(jvms != NULL, "sanity");
3197     int start = jvms->debug_start();
3198     int end   = n->req();
3199     bool is_uncommon = (n->is_CallStaticJava() &&
3200                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3201     for (int j = start; j < end; j++) {
3202       Node* in = n->in(j);
3203       if (in->is_DecodeNarrowPtr()) {
3204         bool safe_to_skip = true;
3205         if (!is_uncommon ) {
3206           // Is it safe to skip?
3207           for (uint i = 0; i < in->outcnt(); i++) {
3208             Node* u = in->raw_out(i);
3209             if (!u->is_SafePoint() ||
3210                  u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3211               safe_to_skip = false;
3212             }
3213           }
3214         }
3215         if (safe_to_skip) {
3216           n->set_req(j, in->in(1));
3217         }
3218         if (in->outcnt() == 0) {
3219           in->disconnect_inputs(NULL, this);
3220         }
3221       }
3222     }
3223   }
3224 }
3225 
3226 //------------------------------final_graph_reshaping--------------------------
3227 // Final Graph Reshaping.
3228 //
3229 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3230 //     and not commoned up and forced early.  Must come after regular
3231 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3232 //     inputs to Loop Phis; these will be split by the allocator anyways.
3233 //     Remove Opaque nodes.
3234 // (2) Move last-uses by commutative operations to the left input to encourage
3235 //     Intel update-in-place two-address operations and better register usage
3236 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3237 //     calls canonicalizing them back.
3238 // (3) Count the number of double-precision FP ops, single-precision FP ops
3239 //     and call sites.  On Intel, we can get correct rounding either by
3240 //     forcing singles to memory (requires extra stores and loads after each
3241 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3242 //     clearing the mode bit around call sites).  The mode bit is only used
3243 //     if the relative frequency of single FP ops to calls is low enough.
3244 //     This is a key transform for SPEC mpeg_audio.
3245 // (4) Detect infinite loops; blobs of code reachable from above but not
3246 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3247 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3248 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3249 //     Detection is by looking for IfNodes where only 1 projection is
3250 //     reachable from below or CatchNodes missing some targets.
3251 // (5) Assert for insane oop offsets in debug mode.
3252 
3253 bool Compile::final_graph_reshaping() {
3254   // an infinite loop may have been eliminated by the optimizer,
3255   // in which case the graph will be empty.
3256   if (root()->req() == 1) {
3257     record_method_not_compilable("trivial infinite loop");
3258     return true;
3259   }
3260 
3261   // Expensive nodes have their control input set to prevent the GVN
3262   // from freely commoning them. There's no GVN beyond this point so
3263   // no need to keep the control input. We want the expensive nodes to
3264   // be freely moved to the least frequent code path by gcm.
3265   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3266   for (int i = 0; i < expensive_count(); i++) {
3267     _expensive_nodes->at(i)->set_req(0, NULL);
3268   }
3269 
3270   Final_Reshape_Counts frc;
3271 
3272   // Visit everybody reachable!
3273   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3274   Node_Stack nstack(live_nodes() >> 1);
3275   final_graph_reshaping_walk(nstack, root(), frc);
3276 
3277   // Check for unreachable (from below) code (i.e., infinite loops).
3278   for( uint i = 0; i < frc._tests.size(); i++ ) {
3279     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3280     // Get number of CFG targets.
3281     // Note that PCTables include exception targets after calls.
3282     uint required_outcnt = n->required_outcnt();
3283     if (n->outcnt() != required_outcnt) {
3284       // Check for a few special cases.  Rethrow Nodes never take the
3285       // 'fall-thru' path, so expected kids is 1 less.
3286       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3287         if (n->in(0)->in(0)->is_Call()) {
3288           CallNode *call = n->in(0)->in(0)->as_Call();
3289           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3290             required_outcnt--;      // Rethrow always has 1 less kid
3291           } else if (call->req() > TypeFunc::Parms &&
3292                      call->is_CallDynamicJava()) {
3293             // Check for null receiver. In such case, the optimizer has
3294             // detected that the virtual call will always result in a null
3295             // pointer exception. The fall-through projection of this CatchNode
3296             // will not be populated.
3297             Node *arg0 = call->in(TypeFunc::Parms);
3298             if (arg0->is_Type() &&
3299                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3300               required_outcnt--;
3301             }
3302           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3303                      call->req() > TypeFunc::Parms+1 &&
3304                      call->is_CallStaticJava()) {
3305             // Check for negative array length. In such case, the optimizer has
3306             // detected that the allocation attempt will always result in an
3307             // exception. There is no fall-through projection of this CatchNode .
3308             Node *arg1 = call->in(TypeFunc::Parms+1);
3309             if (arg1->is_Type() &&
3310                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3311               required_outcnt--;
3312             }
3313           }
3314         }
3315       }
3316       // Recheck with a better notion of 'required_outcnt'
3317       if (n->outcnt() != required_outcnt) {
3318         record_method_not_compilable("malformed control flow");
3319         return true;            // Not all targets reachable!
3320       }
3321     }
3322     // Check that I actually visited all kids.  Unreached kids
3323     // must be infinite loops.
3324     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3325       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3326         record_method_not_compilable("infinite loop");
3327         return true;            // Found unvisited kid; must be unreach
3328       }
3329   }
3330 
3331   // If original bytecodes contained a mixture of floats and doubles
3332   // check if the optimizer has made it homogenous, item (3).
3333   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3334       frc.get_float_count() > 32 &&
3335       frc.get_double_count() == 0 &&
3336       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3337     set_24_bit_selection_and_mode( false,  true );
3338   }
3339 
3340   set_java_calls(frc.get_java_call_count());
3341   set_inner_loops(frc.get_inner_loop_count());
3342 
3343   // No infinite loops, no reason to bail out.
3344   return false;
3345 }
3346 
3347 //-----------------------------too_many_traps----------------------------------
3348 // Report if there are too many traps at the current method and bci.
3349 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3350 bool Compile::too_many_traps(ciMethod* method,
3351                              int bci,
3352                              Deoptimization::DeoptReason reason) {
3353   ciMethodData* md = method->method_data();
3354   if (md->is_empty()) {
3355     // Assume the trap has not occurred, or that it occurred only
3356     // because of a transient condition during start-up in the interpreter.
3357     return false;
3358   }
3359   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3360   if (md->has_trap_at(bci, m, reason) != 0) {
3361     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3362     // Also, if there are multiple reasons, or if there is no per-BCI record,
3363     // assume the worst.
3364     if (log())
3365       log()->elem("observe trap='%s' count='%d'",
3366                   Deoptimization::trap_reason_name(reason),
3367                   md->trap_count(reason));
3368     return true;
3369   } else {
3370     // Ignore method/bci and see if there have been too many globally.
3371     return too_many_traps(reason, md);
3372   }
3373 }
3374 
3375 // Less-accurate variant which does not require a method and bci.
3376 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3377                              ciMethodData* logmd) {
3378   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3379     // Too many traps globally.
3380     // Note that we use cumulative trap_count, not just md->trap_count.
3381     if (log()) {
3382       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3383       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3384                   Deoptimization::trap_reason_name(reason),
3385                   mcount, trap_count(reason));
3386     }
3387     return true;
3388   } else {
3389     // The coast is clear.
3390     return false;
3391   }
3392 }
3393 
3394 //--------------------------too_many_recompiles--------------------------------
3395 // Report if there are too many recompiles at the current method and bci.
3396 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3397 // Is not eager to return true, since this will cause the compiler to use
3398 // Action_none for a trap point, to avoid too many recompilations.
3399 bool Compile::too_many_recompiles(ciMethod* method,
3400                                   int bci,
3401                                   Deoptimization::DeoptReason reason) {
3402   ciMethodData* md = method->method_data();
3403   if (md->is_empty()) {
3404     // Assume the trap has not occurred, or that it occurred only
3405     // because of a transient condition during start-up in the interpreter.
3406     return false;
3407   }
3408   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3409   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3410   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3411   Deoptimization::DeoptReason per_bc_reason
3412     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3413   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3414   if ((per_bc_reason == Deoptimization::Reason_none
3415        || md->has_trap_at(bci, m, reason) != 0)
3416       // The trap frequency measure we care about is the recompile count:
3417       && md->trap_recompiled_at(bci, m)
3418       && md->overflow_recompile_count() >= bc_cutoff) {
3419     // Do not emit a trap here if it has already caused recompilations.
3420     // Also, if there are multiple reasons, or if there is no per-BCI record,
3421     // assume the worst.
3422     if (log())
3423       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3424                   Deoptimization::trap_reason_name(reason),
3425                   md->trap_count(reason),
3426                   md->overflow_recompile_count());
3427     return true;
3428   } else if (trap_count(reason) != 0
3429              && decompile_count() >= m_cutoff) {
3430     // Too many recompiles globally, and we have seen this sort of trap.
3431     // Use cumulative decompile_count, not just md->decompile_count.
3432     if (log())
3433       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3434                   Deoptimization::trap_reason_name(reason),
3435                   md->trap_count(reason), trap_count(reason),
3436                   md->decompile_count(), decompile_count());
3437     return true;
3438   } else {
3439     // The coast is clear.
3440     return false;
3441   }
3442 }
3443 
3444 // Compute when not to trap. Used by matching trap based nodes and
3445 // NullCheck optimization.
3446 void Compile::set_allowed_deopt_reasons() {
3447   _allowed_reasons = 0;
3448   if (is_method_compilation()) {
3449     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3450       assert(rs < BitsPerInt, "recode bit map");
3451       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3452         _allowed_reasons |= nth_bit(rs);
3453       }
3454     }
3455   }
3456 }
3457 
3458 #ifndef PRODUCT
3459 //------------------------------verify_graph_edges---------------------------
3460 // Walk the Graph and verify that there is a one-to-one correspondence
3461 // between Use-Def edges and Def-Use edges in the graph.
3462 void Compile::verify_graph_edges(bool no_dead_code) {
3463   if (VerifyGraphEdges) {
3464     ResourceArea *area = Thread::current()->resource_area();
3465     Unique_Node_List visited(area);
3466     // Call recursive graph walk to check edges
3467     _root->verify_edges(visited);
3468     if (no_dead_code) {
3469       // Now make sure that no visited node is used by an unvisited node.
3470       bool dead_nodes = 0;
3471       Unique_Node_List checked(area);
3472       while (visited.size() > 0) {
3473         Node* n = visited.pop();
3474         checked.push(n);
3475         for (uint i = 0; i < n->outcnt(); i++) {
3476           Node* use = n->raw_out(i);
3477           if (checked.member(use))  continue;  // already checked
3478           if (visited.member(use))  continue;  // already in the graph
3479           if (use->is_Con())        continue;  // a dead ConNode is OK
3480           // At this point, we have found a dead node which is DU-reachable.
3481           if (dead_nodes++ == 0)
3482             tty->print_cr("*** Dead nodes reachable via DU edges:");
3483           use->dump(2);
3484           tty->print_cr("---");
3485           checked.push(use);  // No repeats; pretend it is now checked.
3486         }
3487       }
3488       assert(dead_nodes == 0, "using nodes must be reachable from root");
3489     }
3490   }
3491 }
3492 
3493 // Verify GC barriers consistency
3494 // Currently supported:
3495 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3496 void Compile::verify_barriers() {
3497   if (UseG1GC) {
3498     // Verify G1 pre-barriers
3499     const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3500 
3501     ResourceArea *area = Thread::current()->resource_area();
3502     Unique_Node_List visited(area);
3503     Node_List worklist(area);
3504     // We're going to walk control flow backwards starting from the Root
3505     worklist.push(_root);
3506     while (worklist.size() > 0) {
3507       Node* x = worklist.pop();
3508       if (x == NULL || x == top()) continue;
3509       if (visited.member(x)) {
3510         continue;
3511       } else {
3512         visited.push(x);
3513       }
3514 
3515       if (x->is_Region()) {
3516         for (uint i = 1; i < x->req(); i++) {
3517           worklist.push(x->in(i));
3518         }
3519       } else {
3520         worklist.push(x->in(0));
3521         // We are looking for the pattern:
3522         //                            /->ThreadLocal
3523         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3524         //              \->ConI(0)
3525         // We want to verify that the If and the LoadB have the same control
3526         // See GraphKit::g1_write_barrier_pre()
3527         if (x->is_If()) {
3528           IfNode *iff = x->as_If();
3529           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3530             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3531             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3532                 && cmp->in(1)->is_Load()) {
3533               LoadNode* load = cmp->in(1)->as_Load();
3534               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3535                   && load->in(2)->in(3)->is_Con()
3536                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3537 
3538                 Node* if_ctrl = iff->in(0);
3539                 Node* load_ctrl = load->in(0);
3540 
3541                 if (if_ctrl != load_ctrl) {
3542                   // Skip possible CProj->NeverBranch in infinite loops
3543                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3544                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3545                     if_ctrl = if_ctrl->in(0)->in(0);
3546                   }
3547                 }
3548                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3549               }
3550             }
3551           }
3552         }
3553       }
3554     }
3555   }
3556 }
3557 
3558 #endif
3559 
3560 // The Compile object keeps track of failure reasons separately from the ciEnv.
3561 // This is required because there is not quite a 1-1 relation between the
3562 // ciEnv and its compilation task and the Compile object.  Note that one
3563 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3564 // to backtrack and retry without subsuming loads.  Other than this backtracking
3565 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3566 // by the logic in C2Compiler.
3567 void Compile::record_failure(const char* reason) {
3568   if (log() != NULL) {
3569     log()->elem("failure reason='%s' phase='compile'", reason);
3570   }
3571   if (_failure_reason == NULL) {
3572     // Record the first failure reason.
3573     _failure_reason = reason;
3574   }
3575 
3576   EventCompilerFailure event;
3577   if (event.should_commit()) {
3578     event.set_compileID(Compile::compile_id());
3579     event.set_failure(reason);
3580     event.commit();
3581   }
3582 
3583   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3584     C->print_method(PHASE_FAILURE);
3585   }
3586   _root = NULL;  // flush the graph, too
3587 }
3588 
3589 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3590   : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3591     _phase_name(name), _dolog(dolog)
3592 {
3593   if (dolog) {
3594     C = Compile::current();
3595     _log = C->log();
3596   } else {
3597     C = NULL;
3598     _log = NULL;
3599   }
3600   if (_log != NULL) {
3601     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3602     _log->stamp();
3603     _log->end_head();
3604   }
3605 }
3606 
3607 Compile::TracePhase::~TracePhase() {
3608 
3609   C = Compile::current();
3610   if (_dolog) {
3611     _log = C->log();
3612   } else {
3613     _log = NULL;
3614   }
3615 
3616 #ifdef ASSERT
3617   if (PrintIdealNodeCount) {
3618     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3619                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3620   }
3621 
3622   if (VerifyIdealNodeCount) {
3623     Compile::current()->print_missing_nodes();
3624   }
3625 #endif
3626 
3627   if (_log != NULL) {
3628     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3629   }
3630 }
3631 
3632 //=============================================================================
3633 // Two Constant's are equal when the type and the value are equal.
3634 bool Compile::Constant::operator==(const Constant& other) {
3635   if (type()          != other.type()         )  return false;
3636   if (can_be_reused() != other.can_be_reused())  return false;
3637   // For floating point values we compare the bit pattern.
3638   switch (type()) {
3639   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3640   case T_LONG:
3641   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
3642   case T_OBJECT:
3643   case T_ADDRESS: return (_v._value.l == other._v._value.l);
3644   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
3645   case T_METADATA: return (_v._metadata == other._v._metadata);
3646   default: ShouldNotReachHere();
3647   }
3648   return false;
3649 }
3650 
3651 static int type_to_size_in_bytes(BasicType t) {
3652   switch (t) {
3653   case T_LONG:    return sizeof(jlong  );
3654   case T_FLOAT:   return sizeof(jfloat );
3655   case T_DOUBLE:  return sizeof(jdouble);
3656   case T_METADATA: return sizeof(Metadata*);
3657     // We use T_VOID as marker for jump-table entries (labels) which
3658     // need an internal word relocation.
3659   case T_VOID:
3660   case T_ADDRESS:
3661   case T_OBJECT:  return sizeof(jobject);
3662   }
3663 
3664   ShouldNotReachHere();
3665   return -1;
3666 }
3667 
3668 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3669   // sort descending
3670   if (a->freq() > b->freq())  return -1;
3671   if (a->freq() < b->freq())  return  1;
3672   return 0;
3673 }
3674 
3675 void Compile::ConstantTable::calculate_offsets_and_size() {
3676   // First, sort the array by frequencies.
3677   _constants.sort(qsort_comparator);
3678 
3679 #ifdef ASSERT
3680   // Make sure all jump-table entries were sorted to the end of the
3681   // array (they have a negative frequency).
3682   bool found_void = false;
3683   for (int i = 0; i < _constants.length(); i++) {
3684     Constant con = _constants.at(i);
3685     if (con.type() == T_VOID)
3686       found_void = true;  // jump-tables
3687     else
3688       assert(!found_void, "wrong sorting");
3689   }
3690 #endif
3691 
3692   int offset = 0;
3693   for (int i = 0; i < _constants.length(); i++) {
3694     Constant* con = _constants.adr_at(i);
3695 
3696     // Align offset for type.
3697     int typesize = type_to_size_in_bytes(con->type());
3698     offset = align_size_up(offset, typesize);
3699     con->set_offset(offset);   // set constant's offset
3700 
3701     if (con->type() == T_VOID) {
3702       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3703       offset = offset + typesize * n->outcnt();  // expand jump-table
3704     } else {
3705       offset = offset + typesize;
3706     }
3707   }
3708 
3709   // Align size up to the next section start (which is insts; see
3710   // CodeBuffer::align_at_start).
3711   assert(_size == -1, "already set?");
3712   _size = align_size_up(offset, CodeEntryAlignment);
3713 }
3714 
3715 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3716   MacroAssembler _masm(&cb);
3717   for (int i = 0; i < _constants.length(); i++) {
3718     Constant con = _constants.at(i);
3719     address constant_addr = NULL;
3720     switch (con.type()) {
3721     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
3722     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3723     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3724     case T_OBJECT: {
3725       jobject obj = con.get_jobject();
3726       int oop_index = _masm.oop_recorder()->find_index(obj);
3727       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3728       break;
3729     }
3730     case T_ADDRESS: {
3731       address addr = (address) con.get_jobject();
3732       constant_addr = _masm.address_constant(addr);
3733       break;
3734     }
3735     // We use T_VOID as marker for jump-table entries (labels) which
3736     // need an internal word relocation.
3737     case T_VOID: {
3738       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3739       // Fill the jump-table with a dummy word.  The real value is
3740       // filled in later in fill_jump_table.
3741       address dummy = (address) n;
3742       constant_addr = _masm.address_constant(dummy);
3743       // Expand jump-table
3744       for (uint i = 1; i < n->outcnt(); i++) {
3745         address temp_addr = _masm.address_constant(dummy + i);
3746         assert(temp_addr, "consts section too small");
3747       }
3748       break;
3749     }
3750     case T_METADATA: {
3751       Metadata* obj = con.get_metadata();
3752       int metadata_index = _masm.oop_recorder()->find_index(obj);
3753       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3754       break;
3755     }
3756     default: ShouldNotReachHere();
3757     }
3758     assert(constant_addr, "consts section too small");
3759     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3760             err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3761   }
3762 }
3763 
3764 int Compile::ConstantTable::find_offset(Constant& con) const {
3765   int idx = _constants.find(con);
3766   assert(idx != -1, "constant must be in constant table");
3767   int offset = _constants.at(idx).offset();
3768   assert(offset != -1, "constant table not emitted yet?");
3769   return offset;
3770 }
3771 
3772 void Compile::ConstantTable::add(Constant& con) {
3773   if (con.can_be_reused()) {
3774     int idx = _constants.find(con);
3775     if (idx != -1 && _constants.at(idx).can_be_reused()) {
3776       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
3777       return;
3778     }
3779   }
3780   (void) _constants.append(con);
3781 }
3782 
3783 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3784   Block* b = Compile::current()->cfg()->get_block_for_node(n);
3785   Constant con(type, value, b->_freq);
3786   add(con);
3787   return con;
3788 }
3789 
3790 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3791   Constant con(metadata);
3792   add(con);
3793   return con;
3794 }
3795 
3796 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3797   jvalue value;
3798   BasicType type = oper->type()->basic_type();
3799   switch (type) {
3800   case T_LONG:    value.j = oper->constantL(); break;
3801   case T_FLOAT:   value.f = oper->constantF(); break;
3802   case T_DOUBLE:  value.d = oper->constantD(); break;
3803   case T_OBJECT:
3804   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3805   case T_METADATA: return add((Metadata*)oper->constant()); break;
3806   default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3807   }
3808   return add(n, type, value);
3809 }
3810 
3811 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3812   jvalue value;
3813   // We can use the node pointer here to identify the right jump-table
3814   // as this method is called from Compile::Fill_buffer right before
3815   // the MachNodes are emitted and the jump-table is filled (means the
3816   // MachNode pointers do not change anymore).
3817   value.l = (jobject) n;
3818   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
3819   add(con);
3820   return con;
3821 }
3822 
3823 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3824   // If called from Compile::scratch_emit_size do nothing.
3825   if (Compile::current()->in_scratch_emit_size())  return;
3826 
3827   assert(labels.is_nonempty(), "must be");
3828   assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3829 
3830   // Since MachConstantNode::constant_offset() also contains
3831   // table_base_offset() we need to subtract the table_base_offset()
3832   // to get the plain offset into the constant table.
3833   int offset = n->constant_offset() - table_base_offset();
3834 
3835   MacroAssembler _masm(&cb);
3836   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3837 
3838   for (uint i = 0; i < n->outcnt(); i++) {
3839     address* constant_addr = &jump_table_base[i];
3840     assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
3841     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3842     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3843   }
3844 }
3845 
3846 void Compile::dump_inlining() {
3847   if (print_inlining() || print_intrinsics()) {
3848     // Print inlining message for candidates that we couldn't inline
3849     // for lack of space or non constant receiver
3850     for (int i = 0; i < _late_inlines.length(); i++) {
3851       CallGenerator* cg = _late_inlines.at(i);
3852       cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3853     }
3854     Unique_Node_List useful;
3855     useful.push(root());
3856     for (uint next = 0; next < useful.size(); ++next) {
3857       Node* n  = useful.at(next);
3858       if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3859         CallNode* call = n->as_Call();
3860         CallGenerator* cg = call->generator();
3861         cg->print_inlining_late("receiver not constant");
3862       }
3863       uint max = n->len();
3864       for ( uint i = 0; i < max; ++i ) {
3865         Node *m = n->in(i);
3866         if ( m == NULL ) continue;
3867         useful.push(m);
3868       }
3869     }
3870     for (int i = 0; i < _print_inlining_list->length(); i++) {
3871       tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3872     }
3873   }
3874 }
3875 
3876 // Dump inlining replay data to the stream.
3877 // Don't change thread state and acquire any locks.
3878 void Compile::dump_inline_data(outputStream* out) {
3879   InlineTree* inl_tree = ilt();
3880   if (inl_tree != NULL) {
3881     out->print(" inline %d", inl_tree->count());
3882     inl_tree->dump_replay_data(out);
3883   }
3884 }
3885 
3886 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3887   if (n1->Opcode() < n2->Opcode())      return -1;
3888   else if (n1->Opcode() > n2->Opcode()) return 1;
3889 
3890   assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
3891   for (uint i = 1; i < n1->req(); i++) {
3892     if (n1->in(i) < n2->in(i))      return -1;
3893     else if (n1->in(i) > n2->in(i)) return 1;
3894   }
3895 
3896   return 0;
3897 }
3898 
3899 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
3900   Node* n1 = *n1p;
3901   Node* n2 = *n2p;
3902 
3903   return cmp_expensive_nodes(n1, n2);
3904 }
3905 
3906 void Compile::sort_expensive_nodes() {
3907   if (!expensive_nodes_sorted()) {
3908     _expensive_nodes->sort(cmp_expensive_nodes);
3909   }
3910 }
3911 
3912 bool Compile::expensive_nodes_sorted() const {
3913   for (int i = 1; i < _expensive_nodes->length(); i++) {
3914     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
3915       return false;
3916     }
3917   }
3918   return true;
3919 }
3920 
3921 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
3922   if (_expensive_nodes->length() == 0) {
3923     return false;
3924   }
3925 
3926   assert(OptimizeExpensiveOps, "optimization off?");
3927 
3928   // Take this opportunity to remove dead nodes from the list
3929   int j = 0;
3930   for (int i = 0; i < _expensive_nodes->length(); i++) {
3931     Node* n = _expensive_nodes->at(i);
3932     if (!n->is_unreachable(igvn)) {
3933       assert(n->is_expensive(), "should be expensive");
3934       _expensive_nodes->at_put(j, n);
3935       j++;
3936     }
3937   }
3938   _expensive_nodes->trunc_to(j);
3939 
3940   // Then sort the list so that similar nodes are next to each other
3941   // and check for at least two nodes of identical kind with same data
3942   // inputs.
3943   sort_expensive_nodes();
3944 
3945   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
3946     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
3947       return true;
3948     }
3949   }
3950 
3951   return false;
3952 }
3953 
3954 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
3955   if (_expensive_nodes->length() == 0) {
3956     return;
3957   }
3958 
3959   assert(OptimizeExpensiveOps, "optimization off?");
3960 
3961   // Sort to bring similar nodes next to each other and clear the
3962   // control input of nodes for which there's only a single copy.
3963   sort_expensive_nodes();
3964 
3965   int j = 0;
3966   int identical = 0;
3967   int i = 0;
3968   for (; i < _expensive_nodes->length()-1; i++) {
3969     assert(j <= i, "can't write beyond current index");
3970     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
3971       identical++;
3972       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3973       continue;
3974     }
3975     if (identical > 0) {
3976       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3977       identical = 0;
3978     } else {
3979       Node* n = _expensive_nodes->at(i);
3980       igvn.hash_delete(n);
3981       n->set_req(0, NULL);
3982       igvn.hash_insert(n);
3983     }
3984   }
3985   if (identical > 0) {
3986     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
3987   } else if (_expensive_nodes->length() >= 1) {
3988     Node* n = _expensive_nodes->at(i);
3989     igvn.hash_delete(n);
3990     n->set_req(0, NULL);
3991     igvn.hash_insert(n);
3992   }
3993   _expensive_nodes->trunc_to(j);
3994 }
3995 
3996 void Compile::add_expensive_node(Node * n) {
3997   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
3998   assert(n->is_expensive(), "expensive nodes with non-null control here only");
3999   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4000   if (OptimizeExpensiveOps) {
4001     _expensive_nodes->append(n);
4002   } else {
4003     // Clear control input and let IGVN optimize expensive nodes if
4004     // OptimizeExpensiveOps is off.
4005     n->set_req(0, NULL);
4006   }
4007 }
4008 
4009 /**
4010  * Remove the speculative part of types and clean up the graph
4011  */
4012 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4013   if (UseTypeSpeculation) {
4014     Unique_Node_List worklist;
4015     worklist.push(root());
4016     int modified = 0;
4017     // Go over all type nodes that carry a speculative type, drop the
4018     // speculative part of the type and enqueue the node for an igvn
4019     // which may optimize it out.
4020     for (uint next = 0; next < worklist.size(); ++next) {
4021       Node *n  = worklist.at(next);
4022       if (n->is_Type()) {
4023         TypeNode* tn = n->as_Type();
4024         const Type* t = tn->type();
4025         const Type* t_no_spec = t->remove_speculative();
4026         if (t_no_spec != t) {
4027           bool in_hash = igvn.hash_delete(n);
4028           assert(in_hash, "node should be in igvn hash table");
4029           tn->set_type(t_no_spec);
4030           igvn.hash_insert(n);
4031           igvn._worklist.push(n); // give it a chance to go away
4032           modified++;
4033         }
4034       }
4035       uint max = n->len();
4036       for( uint i = 0; i < max; ++i ) {
4037         Node *m = n->in(i);
4038         if (not_a_node(m))  continue;
4039         worklist.push(m);
4040       }
4041     }
4042     // Drop the speculative part of all types in the igvn's type table
4043     igvn.remove_speculative_types();
4044     if (modified > 0) {
4045       igvn.optimize();
4046     }
4047 #ifdef ASSERT
4048     // Verify that after the IGVN is over no speculative type has resurfaced
4049     worklist.clear();
4050     worklist.push(root());
4051     for (uint next = 0; next < worklist.size(); ++next) {
4052       Node *n  = worklist.at(next);
4053       const Type* t = igvn.type_or_null(n);
4054       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4055       if (n->is_Type()) {
4056         t = n->as_Type()->type();
4057         assert(t == t->remove_speculative(), "no more speculative types");
4058       }
4059       uint max = n->len();
4060       for( uint i = 0; i < max; ++i ) {
4061         Node *m = n->in(i);
4062         if (not_a_node(m))  continue;
4063         worklist.push(m);
4064       }
4065     }
4066     igvn.check_no_speculative_types();
4067 #endif
4068   }
4069 }
4070 
4071 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4072 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4073   if (ctrl != NULL) {
4074     // Express control dependency by a CastII node with a narrow type.
4075     value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
4076     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4077     // node from floating above the range check during loop optimizations. Otherwise, the
4078     // ConvI2L node may be eliminated independently of the range check, causing the data path
4079     // to become TOP while the control path is still there (although it's unreachable).
4080     value->set_req(0, ctrl);
4081     // Save CastII node to remove it after loop optimizations.
4082     phase->C->add_range_check_cast(value);
4083     value = phase->transform(value);
4084   }
4085   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4086   return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
4087 }
4088 
4089 // Auxiliary method to support randomized stressing/fuzzing.
4090 //
4091 // This method can be called the arbitrary number of times, with current count
4092 // as the argument. The logic allows selecting a single candidate from the
4093 // running list of candidates as follows:
4094 //    int count = 0;
4095 //    Cand* selected = null;
4096 //    while(cand = cand->next()) {
4097 //      if (randomized_select(++count)) {
4098 //        selected = cand;
4099 //      }
4100 //    }
4101 //
4102 // Including count equalizes the chances any candidate is "selected".
4103 // This is useful when we don't have the complete list of candidates to choose
4104 // from uniformly. In this case, we need to adjust the randomicity of the
4105 // selection, or else we will end up biasing the selection towards the latter
4106 // candidates.
4107 //
4108 // Quick back-envelope calculation shows that for the list of n candidates
4109 // the equal probability for the candidate to persist as "best" can be
4110 // achieved by replacing it with "next" k-th candidate with the probability
4111 // of 1/k. It can be easily shown that by the end of the run, the
4112 // probability for any candidate is converged to 1/n, thus giving the
4113 // uniform distribution among all the candidates.
4114 //
4115 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4116 #define RANDOMIZED_DOMAIN_POW 29
4117 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4118 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4119 bool Compile::randomized_select(int count) {
4120   assert(count > 0, "only positive");
4121   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4122 }