1 /*
   2  * Copyright (c) 1997, 2018, 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_aarch64
  77 # include "adfiles/ad_aarch64.hpp"
  78 #elif defined TARGET_ARCH_MODEL_sparc
  79 # include "adfiles/ad_sparc.hpp"
  80 #elif defined TARGET_ARCH_MODEL_zero
  81 # include "adfiles/ad_zero.hpp"
  82 #elif defined TARGET_ARCH_MODEL_ppc_64
  83 # include "adfiles/ad_ppc_64.hpp"
  84 #endif
  85 
  86 #if INCLUDE_ALL_GCS
  87 #include "gc_implementation/shenandoah/shenandoahForwarding.hpp"
  88 #include "gc_implementation/shenandoah/shenandoahSupport.hpp"
  89 #endif
  90 
  91 // -------------------- Compile::mach_constant_base_node -----------------------
  92 // Constant table base node singleton.
  93 MachConstantBaseNode* Compile::mach_constant_base_node() {
  94   if (_mach_constant_base_node == NULL) {
  95     _mach_constant_base_node = new (C) MachConstantBaseNode();
  96     _mach_constant_base_node->add_req(C->root());
  97   }
  98   return _mach_constant_base_node;
  99 }
 100 
 101 
 102 /// Support for intrinsics.
 103 
 104 // Return the index at which m must be inserted (or already exists).
 105 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
 106 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
 107 #ifdef ASSERT
 108   for (int i = 1; i < _intrinsics->length(); i++) {
 109     CallGenerator* cg1 = _intrinsics->at(i-1);
 110     CallGenerator* cg2 = _intrinsics->at(i);
 111     assert(cg1->method() != cg2->method()
 112            ? cg1->method()     < cg2->method()
 113            : cg1->is_virtual() < cg2->is_virtual(),
 114            "compiler intrinsics list must stay sorted");
 115   }
 116 #endif
 117   // Binary search sorted list, in decreasing intervals [lo, hi].
 118   int lo = 0, hi = _intrinsics->length()-1;
 119   while (lo <= hi) {
 120     int mid = (uint)(hi + lo) / 2;
 121     ciMethod* mid_m = _intrinsics->at(mid)->method();
 122     if (m < mid_m) {
 123       hi = mid-1;
 124     } else if (m > mid_m) {
 125       lo = mid+1;
 126     } else {
 127       // look at minor sort key
 128       bool mid_virt = _intrinsics->at(mid)->is_virtual();
 129       if (is_virtual < mid_virt) {
 130         hi = mid-1;
 131       } else if (is_virtual > mid_virt) {
 132         lo = mid+1;
 133       } else {
 134         return mid;  // exact match
 135       }
 136     }
 137   }
 138   return lo;  // inexact match
 139 }
 140 
 141 void Compile::register_intrinsic(CallGenerator* cg) {
 142   if (_intrinsics == NULL) {
 143     _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
 144   }
 145   // This code is stolen from ciObjectFactory::insert.
 146   // Really, GrowableArray should have methods for
 147   // insert_at, remove_at, and binary_search.
 148   int len = _intrinsics->length();
 149   int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
 150   if (index == len) {
 151     _intrinsics->append(cg);
 152   } else {
 153 #ifdef ASSERT
 154     CallGenerator* oldcg = _intrinsics->at(index);
 155     assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
 156 #endif
 157     _intrinsics->append(_intrinsics->at(len-1));
 158     int pos;
 159     for (pos = len-2; pos >= index; pos--) {
 160       _intrinsics->at_put(pos+1,_intrinsics->at(pos));
 161     }
 162     _intrinsics->at_put(index, cg);
 163   }
 164   assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
 165 }
 166 
 167 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
 168   assert(m->is_loaded(), "don't try this on unloaded methods");
 169   if (_intrinsics != NULL) {
 170     int index = intrinsic_insertion_index(m, is_virtual);
 171     if (index < _intrinsics->length()
 172         && _intrinsics->at(index)->method() == m
 173         && _intrinsics->at(index)->is_virtual() == is_virtual) {
 174       return _intrinsics->at(index);
 175     }
 176   }
 177   // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
 178   if (m->intrinsic_id() != vmIntrinsics::_none &&
 179       m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
 180     CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
 181     if (cg != NULL) {
 182       // Save it for next time:
 183       register_intrinsic(cg);
 184       return cg;
 185     } else {
 186       gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
 187     }
 188   }
 189   return NULL;
 190 }
 191 
 192 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
 193 // in library_call.cpp.
 194 
 195 
 196 #ifndef PRODUCT
 197 // statistics gathering...
 198 
 199 juint  Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
 200 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
 201 
 202 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
 203   assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
 204   int oflags = _intrinsic_hist_flags[id];
 205   assert(flags != 0, "what happened?");
 206   if (is_virtual) {
 207     flags |= _intrinsic_virtual;
 208   }
 209   bool changed = (flags != oflags);
 210   if ((flags & _intrinsic_worked) != 0) {
 211     juint count = (_intrinsic_hist_count[id] += 1);
 212     if (count == 1) {
 213       changed = true;           // first time
 214     }
 215     // increment the overall count also:
 216     _intrinsic_hist_count[vmIntrinsics::_none] += 1;
 217   }
 218   if (changed) {
 219     if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
 220       // Something changed about the intrinsic's virtuality.
 221       if ((flags & _intrinsic_virtual) != 0) {
 222         // This is the first use of this intrinsic as a virtual call.
 223         if (oflags != 0) {
 224           // We already saw it as a non-virtual, so note both cases.
 225           flags |= _intrinsic_both;
 226         }
 227       } else if ((oflags & _intrinsic_both) == 0) {
 228         // This is the first use of this intrinsic as a non-virtual
 229         flags |= _intrinsic_both;
 230       }
 231     }
 232     _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
 233   }
 234   // update the overall flags also:
 235   _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
 236   return changed;
 237 }
 238 
 239 static char* format_flags(int flags, char* buf) {
 240   buf[0] = 0;
 241   if ((flags & Compile::_intrinsic_worked) != 0)    strcat(buf, ",worked");
 242   if ((flags & Compile::_intrinsic_failed) != 0)    strcat(buf, ",failed");
 243   if ((flags & Compile::_intrinsic_disabled) != 0)  strcat(buf, ",disabled");
 244   if ((flags & Compile::_intrinsic_virtual) != 0)   strcat(buf, ",virtual");
 245   if ((flags & Compile::_intrinsic_both) != 0)      strcat(buf, ",nonvirtual");
 246   if (buf[0] == 0)  strcat(buf, ",");
 247   assert(buf[0] == ',', "must be");
 248   return &buf[1];
 249 }
 250 
 251 void Compile::print_intrinsic_statistics() {
 252   char flagsbuf[100];
 253   ttyLocker ttyl;
 254   if (xtty != NULL)  xtty->head("statistics type='intrinsic'");
 255   tty->print_cr("Compiler intrinsic usage:");
 256   juint total = _intrinsic_hist_count[vmIntrinsics::_none];
 257   if (total == 0)  total = 1;  // avoid div0 in case of no successes
 258   #define PRINT_STAT_LINE(name, c, f) \
 259     tty->print_cr("  %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
 260   for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
 261     vmIntrinsics::ID id = (vmIntrinsics::ID) index;
 262     int   flags = _intrinsic_hist_flags[id];
 263     juint count = _intrinsic_hist_count[id];
 264     if ((flags | count) != 0) {
 265       PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
 266     }
 267   }
 268   PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
 269   if (xtty != NULL)  xtty->tail("statistics");
 270 }
 271 
 272 void Compile::print_statistics() {
 273   { ttyLocker ttyl;
 274     if (xtty != NULL)  xtty->head("statistics type='opto'");
 275     Parse::print_statistics();
 276     PhaseCCP::print_statistics();
 277     PhaseRegAlloc::print_statistics();
 278     Scheduling::print_statistics();
 279     PhasePeephole::print_statistics();
 280     PhaseIdealLoop::print_statistics();
 281     if (xtty != NULL)  xtty->tail("statistics");
 282   }
 283   if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
 284     // put this under its own <statistics> element.
 285     print_intrinsic_statistics();
 286   }
 287 }
 288 #endif //PRODUCT
 289 
 290 // Support for bundling info
 291 Bundle* Compile::node_bundling(const Node *n) {
 292   assert(valid_bundle_info(n), "oob");
 293   return &_node_bundling_base[n->_idx];
 294 }
 295 
 296 bool Compile::valid_bundle_info(const Node *n) {
 297   return (_node_bundling_limit > n->_idx);
 298 }
 299 
 300 
 301 void Compile::gvn_replace_by(Node* n, Node* nn) {
 302   for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
 303     Node* use = n->last_out(i);
 304     bool is_in_table = initial_gvn()->hash_delete(use);
 305     uint uses_found = 0;
 306     for (uint j = 0; j < use->len(); j++) {
 307       if (use->in(j) == n) {
 308         if (j < use->req())
 309           use->set_req(j, nn);
 310         else
 311           use->set_prec(j, nn);
 312         uses_found++;
 313       }
 314     }
 315     if (is_in_table) {
 316       // reinsert into table
 317       initial_gvn()->hash_find_insert(use);
 318     }
 319     record_for_igvn(use);
 320     i -= uses_found;    // we deleted 1 or more copies of this edge
 321   }
 322 }
 323 
 324 
 325 static inline bool not_a_node(const Node* n) {
 326   if (n == NULL)                   return true;
 327   if (((intptr_t)n & 1) != 0)      return true;  // uninitialized, etc.
 328   if (*(address*)n == badAddress)  return true;  // kill by Node::destruct
 329   return false;
 330 }
 331 
 332 // Identify all nodes that are reachable from below, useful.
 333 // Use breadth-first pass that records state in a Unique_Node_List,
 334 // recursive traversal is slower.
 335 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
 336   int estimated_worklist_size = live_nodes();
 337   useful.map( estimated_worklist_size, NULL );  // preallocate space
 338 
 339   // Initialize worklist
 340   if (root() != NULL)     { useful.push(root()); }
 341   // If 'top' is cached, declare it useful to preserve cached node
 342   if( cached_top_node() ) { useful.push(cached_top_node()); }
 343 
 344   // Push all useful nodes onto the list, breadthfirst
 345   for( uint next = 0; next < useful.size(); ++next ) {
 346     assert( next < unique(), "Unique useful nodes < total nodes");
 347     Node *n  = useful.at(next);
 348     uint max = n->len();
 349     for( uint i = 0; i < max; ++i ) {
 350       Node *m = n->in(i);
 351       if (not_a_node(m))  continue;
 352       useful.push(m);
 353     }
 354   }
 355 }
 356 
 357 // Update dead_node_list with any missing dead nodes using useful
 358 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
 359 void Compile::update_dead_node_list(Unique_Node_List &useful) {
 360   uint max_idx = unique();
 361   VectorSet& useful_node_set = useful.member_set();
 362 
 363   for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
 364     // If node with index node_idx is not in useful set,
 365     // mark it as dead in dead node list.
 366     if (! useful_node_set.test(node_idx) ) {
 367       record_dead_node(node_idx);
 368     }
 369   }
 370 }
 371 
 372 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
 373   int shift = 0;
 374   for (int i = 0; i < inlines->length(); i++) {
 375     CallGenerator* cg = inlines->at(i);
 376     CallNode* call = cg->call_node();
 377     if (shift > 0) {
 378       inlines->at_put(i-shift, cg);
 379     }
 380     if (!useful.member(call)) {
 381       shift++;
 382     }
 383   }
 384   inlines->trunc_to(inlines->length()-shift);
 385 }
 386 
 387 // Disconnect all useless nodes by disconnecting those at the boundary.
 388 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
 389   uint next = 0;
 390   while (next < useful.size()) {
 391     Node *n = useful.at(next++);
 392     if (n->is_SafePoint()) {
 393       // We're done with a parsing phase. Replaced nodes are not valid
 394       // beyond that point.
 395       n->as_SafePoint()->delete_replaced_nodes();
 396     }
 397     // Use raw traversal of out edges since this code removes out edges
 398     int max = n->outcnt();
 399     for (int j = 0; j < max; ++j) {
 400       Node* child = n->raw_out(j);
 401       if (! useful.member(child)) {
 402         assert(!child->is_top() || child != top(),
 403                "If top is cached in Compile object it is in useful list");
 404         // Only need to remove this out-edge to the useless node
 405         n->raw_del_out(j);
 406         --j;
 407         --max;
 408       }
 409     }
 410     if (n->outcnt() == 1 && n->has_special_unique_user()) {
 411       record_for_igvn(n->unique_out());
 412     }
 413     if (n->Opcode() == Op_AddP && CallLeafNode::has_only_g1_wb_pre_uses(n)) {
 414       for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
 415         record_for_igvn(n->fast_out(i));
 416       }
 417     }
 418   }
 419   // Remove useless macro and predicate opaq nodes
 420   for (int i = C->macro_count()-1; i >= 0; i--) {
 421     Node* n = C->macro_node(i);
 422     if (!useful.member(n)) {
 423       remove_macro_node(n);
 424     }
 425   }
 426   // Remove useless CastII nodes with range check dependency
 427   for (int i = range_check_cast_count() - 1; i >= 0; i--) {
 428     Node* cast = range_check_cast_node(i);
 429     if (!useful.member(cast)) {
 430       remove_range_check_cast(cast);
 431     }
 432   }
 433   // Remove useless expensive node
 434   for (int i = C->expensive_count()-1; i >= 0; i--) {
 435     Node* n = C->expensive_node(i);
 436     if (!useful.member(n)) {
 437       remove_expensive_node(n);
 438     }
 439   }
 440   for (int i = C->shenandoah_barriers_count()-1; i >= 0; i--) {
 441     ShenandoahLoadReferenceBarrierNode* n = C->shenandoah_barrier(i);
 442     if (!useful.member(n)) {
 443       remove_shenandoah_barrier(n);
 444     }
 445   }
 446   // clean up the late inline lists
 447   remove_useless_late_inlines(&_string_late_inlines, useful);
 448   remove_useless_late_inlines(&_boxing_late_inlines, useful);
 449   remove_useless_late_inlines(&_late_inlines, useful);
 450   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 451 }
 452 
 453 //------------------------------frame_size_in_words-----------------------------
 454 // frame_slots in units of words
 455 int Compile::frame_size_in_words() const {
 456   // shift is 0 in LP32 and 1 in LP64
 457   const int shift = (LogBytesPerWord - LogBytesPerInt);
 458   int words = _frame_slots >> shift;
 459   assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
 460   return words;
 461 }
 462 
 463 // To bang the stack of this compiled method we use the stack size
 464 // that the interpreter would need in case of a deoptimization. This
 465 // removes the need to bang the stack in the deoptimization blob which
 466 // in turn simplifies stack overflow handling.
 467 int Compile::bang_size_in_bytes() const {
 468   return MAX2(_interpreter_frame_size, frame_size_in_bytes());
 469 }
 470 
 471 // ============================================================================
 472 //------------------------------CompileWrapper---------------------------------
 473 class CompileWrapper : public StackObj {
 474   Compile *const _compile;
 475  public:
 476   CompileWrapper(Compile* compile);
 477 
 478   ~CompileWrapper();
 479 };
 480 
 481 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
 482   // the Compile* pointer is stored in the current ciEnv:
 483   ciEnv* env = compile->env();
 484   assert(env == ciEnv::current(), "must already be a ciEnv active");
 485   assert(env->compiler_data() == NULL, "compile already active?");
 486   env->set_compiler_data(compile);
 487   assert(compile == Compile::current(), "sanity");
 488 
 489   compile->set_type_dict(NULL);
 490   compile->set_type_hwm(NULL);
 491   compile->set_type_last_size(0);
 492   compile->set_last_tf(NULL, NULL);
 493   compile->set_indexSet_arena(NULL);
 494   compile->set_indexSet_free_block_list(NULL);
 495   compile->init_type_arena();
 496   Type::Initialize(compile);
 497   _compile->set_scratch_buffer_blob(NULL);
 498   _compile->begin_method();
 499 }
 500 CompileWrapper::~CompileWrapper() {
 501   _compile->end_method();
 502   if (_compile->scratch_buffer_blob() != NULL)
 503     BufferBlob::free(_compile->scratch_buffer_blob());
 504   _compile->env()->set_compiler_data(NULL);
 505 }
 506 
 507 
 508 //----------------------------print_compile_messages---------------------------
 509 void Compile::print_compile_messages() {
 510 #ifndef PRODUCT
 511   // Check if recompiling
 512   if (_subsume_loads == false && PrintOpto) {
 513     // Recompiling without allowing machine instructions to subsume loads
 514     tty->print_cr("*********************************************************");
 515     tty->print_cr("** Bailout: Recompile without subsuming loads          **");
 516     tty->print_cr("*********************************************************");
 517   }
 518   if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
 519     // Recompiling without escape analysis
 520     tty->print_cr("*********************************************************");
 521     tty->print_cr("** Bailout: Recompile without escape analysis          **");
 522     tty->print_cr("*********************************************************");
 523   }
 524   if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
 525     // Recompiling without boxing elimination
 526     tty->print_cr("*********************************************************");
 527     tty->print_cr("** Bailout: Recompile without boxing elimination       **");
 528     tty->print_cr("*********************************************************");
 529   }
 530   if (env()->break_at_compile()) {
 531     // Open the debugger when compiling this method.
 532     tty->print("### Breaking when compiling: ");
 533     method()->print_short_name();
 534     tty->cr();
 535     BREAKPOINT;
 536   }
 537 
 538   if( PrintOpto ) {
 539     if (is_osr_compilation()) {
 540       tty->print("[OSR]%3d", _compile_id);
 541     } else {
 542       tty->print("%3d", _compile_id);
 543     }
 544   }
 545 #endif
 546 }
 547 
 548 
 549 //-----------------------init_scratch_buffer_blob------------------------------
 550 // Construct a temporary BufferBlob and cache it for this compile.
 551 void Compile::init_scratch_buffer_blob(int const_size) {
 552   // If there is already a scratch buffer blob allocated and the
 553   // constant section is big enough, use it.  Otherwise free the
 554   // current and allocate a new one.
 555   BufferBlob* blob = scratch_buffer_blob();
 556   if ((blob != NULL) && (const_size <= _scratch_const_size)) {
 557     // Use the current blob.
 558   } else {
 559     if (blob != NULL) {
 560       BufferBlob::free(blob);
 561     }
 562 
 563     ResourceMark rm;
 564     _scratch_const_size = const_size;
 565     int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
 566     blob = BufferBlob::create("Compile::scratch_buffer", size);
 567     // Record the buffer blob for next time.
 568     set_scratch_buffer_blob(blob);
 569     // Have we run out of code space?
 570     if (scratch_buffer_blob() == NULL) {
 571       // Let CompilerBroker disable further compilations.
 572       record_failure("Not enough space for scratch buffer in CodeCache");
 573       return;
 574     }
 575   }
 576 
 577   // Initialize the relocation buffers
 578   relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
 579   set_scratch_locs_memory(locs_buf);
 580 }
 581 
 582 
 583 //-----------------------scratch_emit_size-------------------------------------
 584 // Helper function that computes size by emitting code
 585 uint Compile::scratch_emit_size(const Node* n) {
 586   // Start scratch_emit_size section.
 587   set_in_scratch_emit_size(true);
 588 
 589   // Emit into a trash buffer and count bytes emitted.
 590   // This is a pretty expensive way to compute a size,
 591   // but it works well enough if seldom used.
 592   // All common fixed-size instructions are given a size
 593   // method by the AD file.
 594   // Note that the scratch buffer blob and locs memory are
 595   // allocated at the beginning of the compile task, and
 596   // may be shared by several calls to scratch_emit_size.
 597   // The allocation of the scratch buffer blob is particularly
 598   // expensive, since it has to grab the code cache lock.
 599   BufferBlob* blob = this->scratch_buffer_blob();
 600   assert(blob != NULL, "Initialize BufferBlob at start");
 601   assert(blob->size() > MAX_inst_size, "sanity");
 602   relocInfo* locs_buf = scratch_locs_memory();
 603   address blob_begin = blob->content_begin();
 604   address blob_end   = (address)locs_buf;
 605   assert(blob->content_contains(blob_end), "sanity");
 606   CodeBuffer buf(blob_begin, blob_end - blob_begin);
 607   buf.initialize_consts_size(_scratch_const_size);
 608   buf.initialize_stubs_size(MAX_stubs_size);
 609   assert(locs_buf != NULL, "sanity");
 610   int lsize = MAX_locs_size / 3;
 611   buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
 612   buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
 613   buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
 614 
 615   // Do the emission.
 616 
 617   Label fakeL; // Fake label for branch instructions.
 618   Label*   saveL = NULL;
 619   uint save_bnum = 0;
 620   bool is_branch = n->is_MachBranch();
 621   if (is_branch) {
 622     MacroAssembler masm(&buf);
 623     masm.bind(fakeL);
 624     n->as_MachBranch()->save_label(&saveL, &save_bnum);
 625     n->as_MachBranch()->label_set(&fakeL, 0);
 626   }
 627   n->emit(buf, this->regalloc());
 628 
 629   // Emitting into the scratch buffer should not fail
 630   assert (!failing(), err_msg_res("Must not have pending failure. Reason is: %s", failure_reason()));
 631 
 632   if (is_branch) // Restore label.
 633     n->as_MachBranch()->label_set(saveL, save_bnum);
 634 
 635   // End scratch_emit_size section.
 636   set_in_scratch_emit_size(false);
 637 
 638   return buf.insts_size();
 639 }
 640 
 641 
 642 // ============================================================================
 643 //------------------------------Compile standard-------------------------------
 644 debug_only( int Compile::_debug_idx = 100000; )
 645 
 646 // Compile a method.  entry_bci is -1 for normal compilations and indicates
 647 // the continuation bci for on stack replacement.
 648 
 649 
 650 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
 651                   bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
 652                 : Phase(Compiler),
 653                   _env(ci_env),
 654                   _log(ci_env->log()),
 655                   _compile_id(ci_env->compile_id()),
 656                   _save_argument_registers(false),
 657                   _stub_name(NULL),
 658                   _stub_function(NULL),
 659                   _stub_entry_point(NULL),
 660                   _method(target),
 661                   _entry_bci(osr_bci),
 662                   _initial_gvn(NULL),
 663                   _for_igvn(NULL),
 664                   _warm_calls(NULL),
 665                   _subsume_loads(subsume_loads),
 666                   _do_escape_analysis(do_escape_analysis),
 667                   _eliminate_boxing(eliminate_boxing),
 668                   _failure_reason(NULL),
 669                   _code_buffer("Compile::Fill_buffer"),
 670                   _orig_pc_slot(0),
 671                   _orig_pc_slot_offset_in_bytes(0),
 672                   _has_method_handle_invokes(false),
 673                   _mach_constant_base_node(NULL),
 674                   _node_bundling_limit(0),
 675                   _node_bundling_base(NULL),
 676                   _java_calls(0),
 677                   _inner_loops(0),
 678                   _scratch_const_size(-1),
 679                   _in_scratch_emit_size(false),
 680                   _dead_node_list(comp_arena()),
 681                   _dead_node_count(0),
 682 #ifndef PRODUCT
 683                   _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
 684                   _in_dump_cnt(0),
 685                   _printer(IdealGraphPrinter::printer()),
 686 #endif
 687                   _congraph(NULL),
 688                   _comp_arena(mtCompiler),
 689                   _node_arena(mtCompiler),
 690                   _old_arena(mtCompiler),
 691                   _Compile_types(mtCompiler),
 692                   _replay_inline_data(NULL),
 693                   _late_inlines(comp_arena(), 2, 0, NULL),
 694                   _string_late_inlines(comp_arena(), 2, 0, NULL),
 695                   _boxing_late_inlines(comp_arena(), 2, 0, NULL),
 696                   _late_inlines_pos(0),
 697                   _number_of_mh_late_inlines(0),
 698                   _inlining_progress(false),
 699                   _inlining_incrementally(false),
 700                   _print_inlining_list(NULL),
 701                   _print_inlining_idx(0),
 702                   _interpreter_frame_size(0),
 703                   _max_node_limit(MaxNodeLimit) {
 704   C = this;
 705 
 706   CompileWrapper cw(this);
 707 #ifndef PRODUCT
 708   if (TimeCompiler2) {
 709     tty->print(" ");
 710     target->holder()->name()->print();
 711     tty->print(".");
 712     target->print_short_name();
 713     tty->print("  ");
 714   }
 715   TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
 716   TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
 717   bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
 718   if (!print_opto_assembly) {
 719     bool print_assembly = (PrintAssembly || _method->should_print_assembly());
 720     if (print_assembly && !Disassembler::can_decode()) {
 721       tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
 722       print_opto_assembly = true;
 723     }
 724   }
 725   set_print_assembly(print_opto_assembly);
 726   set_parsed_irreducible_loop(false);
 727 
 728   if (method()->has_option("ReplayInline")) {
 729     _replay_inline_data = ciReplay::load_inline_data(method(), entry_bci(), ci_env->comp_level());
 730   }
 731 #endif
 732   set_print_inlining(PrintInlining || method()->has_option("PrintInlining") NOT_PRODUCT( || PrintOptoInlining));
 733   set_print_intrinsics(PrintIntrinsics || method()->has_option("PrintIntrinsics"));
 734   set_has_irreducible_loop(true); // conservative until build_loop_tree() reset it
 735 
 736   if (ProfileTraps RTM_OPT_ONLY( || UseRTMLocking )) {
 737     // Make sure the method being compiled gets its own MDO,
 738     // so we can at least track the decompile_count().
 739     // Need MDO to record RTM code generation state.
 740     method()->ensure_method_data();
 741   }
 742 
 743   Init(::AliasLevel);
 744 
 745 
 746   print_compile_messages();
 747 
 748   _ilt = InlineTree::build_inline_tree_root();
 749 
 750   // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
 751   assert(num_alias_types() >= AliasIdxRaw, "");
 752 
 753 #define MINIMUM_NODE_HASH  1023
 754   // Node list that Iterative GVN will start with
 755   Unique_Node_List for_igvn(comp_arena());
 756   set_for_igvn(&for_igvn);
 757 
 758   // GVN that will be run immediately on new nodes
 759   uint estimated_size = method()->code_size()*4+64;
 760   estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
 761   PhaseGVN gvn(node_arena(), estimated_size);
 762   set_initial_gvn(&gvn);
 763 
 764   if (print_inlining() || print_intrinsics()) {
 765     _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
 766   }
 767   { // Scope for timing the parser
 768     TracePhase t3("parse", &_t_parser, true);
 769 
 770     // Put top into the hash table ASAP.
 771     initial_gvn()->transform_no_reclaim(top());
 772 
 773     // Set up tf(), start(), and find a CallGenerator.
 774     CallGenerator* cg = NULL;
 775     if (is_osr_compilation()) {
 776       const TypeTuple *domain = StartOSRNode::osr_domain();
 777       const TypeTuple *range = TypeTuple::make_range(method()->signature());
 778       init_tf(TypeFunc::make(domain, range));
 779       StartNode* s = new (this) StartOSRNode(root(), domain);
 780       initial_gvn()->set_type_bottom(s);
 781       init_start(s);
 782       cg = CallGenerator::for_osr(method(), entry_bci());
 783     } else {
 784       // Normal case.
 785       init_tf(TypeFunc::make(method()));
 786       StartNode* s = new (this) StartNode(root(), tf()->domain());
 787       initial_gvn()->set_type_bottom(s);
 788       init_start(s);
 789       if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && (UseG1GC || UseShenandoahGC)) {
 790         // With java.lang.ref.reference.get() we must go through the
 791         // intrinsic when G1 is enabled - even when get() is the root
 792         // method of the compile - so that, if necessary, the value in
 793         // the referent field of the reference object gets recorded by
 794         // the pre-barrier code.
 795         // Specifically, if G1 is enabled, the value in the referent
 796         // field is recorded by the G1 SATB pre barrier. This will
 797         // result in the referent being marked live and the reference
 798         // object removed from the list of discovered references during
 799         // reference processing.
 800         cg = find_intrinsic(method(), false);
 801       }
 802       if (cg == NULL) {
 803         float past_uses = method()->interpreter_invocation_count();
 804         float expected_uses = past_uses;
 805         cg = CallGenerator::for_inline(method(), expected_uses);
 806       }
 807     }
 808     if (failing())  return;
 809     if (cg == NULL) {
 810       record_method_not_compilable_all_tiers("cannot parse method");
 811       return;
 812     }
 813     JVMState* jvms = build_start_state(start(), tf());
 814     if ((jvms = cg->generate(jvms)) == NULL) {
 815       if (!failure_reason_is(C2Compiler::retry_class_loading_during_parsing())) {
 816         record_method_not_compilable("method parse failed");
 817       }
 818       return;
 819     }
 820     GraphKit kit(jvms);
 821 
 822     if (!kit.stopped()) {
 823       // Accept return values, and transfer control we know not where.
 824       // This is done by a special, unique ReturnNode bound to root.
 825       return_values(kit.jvms());
 826     }
 827 
 828     if (kit.has_exceptions()) {
 829       // Any exceptions that escape from this call must be rethrown
 830       // to whatever caller is dynamically above us on the stack.
 831       // This is done by a special, unique RethrowNode bound to root.
 832       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
 833     }
 834 
 835     assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 836 
 837     if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
 838       inline_string_calls(true);
 839     }
 840 
 841     if (failing())  return;
 842 
 843     print_method(PHASE_BEFORE_REMOVEUSELESS, 3);
 844 
 845     // Remove clutter produced by parsing.
 846     if (!failing()) {
 847       ResourceMark rm;
 848       PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
 849     }
 850   }
 851 
 852   // Note:  Large methods are capped off in do_one_bytecode().
 853   if (failing())  return;
 854 
 855   // After parsing, node notes are no longer automagic.
 856   // They must be propagated by register_new_node_with_optimizer(),
 857   // clone(), or the like.
 858   set_default_node_notes(NULL);
 859 
 860   for (;;) {
 861     int successes = Inline_Warm();
 862     if (failing())  return;
 863     if (successes == 0)  break;
 864   }
 865 
 866   // Drain the list.
 867   Finish_Warm();
 868 #ifndef PRODUCT
 869   if (_printer) {
 870     _printer->print_inlining(this);
 871   }
 872 #endif
 873 
 874   if (failing())  return;
 875   NOT_PRODUCT( verify_graph_edges(); )
 876 
 877   // Now optimize
 878   Optimize();
 879   if (failing())  return;
 880   NOT_PRODUCT( verify_graph_edges(); )
 881 
 882 #ifndef PRODUCT
 883   if (PrintIdeal) {
 884     ttyLocker ttyl;  // keep the following output all in one block
 885     // This output goes directly to the tty, not the compiler log.
 886     // To enable tools to match it up with the compilation activity,
 887     // be sure to tag this tty output with the compile ID.
 888     if (xtty != NULL) {
 889       xtty->head("ideal compile_id='%d'%s", compile_id(),
 890                  is_osr_compilation()    ? " compile_kind='osr'" :
 891                  "");
 892     }
 893     root()->dump(9999);
 894     if (xtty != NULL) {
 895       xtty->tail("ideal");
 896     }
 897   }
 898 #endif
 899 
 900   NOT_PRODUCT( verify_barriers(); )
 901 
 902   // Dump compilation data to replay it.
 903   if (method()->has_option("DumpReplay")) {
 904     env()->dump_replay_data(_compile_id);
 905   }
 906   if (method()->has_option("DumpInline") && (ilt() != NULL)) {
 907     env()->dump_inline_data(_compile_id);
 908   }
 909 
 910   // Now that we know the size of all the monitors we can add a fixed slot
 911   // for the original deopt pc.
 912 
 913   _orig_pc_slot =  fixed_slots();
 914   int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
 915   set_fixed_slots(next_slot);
 916 
 917   // Compute when to use implicit null checks. Used by matching trap based
 918   // nodes and NullCheck optimization.
 919   set_allowed_deopt_reasons();
 920 
 921   // Now generate code
 922   Code_Gen();
 923   if (failing())  return;
 924 
 925   // Check if we want to skip execution of all compiled code.
 926   {
 927 #ifndef PRODUCT
 928     if (OptoNoExecute) {
 929       record_method_not_compilable("+OptoNoExecute");  // Flag as failed
 930       return;
 931     }
 932     TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
 933 #endif
 934 
 935     if (is_osr_compilation()) {
 936       _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
 937       _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
 938     } else {
 939       _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
 940       _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
 941     }
 942 
 943     env()->register_method(_method, _entry_bci,
 944                            &_code_offsets,
 945                            _orig_pc_slot_offset_in_bytes,
 946                            code_buffer(),
 947                            frame_size_in_words(), _oop_map_set,
 948                            &_handler_table, &_inc_table,
 949                            compiler,
 950                            env()->comp_level(),
 951                            has_unsafe_access(),
 952                            SharedRuntime::is_wide_vector(max_vector_size()),
 953                            rtm_state()
 954                            );
 955 
 956     if (log() != NULL) // Print code cache state into compiler log
 957       log()->code_cache_state();
 958   }
 959 }
 960 
 961 //------------------------------Compile----------------------------------------
 962 // Compile a runtime stub
 963 Compile::Compile( ciEnv* ci_env,
 964                   TypeFunc_generator generator,
 965                   address stub_function,
 966                   const char *stub_name,
 967                   int is_fancy_jump,
 968                   bool pass_tls,
 969                   bool save_arg_registers,
 970                   bool return_pc )
 971   : Phase(Compiler),
 972     _env(ci_env),
 973     _log(ci_env->log()),
 974     _compile_id(0),
 975     _save_argument_registers(save_arg_registers),
 976     _method(NULL),
 977     _stub_name(stub_name),
 978     _stub_function(stub_function),
 979     _stub_entry_point(NULL),
 980     _entry_bci(InvocationEntryBci),
 981     _initial_gvn(NULL),
 982     _for_igvn(NULL),
 983     _warm_calls(NULL),
 984     _orig_pc_slot(0),
 985     _orig_pc_slot_offset_in_bytes(0),
 986     _subsume_loads(true),
 987     _do_escape_analysis(false),
 988     _eliminate_boxing(false),
 989     _failure_reason(NULL),
 990     _code_buffer("Compile::Fill_buffer"),
 991     _has_method_handle_invokes(false),
 992     _mach_constant_base_node(NULL),
 993     _node_bundling_limit(0),
 994     _node_bundling_base(NULL),
 995     _java_calls(0),
 996     _inner_loops(0),
 997 #ifndef PRODUCT
 998     _trace_opto_output(TraceOptoOutput),
 999     _in_dump_cnt(0),
1000     _printer(NULL),
1001 #endif
1002     _comp_arena(mtCompiler),
1003     _node_arena(mtCompiler),
1004     _old_arena(mtCompiler),
1005     _Compile_types(mtCompiler),
1006     _dead_node_list(comp_arena()),
1007     _dead_node_count(0),
1008     _congraph(NULL),
1009     _replay_inline_data(NULL),
1010     _number_of_mh_late_inlines(0),
1011     _inlining_progress(false),
1012     _inlining_incrementally(false),
1013     _print_inlining_list(NULL),
1014     _print_inlining_idx(0),
1015     _allowed_reasons(0),
1016     _interpreter_frame_size(0),
1017     _max_node_limit(MaxNodeLimit) {
1018   C = this;
1019 
1020 #ifndef PRODUCT
1021   TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
1022   TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
1023   set_print_assembly(PrintFrameConverterAssembly);
1024   set_parsed_irreducible_loop(false);
1025 #endif
1026   set_has_irreducible_loop(false); // no loops
1027 
1028   CompileWrapper cw(this);
1029   Init(/*AliasLevel=*/ 0);
1030   init_tf((*generator)());
1031 
1032   {
1033     // The following is a dummy for the sake of GraphKit::gen_stub
1034     Unique_Node_List for_igvn(comp_arena());
1035     set_for_igvn(&for_igvn);  // not used, but some GraphKit guys push on this
1036     PhaseGVN gvn(Thread::current()->resource_area(),255);
1037     set_initial_gvn(&gvn);    // not significant, but GraphKit guys use it pervasively
1038     gvn.transform_no_reclaim(top());
1039 
1040     GraphKit kit;
1041     kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
1042   }
1043 
1044   NOT_PRODUCT( verify_graph_edges(); )
1045   Code_Gen();
1046   if (failing())  return;
1047 
1048 
1049   // Entry point will be accessed using compile->stub_entry_point();
1050   if (code_buffer() == NULL) {
1051     Matcher::soft_match_failure();
1052   } else {
1053     if (PrintAssembly && (WizardMode || Verbose))
1054       tty->print_cr("### Stub::%s", stub_name);
1055 
1056     if (!failing()) {
1057       assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
1058 
1059       // Make the NMethod
1060       // For now we mark the frame as never safe for profile stackwalking
1061       RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
1062                                                       code_buffer(),
1063                                                       CodeOffsets::frame_never_safe,
1064                                                       // _code_offsets.value(CodeOffsets::Frame_Complete),
1065                                                       frame_size_in_words(),
1066                                                       _oop_map_set,
1067                                                       save_arg_registers);
1068       assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
1069 
1070       _stub_entry_point = rs->entry_point();
1071     }
1072   }
1073 }
1074 
1075 //------------------------------Init-------------------------------------------
1076 // Prepare for a single compilation
1077 void Compile::Init(int aliaslevel) {
1078   _unique  = 0;
1079   _regalloc = NULL;
1080 
1081   _tf      = NULL;  // filled in later
1082   _top     = NULL;  // cached later
1083   _matcher = NULL;  // filled in later
1084   _cfg     = NULL;  // filled in later
1085 
1086   set_24_bit_selection_and_mode(Use24BitFP, false);
1087 
1088   _node_note_array = NULL;
1089   _default_node_notes = NULL;
1090 
1091   _immutable_memory = NULL; // filled in at first inquiry
1092 
1093   // Globally visible Nodes
1094   // First set TOP to NULL to give safe behavior during creation of RootNode
1095   set_cached_top_node(NULL);
1096   set_root(new (this) RootNode());
1097   // Now that you have a Root to point to, create the real TOP
1098   set_cached_top_node( new (this) ConNode(Type::TOP) );
1099   set_recent_alloc(NULL, NULL);
1100 
1101   // Create Debug Information Recorder to record scopes, oopmaps, etc.
1102   env()->set_oop_recorder(new OopRecorder(env()->arena()));
1103   env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
1104   env()->set_dependencies(new Dependencies(env()));
1105 
1106   _fixed_slots = 0;
1107   set_has_split_ifs(false);
1108   set_has_loops(has_method() && method()->has_loops()); // first approximation
1109   set_has_stringbuilder(false);
1110   set_has_boxed_value(false);
1111   _trap_can_recompile = false;  // no traps emitted yet
1112   _major_progress = true; // start out assuming good things will happen
1113   set_has_unsafe_access(false);
1114   set_max_vector_size(0);
1115   Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
1116   set_decompile_count(0);
1117 
1118   set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
1119   set_num_loop_opts(LoopOptsCount);
1120   set_do_inlining(Inline);
1121   set_max_inline_size(MaxInlineSize);
1122   set_freq_inline_size(FreqInlineSize);
1123   set_do_scheduling(OptoScheduling);
1124   set_do_count_invocations(false);
1125   set_do_method_data_update(false);
1126   set_rtm_state(NoRTM); // No RTM lock eliding by default
1127   method_has_option_value("MaxNodeLimit", _max_node_limit);
1128 #if INCLUDE_RTM_OPT
1129   if (UseRTMLocking && has_method() && (method()->method_data_or_null() != NULL)) {
1130     int rtm_state = method()->method_data()->rtm_state();
1131     if (method_has_option("NoRTMLockEliding") || ((rtm_state & NoRTM) != 0)) {
1132       // Don't generate RTM lock eliding code.
1133       set_rtm_state(NoRTM);
1134     } else if (method_has_option("UseRTMLockEliding") || ((rtm_state & UseRTM) != 0) || !UseRTMDeopt) {
1135       // Generate RTM lock eliding code without abort ratio calculation code.
1136       set_rtm_state(UseRTM);
1137     } else if (UseRTMDeopt) {
1138       // Generate RTM lock eliding code and include abort ratio calculation
1139       // code if UseRTMDeopt is on.
1140       set_rtm_state(ProfileRTM);
1141     }
1142   }
1143 #endif
1144   if (debug_info()->recording_non_safepoints()) {
1145     set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
1146                         (comp_arena(), 8, 0, NULL));
1147     set_default_node_notes(Node_Notes::make(this));
1148   }
1149 
1150   // // -- Initialize types before each compile --
1151   // // Update cached type information
1152   // if( _method && _method->constants() )
1153   //   Type::update_loaded_types(_method, _method->constants());
1154 
1155   // Init alias_type map.
1156   if (!_do_escape_analysis && aliaslevel == 3)
1157     aliaslevel = 2;  // No unique types without escape analysis
1158   _AliasLevel = aliaslevel;
1159   const int grow_ats = 16;
1160   _max_alias_types = grow_ats;
1161   _alias_types   = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
1162   AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType,  grow_ats);
1163   Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
1164   {
1165     for (int i = 0; i < grow_ats; i++)  _alias_types[i] = &ats[i];
1166   }
1167   // Initialize the first few types.
1168   _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
1169   _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
1170   _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
1171   _num_alias_types = AliasIdxRaw+1;
1172   // Zero out the alias type cache.
1173   Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
1174   // A NULL adr_type hits in the cache right away.  Preload the right answer.
1175   probe_alias_cache(NULL)->_index = AliasIdxTop;
1176 
1177   _intrinsics = NULL;
1178   _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1179   _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1180   _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1181   _range_check_casts = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8,  0, NULL);
1182   _shenandoah_barriers = new(comp_arena()) GrowableArray<ShenandoahLoadReferenceBarrierNode*>(comp_arena(), 8,  0, NULL);
1183   register_library_intrinsics();
1184 }
1185 
1186 //---------------------------init_start----------------------------------------
1187 // Install the StartNode on this compile object.
1188 void Compile::init_start(StartNode* s) {
1189   if (failing())
1190     return; // already failing
1191   assert(s == start(), "");
1192 }
1193 
1194 StartNode* Compile::start() const {
1195   assert(!failing(), "");
1196   for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
1197     Node* start = root()->fast_out(i);
1198     if( start->is_Start() )
1199       return start->as_Start();
1200   }
1201   fatal("Did not find Start node!");
1202   return NULL;
1203 }
1204 
1205 //-------------------------------immutable_memory-------------------------------------
1206 // Access immutable memory
1207 Node* Compile::immutable_memory() {
1208   if (_immutable_memory != NULL) {
1209     return _immutable_memory;
1210   }
1211   StartNode* s = start();
1212   for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
1213     Node *p = s->fast_out(i);
1214     if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
1215       _immutable_memory = p;
1216       return _immutable_memory;
1217     }
1218   }
1219   ShouldNotReachHere();
1220   return NULL;
1221 }
1222 
1223 //----------------------set_cached_top_node------------------------------------
1224 // Install the cached top node, and make sure Node::is_top works correctly.
1225 void Compile::set_cached_top_node(Node* tn) {
1226   if (tn != NULL)  verify_top(tn);
1227   Node* old_top = _top;
1228   _top = tn;
1229   // Calling Node::setup_is_top allows the nodes the chance to adjust
1230   // their _out arrays.
1231   if (_top != NULL)     _top->setup_is_top();
1232   if (old_top != NULL)  old_top->setup_is_top();
1233   assert(_top == NULL || top()->is_top(), "");
1234 }
1235 
1236 #ifdef ASSERT
1237 uint Compile::count_live_nodes_by_graph_walk() {
1238   Unique_Node_List useful(comp_arena());
1239   // Get useful node list by walking the graph.
1240   identify_useful_nodes(useful);
1241   return useful.size();
1242 }
1243 
1244 void Compile::print_missing_nodes() {
1245 
1246   // Return if CompileLog is NULL and PrintIdealNodeCount is false.
1247   if ((_log == NULL) && (! PrintIdealNodeCount)) {
1248     return;
1249   }
1250 
1251   // This is an expensive function. It is executed only when the user
1252   // specifies VerifyIdealNodeCount option or otherwise knows the
1253   // additional work that needs to be done to identify reachable nodes
1254   // by walking the flow graph and find the missing ones using
1255   // _dead_node_list.
1256 
1257   Unique_Node_List useful(comp_arena());
1258   // Get useful node list by walking the graph.
1259   identify_useful_nodes(useful);
1260 
1261   uint l_nodes = C->live_nodes();
1262   uint l_nodes_by_walk = useful.size();
1263 
1264   if (l_nodes != l_nodes_by_walk) {
1265     if (_log != NULL) {
1266       _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
1267       _log->stamp();
1268       _log->end_head();
1269     }
1270     VectorSet& useful_member_set = useful.member_set();
1271     int last_idx = l_nodes_by_walk;
1272     for (int i = 0; i < last_idx; i++) {
1273       if (useful_member_set.test(i)) {
1274         if (_dead_node_list.test(i)) {
1275           if (_log != NULL) {
1276             _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
1277           }
1278           if (PrintIdealNodeCount) {
1279             // Print the log message to tty
1280               tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
1281               useful.at(i)->dump();
1282           }
1283         }
1284       }
1285       else if (! _dead_node_list.test(i)) {
1286         if (_log != NULL) {
1287           _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
1288         }
1289         if (PrintIdealNodeCount) {
1290           // Print the log message to tty
1291           tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
1292         }
1293       }
1294     }
1295     if (_log != NULL) {
1296       _log->tail("mismatched_nodes");
1297     }
1298   }
1299 }
1300 #endif
1301 
1302 #ifndef PRODUCT
1303 void Compile::verify_top(Node* tn) const {
1304   if (tn != NULL) {
1305     assert(tn->is_Con(), "top node must be a constant");
1306     assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
1307     assert(tn->in(0) != NULL, "must have live top node");
1308   }
1309 }
1310 #endif
1311 
1312 
1313 ///-------------------Managing Per-Node Debug & Profile Info-------------------
1314 
1315 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
1316   guarantee(arr != NULL, "");
1317   int num_blocks = arr->length();
1318   if (grow_by < num_blocks)  grow_by = num_blocks;
1319   int num_notes = grow_by * _node_notes_block_size;
1320   Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
1321   Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
1322   while (num_notes > 0) {
1323     arr->append(notes);
1324     notes     += _node_notes_block_size;
1325     num_notes -= _node_notes_block_size;
1326   }
1327   assert(num_notes == 0, "exact multiple, please");
1328 }
1329 
1330 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
1331   if (source == NULL || dest == NULL)  return false;
1332 
1333   if (dest->is_Con())
1334     return false;               // Do not push debug info onto constants.
1335 
1336 #ifdef ASSERT
1337   // Leave a bread crumb trail pointing to the original node:
1338   if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
1339     dest->set_debug_orig(source);
1340   }
1341 #endif
1342 
1343   if (node_note_array() == NULL)
1344     return false;               // Not collecting any notes now.
1345 
1346   // This is a copy onto a pre-existing node, which may already have notes.
1347   // If both nodes have notes, do not overwrite any pre-existing notes.
1348   Node_Notes* source_notes = node_notes_at(source->_idx);
1349   if (source_notes == NULL || source_notes->is_clear())  return false;
1350   Node_Notes* dest_notes   = node_notes_at(dest->_idx);
1351   if (dest_notes == NULL || dest_notes->is_clear()) {
1352     return set_node_notes_at(dest->_idx, source_notes);
1353   }
1354 
1355   Node_Notes merged_notes = (*source_notes);
1356   // The order of operations here ensures that dest notes will win...
1357   merged_notes.update_from(dest_notes);
1358   return set_node_notes_at(dest->_idx, &merged_notes);
1359 }
1360 
1361 
1362 //--------------------------allow_range_check_smearing-------------------------
1363 // Gating condition for coalescing similar range checks.
1364 // Sometimes we try 'speculatively' replacing a series of a range checks by a
1365 // single covering check that is at least as strong as any of them.
1366 // If the optimization succeeds, the simplified (strengthened) range check
1367 // will always succeed.  If it fails, we will deopt, and then give up
1368 // on the optimization.
1369 bool Compile::allow_range_check_smearing() const {
1370   // If this method has already thrown a range-check,
1371   // assume it was because we already tried range smearing
1372   // and it failed.
1373   uint already_trapped = trap_count(Deoptimization::Reason_range_check);
1374   return !already_trapped;
1375 }
1376 
1377 
1378 //------------------------------flatten_alias_type-----------------------------
1379 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
1380   int offset = tj->offset();
1381   TypePtr::PTR ptr = tj->ptr();
1382 
1383   // Known instance (scalarizable allocation) alias only with itself.
1384   bool is_known_inst = tj->isa_oopptr() != NULL &&
1385                        tj->is_oopptr()->is_known_instance();
1386 
1387   // Process weird unsafe references.
1388   if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
1389     assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
1390     assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
1391     tj = TypeOopPtr::BOTTOM;
1392     ptr = tj->ptr();
1393     offset = tj->offset();
1394   }
1395 
1396   // Array pointers need some flattening
1397   const TypeAryPtr *ta = tj->isa_aryptr();
1398   if (ta && ta->is_stable()) {
1399     // Erase stability property for alias analysis.
1400     tj = ta = ta->cast_to_stable(false);
1401   }
1402   if( ta && is_known_inst ) {
1403     if ( offset != Type::OffsetBot &&
1404          offset > arrayOopDesc::length_offset_in_bytes() ) {
1405       offset = Type::OffsetBot; // Flatten constant access into array body only
1406       tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
1407     }
1408   } else if( ta && _AliasLevel >= 2 ) {
1409     // For arrays indexed by constant indices, we flatten the alias
1410     // space to include all of the array body.  Only the header, klass
1411     // and array length can be accessed un-aliased.
1412     if( offset != Type::OffsetBot ) {
1413       if( ta->const_oop() ) { // MethodData* or Method*
1414         offset = Type::OffsetBot;   // Flatten constant access into array body
1415         tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
1416       } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
1417         // range is OK as-is.
1418         tj = ta = TypeAryPtr::RANGE;
1419       } else if( offset == oopDesc::klass_offset_in_bytes() ) {
1420         tj = TypeInstPtr::KLASS; // all klass loads look alike
1421         ta = TypeAryPtr::RANGE; // generic ignored junk
1422         ptr = TypePtr::BotPTR;
1423       } else if( offset == oopDesc::mark_offset_in_bytes() ) {
1424         tj = TypeInstPtr::MARK;
1425         ta = TypeAryPtr::RANGE; // generic ignored junk
1426         ptr = TypePtr::BotPTR;
1427       } else {                  // Random constant offset into array body
1428         offset = Type::OffsetBot;   // Flatten constant access into array body
1429         tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
1430       }
1431     }
1432     // Arrays of fixed size alias with arrays of unknown size.
1433     if (ta->size() != TypeInt::POS) {
1434       const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
1435       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
1436     }
1437     // Arrays of known objects become arrays of unknown objects.
1438     if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
1439       const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
1440       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1441     }
1442     if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
1443       const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
1444       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
1445     }
1446     // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
1447     // cannot be distinguished by bytecode alone.
1448     if (ta->elem() == TypeInt::BOOL) {
1449       const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
1450       ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
1451       tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
1452     }
1453     // During the 2nd round of IterGVN, NotNull castings are removed.
1454     // Make sure the Bottom and NotNull variants alias the same.
1455     // Also, make sure exact and non-exact variants alias the same.
1456     if (ptr == TypePtr::NotNull || ta->klass_is_exact() || ta->speculative() != NULL) {
1457       tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
1458     }
1459   }
1460 
1461   // Oop pointers need some flattening
1462   const TypeInstPtr *to = tj->isa_instptr();
1463   if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
1464     ciInstanceKlass *k = to->klass()->as_instance_klass();
1465     if( ptr == TypePtr::Constant ) {
1466       if (to->klass() != ciEnv::current()->Class_klass() ||
1467           offset < k->size_helper() * wordSize) {
1468         // No constant oop pointers (such as Strings); they alias with
1469         // unknown strings.
1470         assert(!is_known_inst, "not scalarizable allocation");
1471         tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1472       }
1473     } else if( is_known_inst ) {
1474       tj = to; // Keep NotNull and klass_is_exact for instance type
1475     } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
1476       // During the 2nd round of IterGVN, NotNull castings are removed.
1477       // Make sure the Bottom and NotNull variants alias the same.
1478       // Also, make sure exact and non-exact variants alias the same.
1479       tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
1480     }
1481     if (to->speculative() != NULL) {
1482       tj = to = TypeInstPtr::make(to->ptr(),to->klass(),to->klass_is_exact(),to->const_oop(),to->offset(), to->instance_id());
1483     }
1484     // Canonicalize the holder of this field
1485     if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
1486       // First handle header references such as a LoadKlassNode, even if the
1487       // object's klass is unloaded at compile time (4965979).
1488       if (!is_known_inst) { // Do it only for non-instance types
1489         tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
1490       }
1491     } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
1492       // Static fields are in the space above the normal instance
1493       // fields in the java.lang.Class instance.
1494       if (to->klass() != ciEnv::current()->Class_klass()) {
1495         to = NULL;
1496         tj = TypeOopPtr::BOTTOM;
1497         offset = tj->offset();
1498       }
1499     } else {
1500       ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
1501       if (!k->equals(canonical_holder) || tj->offset() != offset) {
1502         if( is_known_inst ) {
1503           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
1504         } else {
1505           tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
1506         }
1507       }
1508     }
1509   }
1510 
1511   // Klass pointers to object array klasses need some flattening
1512   const TypeKlassPtr *tk = tj->isa_klassptr();
1513   if( tk ) {
1514     // If we are referencing a field within a Klass, we need
1515     // to assume the worst case of an Object.  Both exact and
1516     // inexact types must flatten to the same alias class so
1517     // use NotNull as the PTR.
1518     if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
1519 
1520       tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
1521                                    TypeKlassPtr::OBJECT->klass(),
1522                                    offset);
1523     }
1524 
1525     ciKlass* klass = tk->klass();
1526     if( klass->is_obj_array_klass() ) {
1527       ciKlass* k = TypeAryPtr::OOPS->klass();
1528       if( !k || !k->is_loaded() )                  // Only fails for some -Xcomp runs
1529         k = TypeInstPtr::BOTTOM->klass();
1530       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
1531     }
1532 
1533     // Check for precise loads from the primary supertype array and force them
1534     // to the supertype cache alias index.  Check for generic array loads from
1535     // the primary supertype array and also force them to the supertype cache
1536     // alias index.  Since the same load can reach both, we need to merge
1537     // these 2 disparate memories into the same alias class.  Since the
1538     // primary supertype array is read-only, there's no chance of confusion
1539     // where we bypass an array load and an array store.
1540     int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
1541     if (offset == Type::OffsetBot ||
1542         (offset >= primary_supers_offset &&
1543          offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
1544         offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
1545       offset = in_bytes(Klass::secondary_super_cache_offset());
1546       tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
1547     }
1548   }
1549 
1550   // Flatten all Raw pointers together.
1551   if (tj->base() == Type::RawPtr)
1552     tj = TypeRawPtr::BOTTOM;
1553 
1554   if (tj->base() == Type::AnyPtr)
1555     tj = TypePtr::BOTTOM;      // An error, which the caller must check for.
1556 
1557   // Flatten all to bottom for now
1558   switch( _AliasLevel ) {
1559   case 0:
1560     tj = TypePtr::BOTTOM;
1561     break;
1562   case 1:                       // Flatten to: oop, static, field or array
1563     switch (tj->base()) {
1564     //case Type::AryPtr: tj = TypeAryPtr::RANGE;    break;
1565     case Type::RawPtr:   tj = TypeRawPtr::BOTTOM;   break;
1566     case Type::AryPtr:   // do not distinguish arrays at all
1567     case Type::InstPtr:  tj = TypeInstPtr::BOTTOM;  break;
1568     case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
1569     case Type::AnyPtr:   tj = TypePtr::BOTTOM;      break;  // caller checks it
1570     default: ShouldNotReachHere();
1571     }
1572     break;
1573   case 2:                       // No collapsing at level 2; keep all splits
1574   case 3:                       // No collapsing at level 3; keep all splits
1575     break;
1576   default:
1577     Unimplemented();
1578   }
1579 
1580   offset = tj->offset();
1581   assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
1582 
1583   assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
1584           (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
1585           (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
1586           (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
1587           (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1588           (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
1589           (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr)  ,
1590           "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
1591   assert( tj->ptr() != TypePtr::TopPTR &&
1592           tj->ptr() != TypePtr::AnyNull &&
1593           tj->ptr() != TypePtr::Null, "No imprecise addresses" );
1594 //    assert( tj->ptr() != TypePtr::Constant ||
1595 //            tj->base() == Type::RawPtr ||
1596 //            tj->base() == Type::KlassPtr, "No constant oop addresses" );
1597 
1598   return tj;
1599 }
1600 
1601 void Compile::AliasType::Init(int i, const TypePtr* at) {
1602   _index = i;
1603   _adr_type = at;
1604   _field = NULL;
1605   _element = NULL;
1606   _is_rewritable = true; // default
1607   const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
1608   if (atoop != NULL && atoop->is_known_instance()) {
1609     const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
1610     _general_index = Compile::current()->get_alias_index(gt);
1611   } else {
1612     _general_index = 0;
1613   }
1614 }
1615 
1616 BasicType Compile::AliasType::basic_type() const {
1617   if (element() != NULL) {
1618     const Type* element = adr_type()->is_aryptr()->elem();
1619     return element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
1620   } if (field() != NULL) {
1621     return field()->layout_type();
1622   } else {
1623     return T_ILLEGAL; // unknown
1624   }
1625 }
1626 
1627 //---------------------------------print_on------------------------------------
1628 #ifndef PRODUCT
1629 void Compile::AliasType::print_on(outputStream* st) {
1630   if (index() < 10)
1631         st->print("@ <%d> ", index());
1632   else  st->print("@ <%d>",  index());
1633   st->print(is_rewritable() ? "   " : " RO");
1634   int offset = adr_type()->offset();
1635   if (offset == Type::OffsetBot)
1636         st->print(" +any");
1637   else  st->print(" +%-3d", offset);
1638   st->print(" in ");
1639   adr_type()->dump_on(st);
1640   const TypeOopPtr* tjp = adr_type()->isa_oopptr();
1641   if (field() != NULL && tjp) {
1642     if (tjp->klass()  != field()->holder() ||
1643         tjp->offset() != field()->offset_in_bytes()) {
1644       st->print(" != ");
1645       field()->print();
1646       st->print(" ***");
1647     }
1648   }
1649 }
1650 
1651 void print_alias_types() {
1652   Compile* C = Compile::current();
1653   tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
1654   for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
1655     C->alias_type(idx)->print_on(tty);
1656     tty->cr();
1657   }
1658 }
1659 #endif
1660 
1661 
1662 //----------------------------probe_alias_cache--------------------------------
1663 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
1664   intptr_t key = (intptr_t) adr_type;
1665   key ^= key >> logAliasCacheSize;
1666   return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
1667 }
1668 
1669 
1670 //-----------------------------grow_alias_types--------------------------------
1671 void Compile::grow_alias_types() {
1672   const int old_ats  = _max_alias_types; // how many before?
1673   const int new_ats  = old_ats;          // how many more?
1674   const int grow_ats = old_ats+new_ats;  // how many now?
1675   _max_alias_types = grow_ats;
1676   _alias_types =  REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
1677   AliasType* ats =    NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
1678   Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
1679   for (int i = 0; i < new_ats; i++)  _alias_types[old_ats+i] = &ats[i];
1680 }
1681 
1682 
1683 //--------------------------------find_alias_type------------------------------
1684 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
1685   if (_AliasLevel == 0)
1686     return alias_type(AliasIdxBot);
1687 
1688   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1689   if (ace->_adr_type == adr_type) {
1690     return alias_type(ace->_index);
1691   }
1692 
1693   // Handle special cases.
1694   if (adr_type == NULL)             return alias_type(AliasIdxTop);
1695   if (adr_type == TypePtr::BOTTOM)  return alias_type(AliasIdxBot);
1696 
1697   // Do it the slow way.
1698   const TypePtr* flat = flatten_alias_type(adr_type);
1699 
1700 #ifdef ASSERT
1701   {
1702     ResourceMark rm;
1703     assert(flat == flatten_alias_type(flat),
1704            err_msg("not idempotent: adr_type = %s; flat = %s => %s", Type::str(adr_type),
1705                    Type::str(flat), Type::str(flatten_alias_type(flat))));
1706     assert(flat != TypePtr::BOTTOM,
1707            err_msg("cannot alias-analyze an untyped ptr: adr_type = %s", Type::str(adr_type)));
1708     if (flat->isa_oopptr() && !flat->isa_klassptr()) {
1709       const TypeOopPtr* foop = flat->is_oopptr();
1710       // Scalarizable allocations have exact klass always.
1711       bool exact = !foop->klass_is_exact() || foop->is_known_instance();
1712       const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
1713       assert(foop == flatten_alias_type(xoop),
1714              err_msg("exactness must not affect alias type: foop = %s; xoop = %s",
1715                      Type::str(foop), Type::str(xoop)));
1716     }
1717   }
1718 #endif
1719 
1720   int idx = AliasIdxTop;
1721   for (int i = 0; i < num_alias_types(); i++) {
1722     if (alias_type(i)->adr_type() == flat) {
1723       idx = i;
1724       break;
1725     }
1726   }
1727 
1728   if (idx == AliasIdxTop) {
1729     if (no_create)  return NULL;
1730     // Grow the array if necessary.
1731     if (_num_alias_types == _max_alias_types)  grow_alias_types();
1732     // Add a new alias type.
1733     idx = _num_alias_types++;
1734     _alias_types[idx]->Init(idx, flat);
1735     if (flat == TypeInstPtr::KLASS)  alias_type(idx)->set_rewritable(false);
1736     if (flat == TypeAryPtr::RANGE)   alias_type(idx)->set_rewritable(false);
1737     if (flat->isa_instptr()) {
1738       if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
1739           && flat->is_instptr()->klass() == env()->Class_klass())
1740         alias_type(idx)->set_rewritable(false);
1741     }
1742     if (flat->isa_aryptr()) {
1743 #ifdef ASSERT
1744       const int header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1745       // (T_BYTE has the weakest alignment and size restrictions...)
1746       assert(flat->offset() < header_size_min, "array body reference must be OffsetBot");
1747 #endif
1748       if (flat->offset() == TypePtr::OffsetBot) {
1749         alias_type(idx)->set_element(flat->is_aryptr()->elem());
1750       }
1751     }
1752     if (flat->isa_klassptr()) {
1753       if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
1754         alias_type(idx)->set_rewritable(false);
1755       if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
1756         alias_type(idx)->set_rewritable(false);
1757       if (flat->offset() == in_bytes(Klass::access_flags_offset()))
1758         alias_type(idx)->set_rewritable(false);
1759       if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
1760         alias_type(idx)->set_rewritable(false);
1761     }
1762     // %%% (We would like to finalize JavaThread::threadObj_offset(),
1763     // but the base pointer type is not distinctive enough to identify
1764     // references into JavaThread.)
1765 
1766     // Check for final fields.
1767     const TypeInstPtr* tinst = flat->isa_instptr();
1768     if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
1769       ciField* field;
1770       if (tinst->const_oop() != NULL &&
1771           tinst->klass() == ciEnv::current()->Class_klass() &&
1772           tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
1773         // static field
1774         ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1775         field = k->get_field_by_offset(tinst->offset(), true);
1776       } else {
1777         ciInstanceKlass *k = tinst->klass()->as_instance_klass();
1778         field = k->get_field_by_offset(tinst->offset(), false);
1779       }
1780       assert(field == NULL ||
1781              original_field == NULL ||
1782              (field->holder() == original_field->holder() &&
1783               field->offset() == original_field->offset() &&
1784               field->is_static() == original_field->is_static()), "wrong field?");
1785       // Set field() and is_rewritable() attributes.
1786       if (field != NULL)  alias_type(idx)->set_field(field);
1787     }
1788   }
1789 
1790   // Fill the cache for next time.
1791   ace->_adr_type = adr_type;
1792   ace->_index    = idx;
1793   assert(alias_type(adr_type) == alias_type(idx),  "type must be installed");
1794 
1795   // Might as well try to fill the cache for the flattened version, too.
1796   AliasCacheEntry* face = probe_alias_cache(flat);
1797   if (face->_adr_type == NULL) {
1798     face->_adr_type = flat;
1799     face->_index    = idx;
1800     assert(alias_type(flat) == alias_type(idx), "flat type must work too");
1801   }
1802 
1803   return alias_type(idx);
1804 }
1805 
1806 
1807 Compile::AliasType* Compile::alias_type(ciField* field) {
1808   const TypeOopPtr* t;
1809   if (field->is_static())
1810     t = TypeInstPtr::make(field->holder()->java_mirror());
1811   else
1812     t = TypeOopPtr::make_from_klass_raw(field->holder());
1813   AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
1814   assert((field->is_final() || field->is_stable()) == !atp->is_rewritable(), "must get the rewritable bits correct");
1815   return atp;
1816 }
1817 
1818 
1819 //------------------------------have_alias_type--------------------------------
1820 bool Compile::have_alias_type(const TypePtr* adr_type) {
1821   AliasCacheEntry* ace = probe_alias_cache(adr_type);
1822   if (ace->_adr_type == adr_type) {
1823     return true;
1824   }
1825 
1826   // Handle special cases.
1827   if (adr_type == NULL)             return true;
1828   if (adr_type == TypePtr::BOTTOM)  return true;
1829 
1830   return find_alias_type(adr_type, true, NULL) != NULL;
1831 }
1832 
1833 //-----------------------------must_alias--------------------------------------
1834 // True if all values of the given address type are in the given alias category.
1835 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
1836   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1837   if (adr_type == NULL)                 return true;  // NULL serves as TypePtr::TOP
1838   if (alias_idx == AliasIdxTop)         return false; // the empty category
1839   if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
1840 
1841   // the only remaining possible overlap is identity
1842   int adr_idx = get_alias_index(adr_type);
1843   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1844   assert(adr_idx == alias_idx ||
1845          (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
1846           && adr_type                       != TypeOopPtr::BOTTOM),
1847          "should not be testing for overlap with an unsafe pointer");
1848   return adr_idx == alias_idx;
1849 }
1850 
1851 //------------------------------can_alias--------------------------------------
1852 // True if any values of the given address type are in the given alias category.
1853 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
1854   if (alias_idx == AliasIdxTop)         return false; // the empty category
1855   if (adr_type == NULL)                 return false; // NULL serves as TypePtr::TOP
1856   if (alias_idx == AliasIdxBot)         return true;  // the universal category
1857   if (adr_type->base() == Type::AnyPtr) return true;  // TypePtr::BOTTOM or its twins
1858 
1859   // the only remaining possible overlap is identity
1860   int adr_idx = get_alias_index(adr_type);
1861   assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
1862   return adr_idx == alias_idx;
1863 }
1864 
1865 
1866 
1867 //---------------------------pop_warm_call-------------------------------------
1868 WarmCallInfo* Compile::pop_warm_call() {
1869   WarmCallInfo* wci = _warm_calls;
1870   if (wci != NULL)  _warm_calls = wci->remove_from(wci);
1871   return wci;
1872 }
1873 
1874 //----------------------------Inline_Warm--------------------------------------
1875 int Compile::Inline_Warm() {
1876   // If there is room, try to inline some more warm call sites.
1877   // %%% Do a graph index compaction pass when we think we're out of space?
1878   if (!InlineWarmCalls)  return 0;
1879 
1880   int calls_made_hot = 0;
1881   int room_to_grow   = NodeCountInliningCutoff - unique();
1882   int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
1883   int amount_grown   = 0;
1884   WarmCallInfo* call;
1885   while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
1886     int est_size = (int)call->size();
1887     if (est_size > (room_to_grow - amount_grown)) {
1888       // This one won't fit anyway.  Get rid of it.
1889       call->make_cold();
1890       continue;
1891     }
1892     call->make_hot();
1893     calls_made_hot++;
1894     amount_grown   += est_size;
1895     amount_to_grow -= est_size;
1896   }
1897 
1898   if (calls_made_hot > 0)  set_major_progress();
1899   return calls_made_hot;
1900 }
1901 
1902 
1903 //----------------------------Finish_Warm--------------------------------------
1904 void Compile::Finish_Warm() {
1905   if (!InlineWarmCalls)  return;
1906   if (failing())  return;
1907   if (warm_calls() == NULL)  return;
1908 
1909   // Clean up loose ends, if we are out of space for inlining.
1910   WarmCallInfo* call;
1911   while ((call = pop_warm_call()) != NULL) {
1912     call->make_cold();
1913   }
1914 }
1915 
1916 //---------------------cleanup_loop_predicates-----------------------
1917 // Remove the opaque nodes that protect the predicates so that all unused
1918 // checks and uncommon_traps will be eliminated from the ideal graph
1919 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
1920   if (predicate_count()==0) return;
1921   for (int i = predicate_count(); i > 0; i--) {
1922     Node * n = predicate_opaque1_node(i-1);
1923     assert(n->Opcode() == Op_Opaque1, "must be");
1924     igvn.replace_node(n, n->in(1));
1925   }
1926   assert(predicate_count()==0, "should be clean!");
1927 }
1928 
1929 void Compile::add_range_check_cast(Node* n) {
1930   assert(n->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1931   assert(!_range_check_casts->contains(n), "duplicate entry in range check casts");
1932   _range_check_casts->append(n);
1933 }
1934 
1935 // Remove all range check dependent CastIINodes.
1936 void Compile::remove_range_check_casts(PhaseIterGVN &igvn) {
1937   for (int i = range_check_cast_count(); i > 0; i--) {
1938     Node* cast = range_check_cast_node(i-1);
1939     assert(cast->isa_CastII()->has_range_check(), "CastII should have range check dependency");
1940     igvn.replace_node(cast, cast->in(1));
1941   }
1942   assert(range_check_cast_count() == 0, "should be empty");
1943 }
1944 
1945 // StringOpts and late inlining of string methods
1946 void Compile::inline_string_calls(bool parse_time) {
1947   {
1948     // remove useless nodes to make the usage analysis simpler
1949     ResourceMark rm;
1950     PhaseRemoveUseless pru(initial_gvn(), for_igvn());
1951   }
1952 
1953   {
1954     ResourceMark rm;
1955     print_method(PHASE_BEFORE_STRINGOPTS, 3);
1956     PhaseStringOpts pso(initial_gvn(), for_igvn());
1957     print_method(PHASE_AFTER_STRINGOPTS, 3);
1958   }
1959 
1960   // now inline anything that we skipped the first time around
1961   if (!parse_time) {
1962     _late_inlines_pos = _late_inlines.length();
1963   }
1964 
1965   while (_string_late_inlines.length() > 0) {
1966     CallGenerator* cg = _string_late_inlines.pop();
1967     cg->do_late_inline();
1968     if (failing())  return;
1969   }
1970   _string_late_inlines.trunc_to(0);
1971 }
1972 
1973 // Late inlining of boxing methods
1974 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
1975   if (_boxing_late_inlines.length() > 0) {
1976     assert(has_boxed_value(), "inconsistent");
1977 
1978     PhaseGVN* gvn = initial_gvn();
1979     set_inlining_incrementally(true);
1980 
1981     assert( igvn._worklist.size() == 0, "should be done with igvn" );
1982     for_igvn()->clear();
1983     gvn->replace_with(&igvn);
1984 
1985     _late_inlines_pos = _late_inlines.length();
1986 
1987     while (_boxing_late_inlines.length() > 0) {
1988       CallGenerator* cg = _boxing_late_inlines.pop();
1989       cg->do_late_inline();
1990       if (failing())  return;
1991     }
1992     _boxing_late_inlines.trunc_to(0);
1993 
1994     {
1995       ResourceMark rm;
1996       PhaseRemoveUseless pru(gvn, for_igvn());
1997     }
1998 
1999     igvn = PhaseIterGVN(gvn);
2000     igvn.optimize();
2001 
2002     set_inlining_progress(false);
2003     set_inlining_incrementally(false);
2004   }
2005 }
2006 
2007 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
2008   assert(IncrementalInline, "incremental inlining should be on");
2009   PhaseGVN* gvn = initial_gvn();
2010 
2011   set_inlining_progress(false);
2012   for_igvn()->clear();
2013   gvn->replace_with(&igvn);
2014 
2015   int i = 0;
2016 
2017   for (; i <_late_inlines.length() && !inlining_progress(); i++) {
2018     CallGenerator* cg = _late_inlines.at(i);
2019     _late_inlines_pos = i+1;
2020     cg->do_late_inline();
2021     if (failing())  return;
2022   }
2023   int j = 0;
2024   for (; i < _late_inlines.length(); i++, j++) {
2025     _late_inlines.at_put(j, _late_inlines.at(i));
2026   }
2027   _late_inlines.trunc_to(j);
2028 
2029   {
2030     ResourceMark rm;
2031     PhaseRemoveUseless pru(gvn, for_igvn());
2032   }
2033 
2034   igvn = PhaseIterGVN(gvn);
2035 }
2036 
2037 // Perform incremental inlining until bound on number of live nodes is reached
2038 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
2039   PhaseGVN* gvn = initial_gvn();
2040 
2041   set_inlining_incrementally(true);
2042   set_inlining_progress(true);
2043   uint low_live_nodes = 0;
2044 
2045   while(inlining_progress() && _late_inlines.length() > 0) {
2046 
2047     if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2048       if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
2049         // PhaseIdealLoop is expensive so we only try it once we are
2050         // out of live nodes and we only try it again if the previous
2051         // helped got the number of nodes down significantly
2052         PhaseIdealLoop ideal_loop( igvn, false, true );
2053         if (failing())  return;
2054         low_live_nodes = live_nodes();
2055         _major_progress = true;
2056       }
2057 
2058       if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
2059         break;
2060       }
2061     }
2062 
2063     inline_incrementally_one(igvn);
2064 
2065     if (failing())  return;
2066 
2067     igvn.optimize();
2068 
2069     if (failing())  return;
2070   }
2071 
2072   assert( igvn._worklist.size() == 0, "should be done with igvn" );
2073 
2074   if (_string_late_inlines.length() > 0) {
2075     assert(has_stringbuilder(), "inconsistent");
2076     for_igvn()->clear();
2077     initial_gvn()->replace_with(&igvn);
2078 
2079     inline_string_calls(false);
2080 
2081     if (failing())  return;
2082 
2083     {
2084       ResourceMark rm;
2085       PhaseRemoveUseless pru(initial_gvn(), for_igvn());
2086     }
2087 
2088     igvn = PhaseIterGVN(gvn);
2089 
2090     igvn.optimize();
2091   }
2092 
2093   set_inlining_incrementally(false);
2094 }
2095 
2096 
2097 //------------------------------Optimize---------------------------------------
2098 // Given a graph, optimize it.
2099 void Compile::Optimize() {
2100   TracePhase t1("optimizer", &_t_optimizer, true);
2101 
2102 #ifndef PRODUCT
2103   if (env()->break_at_compile()) {
2104     BREAKPOINT;
2105   }
2106 
2107 #endif
2108 
2109   ResourceMark rm;
2110   int          loop_opts_cnt;
2111 
2112   NOT_PRODUCT( verify_graph_edges(); )
2113 
2114   print_method(PHASE_AFTER_PARSING);
2115 
2116  {
2117   // Iterative Global Value Numbering, including ideal transforms
2118   // Initialize IterGVN with types and values from parse-time GVN
2119   PhaseIterGVN igvn(initial_gvn());
2120   {
2121     NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
2122     igvn.optimize();
2123   }
2124 
2125   print_method(PHASE_ITER_GVN1, 2);
2126 
2127   if (failing())  return;
2128 
2129   {
2130     NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2131     inline_incrementally(igvn);
2132   }
2133 
2134   print_method(PHASE_INCREMENTAL_INLINE, 2);
2135 
2136   if (failing())  return;
2137 
2138   if (eliminate_boxing()) {
2139     NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
2140     // Inline valueOf() methods now.
2141     inline_boxing_calls(igvn);
2142 
2143     if (AlwaysIncrementalInline) {
2144       inline_incrementally(igvn);
2145     }
2146 
2147     print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2);
2148 
2149     if (failing())  return;
2150   }
2151 
2152   // Remove the speculative part of types and clean up the graph from
2153   // the extra CastPP nodes whose only purpose is to carry them. Do
2154   // that early so that optimizations are not disrupted by the extra
2155   // CastPP nodes.
2156   remove_speculative_types(igvn);
2157 
2158   // No more new expensive nodes will be added to the list from here
2159   // so keep only the actual candidates for optimizations.
2160   cleanup_expensive_nodes(igvn);
2161 
2162   if (!failing() && RenumberLiveNodes && live_nodes() + NodeLimitFudgeFactor < unique()) {
2163     NOT_PRODUCT(Compile::TracePhase t2("", &_t_renumberLive, TimeCompiler);)
2164     initial_gvn()->replace_with(&igvn);
2165     for_igvn()->clear();
2166     Unique_Node_List new_worklist(C->comp_arena());
2167     {
2168       ResourceMark rm;
2169       PhaseRenumberLive prl = PhaseRenumberLive(initial_gvn(), for_igvn(), &new_worklist);
2170     }
2171     set_for_igvn(&new_worklist);
2172     igvn = PhaseIterGVN(initial_gvn());
2173     igvn.optimize();
2174   }
2175 
2176   // Perform escape analysis
2177   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
2178     if (has_loops()) {
2179       // Cleanup graph (remove dead nodes).
2180       TracePhase t2("idealLoop", &_t_idealLoop, true);
2181       PhaseIdealLoop ideal_loop( igvn, false, true );
2182       if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2);
2183       if (failing())  return;
2184     }
2185     ConnectionGraph::do_analysis(this, &igvn);
2186 
2187     if (failing())  return;
2188 
2189     // Optimize out fields loads from scalar replaceable allocations.
2190     igvn.optimize();
2191     print_method(PHASE_ITER_GVN_AFTER_EA, 2);
2192 
2193     if (failing())  return;
2194 
2195     if (congraph() != NULL && macro_count() > 0) {
2196       NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
2197       PhaseMacroExpand mexp(igvn);
2198       mexp.eliminate_macro_nodes();
2199       igvn.set_delay_transform(false);
2200 
2201       igvn.optimize();
2202       print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2);
2203 
2204       if (failing())  return;
2205     }
2206   }
2207 
2208   // Loop transforms on the ideal graph.  Range Check Elimination,
2209   // peeling, unrolling, etc.
2210 
2211   // Set loop opts counter
2212   loop_opts_cnt = num_loop_opts();
2213   if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
2214     {
2215       TracePhase t2("idealLoop", &_t_idealLoop, true);
2216       PhaseIdealLoop ideal_loop( igvn, true );
2217       loop_opts_cnt--;
2218       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2);
2219       if (failing())  return;
2220     }
2221     // Loop opts pass if partial peeling occurred in previous pass
2222     if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
2223       TracePhase t3("idealLoop", &_t_idealLoop, true);
2224       PhaseIdealLoop ideal_loop( igvn, false );
2225       loop_opts_cnt--;
2226       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2);
2227       if (failing())  return;
2228     }
2229     // Loop opts pass for loop-unrolling before CCP
2230     if(major_progress() && (loop_opts_cnt > 0)) {
2231       TracePhase t4("idealLoop", &_t_idealLoop, true);
2232       PhaseIdealLoop ideal_loop( igvn, false );
2233       loop_opts_cnt--;
2234       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2);
2235     }
2236     if (!failing()) {
2237       // Verify that last round of loop opts produced a valid graph
2238       NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2239       PhaseIdealLoop::verify(igvn);
2240     }
2241   }
2242   if (failing())  return;
2243 
2244   // Conditional Constant Propagation;
2245   PhaseCCP ccp( &igvn );
2246   assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
2247   {
2248     TracePhase t2("ccp", &_t_ccp, true);
2249     ccp.do_transform();
2250   }
2251   print_method(PHASE_CPP1, 2);
2252 
2253   assert( true, "Break here to ccp.dump_old2new_map()");
2254 
2255   // Iterative Global Value Numbering, including ideal transforms
2256   {
2257     NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
2258     igvn = ccp;
2259     igvn.optimize();
2260   }
2261 
2262   print_method(PHASE_ITER_GVN2, 2);
2263 
2264   if (failing())  return;
2265 
2266   // Loop transforms on the ideal graph.  Range Check Elimination,
2267   // peeling, unrolling, etc.
2268   if(loop_opts_cnt > 0) {
2269     debug_only( int cnt = 0; );
2270     while(major_progress() && (loop_opts_cnt > 0)) {
2271       TracePhase t2("idealLoop", &_t_idealLoop, true);
2272       assert( cnt++ < 40, "infinite cycle in loop optimization" );
2273       PhaseIdealLoop ideal_loop( igvn, true);
2274       loop_opts_cnt--;
2275       if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2);
2276       if (failing())  return;
2277     }
2278   }
2279 
2280   {
2281     // Verify that all previous optimizations produced a valid graph
2282     // at least to this point, even if no loop optimizations were done.
2283     NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
2284     PhaseIdealLoop::verify(igvn);
2285   }
2286 
2287   if (range_check_cast_count() > 0) {
2288     // No more loop optimizations. Remove all range check dependent CastIINodes.
2289     C->remove_range_check_casts(igvn);
2290     igvn.optimize();
2291   }
2292 
2293 #ifdef ASSERT
2294   if (UseShenandoahGC && ShenandoahVerifyOptoBarriers) {
2295     ShenandoahBarrierC2Support::verify(C->root());
2296   }
2297 #endif
2298 
2299   {
2300     NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
2301     PhaseMacroExpand  mex(igvn);
2302     if (mex.expand_macro_nodes()) {
2303       assert(failing(), "must bail out w/ explicit message");
2304       return;
2305     }
2306   }
2307 
2308 #if INCLUDE_ALL_GCS
2309   if (UseShenandoahGC) {
2310     ShenandoahBarrierC2Support::expand(this, igvn);
2311   }
2312 #endif
2313 
2314  } // (End scope of igvn; run destructor if necessary for asserts.)
2315 
2316   dump_inlining();
2317   // A method with only infinite loops has no edges entering loops from root
2318   {
2319     NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
2320     if (final_graph_reshaping()) {
2321       assert(failing(), "must bail out w/ explicit message");
2322       return;
2323     }
2324   }
2325 
2326   print_method(PHASE_OPTIMIZE_FINISHED, 2);
2327 }
2328 
2329 
2330 //------------------------------Code_Gen---------------------------------------
2331 // Given a graph, generate code for it
2332 void Compile::Code_Gen() {
2333   if (failing()) {
2334     return;
2335   }
2336 
2337   // Perform instruction selection.  You might think we could reclaim Matcher
2338   // memory PDQ, but actually the Matcher is used in generating spill code.
2339   // Internals of the Matcher (including some VectorSets) must remain live
2340   // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
2341   // set a bit in reclaimed memory.
2342 
2343   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2344   // nodes.  Mapping is only valid at the root of each matched subtree.
2345   NOT_PRODUCT( verify_graph_edges(); )
2346 
2347   Matcher matcher;
2348   _matcher = &matcher;
2349   {
2350     TracePhase t2("matcher", &_t_matcher, true);
2351     matcher.match();
2352   }
2353   // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
2354   // nodes.  Mapping is only valid at the root of each matched subtree.
2355   NOT_PRODUCT( verify_graph_edges(); )
2356 
2357   // If you have too many nodes, or if matching has failed, bail out
2358   check_node_count(0, "out of nodes matching instructions");
2359   if (failing()) {
2360     return;
2361   }
2362 
2363   // Build a proper-looking CFG
2364   PhaseCFG cfg(node_arena(), root(), matcher);
2365   _cfg = &cfg;
2366   {
2367     NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
2368     bool success = cfg.do_global_code_motion();
2369     if (!success) {
2370       return;
2371     }
2372 
2373     print_method(PHASE_GLOBAL_CODE_MOTION, 2);
2374     NOT_PRODUCT( verify_graph_edges(); )
2375     debug_only( cfg.verify(); )
2376   }
2377 
2378   PhaseChaitin regalloc(unique(), cfg, matcher);
2379   _regalloc = &regalloc;
2380   {
2381     TracePhase t2("regalloc", &_t_registerAllocation, true);
2382     // Perform register allocation.  After Chaitin, use-def chains are
2383     // no longer accurate (at spill code) and so must be ignored.
2384     // Node->LRG->reg mappings are still accurate.
2385     _regalloc->Register_Allocate();
2386 
2387     // Bail out if the allocator builds too many nodes
2388     if (failing()) {
2389       return;
2390     }
2391   }
2392 
2393   // Prior to register allocation we kept empty basic blocks in case the
2394   // the allocator needed a place to spill.  After register allocation we
2395   // are not adding any new instructions.  If any basic block is empty, we
2396   // can now safely remove it.
2397   {
2398     NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
2399     cfg.remove_empty_blocks();
2400     if (do_freq_based_layout()) {
2401       PhaseBlockLayout layout(cfg);
2402     } else {
2403       cfg.set_loop_alignment();
2404     }
2405     cfg.fixup_flow();
2406   }
2407 
2408   // Apply peephole optimizations
2409   if( OptoPeephole ) {
2410     NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
2411     PhasePeephole peep( _regalloc, cfg);
2412     peep.do_transform();
2413   }
2414 
2415   // Do late expand if CPU requires this.
2416   if (Matcher::require_postalloc_expand) {
2417     NOT_PRODUCT(TracePhase t2c("postalloc_expand", &_t_postalloc_expand, true));
2418     cfg.postalloc_expand(_regalloc);
2419   }
2420 
2421   // Convert Nodes to instruction bits in a buffer
2422   {
2423     // %%%% workspace merge brought two timers together for one job
2424     TracePhase t2a("output", &_t_output, true);
2425     NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
2426     Output();
2427   }
2428 
2429   print_method(PHASE_FINAL_CODE);
2430 
2431   // He's dead, Jim.
2432   _cfg     = (PhaseCFG*)((intptr_t)0xdeadbeef);
2433   _regalloc = (PhaseChaitin*)((intptr_t)0xdeadbeef);
2434 }
2435 
2436 
2437 //------------------------------dump_asm---------------------------------------
2438 // Dump formatted assembly
2439 #ifndef PRODUCT
2440 void Compile::dump_asm(int *pcs, uint pc_limit) {
2441   bool cut_short = false;
2442   tty->print_cr("#");
2443   tty->print("#  ");  _tf->dump();  tty->cr();
2444   tty->print_cr("#");
2445 
2446   // For all blocks
2447   int pc = 0x0;                 // Program counter
2448   char starts_bundle = ' ';
2449   _regalloc->dump_frame();
2450 
2451   Node *n = NULL;
2452   for (uint i = 0; i < _cfg->number_of_blocks(); i++) {
2453     if (VMThread::should_terminate()) {
2454       cut_short = true;
2455       break;
2456     }
2457     Block* block = _cfg->get_block(i);
2458     if (block->is_connector() && !Verbose) {
2459       continue;
2460     }
2461     n = block->head();
2462     if (pcs && n->_idx < pc_limit) {
2463       tty->print("%3.3x   ", pcs[n->_idx]);
2464     } else {
2465       tty->print("      ");
2466     }
2467     block->dump_head(_cfg);
2468     if (block->is_connector()) {
2469       tty->print_cr("        # Empty connector block");
2470     } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
2471       tty->print_cr("        # Block is sole successor of call");
2472     }
2473 
2474     // For all instructions
2475     Node *delay = NULL;
2476     for (uint j = 0; j < block->number_of_nodes(); j++) {
2477       if (VMThread::should_terminate()) {
2478         cut_short = true;
2479         break;
2480       }
2481       n = block->get_node(j);
2482       if (valid_bundle_info(n)) {
2483         Bundle* bundle = node_bundling(n);
2484         if (bundle->used_in_unconditional_delay()) {
2485           delay = n;
2486           continue;
2487         }
2488         if (bundle->starts_bundle()) {
2489           starts_bundle = '+';
2490         }
2491       }
2492 
2493       if (WizardMode) {
2494         n->dump();
2495       }
2496 
2497       if( !n->is_Region() &&    // Dont print in the Assembly
2498           !n->is_Phi() &&       // a few noisely useless nodes
2499           !n->is_Proj() &&
2500           !n->is_MachTemp() &&
2501           !n->is_SafePointScalarObject() &&
2502           !n->is_Catch() &&     // Would be nice to print exception table targets
2503           !n->is_MergeMem() &&  // Not very interesting
2504           !n->is_top() &&       // Debug info table constants
2505           !(n->is_Con() && !n->is_Mach())// Debug info table constants
2506           ) {
2507         if (pcs && n->_idx < pc_limit)
2508           tty->print("%3.3x", pcs[n->_idx]);
2509         else
2510           tty->print("   ");
2511         tty->print(" %c ", starts_bundle);
2512         starts_bundle = ' ';
2513         tty->print("\t");
2514         n->format(_regalloc, tty);
2515         tty->cr();
2516       }
2517 
2518       // If we have an instruction with a delay slot, and have seen a delay,
2519       // then back up and print it
2520       if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
2521         assert(delay != NULL, "no unconditional delay instruction");
2522         if (WizardMode) delay->dump();
2523 
2524         if (node_bundling(delay)->starts_bundle())
2525           starts_bundle = '+';
2526         if (pcs && n->_idx < pc_limit)
2527           tty->print("%3.3x", pcs[n->_idx]);
2528         else
2529           tty->print("   ");
2530         tty->print(" %c ", starts_bundle);
2531         starts_bundle = ' ';
2532         tty->print("\t");
2533         delay->format(_regalloc, tty);
2534         tty->cr();
2535         delay = NULL;
2536       }
2537 
2538       // Dump the exception table as well
2539       if( n->is_Catch() && (Verbose || WizardMode) ) {
2540         // Print the exception table for this offset
2541         _handler_table.print_subtable_for(pc);
2542       }
2543     }
2544 
2545     if (pcs && n->_idx < pc_limit)
2546       tty->print_cr("%3.3x", pcs[n->_idx]);
2547     else
2548       tty->cr();
2549 
2550     assert(cut_short || delay == NULL, "no unconditional delay branch");
2551 
2552   } // End of per-block dump
2553   tty->cr();
2554 
2555   if (cut_short)  tty->print_cr("*** disassembly is cut short ***");
2556 }
2557 #endif
2558 
2559 //------------------------------Final_Reshape_Counts---------------------------
2560 // This class defines counters to help identify when a method
2561 // may/must be executed using hardware with only 24-bit precision.
2562 struct Final_Reshape_Counts : public StackObj {
2563   int  _call_count;             // count non-inlined 'common' calls
2564   int  _float_count;            // count float ops requiring 24-bit precision
2565   int  _double_count;           // count double ops requiring more precision
2566   int  _java_call_count;        // count non-inlined 'java' calls
2567   int  _inner_loop_count;       // count loops which need alignment
2568   VectorSet _visited;           // Visitation flags
2569   Node_List _tests;             // Set of IfNodes & PCTableNodes
2570 
2571   Final_Reshape_Counts() :
2572     _call_count(0), _float_count(0), _double_count(0),
2573     _java_call_count(0), _inner_loop_count(0),
2574     _visited( Thread::current()->resource_area() ) { }
2575 
2576   void inc_call_count  () { _call_count  ++; }
2577   void inc_float_count () { _float_count ++; }
2578   void inc_double_count() { _double_count++; }
2579   void inc_java_call_count() { _java_call_count++; }
2580   void inc_inner_loop_count() { _inner_loop_count++; }
2581 
2582   int  get_call_count  () const { return _call_count  ; }
2583   int  get_float_count () const { return _float_count ; }
2584   int  get_double_count() const { return _double_count; }
2585   int  get_java_call_count() const { return _java_call_count; }
2586   int  get_inner_loop_count() const { return _inner_loop_count; }
2587 };
2588 
2589 #ifdef ASSERT
2590 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
2591   ciInstanceKlass *k = tp->klass()->as_instance_klass();
2592   // Make sure the offset goes inside the instance layout.
2593   return k->contains_field_offset(tp->offset());
2594   // Note that OffsetBot and OffsetTop are very negative.
2595 }
2596 #endif
2597 
2598 // Eliminate trivially redundant StoreCMs and accumulate their
2599 // precedence edges.
2600 void Compile::eliminate_redundant_card_marks(Node* n) {
2601   assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
2602   if (n->in(MemNode::Address)->outcnt() > 1) {
2603     // There are multiple users of the same address so it might be
2604     // possible to eliminate some of the StoreCMs
2605     Node* mem = n->in(MemNode::Memory);
2606     Node* adr = n->in(MemNode::Address);
2607     Node* val = n->in(MemNode::ValueIn);
2608     Node* prev = n;
2609     bool done = false;
2610     // Walk the chain of StoreCMs eliminating ones that match.  As
2611     // long as it's a chain of single users then the optimization is
2612     // safe.  Eliminating partially redundant StoreCMs would require
2613     // cloning copies down the other paths.
2614     while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
2615       if (adr == mem->in(MemNode::Address) &&
2616           val == mem->in(MemNode::ValueIn)) {
2617         // redundant StoreCM
2618         if (mem->req() > MemNode::OopStore) {
2619           // Hasn't been processed by this code yet.
2620           n->add_prec(mem->in(MemNode::OopStore));
2621         } else {
2622           // Already converted to precedence edge
2623           for (uint i = mem->req(); i < mem->len(); i++) {
2624             // Accumulate any precedence edges
2625             if (mem->in(i) != NULL) {
2626               n->add_prec(mem->in(i));
2627             }
2628           }
2629           // Everything above this point has been processed.
2630           done = true;
2631         }
2632         // Eliminate the previous StoreCM
2633         prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
2634         assert(mem->outcnt() == 0, "should be dead");
2635         mem->disconnect_inputs(NULL, this);
2636       } else {
2637         prev = mem;
2638       }
2639       mem = prev->in(MemNode::Memory);
2640     }
2641   }
2642 }
2643 
2644 //------------------------------final_graph_reshaping_impl----------------------
2645 // Implement items 1-5 from final_graph_reshaping below.
2646 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
2647 
2648   if ( n->outcnt() == 0 ) return; // dead node
2649   uint nop = n->Opcode();
2650 
2651   // Check for 2-input instruction with "last use" on right input.
2652   // Swap to left input.  Implements item (2).
2653   if( n->req() == 3 &&          // two-input instruction
2654       n->in(1)->outcnt() > 1 && // left use is NOT a last use
2655       (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
2656       n->in(2)->outcnt() == 1 &&// right use IS a last use
2657       !n->in(2)->is_Con() ) {   // right use is not a constant
2658     // Check for commutative opcode
2659     switch( nop ) {
2660     case Op_AddI:  case Op_AddF:  case Op_AddD:  case Op_AddL:
2661     case Op_MaxI:  case Op_MinI:
2662     case Op_MulI:  case Op_MulF:  case Op_MulD:  case Op_MulL:
2663     case Op_AndL:  case Op_XorL:  case Op_OrL:
2664     case Op_AndI:  case Op_XorI:  case Op_OrI: {
2665       // Move "last use" input to left by swapping inputs
2666       n->swap_edges(1, 2);
2667       break;
2668     }
2669     default:
2670       break;
2671     }
2672   }
2673 
2674 #ifdef ASSERT
2675   if( n->is_Mem() ) {
2676     int alias_idx = get_alias_index(n->as_Mem()->adr_type());
2677     assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
2678             // oop will be recorded in oop map if load crosses safepoint
2679             n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
2680                              LoadNode::is_immutable_value(n->in(MemNode::Address))),
2681             "raw memory operations should have control edge");
2682   }
2683   if (n->is_MemBar()) {
2684     MemBarNode* mb = n->as_MemBar();
2685     if (mb->trailing_store() || mb->trailing_load_store()) {
2686       assert(mb->leading_membar()->trailing_membar() == mb, "bad membar pair");
2687       Node* mem = mb->in(MemBarNode::Precedent);
2688       assert((mb->trailing_store() && mem->is_Store() && mem->as_Store()->is_release()) ||
2689              (mb->trailing_load_store() && mem->is_LoadStore()), "missing mem op");
2690     } else if (mb->leading()) {
2691       assert(mb->trailing_membar()->leading_membar() == mb, "bad membar pair");
2692     }
2693   }
2694 #endif
2695   // Count FPU ops and common calls, implements item (3)
2696   switch( nop ) {
2697   // Count all float operations that may use FPU
2698   case Op_AddF:
2699   case Op_SubF:
2700   case Op_MulF:
2701   case Op_DivF:
2702   case Op_NegF:
2703   case Op_ModF:
2704   case Op_ConvI2F:
2705   case Op_ConF:
2706   case Op_CmpF:
2707   case Op_CmpF3:
2708   // case Op_ConvL2F: // longs are split into 32-bit halves
2709     frc.inc_float_count();
2710     break;
2711 
2712   case Op_ConvF2D:
2713   case Op_ConvD2F:
2714     frc.inc_float_count();
2715     frc.inc_double_count();
2716     break;
2717 
2718   // Count all double operations that may use FPU
2719   case Op_AddD:
2720   case Op_SubD:
2721   case Op_MulD:
2722   case Op_DivD:
2723   case Op_NegD:
2724   case Op_ModD:
2725   case Op_ConvI2D:
2726   case Op_ConvD2I:
2727   // case Op_ConvL2D: // handled by leaf call
2728   // case Op_ConvD2L: // handled by leaf call
2729   case Op_ConD:
2730   case Op_CmpD:
2731   case Op_CmpD3:
2732     frc.inc_double_count();
2733     break;
2734   case Op_Opaque1:              // Remove Opaque Nodes before matching
2735   case Op_Opaque2:              // Remove Opaque Nodes before matching
2736   case Op_Opaque3:
2737     n->subsume_by(n->in(1), this);
2738     break;
2739   case Op_CallStaticJava:
2740   case Op_CallJava:
2741   case Op_CallDynamicJava:
2742     frc.inc_java_call_count(); // Count java call site;
2743   case Op_CallRuntime:
2744   case Op_CallLeaf:
2745   case Op_CallLeafNoFP: {
2746     assert( n->is_Call(), "" );
2747     CallNode *call = n->as_Call();
2748     if (UseShenandoahGC && call->is_g1_wb_pre_call()) {
2749       uint cnt = OptoRuntime::g1_wb_pre_Type()->domain()->cnt();
2750       if (call->req() > cnt) {
2751         assert(call->req() == cnt+1, "only one extra input");
2752         Node* addp = call->in(cnt);
2753         assert(!CallLeafNode::has_only_g1_wb_pre_uses(addp), "useless address computation?");
2754         call->del_req(cnt);
2755       }
2756     }
2757     // Count call sites where the FP mode bit would have to be flipped.
2758     // Do not count uncommon runtime calls:
2759     // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
2760     // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
2761     if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
2762       frc.inc_call_count();   // Count the call site
2763     } else {                  // See if uncommon argument is shared
2764       Node *n = call->in(TypeFunc::Parms);
2765       int nop = n->Opcode();
2766       // Clone shared simple arguments to uncommon calls, item (1).
2767       if( n->outcnt() > 1 &&
2768           !n->is_Proj() &&
2769           nop != Op_CreateEx &&
2770           nop != Op_CheckCastPP &&
2771           nop != Op_DecodeN &&
2772           nop != Op_DecodeNKlass &&
2773           !n->is_Mem() ) {
2774         Node *x = n->clone();
2775         call->set_req( TypeFunc::Parms, x );
2776       }
2777     }
2778     break;
2779   }
2780 
2781   case Op_StoreD:
2782   case Op_LoadD:
2783   case Op_LoadD_unaligned:
2784     frc.inc_double_count();
2785     goto handle_mem;
2786   case Op_StoreF:
2787   case Op_LoadF:
2788     frc.inc_float_count();
2789     goto handle_mem;
2790 
2791   case Op_StoreCM:
2792     {
2793       // Convert OopStore dependence into precedence edge
2794       Node* prec = n->in(MemNode::OopStore);
2795       n->del_req(MemNode::OopStore);
2796       n->add_prec(prec);
2797       eliminate_redundant_card_marks(n);
2798     }
2799 
2800     // fall through
2801 
2802   case Op_StoreB:
2803   case Op_StoreC:
2804   case Op_StorePConditional:
2805   case Op_StoreI:
2806   case Op_StoreL:
2807   case Op_StoreIConditional:
2808   case Op_StoreLConditional:
2809   case Op_CompareAndSwapI:
2810   case Op_CompareAndSwapL:
2811   case Op_CompareAndSwapP:
2812   case Op_CompareAndSwapN:
2813   case Op_GetAndAddI:
2814   case Op_GetAndAddL:
2815   case Op_GetAndSetI:
2816   case Op_GetAndSetL:
2817   case Op_GetAndSetP:
2818   case Op_GetAndSetN:
2819   case Op_StoreP:
2820   case Op_StoreN:
2821   case Op_StoreNKlass:
2822   case Op_LoadB:
2823   case Op_LoadUB:
2824   case Op_LoadUS:
2825   case Op_LoadI:
2826   case Op_LoadKlass:
2827   case Op_LoadNKlass:
2828   case Op_LoadL:
2829   case Op_LoadL_unaligned:
2830   case Op_LoadPLocked:
2831   case Op_LoadP:
2832   case Op_LoadN:
2833   case Op_LoadRange:
2834   case Op_LoadS: {
2835   handle_mem:
2836 #ifdef ASSERT
2837     if( VerifyOptoOopOffsets ) {
2838       assert( n->is_Mem(), "" );
2839       MemNode *mem  = (MemNode*)n;
2840       // Check to see if address types have grounded out somehow.
2841       const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
2842       assert( !tp || oop_offset_is_sane(tp), "" );
2843     }
2844 #endif
2845     break;
2846   }
2847 
2848   case Op_AddP: {               // Assert sane base pointers
2849     Node *addp = n->in(AddPNode::Address);
2850     assert( !addp->is_AddP() ||
2851             addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
2852             addp->in(AddPNode::Base) == n->in(AddPNode::Base),
2853             "Base pointers must match" );
2854 #ifdef _LP64
2855     if ((UseCompressedOops || UseCompressedClassPointers) &&
2856         addp->Opcode() == Op_ConP &&
2857         addp == n->in(AddPNode::Base) &&
2858         n->in(AddPNode::Offset)->is_Con()) {
2859       // Use addressing with narrow klass to load with offset on x86.
2860       // On sparc loading 32-bits constant and decoding it have less
2861       // instructions (4) then load 64-bits constant (7).
2862       // Do this transformation here since IGVN will convert ConN back to ConP.
2863       const Type* t = addp->bottom_type();
2864       if (t->isa_oopptr() || t->isa_klassptr()) {
2865         Node* nn = NULL;
2866 
2867         int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
2868 
2869         // Look for existing ConN node of the same exact type.
2870         Node* r  = root();
2871         uint cnt = r->outcnt();
2872         for (uint i = 0; i < cnt; i++) {
2873           Node* m = r->raw_out(i);
2874           if (m!= NULL && m->Opcode() == op &&
2875               m->bottom_type()->make_ptr() == t) {
2876             nn = m;
2877             break;
2878           }
2879         }
2880         if (nn != NULL) {
2881           // Decode a narrow oop to match address
2882           // [R12 + narrow_oop_reg<<3 + offset]
2883           if (t->isa_oopptr()) {
2884             nn = new (this) DecodeNNode(nn, t);
2885           } else {
2886             nn = new (this) DecodeNKlassNode(nn, t);
2887           }
2888           n->set_req(AddPNode::Base, nn);
2889           n->set_req(AddPNode::Address, nn);
2890           if (addp->outcnt() == 0) {
2891             addp->disconnect_inputs(NULL, this);
2892           }
2893         }
2894       }
2895     }
2896 #endif
2897     break;
2898   }
2899 
2900   case Op_CastPP: {
2901     // Remove CastPP nodes to gain more freedom during scheduling but
2902     // keep the dependency they encode as control or precedence edges
2903     // (if control is set already) on memory operations. Some CastPP
2904     // nodes don't have a control (don't carry a dependency): skip
2905     // those.
2906     if (n->in(0) != NULL) {
2907       ResourceMark rm;
2908       Unique_Node_List wq;
2909       wq.push(n);
2910       for (uint next = 0; next < wq.size(); ++next) {
2911         Node *m = wq.at(next);
2912         for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2913           Node* use = m->fast_out(i);
2914           if (use->is_Mem() || use->is_EncodeNarrowPtr() || use->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
2915             use->ensure_control_or_add_prec(n->in(0));
2916           } else if (use->in(0) == NULL) {
2917             switch(use->Opcode()) {
2918             case Op_AddP:
2919             case Op_DecodeN:
2920             case Op_DecodeNKlass:
2921             case Op_CheckCastPP:
2922             case Op_CastPP:
2923               wq.push(use);
2924               break;
2925             }
2926           }
2927         }
2928       }
2929     }
2930     const bool is_LP64 = LP64_ONLY(true) NOT_LP64(false);
2931     if (is_LP64 && n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
2932       Node* in1 = n->in(1);
2933       const Type* t = n->bottom_type();
2934       Node* new_in1 = in1->clone();
2935       new_in1->as_DecodeN()->set_type(t);
2936 
2937       if (!Matcher::narrow_oop_use_complex_address()) {
2938         //
2939         // x86, ARM and friends can handle 2 adds in addressing mode
2940         // and Matcher can fold a DecodeN node into address by using
2941         // a narrow oop directly and do implicit NULL check in address:
2942         //
2943         // [R12 + narrow_oop_reg<<3 + offset]
2944         // NullCheck narrow_oop_reg
2945         //
2946         // On other platforms (Sparc) we have to keep new DecodeN node and
2947         // use it to do implicit NULL check in address:
2948         //
2949         // decode_not_null narrow_oop_reg, base_reg
2950         // [base_reg + offset]
2951         // NullCheck base_reg
2952         //
2953         // Pin the new DecodeN node to non-null path on these platform (Sparc)
2954         // to keep the information to which NULL check the new DecodeN node
2955         // corresponds to use it as value in implicit_null_check().
2956         //
2957         new_in1->set_req(0, n->in(0));
2958       }
2959 
2960       n->subsume_by(new_in1, this);
2961       if (in1->outcnt() == 0) {
2962         in1->disconnect_inputs(NULL, this);
2963       }
2964     } else {
2965       n->subsume_by(n->in(1), this);
2966       if (n->outcnt() == 0) {
2967         n->disconnect_inputs(NULL, this);
2968       }
2969     }
2970     break;
2971   }
2972 #ifdef _LP64
2973   case Op_CmpP:
2974     // Do this transformation here to preserve CmpPNode::sub() and
2975     // other TypePtr related Ideal optimizations (for example, ptr nullness).
2976     if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
2977       Node* in1 = n->in(1);
2978       Node* in2 = n->in(2);
2979       if (!in1->is_DecodeNarrowPtr()) {
2980         in2 = in1;
2981         in1 = n->in(2);
2982       }
2983       assert(in1->is_DecodeNarrowPtr(), "sanity");
2984 
2985       Node* new_in2 = NULL;
2986       if (in2->is_DecodeNarrowPtr()) {
2987         assert(in2->Opcode() == in1->Opcode(), "must be same node type");
2988         new_in2 = in2->in(1);
2989       } else if (in2->Opcode() == Op_ConP) {
2990         const Type* t = in2->bottom_type();
2991         if (t == TypePtr::NULL_PTR) {
2992           assert(in1->is_DecodeN(), "compare klass to null?");
2993           // Don't convert CmpP null check into CmpN if compressed
2994           // oops implicit null check is not generated.
2995           // This will allow to generate normal oop implicit null check.
2996           if (Matcher::gen_narrow_oop_implicit_null_checks())
2997             new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
2998           //
2999           // This transformation together with CastPP transformation above
3000           // will generated code for implicit NULL checks for compressed oops.
3001           //
3002           // The original code after Optimize()
3003           //
3004           //    LoadN memory, narrow_oop_reg
3005           //    decode narrow_oop_reg, base_reg
3006           //    CmpP base_reg, NULL
3007           //    CastPP base_reg // NotNull
3008           //    Load [base_reg + offset], val_reg
3009           //
3010           // after these transformations will be
3011           //
3012           //    LoadN memory, narrow_oop_reg
3013           //    CmpN narrow_oop_reg, NULL
3014           //    decode_not_null narrow_oop_reg, base_reg
3015           //    Load [base_reg + offset], val_reg
3016           //
3017           // and the uncommon path (== NULL) will use narrow_oop_reg directly
3018           // since narrow oops can be used in debug info now (see the code in
3019           // final_graph_reshaping_walk()).
3020           //
3021           // At the end the code will be matched to
3022           // on x86:
3023           //
3024           //    Load_narrow_oop memory, narrow_oop_reg
3025           //    Load [R12 + narrow_oop_reg<<3 + offset], val_reg
3026           //    NullCheck narrow_oop_reg
3027           //
3028           // and on sparc:
3029           //
3030           //    Load_narrow_oop memory, narrow_oop_reg
3031           //    decode_not_null narrow_oop_reg, base_reg
3032           //    Load [base_reg + offset], val_reg
3033           //    NullCheck base_reg
3034           //
3035         } else if (t->isa_oopptr()) {
3036           new_in2 = ConNode::make(this, t->make_narrowoop());
3037         } else if (t->isa_klassptr()) {
3038           new_in2 = ConNode::make(this, t->make_narrowklass());
3039         }
3040       }
3041       if (new_in2 != NULL) {
3042         Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
3043         n->subsume_by(cmpN, this);
3044         if (in1->outcnt() == 0) {
3045           in1->disconnect_inputs(NULL, this);
3046         }
3047         if (in2->outcnt() == 0) {
3048           in2->disconnect_inputs(NULL, this);
3049         }
3050       }
3051     }
3052     break;
3053 
3054   case Op_DecodeN:
3055   case Op_DecodeNKlass:
3056     assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
3057     // DecodeN could be pinned when it can't be fold into
3058     // an address expression, see the code for Op_CastPP above.
3059     assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
3060     break;
3061 
3062   case Op_EncodeP:
3063   case Op_EncodePKlass: {
3064     Node* in1 = n->in(1);
3065     if (in1->is_DecodeNarrowPtr()) {
3066       n->subsume_by(in1->in(1), this);
3067     } else if (in1->Opcode() == Op_ConP) {
3068       const Type* t = in1->bottom_type();
3069       if (t == TypePtr::NULL_PTR) {
3070         assert(t->isa_oopptr(), "null klass?");
3071         n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
3072       } else if (t->isa_oopptr()) {
3073         n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
3074       } else if (t->isa_klassptr()) {
3075         n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
3076       }
3077     }
3078     if (in1->outcnt() == 0) {
3079       in1->disconnect_inputs(NULL, this);
3080     }
3081     break;
3082   }
3083 
3084   case Op_Proj: {
3085     if (OptimizeStringConcat) {
3086       ProjNode* p = n->as_Proj();
3087       if (p->_is_io_use) {
3088         // Separate projections were used for the exception path which
3089         // are normally removed by a late inline.  If it wasn't inlined
3090         // then they will hang around and should just be replaced with
3091         // the original one.
3092         Node* proj = NULL;
3093         // Replace with just one
3094         for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
3095           Node *use = i.get();
3096           if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
3097             proj = use;
3098             break;
3099           }
3100         }
3101         assert(proj != NULL, "must be found");
3102         p->subsume_by(proj, this);
3103       }
3104     }
3105     break;
3106   }
3107 
3108   case Op_Phi:
3109     if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
3110       // The EncodeP optimization may create Phi with the same edges
3111       // for all paths. It is not handled well by Register Allocator.
3112       Node* unique_in = n->in(1);
3113       assert(unique_in != NULL, "");
3114       uint cnt = n->req();
3115       for (uint i = 2; i < cnt; i++) {
3116         Node* m = n->in(i);
3117         assert(m != NULL, "");
3118         if (unique_in != m)
3119           unique_in = NULL;
3120       }
3121       if (unique_in != NULL) {
3122         n->subsume_by(unique_in, this);
3123       }
3124     }
3125     break;
3126 
3127 #endif
3128 
3129 #ifdef ASSERT
3130   case Op_CastII:
3131     // Verify that all range check dependent CastII nodes were removed.
3132     if (n->isa_CastII()->has_range_check()) {
3133       n->dump(3);
3134       assert(false, "Range check dependent CastII node was not removed");
3135     }
3136     break;
3137 #endif
3138 
3139   case Op_ModI:
3140     if (UseDivMod) {
3141       // Check if a%b and a/b both exist
3142       Node* d = n->find_similar(Op_DivI);
3143       if (d) {
3144         // Replace them with a fused divmod if supported
3145         if (Matcher::has_match_rule(Op_DivModI)) {
3146           DivModINode* divmod = DivModINode::make(this, n);
3147           d->subsume_by(divmod->div_proj(), this);
3148           n->subsume_by(divmod->mod_proj(), this);
3149         } else {
3150           // replace a%b with a-((a/b)*b)
3151           Node* mult = new (this) MulINode(d, d->in(2));
3152           Node* sub  = new (this) SubINode(d->in(1), mult);
3153           n->subsume_by(sub, this);
3154         }
3155       }
3156     }
3157     break;
3158 
3159   case Op_ModL:
3160     if (UseDivMod) {
3161       // Check if a%b and a/b both exist
3162       Node* d = n->find_similar(Op_DivL);
3163       if (d) {
3164         // Replace them with a fused divmod if supported
3165         if (Matcher::has_match_rule(Op_DivModL)) {
3166           DivModLNode* divmod = DivModLNode::make(this, n);
3167           d->subsume_by(divmod->div_proj(), this);
3168           n->subsume_by(divmod->mod_proj(), this);
3169         } else {
3170           // replace a%b with a-((a/b)*b)
3171           Node* mult = new (this) MulLNode(d, d->in(2));
3172           Node* sub  = new (this) SubLNode(d->in(1), mult);
3173           n->subsume_by(sub, this);
3174         }
3175       }
3176     }
3177     break;
3178 
3179   case Op_LoadVector:
3180   case Op_StoreVector:
3181     break;
3182 
3183   case Op_PackB:
3184   case Op_PackS:
3185   case Op_PackI:
3186   case Op_PackF:
3187   case Op_PackL:
3188   case Op_PackD:
3189     if (n->req()-1 > 2) {
3190       // Replace many operand PackNodes with a binary tree for matching
3191       PackNode* p = (PackNode*) n;
3192       Node* btp = p->binary_tree_pack(this, 1, n->req());
3193       n->subsume_by(btp, this);
3194     }
3195     break;
3196   case Op_Loop:
3197   case Op_CountedLoop:
3198     if (n->as_Loop()->is_inner_loop()) {
3199       frc.inc_inner_loop_count();
3200     }
3201     break;
3202   case Op_LShiftI:
3203   case Op_RShiftI:
3204   case Op_URShiftI:
3205   case Op_LShiftL:
3206   case Op_RShiftL:
3207   case Op_URShiftL:
3208     if (Matcher::need_masked_shift_count) {
3209       // The cpu's shift instructions don't restrict the count to the
3210       // lower 5/6 bits. We need to do the masking ourselves.
3211       Node* in2 = n->in(2);
3212       juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
3213       const TypeInt* t = in2->find_int_type();
3214       if (t != NULL && t->is_con()) {
3215         juint shift = t->get_con();
3216         if (shift > mask) { // Unsigned cmp
3217           n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
3218         }
3219       } else {
3220         if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
3221           Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
3222           n->set_req(2, shift);
3223         }
3224       }
3225       if (in2->outcnt() == 0) { // Remove dead node
3226         in2->disconnect_inputs(NULL, this);
3227       }
3228     }
3229     break;
3230   case Op_MemBarStoreStore:
3231   case Op_MemBarRelease:
3232     // Break the link with AllocateNode: it is no longer useful and
3233     // confuses register allocation.
3234     if (n->req() > MemBarNode::Precedent) {
3235       n->set_req(MemBarNode::Precedent, top());
3236     }
3237     break;
3238   case Op_ShenandoahLoadReferenceBarrier:
3239     assert(false, "should have been expanded already");
3240     break;
3241   default:
3242     assert( !n->is_Call(), "" );
3243     assert( !n->is_Mem(), "" );
3244     assert( nop != Op_ProfileBoolean, "should be eliminated during IGVN");
3245     break;
3246   }
3247 
3248   // Collect CFG split points
3249   if (n->is_MultiBranch())
3250     frc._tests.push(n);
3251 }
3252 
3253 //------------------------------final_graph_reshaping_walk---------------------
3254 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
3255 // requires that the walk visits a node's inputs before visiting the node.
3256 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
3257   ResourceArea *area = Thread::current()->resource_area();
3258   Unique_Node_List sfpt(area);
3259 
3260   frc._visited.set(root->_idx); // first, mark node as visited
3261   uint cnt = root->req();
3262   Node *n = root;
3263   uint  i = 0;
3264   while (true) {
3265     if (i < cnt) {
3266       // Place all non-visited non-null inputs onto stack
3267       Node* m = n->in(i);
3268       ++i;
3269       if (m != NULL && !frc._visited.test_set(m->_idx)) {
3270         if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) {
3271           // compute worst case interpreter size in case of a deoptimization
3272           update_interpreter_frame_size(m->as_SafePoint()->jvms()->interpreter_frame_size());
3273 
3274           sfpt.push(m);
3275         }
3276         cnt = m->req();
3277         nstack.push(n, i); // put on stack parent and next input's index
3278         n = m;
3279         i = 0;
3280       }
3281     } else {
3282       // Now do post-visit work
3283       final_graph_reshaping_impl( n, frc );
3284       if (nstack.is_empty())
3285         break;             // finished
3286       n = nstack.node();   // Get node from stack
3287       cnt = n->req();
3288       i = nstack.index();
3289       nstack.pop();        // Shift to the next node on stack
3290     }
3291   }
3292 
3293   // Skip next transformation if compressed oops are not used.
3294   if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
3295       (!UseCompressedOops && !UseCompressedClassPointers))
3296     return;
3297 
3298   // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
3299   // It could be done for an uncommon traps or any safepoints/calls
3300   // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
3301   while (sfpt.size() > 0) {
3302     n = sfpt.pop();
3303     JVMState *jvms = n->as_SafePoint()->jvms();
3304     assert(jvms != NULL, "sanity");
3305     int start = jvms->debug_start();
3306     int end   = n->req();
3307     bool is_uncommon = (n->is_CallStaticJava() &&
3308                         n->as_CallStaticJava()->uncommon_trap_request() != 0);
3309     for (int j = start; j < end; j++) {
3310       Node* in = n->in(j);
3311       if (in->is_DecodeNarrowPtr()) {
3312         bool safe_to_skip = true;
3313         if (!is_uncommon ) {
3314           // Is it safe to skip?
3315           for (uint i = 0; i < in->outcnt(); i++) {
3316             Node* u = in->raw_out(i);
3317             if (!u->is_SafePoint() ||
3318                  u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
3319               safe_to_skip = false;
3320             }
3321           }
3322         }
3323         if (safe_to_skip) {
3324           n->set_req(j, in->in(1));
3325         }
3326         if (in->outcnt() == 0) {
3327           in->disconnect_inputs(NULL, this);
3328         }
3329       }
3330     }
3331   }
3332 }
3333 
3334 //------------------------------final_graph_reshaping--------------------------
3335 // Final Graph Reshaping.
3336 //
3337 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
3338 //     and not commoned up and forced early.  Must come after regular
3339 //     optimizations to avoid GVN undoing the cloning.  Clone constant
3340 //     inputs to Loop Phis; these will be split by the allocator anyways.
3341 //     Remove Opaque nodes.
3342 // (2) Move last-uses by commutative operations to the left input to encourage
3343 //     Intel update-in-place two-address operations and better register usage
3344 //     on RISCs.  Must come after regular optimizations to avoid GVN Ideal
3345 //     calls canonicalizing them back.
3346 // (3) Count the number of double-precision FP ops, single-precision FP ops
3347 //     and call sites.  On Intel, we can get correct rounding either by
3348 //     forcing singles to memory (requires extra stores and loads after each
3349 //     FP bytecode) or we can set a rounding mode bit (requires setting and
3350 //     clearing the mode bit around call sites).  The mode bit is only used
3351 //     if the relative frequency of single FP ops to calls is low enough.
3352 //     This is a key transform for SPEC mpeg_audio.
3353 // (4) Detect infinite loops; blobs of code reachable from above but not
3354 //     below.  Several of the Code_Gen algorithms fail on such code shapes,
3355 //     so we simply bail out.  Happens a lot in ZKM.jar, but also happens
3356 //     from time to time in other codes (such as -Xcomp finalizer loops, etc).
3357 //     Detection is by looking for IfNodes where only 1 projection is
3358 //     reachable from below or CatchNodes missing some targets.
3359 // (5) Assert for insane oop offsets in debug mode.
3360 
3361 bool Compile::final_graph_reshaping() {
3362   // an infinite loop may have been eliminated by the optimizer,
3363   // in which case the graph will be empty.
3364   if (root()->req() == 1) {
3365     record_method_not_compilable("trivial infinite loop");
3366     return true;
3367   }
3368 
3369   // Expensive nodes have their control input set to prevent the GVN
3370   // from freely commoning them. There's no GVN beyond this point so
3371   // no need to keep the control input. We want the expensive nodes to
3372   // be freely moved to the least frequent code path by gcm.
3373   assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
3374   for (int i = 0; i < expensive_count(); i++) {
3375     _expensive_nodes->at(i)->set_req(0, NULL);
3376   }
3377 
3378   Final_Reshape_Counts frc;
3379 
3380   // Visit everybody reachable!
3381   // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
3382   Node_Stack nstack(live_nodes() >> 1);
3383   final_graph_reshaping_walk(nstack, root(), frc);
3384 
3385   // Check for unreachable (from below) code (i.e., infinite loops).
3386   for( uint i = 0; i < frc._tests.size(); i++ ) {
3387     MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
3388     // Get number of CFG targets.
3389     // Note that PCTables include exception targets after calls.
3390     uint required_outcnt = n->required_outcnt();
3391     if (n->outcnt() != required_outcnt) {
3392       // Check for a few special cases.  Rethrow Nodes never take the
3393       // 'fall-thru' path, so expected kids is 1 less.
3394       if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
3395         if (n->in(0)->in(0)->is_Call()) {
3396           CallNode *call = n->in(0)->in(0)->as_Call();
3397           if (call->entry_point() == OptoRuntime::rethrow_stub()) {
3398             required_outcnt--;      // Rethrow always has 1 less kid
3399           } else if (call->req() > TypeFunc::Parms &&
3400                      call->is_CallDynamicJava()) {
3401             // Check for null receiver. In such case, the optimizer has
3402             // detected that the virtual call will always result in a null
3403             // pointer exception. The fall-through projection of this CatchNode
3404             // will not be populated.
3405             Node *arg0 = call->in(TypeFunc::Parms);
3406             if (arg0->is_Type() &&
3407                 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
3408               required_outcnt--;
3409             }
3410           } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
3411                      call->req() > TypeFunc::Parms+1 &&
3412                      call->is_CallStaticJava()) {
3413             // Check for negative array length. In such case, the optimizer has
3414             // detected that the allocation attempt will always result in an
3415             // exception. There is no fall-through projection of this CatchNode .
3416             Node *arg1 = call->in(TypeFunc::Parms+1);
3417             if (arg1->is_Type() &&
3418                 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
3419               required_outcnt--;
3420             }
3421           }
3422         }
3423       }
3424       // Recheck with a better notion of 'required_outcnt'
3425       if (n->outcnt() != required_outcnt) {
3426         record_method_not_compilable("malformed control flow");
3427         return true;            // Not all targets reachable!
3428       }
3429     }
3430     // Check that I actually visited all kids.  Unreached kids
3431     // must be infinite loops.
3432     for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
3433       if (!frc._visited.test(n->fast_out(j)->_idx)) {
3434         record_method_not_compilable("infinite loop");
3435         return true;            // Found unvisited kid; must be unreach
3436       }
3437   }
3438 
3439   // If original bytecodes contained a mixture of floats and doubles
3440   // check if the optimizer has made it homogenous, item (3).
3441   if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
3442       frc.get_float_count() > 32 &&
3443       frc.get_double_count() == 0 &&
3444       (10 * frc.get_call_count() < frc.get_float_count()) ) {
3445     set_24_bit_selection_and_mode( false,  true );
3446   }
3447 
3448   set_java_calls(frc.get_java_call_count());
3449   set_inner_loops(frc.get_inner_loop_count());
3450 
3451   // No infinite loops, no reason to bail out.
3452   return false;
3453 }
3454 
3455 //-----------------------------too_many_traps----------------------------------
3456 // Report if there are too many traps at the current method and bci.
3457 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
3458 bool Compile::too_many_traps(ciMethod* method,
3459                              int bci,
3460                              Deoptimization::DeoptReason reason) {
3461   ciMethodData* md = method->method_data();
3462   if (md->is_empty()) {
3463     // Assume the trap has not occurred, or that it occurred only
3464     // because of a transient condition during start-up in the interpreter.
3465     return false;
3466   }
3467   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3468   if (md->has_trap_at(bci, m, reason) != 0) {
3469     // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
3470     // Also, if there are multiple reasons, or if there is no per-BCI record,
3471     // assume the worst.
3472     if (log())
3473       log()->elem("observe trap='%s' count='%d'",
3474                   Deoptimization::trap_reason_name(reason),
3475                   md->trap_count(reason));
3476     return true;
3477   } else {
3478     // Ignore method/bci and see if there have been too many globally.
3479     return too_many_traps(reason, md);
3480   }
3481 }
3482 
3483 // Less-accurate variant which does not require a method and bci.
3484 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
3485                              ciMethodData* logmd) {
3486   if (trap_count(reason) >= Deoptimization::per_method_trap_limit(reason)) {
3487     // Too many traps globally.
3488     // Note that we use cumulative trap_count, not just md->trap_count.
3489     if (log()) {
3490       int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
3491       log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
3492                   Deoptimization::trap_reason_name(reason),
3493                   mcount, trap_count(reason));
3494     }
3495     return true;
3496   } else {
3497     // The coast is clear.
3498     return false;
3499   }
3500 }
3501 
3502 //--------------------------too_many_recompiles--------------------------------
3503 // Report if there are too many recompiles at the current method and bci.
3504 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
3505 // Is not eager to return true, since this will cause the compiler to use
3506 // Action_none for a trap point, to avoid too many recompilations.
3507 bool Compile::too_many_recompiles(ciMethod* method,
3508                                   int bci,
3509                                   Deoptimization::DeoptReason reason) {
3510   ciMethodData* md = method->method_data();
3511   if (md->is_empty()) {
3512     // Assume the trap has not occurred, or that it occurred only
3513     // because of a transient condition during start-up in the interpreter.
3514     return false;
3515   }
3516   // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
3517   uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
3518   uint m_cutoff  = (uint) PerMethodRecompilationCutoff / 2 + 1;  // not zero
3519   Deoptimization::DeoptReason per_bc_reason
3520     = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
3521   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? this->method() : NULL;
3522   if ((per_bc_reason == Deoptimization::Reason_none
3523        || md->has_trap_at(bci, m, reason) != 0)
3524       // The trap frequency measure we care about is the recompile count:
3525       && md->trap_recompiled_at(bci, m)
3526       && md->overflow_recompile_count() >= bc_cutoff) {
3527     // Do not emit a trap here if it has already caused recompilations.
3528     // Also, if there are multiple reasons, or if there is no per-BCI record,
3529     // assume the worst.
3530     if (log())
3531       log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
3532                   Deoptimization::trap_reason_name(reason),
3533                   md->trap_count(reason),
3534                   md->overflow_recompile_count());
3535     return true;
3536   } else if (trap_count(reason) != 0
3537              && decompile_count() >= m_cutoff) {
3538     // Too many recompiles globally, and we have seen this sort of trap.
3539     // Use cumulative decompile_count, not just md->decompile_count.
3540     if (log())
3541       log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
3542                   Deoptimization::trap_reason_name(reason),
3543                   md->trap_count(reason), trap_count(reason),
3544                   md->decompile_count(), decompile_count());
3545     return true;
3546   } else {
3547     // The coast is clear.
3548     return false;
3549   }
3550 }
3551 
3552 // Compute when not to trap. Used by matching trap based nodes and
3553 // NullCheck optimization.
3554 void Compile::set_allowed_deopt_reasons() {
3555   _allowed_reasons = 0;
3556   if (is_method_compilation()) {
3557     for (int rs = (int)Deoptimization::Reason_none+1; rs < Compile::trapHistLength; rs++) {
3558       assert(rs < BitsPerInt, "recode bit map");
3559       if (!too_many_traps((Deoptimization::DeoptReason) rs)) {
3560         _allowed_reasons |= nth_bit(rs);
3561       }
3562     }
3563   }
3564 }
3565 
3566 #ifndef PRODUCT
3567 //------------------------------verify_graph_edges---------------------------
3568 // Walk the Graph and verify that there is a one-to-one correspondence
3569 // between Use-Def edges and Def-Use edges in the graph.
3570 void Compile::verify_graph_edges(bool no_dead_code) {
3571   if (VerifyGraphEdges) {
3572     ResourceArea *area = Thread::current()->resource_area();
3573     Unique_Node_List visited(area);
3574     // Call recursive graph walk to check edges
3575     _root->verify_edges(visited);
3576     if (no_dead_code) {
3577       // Now make sure that no visited node is used by an unvisited node.
3578       bool dead_nodes = false;
3579       Unique_Node_List checked(area);
3580       while (visited.size() > 0) {
3581         Node* n = visited.pop();
3582         checked.push(n);
3583         for (uint i = 0; i < n->outcnt(); i++) {
3584           Node* use = n->raw_out(i);
3585           if (checked.member(use))  continue;  // already checked
3586           if (visited.member(use))  continue;  // already in the graph
3587           if (use->is_Con())        continue;  // a dead ConNode is OK
3588           // At this point, we have found a dead node which is DU-reachable.
3589           if (!dead_nodes) {
3590             tty->print_cr("*** Dead nodes reachable via DU edges:");
3591             dead_nodes = true;
3592           }
3593           use->dump(2);
3594           tty->print_cr("---");
3595           checked.push(use);  // No repeats; pretend it is now checked.
3596         }
3597       }
3598       assert(!dead_nodes, "using nodes must be reachable from root");
3599     }
3600   }
3601 }
3602 
3603 // Verify GC barriers consistency
3604 // Currently supported:
3605 // - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
3606 void Compile::verify_barriers() {
3607   if (UseG1GC || UseShenandoahGC) {
3608     // Verify G1 pre-barriers
3609     const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
3610 
3611     ResourceArea *area = Thread::current()->resource_area();
3612     Unique_Node_List visited(area);
3613     Node_List worklist(area);
3614     // We're going to walk control flow backwards starting from the Root
3615     worklist.push(_root);
3616     while (worklist.size() > 0) {
3617       Node* x = worklist.pop();
3618       if (x == NULL || x == top()) continue;
3619       if (visited.member(x)) {
3620         continue;
3621       } else {
3622         visited.push(x);
3623       }
3624 
3625       if (x->is_Region()) {
3626         for (uint i = 1; i < x->req(); i++) {
3627           worklist.push(x->in(i));
3628         }
3629       } else {
3630         worklist.push(x->in(0));
3631         // We are looking for the pattern:
3632         //                            /->ThreadLocal
3633         // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
3634         //              \->ConI(0)
3635         // We want to verify that the If and the LoadB have the same control
3636         // See GraphKit::g1_write_barrier_pre()
3637         if (x->is_If()) {
3638           IfNode *iff = x->as_If();
3639           if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
3640             CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
3641             if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
3642                 && cmp->in(1)->is_Load()) {
3643               LoadNode* load = cmp->in(1)->as_Load();
3644               if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
3645                   && load->in(2)->in(3)->is_Con()
3646                   && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
3647 
3648                 Node* if_ctrl = iff->in(0);
3649                 Node* load_ctrl = load->in(0);
3650 
3651                 if (if_ctrl != load_ctrl) {
3652                   // Skip possible CProj->NeverBranch in infinite loops
3653                   if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
3654                       && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
3655                     if_ctrl = if_ctrl->in(0)->in(0);
3656                   }
3657                 }
3658                 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
3659               }
3660             }
3661           }
3662         }
3663       }
3664     }
3665   }
3666 }
3667 
3668 #endif
3669 
3670 // The Compile object keeps track of failure reasons separately from the ciEnv.
3671 // This is required because there is not quite a 1-1 relation between the
3672 // ciEnv and its compilation task and the Compile object.  Note that one
3673 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
3674 // to backtrack and retry without subsuming loads.  Other than this backtracking
3675 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
3676 // by the logic in C2Compiler.
3677 void Compile::record_failure(const char* reason) {
3678   if (log() != NULL) {
3679     log()->elem("failure reason='%s' phase='compile'", reason);
3680   }
3681   if (_failure_reason == NULL) {
3682     // Record the first failure reason.
3683     _failure_reason = reason;
3684   }
3685 
3686   EventCompilerFailure event;
3687   if (event.should_commit()) {
3688     event.set_compileID(Compile::compile_id());
3689     event.set_failure(reason);
3690     event.commit();
3691   }
3692 
3693   if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
3694     C->print_method(PHASE_FAILURE);
3695   }
3696   _root = NULL;  // flush the graph, too
3697 }
3698 
3699 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
3700   : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
3701     _phase_name(name), _dolog(dolog)
3702 {
3703   if (dolog) {
3704     C = Compile::current();
3705     _log = C->log();
3706   } else {
3707     C = NULL;
3708     _log = NULL;
3709   }
3710   if (_log != NULL) {
3711     _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3712     _log->stamp();
3713     _log->end_head();
3714   }
3715 }
3716 
3717 Compile::TracePhase::~TracePhase() {
3718 
3719   C = Compile::current();
3720   if (_dolog) {
3721     _log = C->log();
3722   } else {
3723     _log = NULL;
3724   }
3725 
3726 #ifdef ASSERT
3727   if (PrintIdealNodeCount) {
3728     tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
3729                   _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
3730   }
3731 
3732   if (VerifyIdealNodeCount) {
3733     Compile::current()->print_missing_nodes();
3734   }
3735 #endif
3736 
3737   if (_log != NULL) {
3738     _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
3739   }
3740 }
3741 
3742 //=============================================================================
3743 // Two Constant's are equal when the type and the value are equal.
3744 bool Compile::Constant::operator==(const Constant& other) {
3745   if (type()          != other.type()         )  return false;
3746   if (can_be_reused() != other.can_be_reused())  return false;
3747   // For floating point values we compare the bit pattern.
3748   switch (type()) {
3749   case T_FLOAT:   return (_v._value.i == other._v._value.i);
3750   case T_LONG:
3751   case T_DOUBLE:  return (_v._value.j == other._v._value.j);
3752   case T_OBJECT:
3753   case T_ADDRESS: return (_v._value.l == other._v._value.l);
3754   case T_VOID:    return (_v._value.l == other._v._value.l);  // jump-table entries
3755   case T_METADATA: return (_v._metadata == other._v._metadata);
3756   default: ShouldNotReachHere();
3757   }
3758   return false;
3759 }
3760 
3761 static int type_to_size_in_bytes(BasicType t) {
3762   switch (t) {
3763   case T_LONG:    return sizeof(jlong  );
3764   case T_FLOAT:   return sizeof(jfloat );
3765   case T_DOUBLE:  return sizeof(jdouble);
3766   case T_METADATA: return sizeof(Metadata*);
3767     // We use T_VOID as marker for jump-table entries (labels) which
3768     // need an internal word relocation.
3769   case T_VOID:
3770   case T_ADDRESS:
3771   case T_OBJECT:  return sizeof(jobject);
3772   }
3773 
3774   ShouldNotReachHere();
3775   return -1;
3776 }
3777 
3778 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
3779   // sort descending
3780   if (a->freq() > b->freq())  return -1;
3781   if (a->freq() < b->freq())  return  1;
3782   return 0;
3783 }
3784 
3785 void Compile::ConstantTable::calculate_offsets_and_size() {
3786   // First, sort the array by frequencies.
3787   _constants.sort(qsort_comparator);
3788 
3789 #ifdef ASSERT
3790   // Make sure all jump-table entries were sorted to the end of the
3791   // array (they have a negative frequency).
3792   bool found_void = false;
3793   for (int i = 0; i < _constants.length(); i++) {
3794     Constant con = _constants.at(i);
3795     if (con.type() == T_VOID)
3796       found_void = true;  // jump-tables
3797     else
3798       assert(!found_void, "wrong sorting");
3799   }
3800 #endif
3801 
3802   int offset = 0;
3803   for (int i = 0; i < _constants.length(); i++) {
3804     Constant* con = _constants.adr_at(i);
3805 
3806     // Align offset for type.
3807     int typesize = type_to_size_in_bytes(con->type());
3808     offset = align_size_up(offset, typesize);
3809     con->set_offset(offset);   // set constant's offset
3810 
3811     if (con->type() == T_VOID) {
3812       MachConstantNode* n = (MachConstantNode*) con->get_jobject();
3813       offset = offset + typesize * n->outcnt();  // expand jump-table
3814     } else {
3815       offset = offset + typesize;
3816     }
3817   }
3818 
3819   // Align size up to the next section start (which is insts; see
3820   // CodeBuffer::align_at_start).
3821   assert(_size == -1, "already set?");
3822   _size = align_size_up(offset, CodeEntryAlignment);
3823 }
3824 
3825 void Compile::ConstantTable::emit(CodeBuffer& cb) {
3826   MacroAssembler _masm(&cb);
3827   for (int i = 0; i < _constants.length(); i++) {
3828     Constant con = _constants.at(i);
3829     address constant_addr = NULL;
3830     switch (con.type()) {
3831     case T_LONG:   constant_addr = _masm.long_constant(  con.get_jlong()  ); break;
3832     case T_FLOAT:  constant_addr = _masm.float_constant( con.get_jfloat() ); break;
3833     case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
3834     case T_OBJECT: {
3835       jobject obj = con.get_jobject();
3836       int oop_index = _masm.oop_recorder()->find_index(obj);
3837       constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
3838       break;
3839     }
3840     case T_ADDRESS: {
3841       address addr = (address) con.get_jobject();
3842       constant_addr = _masm.address_constant(addr);
3843       break;
3844     }
3845     // We use T_VOID as marker for jump-table entries (labels) which
3846     // need an internal word relocation.
3847     case T_VOID: {
3848       MachConstantNode* n = (MachConstantNode*) con.get_jobject();
3849       // Fill the jump-table with a dummy word.  The real value is
3850       // filled in later in fill_jump_table.
3851       address dummy = (address) n;
3852       constant_addr = _masm.address_constant(dummy);
3853       // Expand jump-table
3854       for (uint i = 1; i < n->outcnt(); i++) {
3855         address temp_addr = _masm.address_constant(dummy + i);
3856         assert(temp_addr, "consts section too small");
3857       }
3858       break;
3859     }
3860     case T_METADATA: {
3861       Metadata* obj = con.get_metadata();
3862       int metadata_index = _masm.oop_recorder()->find_index(obj);
3863       constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
3864       break;
3865     }
3866     default: ShouldNotReachHere();
3867     }
3868     assert(constant_addr, "consts section too small");
3869     assert((constant_addr - _masm.code()->consts()->start()) == con.offset(),
3870             err_msg_res("must be: %d == %d", (int) (constant_addr - _masm.code()->consts()->start()), (int)(con.offset())));
3871   }
3872 }
3873 
3874 int Compile::ConstantTable::find_offset(Constant& con) const {
3875   int idx = _constants.find(con);
3876   assert(idx != -1, "constant must be in constant table");
3877   int offset = _constants.at(idx).offset();
3878   assert(offset != -1, "constant table not emitted yet?");
3879   return offset;
3880 }
3881 
3882 void Compile::ConstantTable::add(Constant& con) {
3883   if (con.can_be_reused()) {
3884     int idx = _constants.find(con);
3885     if (idx != -1 && _constants.at(idx).can_be_reused()) {
3886       _constants.adr_at(idx)->inc_freq(con.freq());  // increase the frequency by the current value
3887       return;
3888     }
3889   }
3890   (void) _constants.append(con);
3891 }
3892 
3893 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
3894   Block* b = Compile::current()->cfg()->get_block_for_node(n);
3895   Constant con(type, value, b->_freq);
3896   add(con);
3897   return con;
3898 }
3899 
3900 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
3901   Constant con(metadata);
3902   add(con);
3903   return con;
3904 }
3905 
3906 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
3907   jvalue value;
3908   BasicType type = oper->type()->basic_type();
3909   switch (type) {
3910   case T_LONG:    value.j = oper->constantL(); break;
3911   case T_FLOAT:   value.f = oper->constantF(); break;
3912   case T_DOUBLE:  value.d = oper->constantD(); break;
3913   case T_OBJECT:
3914   case T_ADDRESS: value.l = (jobject) oper->constant(); break;
3915   case T_METADATA: return add((Metadata*)oper->constant()); break;
3916   default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
3917   }
3918   return add(n, type, value);
3919 }
3920 
3921 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
3922   jvalue value;
3923   // We can use the node pointer here to identify the right jump-table
3924   // as this method is called from Compile::Fill_buffer right before
3925   // the MachNodes are emitted and the jump-table is filled (means the
3926   // MachNode pointers do not change anymore).
3927   value.l = (jobject) n;
3928   Constant con(T_VOID, value, next_jump_table_freq(), false);  // Labels of a jump-table cannot be reused.
3929   add(con);
3930   return con;
3931 }
3932 
3933 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
3934   // If called from Compile::scratch_emit_size do nothing.
3935   if (Compile::current()->in_scratch_emit_size())  return;
3936 
3937   assert(labels.is_nonempty(), "must be");
3938   assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
3939 
3940   // Since MachConstantNode::constant_offset() also contains
3941   // table_base_offset() we need to subtract the table_base_offset()
3942   // to get the plain offset into the constant table.
3943   int offset = n->constant_offset() - table_base_offset();
3944 
3945   MacroAssembler _masm(&cb);
3946   address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
3947 
3948   for (uint i = 0; i < n->outcnt(); i++) {
3949     address* constant_addr = &jump_table_base[i];
3950     assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, p2i(*constant_addr), p2i(((address) n) + i)));
3951     *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
3952     cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
3953   }
3954 }
3955 
3956 void Compile::dump_inlining() {
3957   if (print_inlining() || print_intrinsics()) {
3958     // Print inlining message for candidates that we couldn't inline
3959     // for lack of space or non constant receiver
3960     for (int i = 0; i < _late_inlines.length(); i++) {
3961       CallGenerator* cg = _late_inlines.at(i);
3962       cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
3963     }
3964     Unique_Node_List useful;
3965     useful.push(root());
3966     for (uint next = 0; next < useful.size(); ++next) {
3967       Node* n  = useful.at(next);
3968       if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
3969         CallNode* call = n->as_Call();
3970         CallGenerator* cg = call->generator();
3971         cg->print_inlining_late("receiver not constant");
3972       }
3973       uint max = n->len();
3974       for ( uint i = 0; i < max; ++i ) {
3975         Node *m = n->in(i);
3976         if ( m == NULL ) continue;
3977         useful.push(m);
3978       }
3979     }
3980     for (int i = 0; i < _print_inlining_list->length(); i++) {
3981       tty->print("%s", _print_inlining_list->adr_at(i)->ss()->as_string());
3982     }
3983   }
3984 }
3985 
3986 // Dump inlining replay data to the stream.
3987 // Don't change thread state and acquire any locks.
3988 void Compile::dump_inline_data(outputStream* out) {
3989   InlineTree* inl_tree = ilt();
3990   if (inl_tree != NULL) {
3991     out->print(" inline %d", inl_tree->count());
3992     inl_tree->dump_replay_data(out);
3993   }
3994 }
3995 
3996 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
3997   if (n1->Opcode() < n2->Opcode())      return -1;
3998   else if (n1->Opcode() > n2->Opcode()) return 1;
3999 
4000   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()));
4001   for (uint i = 1; i < n1->req(); i++) {
4002     if (n1->in(i) < n2->in(i))      return -1;
4003     else if (n1->in(i) > n2->in(i)) return 1;
4004   }
4005 
4006   return 0;
4007 }
4008 
4009 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
4010   Node* n1 = *n1p;
4011   Node* n2 = *n2p;
4012 
4013   return cmp_expensive_nodes(n1, n2);
4014 }
4015 
4016 void Compile::sort_expensive_nodes() {
4017   if (!expensive_nodes_sorted()) {
4018     _expensive_nodes->sort(cmp_expensive_nodes);
4019   }
4020 }
4021 
4022 bool Compile::expensive_nodes_sorted() const {
4023   for (int i = 1; i < _expensive_nodes->length(); i++) {
4024     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
4025       return false;
4026     }
4027   }
4028   return true;
4029 }
4030 
4031 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
4032   if (_expensive_nodes->length() == 0) {
4033     return false;
4034   }
4035 
4036   assert(OptimizeExpensiveOps, "optimization off?");
4037 
4038   // Take this opportunity to remove dead nodes from the list
4039   int j = 0;
4040   for (int i = 0; i < _expensive_nodes->length(); i++) {
4041     Node* n = _expensive_nodes->at(i);
4042     if (!n->is_unreachable(igvn)) {
4043       assert(n->is_expensive(), "should be expensive");
4044       _expensive_nodes->at_put(j, n);
4045       j++;
4046     }
4047   }
4048   _expensive_nodes->trunc_to(j);
4049 
4050   // Then sort the list so that similar nodes are next to each other
4051   // and check for at least two nodes of identical kind with same data
4052   // inputs.
4053   sort_expensive_nodes();
4054 
4055   for (int i = 0; i < _expensive_nodes->length()-1; i++) {
4056     if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
4057       return true;
4058     }
4059   }
4060 
4061   return false;
4062 }
4063 
4064 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
4065   if (_expensive_nodes->length() == 0) {
4066     return;
4067   }
4068 
4069   assert(OptimizeExpensiveOps, "optimization off?");
4070 
4071   // Sort to bring similar nodes next to each other and clear the
4072   // control input of nodes for which there's only a single copy.
4073   sort_expensive_nodes();
4074 
4075   int j = 0;
4076   int identical = 0;
4077   int i = 0;
4078   for (; i < _expensive_nodes->length()-1; i++) {
4079     assert(j <= i, "can't write beyond current index");
4080     if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
4081       identical++;
4082       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4083       continue;
4084     }
4085     if (identical > 0) {
4086       _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4087       identical = 0;
4088     } else {
4089       Node* n = _expensive_nodes->at(i);
4090       igvn.hash_delete(n);
4091       n->set_req(0, NULL);
4092       igvn.hash_insert(n);
4093     }
4094   }
4095   if (identical > 0) {
4096     _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
4097   } else if (_expensive_nodes->length() >= 1) {
4098     Node* n = _expensive_nodes->at(i);
4099     igvn.hash_delete(n);
4100     n->set_req(0, NULL);
4101     igvn.hash_insert(n);
4102   }
4103   _expensive_nodes->trunc_to(j);
4104 }
4105 
4106 void Compile::add_expensive_node(Node * n) {
4107   assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
4108   assert(n->is_expensive(), "expensive nodes with non-null control here only");
4109   assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
4110   if (OptimizeExpensiveOps) {
4111     _expensive_nodes->append(n);
4112   } else {
4113     // Clear control input and let IGVN optimize expensive nodes if
4114     // OptimizeExpensiveOps is off.
4115     n->set_req(0, NULL);
4116   }
4117 }
4118 
4119 /**
4120  * Remove the speculative part of types and clean up the graph
4121  */
4122 void Compile::remove_speculative_types(PhaseIterGVN &igvn) {
4123   if (UseTypeSpeculation) {
4124     Unique_Node_List worklist;
4125     worklist.push(root());
4126     int modified = 0;
4127     // Go over all type nodes that carry a speculative type, drop the
4128     // speculative part of the type and enqueue the node for an igvn
4129     // which may optimize it out.
4130     for (uint next = 0; next < worklist.size(); ++next) {
4131       Node *n  = worklist.at(next);
4132       if (n->is_Type()) {
4133         TypeNode* tn = n->as_Type();
4134         const Type* t = tn->type();
4135         const Type* t_no_spec = t->remove_speculative();
4136         if (t_no_spec != t) {
4137           bool in_hash = igvn.hash_delete(n);
4138           assert(in_hash || n->hash() == Node::NO_HASH, "node should be in igvn hash table");
4139           tn->set_type(t_no_spec);
4140           igvn.hash_insert(n);
4141           igvn._worklist.push(n); // give it a chance to go away
4142           modified++;
4143         }
4144       }
4145       uint max = n->len();
4146       for( uint i = 0; i < max; ++i ) {
4147         Node *m = n->in(i);
4148         if (not_a_node(m))  continue;
4149         worklist.push(m);
4150       }
4151     }
4152     // Drop the speculative part of all types in the igvn's type table
4153     igvn.remove_speculative_types();
4154     if (modified > 0) {
4155       igvn.optimize();
4156     }
4157 #ifdef ASSERT
4158     // Verify that after the IGVN is over no speculative type has resurfaced
4159     worklist.clear();
4160     worklist.push(root());
4161     for (uint next = 0; next < worklist.size(); ++next) {
4162       Node *n  = worklist.at(next);
4163       const Type* t = igvn.type_or_null(n);
4164       assert((t == NULL) || (t == t->remove_speculative()), "no more speculative types");
4165       if (n->is_Type()) {
4166         t = n->as_Type()->type();
4167         assert(t == t->remove_speculative(), "no more speculative types");
4168       }
4169       uint max = n->len();
4170       for( uint i = 0; i < max; ++i ) {
4171         Node *m = n->in(i);
4172         if (not_a_node(m))  continue;
4173         worklist.push(m);
4174       }
4175     }
4176     igvn.check_no_speculative_types();
4177 #endif
4178   }
4179 }
4180 
4181 // Convert integer value to a narrowed long type dependent on ctrl (for example, a range check)
4182 Node* Compile::constrained_convI2L(PhaseGVN* phase, Node* value, const TypeInt* itype, Node* ctrl) {
4183   if (ctrl != NULL) {
4184     // Express control dependency by a CastII node with a narrow type.
4185     value = new (phase->C) CastIINode(value, itype, false, true /* range check dependency */);
4186     // Make the CastII node dependent on the control input to prevent the narrowed ConvI2L
4187     // node from floating above the range check during loop optimizations. Otherwise, the
4188     // ConvI2L node may be eliminated independently of the range check, causing the data path
4189     // to become TOP while the control path is still there (although it's unreachable).
4190     value->set_req(0, ctrl);
4191     // Save CastII node to remove it after loop optimizations.
4192     phase->C->add_range_check_cast(value);
4193     value = phase->transform(value);
4194   }
4195   const TypeLong* ltype = TypeLong::make(itype->_lo, itype->_hi, itype->_widen);
4196   return phase->transform(new (phase->C) ConvI2LNode(value, ltype));
4197 }
4198 
4199 // Auxiliary method to support randomized stressing/fuzzing.
4200 //
4201 // This method can be called the arbitrary number of times, with current count
4202 // as the argument. The logic allows selecting a single candidate from the
4203 // running list of candidates as follows:
4204 //    int count = 0;
4205 //    Cand* selected = null;
4206 //    while(cand = cand->next()) {
4207 //      if (randomized_select(++count)) {
4208 //        selected = cand;
4209 //      }
4210 //    }
4211 //
4212 // Including count equalizes the chances any candidate is "selected".
4213 // This is useful when we don't have the complete list of candidates to choose
4214 // from uniformly. In this case, we need to adjust the randomicity of the
4215 // selection, or else we will end up biasing the selection towards the latter
4216 // candidates.
4217 //
4218 // Quick back-envelope calculation shows that for the list of n candidates
4219 // the equal probability for the candidate to persist as "best" can be
4220 // achieved by replacing it with "next" k-th candidate with the probability
4221 // of 1/k. It can be easily shown that by the end of the run, the
4222 // probability for any candidate is converged to 1/n, thus giving the
4223 // uniform distribution among all the candidates.
4224 //
4225 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
4226 #define RANDOMIZED_DOMAIN_POW 29
4227 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
4228 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
4229 bool Compile::randomized_select(int count) {
4230   assert(count > 0, "only positive");
4231   return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
4232 }
4233 
4234 void Compile::shenandoah_eliminate_g1_wb_pre(Node* call, PhaseIterGVN* igvn) {
4235   assert(UseShenandoahGC && call->is_g1_wb_pre_call(), "");
4236   Node* c = call->as_Call()->proj_out(TypeFunc::Control);
4237   c = c->unique_ctrl_out();
4238   assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4239   c = c->unique_ctrl_out();
4240   assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4241   Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
4242   assert(iff->is_If(), "expect test");
4243   if (!iff->is_shenandoah_marking_if(igvn)) {
4244     c = c->unique_ctrl_out();
4245     assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
4246     iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
4247     assert(iff->is_shenandoah_marking_if(igvn), "expect marking test");
4248   }
4249   Node* cmpx = iff->in(1)->in(1);
4250   igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
4251   igvn->rehash_node_delayed(call);
4252   call->del_req(call->req()-1);
4253 }