1 /*
   2  * Copyright (c) 2001, 2010, 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 "incls/_precompiled.incl"
  26 #include "incls/_graphKit.cpp.incl"
  27 
  28 //----------------------------GraphKit-----------------------------------------
  29 // Main utility constructor.
  30 GraphKit::GraphKit(JVMState* jvms)
  31   : Phase(Phase::Parser),
  32     _env(C->env()),
  33     _gvn(*C->initial_gvn())
  34 {
  35   _exceptions = jvms->map()->next_exception();
  36   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
  37   set_jvms(jvms);
  38 }
  39 
  40 // Private constructor for parser.
  41 GraphKit::GraphKit()
  42   : Phase(Phase::Parser),
  43     _env(C->env()),
  44     _gvn(*C->initial_gvn())
  45 {
  46   _exceptions = NULL;
  47   set_map(NULL);
  48   debug_only(_sp = -99);
  49   debug_only(set_bci(-99));
  50 }
  51 
  52 
  53 
  54 //---------------------------clean_stack---------------------------------------
  55 // Clear away rubbish from the stack area of the JVM state.
  56 // This destroys any arguments that may be waiting on the stack.
  57 void GraphKit::clean_stack(int from_sp) {
  58   SafePointNode* map      = this->map();
  59   JVMState*      jvms     = this->jvms();
  60   int            stk_size = jvms->stk_size();
  61   int            stkoff   = jvms->stkoff();
  62   Node*          top      = this->top();
  63   for (int i = from_sp; i < stk_size; i++) {
  64     if (map->in(stkoff + i) != top) {
  65       map->set_req(stkoff + i, top);
  66     }
  67   }
  68 }
  69 
  70 
  71 //--------------------------------sync_jvms-----------------------------------
  72 // Make sure our current jvms agrees with our parse state.
  73 JVMState* GraphKit::sync_jvms() const {
  74   JVMState* jvms = this->jvms();
  75   jvms->set_bci(bci());       // Record the new bci in the JVMState
  76   jvms->set_sp(sp());         // Record the new sp in the JVMState
  77   assert(jvms_in_sync(), "jvms is now in sync");
  78   return jvms;
  79 }
  80 
  81 #ifdef ASSERT
  82 bool GraphKit::jvms_in_sync() const {
  83   Parse* parse = is_Parse();
  84   if (parse == NULL) {
  85     if (bci() !=      jvms()->bci())          return false;
  86     if (sp()  != (int)jvms()->sp())           return false;
  87     return true;
  88   }
  89   if (jvms()->method() != parse->method())    return false;
  90   if (jvms()->bci()    != parse->bci())       return false;
  91   int jvms_sp = jvms()->sp();
  92   if (jvms_sp          != parse->sp())        return false;
  93   int jvms_depth = jvms()->depth();
  94   if (jvms_depth       != parse->depth())     return false;
  95   return true;
  96 }
  97 
  98 // Local helper checks for special internal merge points
  99 // used to accumulate and merge exception states.
 100 // They are marked by the region's in(0) edge being the map itself.
 101 // Such merge points must never "escape" into the parser at large,
 102 // until they have been handed to gvn.transform.
 103 static bool is_hidden_merge(Node* reg) {
 104   if (reg == NULL)  return false;
 105   if (reg->is_Phi()) {
 106     reg = reg->in(0);
 107     if (reg == NULL)  return false;
 108   }
 109   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
 110 }
 111 
 112 void GraphKit::verify_map() const {
 113   if (map() == NULL)  return;  // null map is OK
 114   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
 115   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
 116   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
 117 }
 118 
 119 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
 120   assert(ex_map->next_exception() == NULL, "not already part of a chain");
 121   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
 122 }
 123 #endif
 124 
 125 //---------------------------stop_and_kill_map---------------------------------
 126 // Set _map to NULL, signalling a stop to further bytecode execution.
 127 // First smash the current map's control to a constant, to mark it dead.
 128 void GraphKit::stop_and_kill_map() {
 129   SafePointNode* dead_map = stop();
 130   if (dead_map != NULL) {
 131     dead_map->disconnect_inputs(NULL); // Mark the map as killed.
 132     assert(dead_map->is_killed(), "must be so marked");
 133   }
 134 }
 135 
 136 
 137 //--------------------------------stopped--------------------------------------
 138 // Tell if _map is NULL, or control is top.
 139 bool GraphKit::stopped() {
 140   if (map() == NULL)           return true;
 141   else if (control() == top()) return true;
 142   else                         return false;
 143 }
 144 
 145 
 146 //-----------------------------has_ex_handler----------------------------------
 147 // Tell if this method or any caller method has exception handlers.
 148 bool GraphKit::has_ex_handler() {
 149   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
 150     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
 151       return true;
 152     }
 153   }
 154   return false;
 155 }
 156 
 157 //------------------------------save_ex_oop------------------------------------
 158 // Save an exception without blowing stack contents or other JVM state.
 159 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
 160   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
 161   ex_map->add_req(ex_oop);
 162   debug_only(verify_exception_state(ex_map));
 163 }
 164 
 165 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
 166   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
 167   Node* ex_oop = ex_map->in(ex_map->req()-1);
 168   if (clear_it)  ex_map->del_req(ex_map->req()-1);
 169   return ex_oop;
 170 }
 171 
 172 //-----------------------------saved_ex_oop------------------------------------
 173 // Recover a saved exception from its map.
 174 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
 175   return common_saved_ex_oop(ex_map, false);
 176 }
 177 
 178 //--------------------------clear_saved_ex_oop---------------------------------
 179 // Erase a previously saved exception from its map.
 180 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
 181   return common_saved_ex_oop(ex_map, true);
 182 }
 183 
 184 #ifdef ASSERT
 185 //---------------------------has_saved_ex_oop----------------------------------
 186 // Erase a previously saved exception from its map.
 187 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
 188   return ex_map->req() == ex_map->jvms()->endoff()+1;
 189 }
 190 #endif
 191 
 192 //-------------------------make_exception_state--------------------------------
 193 // Turn the current JVM state into an exception state, appending the ex_oop.
 194 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
 195   sync_jvms();
 196   SafePointNode* ex_map = stop();  // do not manipulate this map any more
 197   set_saved_ex_oop(ex_map, ex_oop);
 198   return ex_map;
 199 }
 200 
 201 
 202 //--------------------------add_exception_state--------------------------------
 203 // Add an exception to my list of exceptions.
 204 void GraphKit::add_exception_state(SafePointNode* ex_map) {
 205   if (ex_map == NULL || ex_map->control() == top()) {
 206     return;
 207   }
 208 #ifdef ASSERT
 209   verify_exception_state(ex_map);
 210   if (has_exceptions()) {
 211     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
 212   }
 213 #endif
 214 
 215   // If there is already an exception of exactly this type, merge with it.
 216   // In particular, null-checks and other low-level exceptions common up here.
 217   Node*       ex_oop  = saved_ex_oop(ex_map);
 218   const Type* ex_type = _gvn.type(ex_oop);
 219   if (ex_oop == top()) {
 220     // No action needed.
 221     return;
 222   }
 223   assert(ex_type->isa_instptr(), "exception must be an instance");
 224   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
 225     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
 226     // We check sp also because call bytecodes can generate exceptions
 227     // both before and after arguments are popped!
 228     if (ex_type2 == ex_type
 229         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
 230       combine_exception_states(ex_map, e2);
 231       return;
 232     }
 233   }
 234 
 235   // No pre-existing exception of the same type.  Chain it on the list.
 236   push_exception_state(ex_map);
 237 }
 238 
 239 //-----------------------add_exception_states_from-----------------------------
 240 void GraphKit::add_exception_states_from(JVMState* jvms) {
 241   SafePointNode* ex_map = jvms->map()->next_exception();
 242   if (ex_map != NULL) {
 243     jvms->map()->set_next_exception(NULL);
 244     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
 245       next_map = ex_map->next_exception();
 246       ex_map->set_next_exception(NULL);
 247       add_exception_state(ex_map);
 248     }
 249   }
 250 }
 251 
 252 //-----------------------transfer_exceptions_into_jvms-------------------------
 253 JVMState* GraphKit::transfer_exceptions_into_jvms() {
 254   if (map() == NULL) {
 255     // We need a JVMS to carry the exceptions, but the map has gone away.
 256     // Create a scratch JVMS, cloned from any of the exception states...
 257     if (has_exceptions()) {
 258       _map = _exceptions;
 259       _map = clone_map();
 260       _map->set_next_exception(NULL);
 261       clear_saved_ex_oop(_map);
 262       debug_only(verify_map());
 263     } else {
 264       // ...or created from scratch
 265       JVMState* jvms = new (C) JVMState(_method, NULL);
 266       jvms->set_bci(_bci);
 267       jvms->set_sp(_sp);
 268       jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms));
 269       set_jvms(jvms);
 270       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
 271       set_all_memory(top());
 272       while (map()->req() < jvms->endoff())  map()->add_req(top());
 273     }
 274     // (This is a kludge, in case you didn't notice.)
 275     set_control(top());
 276   }
 277   JVMState* jvms = sync_jvms();
 278   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
 279   jvms->map()->set_next_exception(_exceptions);
 280   _exceptions = NULL;   // done with this set of exceptions
 281   return jvms;
 282 }
 283 
 284 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
 285   assert(is_hidden_merge(dstphi), "must be a special merge node");
 286   assert(is_hidden_merge(srcphi), "must be a special merge node");
 287   uint limit = srcphi->req();
 288   for (uint i = PhiNode::Input; i < limit; i++) {
 289     dstphi->add_req(srcphi->in(i));
 290   }
 291 }
 292 static inline void add_one_req(Node* dstphi, Node* src) {
 293   assert(is_hidden_merge(dstphi), "must be a special merge node");
 294   assert(!is_hidden_merge(src), "must not be a special merge node");
 295   dstphi->add_req(src);
 296 }
 297 
 298 //-----------------------combine_exception_states------------------------------
 299 // This helper function combines exception states by building phis on a
 300 // specially marked state-merging region.  These regions and phis are
 301 // untransformed, and can build up gradually.  The region is marked by
 302 // having a control input of its exception map, rather than NULL.  Such
 303 // regions do not appear except in this function, and in use_exception_state.
 304 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
 305   if (failing())  return;  // dying anyway...
 306   JVMState* ex_jvms = ex_map->_jvms;
 307   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
 308   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
 309   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
 310   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
 311   assert(ex_map->req() == phi_map->req(), "matching maps");
 312   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
 313   Node*         hidden_merge_mark = root();
 314   Node*         region  = phi_map->control();
 315   MergeMemNode* phi_mem = phi_map->merged_memory();
 316   MergeMemNode* ex_mem  = ex_map->merged_memory();
 317   if (region->in(0) != hidden_merge_mark) {
 318     // The control input is not (yet) a specially-marked region in phi_map.
 319     // Make it so, and build some phis.
 320     region = new (C, 2) RegionNode(2);
 321     _gvn.set_type(region, Type::CONTROL);
 322     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
 323     region->init_req(1, phi_map->control());
 324     phi_map->set_control(region);
 325     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
 326     record_for_igvn(io_phi);
 327     _gvn.set_type(io_phi, Type::ABIO);
 328     phi_map->set_i_o(io_phi);
 329     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
 330       Node* m = mms.memory();
 331       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
 332       record_for_igvn(m_phi);
 333       _gvn.set_type(m_phi, Type::MEMORY);
 334       mms.set_memory(m_phi);
 335     }
 336   }
 337 
 338   // Either or both of phi_map and ex_map might already be converted into phis.
 339   Node* ex_control = ex_map->control();
 340   // if there is special marking on ex_map also, we add multiple edges from src
 341   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
 342   // how wide was the destination phi_map, originally?
 343   uint orig_width = region->req();
 344 
 345   if (add_multiple) {
 346     add_n_reqs(region, ex_control);
 347     add_n_reqs(phi_map->i_o(), ex_map->i_o());
 348   } else {
 349     // ex_map has no merges, so we just add single edges everywhere
 350     add_one_req(region, ex_control);
 351     add_one_req(phi_map->i_o(), ex_map->i_o());
 352   }
 353   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
 354     if (mms.is_empty()) {
 355       // get a copy of the base memory, and patch some inputs into it
 356       const TypePtr* adr_type = mms.adr_type(C);
 357       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
 358       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
 359       mms.set_memory(phi);
 360       // Prepare to append interesting stuff onto the newly sliced phi:
 361       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
 362     }
 363     // Append stuff from ex_map:
 364     if (add_multiple) {
 365       add_n_reqs(mms.memory(), mms.memory2());
 366     } else {
 367       add_one_req(mms.memory(), mms.memory2());
 368     }
 369   }
 370   uint limit = ex_map->req();
 371   for (uint i = TypeFunc::Parms; i < limit; i++) {
 372     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
 373     if (i == tos)  i = ex_jvms->monoff();
 374     Node* src = ex_map->in(i);
 375     Node* dst = phi_map->in(i);
 376     if (src != dst) {
 377       PhiNode* phi;
 378       if (dst->in(0) != region) {
 379         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
 380         record_for_igvn(phi);
 381         _gvn.set_type(phi, phi->type());
 382         phi_map->set_req(i, dst);
 383         // Prepare to append interesting stuff onto the new phi:
 384         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
 385       } else {
 386         assert(dst->is_Phi(), "nobody else uses a hidden region");
 387         phi = (PhiNode*)dst;
 388       }
 389       if (add_multiple && src->in(0) == ex_control) {
 390         // Both are phis.
 391         add_n_reqs(dst, src);
 392       } else {
 393         while (dst->req() < region->req())  add_one_req(dst, src);
 394       }
 395       const Type* srctype = _gvn.type(src);
 396       if (phi->type() != srctype) {
 397         const Type* dsttype = phi->type()->meet(srctype);
 398         if (phi->type() != dsttype) {
 399           phi->set_type(dsttype);
 400           _gvn.set_type(phi, dsttype);
 401         }
 402       }
 403     }
 404   }
 405 }
 406 
 407 //--------------------------use_exception_state--------------------------------
 408 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
 409   if (failing()) { stop(); return top(); }
 410   Node* region = phi_map->control();
 411   Node* hidden_merge_mark = root();
 412   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
 413   Node* ex_oop = clear_saved_ex_oop(phi_map);
 414   if (region->in(0) == hidden_merge_mark) {
 415     // Special marking for internal ex-states.  Process the phis now.
 416     region->set_req(0, region);  // now it's an ordinary region
 417     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
 418     // Note: Setting the jvms also sets the bci and sp.
 419     set_control(_gvn.transform(region));
 420     uint tos = jvms()->stkoff() + sp();
 421     for (uint i = 1; i < tos; i++) {
 422       Node* x = phi_map->in(i);
 423       if (x->in(0) == region) {
 424         assert(x->is_Phi(), "expected a special phi");
 425         phi_map->set_req(i, _gvn.transform(x));
 426       }
 427     }
 428     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
 429       Node* x = mms.memory();
 430       if (x->in(0) == region) {
 431         assert(x->is_Phi(), "nobody else uses a hidden region");
 432         mms.set_memory(_gvn.transform(x));
 433       }
 434     }
 435     if (ex_oop->in(0) == region) {
 436       assert(ex_oop->is_Phi(), "expected a special phi");
 437       ex_oop = _gvn.transform(ex_oop);
 438     }
 439   } else {
 440     set_jvms(phi_map->jvms());
 441   }
 442 
 443   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
 444   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
 445   return ex_oop;
 446 }
 447 
 448 //---------------------------------java_bc-------------------------------------
 449 Bytecodes::Code GraphKit::java_bc() const {
 450   ciMethod* method = this->method();
 451   int       bci    = this->bci();
 452   if (method != NULL && bci != InvocationEntryBci)
 453     return method->java_code_at_bci(bci);
 454   else
 455     return Bytecodes::_illegal;
 456 }
 457 
 458 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
 459                                                           bool must_throw) {
 460     // if the exception capability is set, then we will generate code
 461     // to check the JavaThread.should_post_on_exceptions flag to see
 462     // if we actually need to report exception events (for this
 463     // thread).  If we don't need to report exception events, we will
 464     // take the normal fast path provided by add_exception_events.  If
 465     // exception event reporting is enabled for this thread, we will
 466     // take the uncommon_trap in the BuildCutout below.
 467 
 468     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
 469     Node* jthread = _gvn.transform(new (C, 1) ThreadLocalNode());
 470     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
 471     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false);
 472 
 473     // Test the should_post_on_exceptions_flag vs. 0
 474     Node* chk = _gvn.transform( new (C, 3) CmpINode(should_post_flag, intcon(0)) );
 475     Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, BoolTest::eq) );
 476 
 477     // Branch to slow_path if should_post_on_exceptions_flag was true
 478     { BuildCutout unless(this, tst, PROB_MAX);
 479       // Do not try anything fancy if we're notifying the VM on every throw.
 480       // Cf. case Bytecodes::_athrow in parse2.cpp.
 481       uncommon_trap(reason, Deoptimization::Action_none,
 482                     (ciKlass*)NULL, (char*)NULL, must_throw);
 483     }
 484 
 485 }
 486 
 487 //------------------------------builtin_throw----------------------------------
 488 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
 489   bool must_throw = true;
 490 
 491   if (env()->jvmti_can_post_on_exceptions()) {
 492     // check if we must post exception events, take uncommon trap if so
 493     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
 494     // here if should_post_on_exceptions is false
 495     // continue on with the normal codegen
 496   }
 497 
 498   // If this particular condition has not yet happened at this
 499   // bytecode, then use the uncommon trap mechanism, and allow for
 500   // a future recompilation if several traps occur here.
 501   // If the throw is hot, try to use a more complicated inline mechanism
 502   // which keeps execution inside the compiled code.
 503   bool treat_throw_as_hot = false;
 504   ciMethodData* md = method()->method_data();
 505 
 506   if (ProfileTraps) {
 507     if (too_many_traps(reason)) {
 508       treat_throw_as_hot = true;
 509     }
 510     // (If there is no MDO at all, assume it is early in
 511     // execution, and that any deopts are part of the
 512     // startup transient, and don't need to be remembered.)
 513 
 514     // Also, if there is a local exception handler, treat all throws
 515     // as hot if there has been at least one in this method.
 516     if (C->trap_count(reason) != 0
 517         && method()->method_data()->trap_count(reason) != 0
 518         && has_ex_handler()) {
 519         treat_throw_as_hot = true;
 520     }
 521   }
 522 
 523   // If this throw happens frequently, an uncommon trap might cause
 524   // a performance pothole.  If there is a local exception handler,
 525   // and if this particular bytecode appears to be deoptimizing often,
 526   // let us handle the throw inline, with a preconstructed instance.
 527   // Note:   If the deopt count has blown up, the uncommon trap
 528   // runtime is going to flush this nmethod, not matter what.
 529   if (treat_throw_as_hot
 530       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
 531     // If the throw is local, we use a pre-existing instance and
 532     // punt on the backtrace.  This would lead to a missing backtrace
 533     // (a repeat of 4292742) if the backtrace object is ever asked
 534     // for its backtrace.
 535     // Fixing this remaining case of 4292742 requires some flavor of
 536     // escape analysis.  Leave that for the future.
 537     ciInstance* ex_obj = NULL;
 538     switch (reason) {
 539     case Deoptimization::Reason_null_check:
 540       ex_obj = env()->NullPointerException_instance();
 541       break;
 542     case Deoptimization::Reason_div0_check:
 543       ex_obj = env()->ArithmeticException_instance();
 544       break;
 545     case Deoptimization::Reason_range_check:
 546       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
 547       break;
 548     case Deoptimization::Reason_class_check:
 549       if (java_bc() == Bytecodes::_aastore) {
 550         ex_obj = env()->ArrayStoreException_instance();
 551       } else {
 552         ex_obj = env()->ClassCastException_instance();
 553       }
 554       break;
 555     }
 556     if (failing()) { stop(); return; }  // exception allocation might fail
 557     if (ex_obj != NULL) {
 558       // Cheat with a preallocated exception object.
 559       if (C->log() != NULL)
 560         C->log()->elem("hot_throw preallocated='1' reason='%s'",
 561                        Deoptimization::trap_reason_name(reason));
 562       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
 563       Node*              ex_node = _gvn.transform( ConNode::make(C, ex_con) );
 564 
 565       // Clear the detail message of the preallocated exception object.
 566       // Weblogic sometimes mutates the detail message of exceptions
 567       // using reflection.
 568       int offset = java_lang_Throwable::get_detailMessage_offset();
 569       const TypePtr* adr_typ = ex_con->add_offset(offset);
 570 
 571       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
 572       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
 573       Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT);
 574 
 575       add_exception_state(make_exception_state(ex_node));
 576       return;
 577     }
 578   }
 579 
 580   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
 581   // It won't be much cheaper than bailing to the interp., since we'll
 582   // have to pass up all the debug-info, and the runtime will have to
 583   // create the stack trace.
 584 
 585   // Usual case:  Bail to interpreter.
 586   // Reserve the right to recompile if we haven't seen anything yet.
 587 
 588   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
 589   if (treat_throw_as_hot
 590       && (method()->method_data()->trap_recompiled_at(bci())
 591           || C->too_many_traps(reason))) {
 592     // We cannot afford to take more traps here.  Suffer in the interpreter.
 593     if (C->log() != NULL)
 594       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
 595                      Deoptimization::trap_reason_name(reason),
 596                      C->trap_count(reason));
 597     action = Deoptimization::Action_none;
 598   }
 599 
 600   // "must_throw" prunes the JVM state to include only the stack, if there
 601   // are no local exception handlers.  This should cut down on register
 602   // allocation time and code size, by drastically reducing the number
 603   // of in-edges on the call to the uncommon trap.
 604 
 605   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
 606 }
 607 
 608 
 609 //----------------------------PreserveJVMState---------------------------------
 610 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
 611   debug_only(kit->verify_map());
 612   _kit    = kit;
 613   _map    = kit->map();   // preserve the map
 614   _sp     = kit->sp();
 615   kit->set_map(clone_map ? kit->clone_map() : NULL);
 616 #ifdef ASSERT
 617   _bci    = kit->bci();
 618   Parse* parser = kit->is_Parse();
 619   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 620   _block  = block;
 621 #endif
 622 }
 623 PreserveJVMState::~PreserveJVMState() {
 624   GraphKit* kit = _kit;
 625 #ifdef ASSERT
 626   assert(kit->bci() == _bci, "bci must not shift");
 627   Parse* parser = kit->is_Parse();
 628   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
 629   assert(block == _block,    "block must not shift");
 630 #endif
 631   kit->set_map(_map);
 632   kit->set_sp(_sp);
 633 }
 634 
 635 
 636 //-----------------------------BuildCutout-------------------------------------
 637 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
 638   : PreserveJVMState(kit)
 639 {
 640   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
 641   SafePointNode* outer_map = _map;   // preserved map is caller's
 642   SafePointNode* inner_map = kit->map();
 643   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
 644   outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
 645   inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
 646 }
 647 BuildCutout::~BuildCutout() {
 648   GraphKit* kit = _kit;
 649   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
 650 }
 651 
 652 //---------------------------PreserveReexecuteState----------------------------
 653 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
 654   assert(!kit->stopped(), "must call stopped() before");
 655   _kit    =    kit;
 656   _sp     =    kit->sp();
 657   _reexecute = kit->jvms()->_reexecute;
 658 }
 659 PreserveReexecuteState::~PreserveReexecuteState() {
 660   if (_kit->stopped()) return;
 661   _kit->jvms()->_reexecute = _reexecute;
 662   _kit->set_sp(_sp);
 663 }
 664 
 665 //------------------------------clone_map--------------------------------------
 666 // Implementation of PreserveJVMState
 667 //
 668 // Only clone_map(...) here. If this function is only used in the
 669 // PreserveJVMState class we may want to get rid of this extra
 670 // function eventually and do it all there.
 671 
 672 SafePointNode* GraphKit::clone_map() {
 673   if (map() == NULL)  return NULL;
 674 
 675   // Clone the memory edge first
 676   Node* mem = MergeMemNode::make(C, map()->memory());
 677   gvn().set_type_bottom(mem);
 678 
 679   SafePointNode *clonemap = (SafePointNode*)map()->clone();
 680   JVMState* jvms = this->jvms();
 681   JVMState* clonejvms = jvms->clone_shallow(C);
 682   clonemap->set_memory(mem);
 683   clonemap->set_jvms(clonejvms);
 684   clonejvms->set_map(clonemap);
 685   record_for_igvn(clonemap);
 686   gvn().set_type_bottom(clonemap);
 687   return clonemap;
 688 }
 689 
 690 
 691 //-----------------------------set_map_clone-----------------------------------
 692 void GraphKit::set_map_clone(SafePointNode* m) {
 693   _map = m;
 694   _map = clone_map();
 695   _map->set_next_exception(NULL);
 696   debug_only(verify_map());
 697 }
 698 
 699 
 700 //----------------------------kill_dead_locals---------------------------------
 701 // Detect any locals which are known to be dead, and force them to top.
 702 void GraphKit::kill_dead_locals() {
 703   // Consult the liveness information for the locals.  If any
 704   // of them are unused, then they can be replaced by top().  This
 705   // should help register allocation time and cut down on the size
 706   // of the deoptimization information.
 707 
 708   // This call is made from many of the bytecode handling
 709   // subroutines called from the Big Switch in do_one_bytecode.
 710   // Every bytecode which might include a slow path is responsible
 711   // for killing its dead locals.  The more consistent we
 712   // are about killing deads, the fewer useless phis will be
 713   // constructed for them at various merge points.
 714 
 715   // bci can be -1 (InvocationEntryBci).  We return the entry
 716   // liveness for the method.
 717 
 718   if (method() == NULL || method()->code_size() == 0) {
 719     // We are building a graph for a call to a native method.
 720     // All locals are live.
 721     return;
 722   }
 723 
 724   ResourceMark rm;
 725 
 726   // Consult the liveness information for the locals.  If any
 727   // of them are unused, then they can be replaced by top().  This
 728   // should help register allocation time and cut down on the size
 729   // of the deoptimization information.
 730   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
 731 
 732   int len = (int)live_locals.size();
 733   assert(len <= jvms()->loc_size(), "too many live locals");
 734   for (int local = 0; local < len; local++) {
 735     if (!live_locals.at(local)) {
 736       set_local(local, top());
 737     }
 738   }
 739 }
 740 
 741 #ifdef ASSERT
 742 //-------------------------dead_locals_are_killed------------------------------
 743 // Return true if all dead locals are set to top in the map.
 744 // Used to assert "clean" debug info at various points.
 745 bool GraphKit::dead_locals_are_killed() {
 746   if (method() == NULL || method()->code_size() == 0) {
 747     // No locals need to be dead, so all is as it should be.
 748     return true;
 749   }
 750 
 751   // Make sure somebody called kill_dead_locals upstream.
 752   ResourceMark rm;
 753   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
 754     if (jvms->loc_size() == 0)  continue;  // no locals to consult
 755     SafePointNode* map = jvms->map();
 756     ciMethod* method = jvms->method();
 757     int       bci    = jvms->bci();
 758     if (jvms == this->jvms()) {
 759       bci = this->bci();  // it might not yet be synched
 760     }
 761     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
 762     int len = (int)live_locals.size();
 763     if (!live_locals.is_valid() || len == 0)
 764       // This method is trivial, or is poisoned by a breakpoint.
 765       return true;
 766     assert(len == jvms->loc_size(), "live map consistent with locals map");
 767     for (int local = 0; local < len; local++) {
 768       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
 769         if (PrintMiscellaneous && (Verbose || WizardMode)) {
 770           tty->print_cr("Zombie local %d: ", local);
 771           jvms->dump();
 772         }
 773         return false;
 774       }
 775     }
 776   }
 777   return true;
 778 }
 779 
 780 #endif //ASSERT
 781 
 782 // Helper function for enforcing certain bytecodes to reexecute if
 783 // deoptimization happens
 784 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
 785   ciMethod* cur_method = jvms->method();
 786   int       cur_bci   = jvms->bci();
 787   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
 788     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
 789     return Interpreter::bytecode_should_reexecute(code) ||
 790            is_anewarray && code == Bytecodes::_multianewarray;
 791     // Reexecute _multianewarray bytecode which was replaced with
 792     // sequence of [a]newarray. See Parse::do_multianewarray().
 793     //
 794     // Note: interpreter should not have it set since this optimization
 795     // is limited by dimensions and guarded by flag so in some cases
 796     // multianewarray() runtime calls will be generated and
 797     // the bytecode should not be reexecutes (stack will not be reset).
 798   } else
 799     return false;
 800 }
 801 
 802 // Helper function for adding JVMState and debug information to node
 803 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
 804   // Add the safepoint edges to the call (or other safepoint).
 805 
 806   // Make sure dead locals are set to top.  This
 807   // should help register allocation time and cut down on the size
 808   // of the deoptimization information.
 809   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
 810 
 811   // Walk the inline list to fill in the correct set of JVMState's
 812   // Also fill in the associated edges for each JVMState.
 813 
 814   JVMState* youngest_jvms = sync_jvms();
 815 
 816   // If we are guaranteed to throw, we can prune everything but the
 817   // input to the current bytecode.
 818   bool can_prune_locals = false;
 819   uint stack_slots_not_pruned = 0;
 820   int inputs = 0, depth = 0;
 821   if (must_throw) {
 822     assert(method() == youngest_jvms->method(), "sanity");
 823     if (compute_stack_effects(inputs, depth)) {
 824       can_prune_locals = true;
 825       stack_slots_not_pruned = inputs;
 826     }
 827   }
 828 
 829   if (env()->jvmti_can_access_local_variables()) {
 830     // At any safepoint, this method can get breakpointed, which would
 831     // then require an immediate deoptimization.
 832     can_prune_locals = false;  // do not prune locals
 833     stack_slots_not_pruned = 0;
 834   }
 835 
 836   // do not scribble on the input jvms
 837   JVMState* out_jvms = youngest_jvms->clone_deep(C);
 838   call->set_jvms(out_jvms); // Start jvms list for call node
 839 
 840   // For a known set of bytecodes, the interpreter should reexecute them if
 841   // deoptimization happens. We set the reexecute state for them here
 842   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
 843       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
 844     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
 845   }
 846 
 847   // Presize the call:
 848   debug_only(uint non_debug_edges = call->req());
 849   call->add_req_batch(top(), youngest_jvms->debug_depth());
 850   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
 851 
 852   // Set up edges so that the call looks like this:
 853   //  Call [state:] ctl io mem fptr retadr
 854   //       [parms:] parm0 ... parmN
 855   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 856   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
 857   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
 858   // Note that caller debug info precedes callee debug info.
 859 
 860   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
 861   uint debug_ptr = call->req();
 862 
 863   // Loop over the map input edges associated with jvms, add them
 864   // to the call node, & reset all offsets to match call node array.
 865   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
 866     uint debug_end   = debug_ptr;
 867     uint debug_start = debug_ptr - in_jvms->debug_size();
 868     debug_ptr = debug_start;  // back up the ptr
 869 
 870     uint p = debug_start;  // walks forward in [debug_start, debug_end)
 871     uint j, k, l;
 872     SafePointNode* in_map = in_jvms->map();
 873     out_jvms->set_map(call);
 874 
 875     if (can_prune_locals) {
 876       assert(in_jvms->method() == out_jvms->method(), "sanity");
 877       // If the current throw can reach an exception handler in this JVMS,
 878       // then we must keep everything live that can reach that handler.
 879       // As a quick and dirty approximation, we look for any handlers at all.
 880       if (in_jvms->method()->has_exception_handlers()) {
 881         can_prune_locals = false;
 882       }
 883     }
 884 
 885     // Add the Locals
 886     k = in_jvms->locoff();
 887     l = in_jvms->loc_size();
 888     out_jvms->set_locoff(p);
 889     if (!can_prune_locals) {
 890       for (j = 0; j < l; j++)
 891         call->set_req(p++, in_map->in(k+j));
 892     } else {
 893       p += l;  // already set to top above by add_req_batch
 894     }
 895 
 896     // Add the Expression Stack
 897     k = in_jvms->stkoff();
 898     l = in_jvms->sp();
 899     out_jvms->set_stkoff(p);
 900     if (!can_prune_locals) {
 901       for (j = 0; j < l; j++)
 902         call->set_req(p++, in_map->in(k+j));
 903     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
 904       // Divide stack into {S0,...,S1}, where S0 is set to top.
 905       uint s1 = stack_slots_not_pruned;
 906       stack_slots_not_pruned = 0;  // for next iteration
 907       if (s1 > l)  s1 = l;
 908       uint s0 = l - s1;
 909       p += s0;  // skip the tops preinstalled by add_req_batch
 910       for (j = s0; j < l; j++)
 911         call->set_req(p++, in_map->in(k+j));
 912     } else {
 913       p += l;  // already set to top above by add_req_batch
 914     }
 915 
 916     // Add the Monitors
 917     k = in_jvms->monoff();
 918     l = in_jvms->mon_size();
 919     out_jvms->set_monoff(p);
 920     for (j = 0; j < l; j++)
 921       call->set_req(p++, in_map->in(k+j));
 922 
 923     // Copy any scalar object fields.
 924     k = in_jvms->scloff();
 925     l = in_jvms->scl_size();
 926     out_jvms->set_scloff(p);
 927     for (j = 0; j < l; j++)
 928       call->set_req(p++, in_map->in(k+j));
 929 
 930     // Finish the new jvms.
 931     out_jvms->set_endoff(p);
 932 
 933     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
 934     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
 935     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
 936     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
 937     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
 938     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
 939 
 940     // Update the two tail pointers in parallel.
 941     out_jvms = out_jvms->caller();
 942     in_jvms  = in_jvms->caller();
 943   }
 944 
 945   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
 946 
 947   // Test the correctness of JVMState::debug_xxx accessors:
 948   assert(call->jvms()->debug_start() == non_debug_edges, "");
 949   assert(call->jvms()->debug_end()   == call->req(), "");
 950   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
 951 }
 952 
 953 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
 954   Bytecodes::Code code = java_bc();
 955   if (code == Bytecodes::_wide) {
 956     code = method()->java_code_at_bci(bci() + 1);
 957   }
 958 
 959   BasicType rtype = T_ILLEGAL;
 960   int       rsize = 0;
 961 
 962   if (code != Bytecodes::_illegal) {
 963     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
 964     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
 965     if (rtype < T_CONFLICT)
 966       rsize = type2size[rtype];
 967   }
 968 
 969   switch (code) {
 970   case Bytecodes::_illegal:
 971     return false;
 972 
 973   case Bytecodes::_ldc:
 974   case Bytecodes::_ldc_w:
 975   case Bytecodes::_ldc2_w:
 976     inputs = 0;
 977     break;
 978 
 979   case Bytecodes::_dup:         inputs = 1;  break;
 980   case Bytecodes::_dup_x1:      inputs = 2;  break;
 981   case Bytecodes::_dup_x2:      inputs = 3;  break;
 982   case Bytecodes::_dup2:        inputs = 2;  break;
 983   case Bytecodes::_dup2_x1:     inputs = 3;  break;
 984   case Bytecodes::_dup2_x2:     inputs = 4;  break;
 985   case Bytecodes::_swap:        inputs = 2;  break;
 986   case Bytecodes::_arraylength: inputs = 1;  break;
 987 
 988   case Bytecodes::_getstatic:
 989   case Bytecodes::_putstatic:
 990   case Bytecodes::_getfield:
 991   case Bytecodes::_putfield:
 992     {
 993       bool is_get = (depth >= 0), is_static = (depth & 1);
 994       bool ignore;
 995       ciBytecodeStream iter(method());
 996       iter.reset_to_bci(bci());
 997       iter.next();
 998       ciField* field = iter.get_field(ignore);
 999       int      size  = field->type()->size();
1000       inputs  = (is_static ? 0 : 1);
1001       if (is_get) {
1002         depth = size - inputs;
1003       } else {
1004         inputs += size;        // putxxx pops the value from the stack
1005         depth = - inputs;
1006       }
1007     }
1008     break;
1009 
1010   case Bytecodes::_invokevirtual:
1011   case Bytecodes::_invokespecial:
1012   case Bytecodes::_invokestatic:
1013   case Bytecodes::_invokedynamic:
1014   case Bytecodes::_invokeinterface:
1015     {
1016       bool ignore;
1017       ciBytecodeStream iter(method());
1018       iter.reset_to_bci(bci());
1019       iter.next();
1020       ciMethod* method = iter.get_method(ignore);
1021       inputs = method->arg_size_no_receiver();
1022       // Add a receiver argument, maybe:
1023       if (code != Bytecodes::_invokestatic &&
1024           code != Bytecodes::_invokedynamic)
1025         inputs += 1;
1026       // (Do not use ciMethod::arg_size(), because
1027       // it might be an unloaded method, which doesn't
1028       // know whether it is static or not.)
1029       int size = method->return_type()->size();
1030       depth = size - inputs;
1031     }
1032     break;
1033 
1034   case Bytecodes::_multianewarray:
1035     {
1036       ciBytecodeStream iter(method());
1037       iter.reset_to_bci(bci());
1038       iter.next();
1039       inputs = iter.get_dimensions();
1040       assert(rsize == 1, "");
1041       depth = rsize - inputs;
1042     }
1043     break;
1044 
1045   case Bytecodes::_ireturn:
1046   case Bytecodes::_lreturn:
1047   case Bytecodes::_freturn:
1048   case Bytecodes::_dreturn:
1049   case Bytecodes::_areturn:
1050     assert(rsize = -depth, "");
1051     inputs = rsize;
1052     break;
1053 
1054   case Bytecodes::_jsr:
1055   case Bytecodes::_jsr_w:
1056     inputs = 0;
1057     depth  = 1;                  // S.B. depth=1, not zero
1058     break;
1059 
1060   default:
1061     // bytecode produces a typed result
1062     inputs = rsize - depth;
1063     assert(inputs >= 0, "");
1064     break;
1065   }
1066 
1067 #ifdef ASSERT
1068   // spot check
1069   int outputs = depth + inputs;
1070   assert(outputs >= 0, "sanity");
1071   switch (code) {
1072   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1073   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1074   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1075   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1076   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1077   }
1078 #endif //ASSERT
1079 
1080   return true;
1081 }
1082 
1083 
1084 
1085 //------------------------------basic_plus_adr---------------------------------
1086 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1087   // short-circuit a common case
1088   if (offset == intcon(0))  return ptr;
1089   return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1090 }
1091 
1092 Node* GraphKit::ConvI2L(Node* offset) {
1093   // short-circuit a common case
1094   jint offset_con = find_int_con(offset, Type::OffsetBot);
1095   if (offset_con != Type::OffsetBot) {
1096     return longcon((long) offset_con);
1097   }
1098   return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1099 }
1100 Node* GraphKit::ConvL2I(Node* offset) {
1101   // short-circuit a common case
1102   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1103   if (offset_con != (jlong)Type::OffsetBot) {
1104     return intcon((int) offset_con);
1105   }
1106   return _gvn.transform( new (C, 2) ConvL2INode(offset));
1107 }
1108 
1109 //-------------------------load_object_klass-----------------------------------
1110 Node* GraphKit::load_object_klass(Node* obj) {
1111   // Special-case a fresh allocation to avoid building nodes:
1112   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1113   if (akls != NULL)  return akls;
1114   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1115   return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1116 }
1117 
1118 //-------------------------load_array_length-----------------------------------
1119 Node* GraphKit::load_array_length(Node* array) {
1120   // Special-case a fresh allocation to avoid building nodes:
1121   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1122   Node *alen;
1123   if (alloc == NULL) {
1124     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1125     alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1126   } else {
1127     alen = alloc->Ideal_length();
1128     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1129     if (ccast != alen) {
1130       alen = _gvn.transform(ccast);
1131     }
1132   }
1133   return alen;
1134 }
1135 
1136 //------------------------------do_null_check----------------------------------
1137 // Helper function to do a NULL pointer check.  Returned value is
1138 // the incoming address with NULL casted away.  You are allowed to use the
1139 // not-null value only if you are control dependent on the test.
1140 extern int explicit_null_checks_inserted,
1141            explicit_null_checks_elided;
1142 Node* GraphKit::null_check_common(Node* value, BasicType type,
1143                                   // optional arguments for variations:
1144                                   bool assert_null,
1145                                   Node* *null_control) {
1146   assert(!assert_null || null_control == NULL, "not both at once");
1147   if (stopped())  return top();
1148   if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1149     // For some performance testing, we may wish to suppress null checking.
1150     value = cast_not_null(value);   // Make it appear to be non-null (4962416).
1151     return value;
1152   }
1153   explicit_null_checks_inserted++;
1154 
1155   // Construct NULL check
1156   Node *chk = NULL;
1157   switch(type) {
1158     case T_LONG   : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1159     case T_INT    : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1160     case T_ARRAY  : // fall through
1161       type = T_OBJECT;  // simplify further tests
1162     case T_OBJECT : {
1163       const Type *t = _gvn.type( value );
1164 
1165       const TypeOopPtr* tp = t->isa_oopptr();
1166       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1167           // Only for do_null_check, not any of its siblings:
1168           && !assert_null && null_control == NULL) {
1169         // Usually, any field access or invocation on an unloaded oop type
1170         // will simply fail to link, since the statically linked class is
1171         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1172         // the static class is loaded but the sharper oop type is not.
1173         // Rather than checking for this obscure case in lots of places,
1174         // we simply observe that a null check on an unloaded class
1175         // will always be followed by a nonsense operation, so we
1176         // can just issue the uncommon trap here.
1177         // Our access to the unloaded class will only be correct
1178         // after it has been loaded and initialized, which requires
1179         // a trip through the interpreter.
1180 #ifndef PRODUCT
1181         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1182 #endif
1183         uncommon_trap(Deoptimization::Reason_unloaded,
1184                       Deoptimization::Action_reinterpret,
1185                       tp->klass(), "!loaded");
1186         return top();
1187       }
1188 
1189       if (assert_null) {
1190         // See if the type is contained in NULL_PTR.
1191         // If so, then the value is already null.
1192         if (t->higher_equal(TypePtr::NULL_PTR)) {
1193           explicit_null_checks_elided++;
1194           return value;           // Elided null assert quickly!
1195         }
1196       } else {
1197         // See if mixing in the NULL pointer changes type.
1198         // If so, then the NULL pointer was not allowed in the original
1199         // type.  In other words, "value" was not-null.
1200         if (t->meet(TypePtr::NULL_PTR) != t) {
1201           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1202           explicit_null_checks_elided++;
1203           return value;           // Elided null check quickly!
1204         }
1205       }
1206       chk = new (C, 3) CmpPNode( value, null() );
1207       break;
1208     }
1209 
1210     default      : ShouldNotReachHere();
1211   }
1212   assert(chk != NULL, "sanity check");
1213   chk = _gvn.transform(chk);
1214 
1215   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1216   BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1217   Node   *tst = _gvn.transform( btst );
1218 
1219   //-----------
1220   // if peephole optimizations occurred, a prior test existed.
1221   // If a prior test existed, maybe it dominates as we can avoid this test.
1222   if (tst != btst && type == T_OBJECT) {
1223     // At this point we want to scan up the CFG to see if we can
1224     // find an identical test (and so avoid this test altogether).
1225     Node *cfg = control();
1226     int depth = 0;
1227     while( depth < 16 ) {       // Limit search depth for speed
1228       if( cfg->Opcode() == Op_IfTrue &&
1229           cfg->in(0)->in(1) == tst ) {
1230         // Found prior test.  Use "cast_not_null" to construct an identical
1231         // CastPP (and hence hash to) as already exists for the prior test.
1232         // Return that casted value.
1233         if (assert_null) {
1234           replace_in_map(value, null());
1235           return null();  // do not issue the redundant test
1236         }
1237         Node *oldcontrol = control();
1238         set_control(cfg);
1239         Node *res = cast_not_null(value);
1240         set_control(oldcontrol);
1241         explicit_null_checks_elided++;
1242         return res;
1243       }
1244       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1245       if (cfg == NULL)  break;  // Quit at region nodes
1246       depth++;
1247     }
1248   }
1249 
1250   //-----------
1251   // Branch to failure if null
1252   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1253   Deoptimization::DeoptReason reason;
1254   if (assert_null)
1255     reason = Deoptimization::Reason_null_assert;
1256   else if (type == T_OBJECT)
1257     reason = Deoptimization::Reason_null_check;
1258   else
1259     reason = Deoptimization::Reason_div0_check;
1260 
1261   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1262   // ciMethodData::has_trap_at will return a conservative -1 if any
1263   // must-be-null assertion has failed.  This could cause performance
1264   // problems for a method after its first do_null_assert failure.
1265   // Consider using 'Reason_class_check' instead?
1266 
1267   // To cause an implicit null check, we set the not-null probability
1268   // to the maximum (PROB_MAX).  For an explicit check the probability
1269   // is set to a smaller value.
1270   if (null_control != NULL || too_many_traps(reason)) {
1271     // probability is less likely
1272     ok_prob =  PROB_LIKELY_MAG(3);
1273   } else if (!assert_null &&
1274              (ImplicitNullCheckThreshold > 0) &&
1275              method() != NULL &&
1276              (method()->method_data()->trap_count(reason)
1277               >= (uint)ImplicitNullCheckThreshold)) {
1278     ok_prob =  PROB_LIKELY_MAG(3);
1279   }
1280 
1281   if (null_control != NULL) {
1282     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1283     Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1284     set_control(      _gvn.transform( new (C, 1) IfTrueNode(iff)));
1285     if (null_true == top())
1286       explicit_null_checks_elided++;
1287     (*null_control) = null_true;
1288   } else {
1289     BuildCutout unless(this, tst, ok_prob);
1290     // Check for optimizer eliding test at parse time
1291     if (stopped()) {
1292       // Failure not possible; do not bother making uncommon trap.
1293       explicit_null_checks_elided++;
1294     } else if (assert_null) {
1295       uncommon_trap(reason,
1296                     Deoptimization::Action_make_not_entrant,
1297                     NULL, "assert_null");
1298     } else {
1299       replace_in_map(value, zerocon(type));
1300       builtin_throw(reason);
1301     }
1302   }
1303 
1304   // Must throw exception, fall-thru not possible?
1305   if (stopped()) {
1306     return top();               // No result
1307   }
1308 
1309   if (assert_null) {
1310     // Cast obj to null on this path.
1311     replace_in_map(value, zerocon(type));
1312     return zerocon(type);
1313   }
1314 
1315   // Cast obj to not-null on this path, if there is no null_control.
1316   // (If there is a null_control, a non-null value may come back to haunt us.)
1317   if (type == T_OBJECT) {
1318     Node* cast = cast_not_null(value, false);
1319     if (null_control == NULL || (*null_control) == top())
1320       replace_in_map(value, cast);
1321     value = cast;
1322   }
1323 
1324   return value;
1325 }
1326 
1327 
1328 //------------------------------cast_not_null----------------------------------
1329 // Cast obj to not-null on this path
1330 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1331   const Type *t = _gvn.type(obj);
1332   const Type *t_not_null = t->join(TypePtr::NOTNULL);
1333   // Object is already not-null?
1334   if( t == t_not_null ) return obj;
1335 
1336   Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1337   cast->init_req(0, control());
1338   cast = _gvn.transform( cast );
1339 
1340   // Scan for instances of 'obj' in the current JVM mapping.
1341   // These instances are known to be not-null after the test.
1342   if (do_replace_in_map)
1343     replace_in_map(obj, cast);
1344 
1345   return cast;                  // Return casted value
1346 }
1347 
1348 
1349 //--------------------------replace_in_map-------------------------------------
1350 void GraphKit::replace_in_map(Node* old, Node* neww) {
1351   this->map()->replace_edge(old, neww);
1352 
1353   // Note: This operation potentially replaces any edge
1354   // on the map.  This includes locals, stack, and monitors
1355   // of the current (innermost) JVM state.
1356 
1357   // We can consider replacing in caller maps.
1358   // The idea would be that an inlined function's null checks
1359   // can be shared with the entire inlining tree.
1360   // The expense of doing this is that the PreserveJVMState class
1361   // would have to preserve caller states too, with a deep copy.
1362 }
1363 
1364 
1365 
1366 //=============================================================================
1367 //--------------------------------memory---------------------------------------
1368 Node* GraphKit::memory(uint alias_idx) {
1369   MergeMemNode* mem = merged_memory();
1370   Node* p = mem->memory_at(alias_idx);
1371   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1372   return p;
1373 }
1374 
1375 //-----------------------------reset_memory------------------------------------
1376 Node* GraphKit::reset_memory() {
1377   Node* mem = map()->memory();
1378   // do not use this node for any more parsing!
1379   debug_only( map()->set_memory((Node*)NULL) );
1380   return _gvn.transform( mem );
1381 }
1382 
1383 //------------------------------set_all_memory---------------------------------
1384 void GraphKit::set_all_memory(Node* newmem) {
1385   Node* mergemem = MergeMemNode::make(C, newmem);
1386   gvn().set_type_bottom(mergemem);
1387   map()->set_memory(mergemem);
1388 }
1389 
1390 //------------------------------set_all_memory_call----------------------------
1391 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1392   Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1393   set_all_memory(newmem);
1394 }
1395 
1396 //=============================================================================
1397 //
1398 // parser factory methods for MemNodes
1399 //
1400 // These are layered on top of the factory methods in LoadNode and StoreNode,
1401 // and integrate with the parser's memory state and _gvn engine.
1402 //
1403 
1404 // factory methods in "int adr_idx"
1405 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1406                           int adr_idx,
1407                           bool require_atomic_access) {
1408   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1409   const TypePtr* adr_type = NULL; // debug-mode-only argument
1410   debug_only(adr_type = C->get_adr_type(adr_idx));
1411   Node* mem = memory(adr_idx);
1412   Node* ld;
1413   if (require_atomic_access && bt == T_LONG) {
1414     ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1415   } else {
1416     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1417   }
1418   return _gvn.transform(ld);
1419 }
1420 
1421 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1422                                 int adr_idx,
1423                                 bool require_atomic_access) {
1424   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1425   const TypePtr* adr_type = NULL;
1426   debug_only(adr_type = C->get_adr_type(adr_idx));
1427   Node *mem = memory(adr_idx);
1428   Node* st;
1429   if (require_atomic_access && bt == T_LONG) {
1430     st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1431   } else {
1432     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1433   }
1434   st = _gvn.transform(st);
1435   set_memory(st, adr_idx);
1436   // Back-to-back stores can only remove intermediate store with DU info
1437   // so push on worklist for optimizer.
1438   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1439     record_for_igvn(st);
1440 
1441   return st;
1442 }
1443 
1444 
1445 void GraphKit::pre_barrier(Node* ctl,
1446                            Node* obj,
1447                            Node* adr,
1448                            uint  adr_idx,
1449                            Node* val,
1450                            const TypeOopPtr* val_type,
1451                            BasicType bt) {
1452   BarrierSet* bs = Universe::heap()->barrier_set();
1453   set_control(ctl);
1454   switch (bs->kind()) {
1455     case BarrierSet::G1SATBCT:
1456     case BarrierSet::G1SATBCTLogging:
1457       g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt);
1458       break;
1459 
1460     case BarrierSet::CardTableModRef:
1461     case BarrierSet::CardTableExtension:
1462     case BarrierSet::ModRef:
1463       break;
1464 
1465     case BarrierSet::Other:
1466     default      :
1467       ShouldNotReachHere();
1468 
1469   }
1470 }
1471 
1472 void GraphKit::post_barrier(Node* ctl,
1473                             Node* store,
1474                             Node* obj,
1475                             Node* adr,
1476                             uint  adr_idx,
1477                             Node* val,
1478                             BasicType bt,
1479                             bool use_precise) {
1480   BarrierSet* bs = Universe::heap()->barrier_set();
1481   set_control(ctl);
1482   switch (bs->kind()) {
1483     case BarrierSet::G1SATBCT:
1484     case BarrierSet::G1SATBCTLogging:
1485       g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1486       break;
1487 
1488     case BarrierSet::CardTableModRef:
1489     case BarrierSet::CardTableExtension:
1490       write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1491       break;
1492 
1493     case BarrierSet::ModRef:
1494       break;
1495 
1496     case BarrierSet::Other:
1497     default      :
1498       ShouldNotReachHere();
1499 
1500   }
1501 }
1502 
1503 Node* GraphKit::store_oop(Node* ctl,
1504                           Node* obj,
1505                           Node* adr,
1506                           const TypePtr* adr_type,
1507                           Node* val,
1508                           const TypeOopPtr* val_type,
1509                           BasicType bt,
1510                           bool use_precise) {
1511 
1512   set_control(ctl);
1513   if (stopped()) return top(); // Dead path ?
1514 
1515   assert(bt == T_OBJECT, "sanity");
1516   assert(val != NULL, "not dead path");
1517   uint adr_idx = C->get_alias_index(adr_type);
1518   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1519 
1520   pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt);
1521   Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1522   post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1523   return store;
1524 }
1525 
1526 // Could be an array or object we don't know at compile time (unsafe ref.)
1527 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1528                              Node* obj,   // containing obj
1529                              Node* adr,  // actual adress to store val at
1530                              const TypePtr* adr_type,
1531                              Node* val,
1532                              BasicType bt) {
1533   Compile::AliasType* at = C->alias_type(adr_type);
1534   const TypeOopPtr* val_type = NULL;
1535   if (adr_type->isa_instptr()) {
1536     if (at->field() != NULL) {
1537       // known field.  This code is a copy of the do_put_xxx logic.
1538       ciField* field = at->field();
1539       if (!field->type()->is_loaded()) {
1540         val_type = TypeInstPtr::BOTTOM;
1541       } else {
1542         val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1543       }
1544     }
1545   } else if (adr_type->isa_aryptr()) {
1546     val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1547   }
1548   if (val_type == NULL) {
1549     val_type = TypeInstPtr::BOTTOM;
1550   }
1551   return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1552 }
1553 
1554 
1555 //-------------------------array_element_address-------------------------
1556 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1557                                       const TypeInt* sizetype) {
1558   uint shift  = exact_log2(type2aelembytes(elembt));
1559   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1560 
1561   // short-circuit a common case (saves lots of confusing waste motion)
1562   jint idx_con = find_int_con(idx, -1);
1563   if (idx_con >= 0) {
1564     intptr_t offset = header + ((intptr_t)idx_con << shift);
1565     return basic_plus_adr(ary, offset);
1566   }
1567 
1568   // must be correct type for alignment purposes
1569   Node* base  = basic_plus_adr(ary, header);
1570 #ifdef _LP64
1571   // The scaled index operand to AddP must be a clean 64-bit value.
1572   // Java allows a 32-bit int to be incremented to a negative
1573   // value, which appears in a 64-bit register as a large
1574   // positive number.  Using that large positive number as an
1575   // operand in pointer arithmetic has bad consequences.
1576   // On the other hand, 32-bit overflow is rare, and the possibility
1577   // can often be excluded, if we annotate the ConvI2L node with
1578   // a type assertion that its value is known to be a small positive
1579   // number.  (The prior range check has ensured this.)
1580   // This assertion is used by ConvI2LNode::Ideal.
1581   int index_max = max_jint - 1;  // array size is max_jint, index is one less
1582   if (sizetype != NULL)  index_max = sizetype->_hi - 1;
1583   const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1584   idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1585 #endif
1586   Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1587   return basic_plus_adr(ary, base, scale);
1588 }
1589 
1590 //-------------------------load_array_element-------------------------
1591 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1592   const Type* elemtype = arytype->elem();
1593   BasicType elembt = elemtype->array_element_basic_type();
1594   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1595   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1596   return ld;
1597 }
1598 
1599 //-------------------------set_arguments_for_java_call-------------------------
1600 // Arguments (pre-popped from the stack) are taken from the JVMS.
1601 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1602   // Add the call arguments:
1603   uint nargs = call->method()->arg_size();
1604   for (uint i = 0; i < nargs; i++) {
1605     Node* arg = argument(i);
1606     call->init_req(i + TypeFunc::Parms, arg);
1607   }
1608 }
1609 
1610 //---------------------------set_edges_for_java_call---------------------------
1611 // Connect a newly created call into the current JVMS.
1612 // A return value node (if any) is returned from set_edges_for_java_call.
1613 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1614 
1615   // Add the predefined inputs:
1616   call->init_req( TypeFunc::Control, control() );
1617   call->init_req( TypeFunc::I_O    , i_o() );
1618   call->init_req( TypeFunc::Memory , reset_memory() );
1619   call->init_req( TypeFunc::FramePtr, frameptr() );
1620   call->init_req( TypeFunc::ReturnAdr, top() );
1621 
1622   add_safepoint_edges(call, must_throw);
1623 
1624   Node* xcall = _gvn.transform(call);
1625 
1626   if (xcall == top()) {
1627     set_control(top());
1628     return;
1629   }
1630   assert(xcall == call, "call identity is stable");
1631 
1632   // Re-use the current map to produce the result.
1633 
1634   set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1635   set_i_o(    _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1636   set_all_memory_call(xcall, separate_io_proj);
1637 
1638   //return xcall;   // no need, caller already has it
1639 }
1640 
1641 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1642   if (stopped())  return top();  // maybe the call folded up?
1643 
1644   // Capture the return value, if any.
1645   Node* ret;
1646   if (call->method() == NULL ||
1647       call->method()->return_type()->basic_type() == T_VOID)
1648         ret = top();
1649   else  ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1650 
1651   // Note:  Since any out-of-line call can produce an exception,
1652   // we always insert an I_O projection from the call into the result.
1653 
1654   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1655 
1656   if (separate_io_proj) {
1657     // The caller requested separate projections be used by the fall
1658     // through and exceptional paths, so replace the projections for
1659     // the fall through path.
1660     set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) ));
1661     set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ));
1662   }
1663   return ret;
1664 }
1665 
1666 //--------------------set_predefined_input_for_runtime_call--------------------
1667 // Reading and setting the memory state is way conservative here.
1668 // The real problem is that I am not doing real Type analysis on memory,
1669 // so I cannot distinguish card mark stores from other stores.  Across a GC
1670 // point the Store Barrier and the card mark memory has to agree.  I cannot
1671 // have a card mark store and its barrier split across the GC point from
1672 // either above or below.  Here I get that to happen by reading ALL of memory.
1673 // A better answer would be to separate out card marks from other memory.
1674 // For now, return the input memory state, so that it can be reused
1675 // after the call, if this call has restricted memory effects.
1676 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1677   // Set fixed predefined input arguments
1678   Node* memory = reset_memory();
1679   call->init_req( TypeFunc::Control,   control()  );
1680   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1681   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
1682   call->init_req( TypeFunc::FramePtr,  frameptr() );
1683   call->init_req( TypeFunc::ReturnAdr, top()      );
1684   return memory;
1685 }
1686 
1687 //-------------------set_predefined_output_for_runtime_call--------------------
1688 // Set control and memory (not i_o) from the call.
1689 // If keep_mem is not NULL, use it for the output state,
1690 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1691 // If hook_mem is NULL, this call produces no memory effects at all.
1692 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1693 // then only that memory slice is taken from the call.
1694 // In the last case, we must put an appropriate memory barrier before
1695 // the call, so as to create the correct anti-dependencies on loads
1696 // preceding the call.
1697 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1698                                                       Node* keep_mem,
1699                                                       const TypePtr* hook_mem) {
1700   // no i/o
1701   set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1702   if (keep_mem) {
1703     // First clone the existing memory state
1704     set_all_memory(keep_mem);
1705     if (hook_mem != NULL) {
1706       // Make memory for the call
1707       Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1708       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1709       // We also use hook_mem to extract specific effects from arraycopy stubs.
1710       set_memory(mem, hook_mem);
1711     }
1712     // ...else the call has NO memory effects.
1713 
1714     // Make sure the call advertises its memory effects precisely.
1715     // This lets us build accurate anti-dependences in gcm.cpp.
1716     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1717            "call node must be constructed correctly");
1718   } else {
1719     assert(hook_mem == NULL, "");
1720     // This is not a "slow path" call; all memory comes from the call.
1721     set_all_memory_call(call);
1722   }
1723 }
1724 
1725 
1726 // Replace the call with the current state of the kit.
1727 void GraphKit::replace_call(CallNode* call, Node* result) {
1728   JVMState* ejvms = NULL;
1729   if (has_exceptions()) {
1730     ejvms = transfer_exceptions_into_jvms();
1731   }
1732 
1733   SafePointNode* final_state = stop();
1734 
1735   // Find all the needed outputs of this call
1736   CallProjections callprojs;
1737   call->extract_projections(&callprojs, true);
1738 
1739   // Replace all the old call edges with the edges from the inlining result
1740   C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control));
1741   C->gvn_replace_by(callprojs.fallthrough_memproj,   final_state->in(TypeFunc::Memory));
1742   C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_state->in(TypeFunc::I_O));
1743   Node* final_mem = final_state->in(TypeFunc::Memory);
1744 
1745   // Replace the result with the new result if it exists and is used
1746   if (callprojs.resproj != NULL && result != NULL) {
1747     C->gvn_replace_by(callprojs.resproj, result);
1748   }
1749 
1750   if (ejvms == NULL) {
1751     // No exception edges to simply kill off those paths
1752     C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1753     C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
1754     C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
1755 
1756     // Replace the old exception object with top
1757     if (callprojs.exobj != NULL) {
1758       C->gvn_replace_by(callprojs.exobj, C->top());
1759     }
1760   } else {
1761     GraphKit ekit(ejvms);
1762 
1763     // Load my combined exception state into the kit, with all phis transformed:
1764     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1765 
1766     Node* ex_oop = ekit.use_exception_state(ex_map);
1767 
1768     C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1769     C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
1770     C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
1771 
1772     // Replace the old exception object with the newly created one
1773     if (callprojs.exobj != NULL) {
1774       C->gvn_replace_by(callprojs.exobj, ex_oop);
1775     }
1776   }
1777 
1778   // Disconnect the call from the graph
1779   call->disconnect_inputs(NULL);
1780   C->gvn_replace_by(call, C->top());
1781 
1782   // Clean up any MergeMems that feed other MergeMems since the
1783   // optimizer doesn't like that.
1784   if (final_mem->is_MergeMem()) {
1785     Node_List wl;
1786     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1787       Node* m = i.get();
1788       if (m->is_MergeMem() && !wl.contains(m)) {
1789         wl.push(m);
1790       }
1791     }
1792     while (wl.size()  > 0) {
1793       _gvn.transform(wl.pop());
1794     }
1795   }
1796 }
1797 
1798 
1799 //------------------------------increment_counter------------------------------
1800 // for statistics: increment a VM counter by 1
1801 
1802 void GraphKit::increment_counter(address counter_addr) {
1803   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1804   increment_counter(adr1);
1805 }
1806 
1807 void GraphKit::increment_counter(Node* counter_addr) {
1808   int adr_type = Compile::AliasIdxRaw;
1809   Node* ctrl = control();
1810   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type);
1811   Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1812   store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type );
1813 }
1814 
1815 
1816 //------------------------------uncommon_trap----------------------------------
1817 // Bail out to the interpreter in mid-method.  Implemented by calling the
1818 // uncommon_trap blob.  This helper function inserts a runtime call with the
1819 // right debug info.
1820 void GraphKit::uncommon_trap(int trap_request,
1821                              ciKlass* klass, const char* comment,
1822                              bool must_throw,
1823                              bool keep_exact_action) {
1824   if (failing())  stop();
1825   if (stopped())  return; // trap reachable?
1826 
1827   // Note:  If ProfileTraps is true, and if a deopt. actually
1828   // occurs here, the runtime will make sure an MDO exists.  There is
1829   // no need to call method()->build_method_data() at this point.
1830 
1831 #ifdef ASSERT
1832   if (!must_throw) {
1833     // Make sure the stack has at least enough depth to execute
1834     // the current bytecode.
1835     int inputs, ignore;
1836     if (compute_stack_effects(inputs, ignore)) {
1837       assert(sp() >= inputs, "must have enough JVMS stack to execute");
1838       // It is a frequent error in library_call.cpp to issue an
1839       // uncommon trap with the _sp value already popped.
1840     }
1841   }
1842 #endif
1843 
1844   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1845   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1846 
1847   switch (action) {
1848   case Deoptimization::Action_maybe_recompile:
1849   case Deoptimization::Action_reinterpret:
1850     // Temporary fix for 6529811 to allow virtual calls to be sure they
1851     // get the chance to go from mono->bi->mega
1852     if (!keep_exact_action &&
1853         Deoptimization::trap_request_index(trap_request) < 0 &&
1854         too_many_recompiles(reason)) {
1855       // This BCI is causing too many recompilations.
1856       action = Deoptimization::Action_none;
1857       trap_request = Deoptimization::make_trap_request(reason, action);
1858     } else {
1859       C->set_trap_can_recompile(true);
1860     }
1861     break;
1862   case Deoptimization::Action_make_not_entrant:
1863     C->set_trap_can_recompile(true);
1864     break;
1865 #ifdef ASSERT
1866   case Deoptimization::Action_none:
1867   case Deoptimization::Action_make_not_compilable:
1868     break;
1869   default:
1870     assert(false, "bad action");
1871 #endif
1872   }
1873 
1874   if (TraceOptoParse) {
1875     char buf[100];
1876     tty->print_cr("Uncommon trap %s at bci:%d",
1877                   Deoptimization::format_trap_request(buf, sizeof(buf),
1878                                                       trap_request), bci());
1879   }
1880 
1881   CompileLog* log = C->log();
1882   if (log != NULL) {
1883     int kid = (klass == NULL)? -1: log->identify(klass);
1884     log->begin_elem("uncommon_trap bci='%d'", bci());
1885     char buf[100];
1886     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1887                                                           trap_request));
1888     if (kid >= 0)         log->print(" klass='%d'", kid);
1889     if (comment != NULL)  log->print(" comment='%s'", comment);
1890     log->end_elem();
1891   }
1892 
1893   // Make sure any guarding test views this path as very unlikely
1894   Node *i0 = control()->in(0);
1895   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
1896     IfNode *iff = i0->as_If();
1897     float f = iff->_prob;   // Get prob
1898     if (control()->Opcode() == Op_IfTrue) {
1899       if (f > PROB_UNLIKELY_MAG(4))
1900         iff->_prob = PROB_MIN;
1901     } else {
1902       if (f < PROB_LIKELY_MAG(4))
1903         iff->_prob = PROB_MAX;
1904     }
1905   }
1906 
1907   // Clear out dead values from the debug info.
1908   kill_dead_locals();
1909 
1910   // Now insert the uncommon trap subroutine call
1911   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
1912   const TypePtr* no_memory_effects = NULL;
1913   // Pass the index of the class to be loaded
1914   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1915                                  (must_throw ? RC_MUST_THROW : 0),
1916                                  OptoRuntime::uncommon_trap_Type(),
1917                                  call_addr, "uncommon_trap", no_memory_effects,
1918                                  intcon(trap_request));
1919   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1920          "must extract request correctly from the graph");
1921   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1922 
1923   call->set_req(TypeFunc::ReturnAdr, returnadr());
1924   // The debug info is the only real input to this call.
1925 
1926   // Halt-and-catch fire here.  The above call should never return!
1927   HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1928   _gvn.set_type_bottom(halt);
1929   root()->add_req(halt);
1930 
1931   stop_and_kill_map();
1932 }
1933 
1934 
1935 //--------------------------just_allocated_object------------------------------
1936 // Report the object that was just allocated.
1937 // It must be the case that there are no intervening safepoints.
1938 // We use this to determine if an object is so "fresh" that
1939 // it does not require card marks.
1940 Node* GraphKit::just_allocated_object(Node* current_control) {
1941   if (C->recent_alloc_ctl() == current_control)
1942     return C->recent_alloc_obj();
1943   return NULL;
1944 }
1945 
1946 
1947 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1948   // (Note:  TypeFunc::make has a cache that makes this fast.)
1949   const TypeFunc* tf    = TypeFunc::make(dest_method);
1950   int             nargs = tf->_domain->_cnt - TypeFunc::Parms;
1951   for (int j = 0; j < nargs; j++) {
1952     const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1953     if( targ->basic_type() == T_DOUBLE ) {
1954       // If any parameters are doubles, they must be rounded before
1955       // the call, dstore_rounding does gvn.transform
1956       Node *arg = argument(j);
1957       arg = dstore_rounding(arg);
1958       set_argument(j, arg);
1959     }
1960   }
1961 }
1962 
1963 void GraphKit::round_double_result(ciMethod* dest_method) {
1964   // A non-strict method may return a double value which has an extended
1965   // exponent, but this must not be visible in a caller which is 'strict'
1966   // If a strict caller invokes a non-strict callee, round a double result
1967 
1968   BasicType result_type = dest_method->return_type()->basic_type();
1969   assert( method() != NULL, "must have caller context");
1970   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
1971     // Destination method's return value is on top of stack
1972     // dstore_rounding() does gvn.transform
1973     Node *result = pop_pair();
1974     result = dstore_rounding(result);
1975     push_pair(result);
1976   }
1977 }
1978 
1979 // rounding for strict float precision conformance
1980 Node* GraphKit::precision_rounding(Node* n) {
1981   return UseStrictFP && _method->flags().is_strict()
1982     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
1983     ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
1984     : n;
1985 }
1986 
1987 // rounding for strict double precision conformance
1988 Node* GraphKit::dprecision_rounding(Node *n) {
1989   return UseStrictFP && _method->flags().is_strict()
1990     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
1991     ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1992     : n;
1993 }
1994 
1995 // rounding for non-strict double stores
1996 Node* GraphKit::dstore_rounding(Node* n) {
1997   return Matcher::strict_fp_requires_explicit_rounding
1998     && UseSSE <= 1
1999     ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
2000     : n;
2001 }
2002 
2003 //=============================================================================
2004 // Generate a fast path/slow path idiom.  Graph looks like:
2005 // [foo] indicates that 'foo' is a parameter
2006 //
2007 //              [in]     NULL
2008 //                 \    /
2009 //                  CmpP
2010 //                  Bool ne
2011 //                   If
2012 //                  /  \
2013 //              True    False-<2>
2014 //              / |
2015 //             /  cast_not_null
2016 //           Load  |    |   ^
2017 //        [fast_test]   |   |
2018 // gvn to   opt_test    |   |
2019 //          /    \      |  <1>
2020 //      True     False  |
2021 //        |         \\  |
2022 //   [slow_call]     \[fast_result]
2023 //    Ctl   Val       \      \
2024 //     |               \      \
2025 //    Catch       <1>   \      \
2026 //   /    \        ^     \      \
2027 //  Ex    No_Ex    |      \      \
2028 //  |       \   \  |       \ <2>  \
2029 //  ...      \  [slow_res] |  |    \   [null_result]
2030 //            \         \--+--+---  |  |
2031 //             \           | /    \ | /
2032 //              --------Region     Phi
2033 //
2034 //=============================================================================
2035 // Code is structured as a series of driver functions all called 'do_XXX' that
2036 // call a set of helper functions.  Helper functions first, then drivers.
2037 
2038 //------------------------------null_check_oop---------------------------------
2039 // Null check oop.  Set null-path control into Region in slot 3.
2040 // Make a cast-not-nullness use the other not-null control.  Return cast.
2041 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2042                                bool never_see_null) {
2043   // Initial NULL check taken path
2044   (*null_control) = top();
2045   Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2046 
2047   // Generate uncommon_trap:
2048   if (never_see_null && (*null_control) != top()) {
2049     // If we see an unexpected null at a check-cast we record it and force a
2050     // recompile; the offending check-cast will be compiled to handle NULLs.
2051     // If we see more than one offending BCI, then all checkcasts in the
2052     // method will be compiled to handle NULLs.
2053     PreserveJVMState pjvms(this);
2054     set_control(*null_control);
2055     replace_in_map(value, null());
2056     uncommon_trap(Deoptimization::Reason_null_check,
2057                   Deoptimization::Action_make_not_entrant);
2058     (*null_control) = top();    // NULL path is dead
2059   }
2060 
2061   // Cast away null-ness on the result
2062   return cast;
2063 }
2064 
2065 //------------------------------opt_iff----------------------------------------
2066 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2067 // Return slow-path control.
2068 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2069   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2070 
2071   // Fast path taken; set region slot 2
2072   Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2073   region->init_req(2,fast_taken); // Capture fast-control
2074 
2075   // Fast path not-taken, i.e. slow path
2076   Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2077   return slow_taken;
2078 }
2079 
2080 //-----------------------------make_runtime_call-------------------------------
2081 Node* GraphKit::make_runtime_call(int flags,
2082                                   const TypeFunc* call_type, address call_addr,
2083                                   const char* call_name,
2084                                   const TypePtr* adr_type,
2085                                   // The following parms are all optional.
2086                                   // The first NULL ends the list.
2087                                   Node* parm0, Node* parm1,
2088                                   Node* parm2, Node* parm3,
2089                                   Node* parm4, Node* parm5,
2090                                   Node* parm6, Node* parm7) {
2091   // Slow-path call
2092   int size = call_type->domain()->cnt();
2093   bool is_leaf = !(flags & RC_NO_LEAF);
2094   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2095   if (call_name == NULL) {
2096     assert(!is_leaf, "must supply name for leaf");
2097     call_name = OptoRuntime::stub_name(call_addr);
2098   }
2099   CallNode* call;
2100   if (!is_leaf) {
2101     call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2102                                            bci(), adr_type);
2103   } else if (flags & RC_NO_FP) {
2104     call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2105   } else {
2106     call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2107   }
2108 
2109   // The following is similar to set_edges_for_java_call,
2110   // except that the memory effects of the call are restricted to AliasIdxRaw.
2111 
2112   // Slow path call has no side-effects, uses few values
2113   bool wide_in  = !(flags & RC_NARROW_MEM);
2114   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2115 
2116   Node* prev_mem = NULL;
2117   if (wide_in) {
2118     prev_mem = set_predefined_input_for_runtime_call(call);
2119   } else {
2120     assert(!wide_out, "narrow in => narrow out");
2121     Node* narrow_mem = memory(adr_type);
2122     prev_mem = reset_memory();
2123     map()->set_memory(narrow_mem);
2124     set_predefined_input_for_runtime_call(call);
2125   }
2126 
2127   // Hook each parm in order.  Stop looking at the first NULL.
2128   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2129   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2130   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2131   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2132   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2133   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2134   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2135   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2136     /* close each nested if ===> */  } } } } } } } }
2137   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2138 
2139   if (!is_leaf) {
2140     // Non-leaves can block and take safepoints:
2141     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2142   }
2143   // Non-leaves can throw exceptions:
2144   if (has_io) {
2145     call->set_req(TypeFunc::I_O, i_o());
2146   }
2147 
2148   if (flags & RC_UNCOMMON) {
2149     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2150     // (An "if" probability corresponds roughly to an unconditional count.
2151     // Sort of.)
2152     call->set_cnt(PROB_UNLIKELY_MAG(4));
2153   }
2154 
2155   Node* c = _gvn.transform(call);
2156   assert(c == call, "cannot disappear");
2157 
2158   if (wide_out) {
2159     // Slow path call has full side-effects.
2160     set_predefined_output_for_runtime_call(call);
2161   } else {
2162     // Slow path call has few side-effects, and/or sets few values.
2163     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2164   }
2165 
2166   if (has_io) {
2167     set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2168   }
2169   return call;
2170 
2171 }
2172 
2173 //------------------------------merge_memory-----------------------------------
2174 // Merge memory from one path into the current memory state.
2175 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2176   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2177     Node* old_slice = mms.force_memory();
2178     Node* new_slice = mms.memory2();
2179     if (old_slice != new_slice) {
2180       PhiNode* phi;
2181       if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2182         phi = new_slice->as_Phi();
2183         #ifdef ASSERT
2184         if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2185           old_slice = old_slice->in(new_path);
2186         // Caller is responsible for ensuring that any pre-existing
2187         // phis are already aware of old memory.
2188         int old_path = (new_path > 1) ? 1 : 2;  // choose old_path != new_path
2189         assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2190         #endif
2191         mms.set_memory(phi);
2192       } else {
2193         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2194         _gvn.set_type(phi, Type::MEMORY);
2195         phi->set_req(new_path, new_slice);
2196         mms.set_memory(_gvn.transform(phi));  // assume it is complete
2197       }
2198     }
2199   }
2200 }
2201 
2202 //------------------------------make_slow_call_ex------------------------------
2203 // Make the exception handler hookups for the slow call
2204 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2205   if (stopped())  return;
2206 
2207   // Make a catch node with just two handlers:  fall-through and catch-all
2208   Node* i_o  = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2209   Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2210   Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2211   Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2212 
2213   { PreserveJVMState pjvms(this);
2214     set_control(excp);
2215     set_i_o(i_o);
2216 
2217     if (excp != top()) {
2218       // Create an exception state also.
2219       // Use an exact type if the caller has specified a specific exception.
2220       const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2221       Node*       ex_oop  = new (C, 2) CreateExNode(ex_type, control(), i_o);
2222       add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2223     }
2224   }
2225 
2226   // Get the no-exception control from the CatchNode.
2227   set_control(norm);
2228 }
2229 
2230 
2231 //-------------------------------gen_subtype_check-----------------------------
2232 // Generate a subtyping check.  Takes as input the subtype and supertype.
2233 // Returns 2 values: sets the default control() to the true path and returns
2234 // the false path.  Only reads invariant memory; sets no (visible) memory.
2235 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2236 // but that's not exposed to the optimizer.  This call also doesn't take in an
2237 // Object; if you wish to check an Object you need to load the Object's class
2238 // prior to coming here.
2239 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2240   // Fast check for identical types, perhaps identical constants.
2241   // The types can even be identical non-constants, in cases
2242   // involving Array.newInstance, Object.clone, etc.
2243   if (subklass == superklass)
2244     return top();             // false path is dead; no test needed.
2245 
2246   if (_gvn.type(superklass)->singleton()) {
2247     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2248     ciKlass* subk   = _gvn.type(subklass)->is_klassptr()->klass();
2249 
2250     // In the common case of an exact superklass, try to fold up the
2251     // test before generating code.  You may ask, why not just generate
2252     // the code and then let it fold up?  The answer is that the generated
2253     // code will necessarily include null checks, which do not always
2254     // completely fold away.  If they are also needless, then they turn
2255     // into a performance loss.  Example:
2256     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2257     // Here, the type of 'fa' is often exact, so the store check
2258     // of fa[1]=x will fold up, without testing the nullness of x.
2259     switch (static_subtype_check(superk, subk)) {
2260     case SSC_always_false:
2261       {
2262         Node* always_fail = control();
2263         set_control(top());
2264         return always_fail;
2265       }
2266     case SSC_always_true:
2267       return top();
2268     case SSC_easy_test:
2269       {
2270         // Just do a direct pointer compare and be done.
2271         Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2272         Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2273         IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2274         set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2275         return       _gvn.transform( new(C, 1) IfFalseNode(iff) );
2276       }
2277     case SSC_full_test:
2278       break;
2279     default:
2280       ShouldNotReachHere();
2281     }
2282   }
2283 
2284   // %%% Possible further optimization:  Even if the superklass is not exact,
2285   // if the subklass is the unique subtype of the superklass, the check
2286   // will always succeed.  We could leave a dependency behind to ensure this.
2287 
2288   // First load the super-klass's check-offset
2289   Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
2290   Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2291   int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
2292   bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2293 
2294   // Load from the sub-klass's super-class display list, or a 1-word cache of
2295   // the secondary superclass list, or a failing value with a sentinel offset
2296   // if the super-klass is an interface or exceptionally deep in the Java
2297   // hierarchy and we have to scan the secondary superclass list the hard way.
2298   // Worst-case type is a little odd: NULL is allowed as a result (usually
2299   // klass loads can never produce a NULL).
2300   Node *chk_off_X = ConvI2X(chk_off);
2301   Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2302   // For some types like interfaces the following loadKlass is from a 1-word
2303   // cache which is mutable so can't use immutable memory.  Other
2304   // types load from the super-class display table which is immutable.
2305   Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2306   Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2307 
2308   // Compile speed common case: ARE a subtype and we canNOT fail
2309   if( superklass == nkls )
2310     return top();             // false path is dead; no test needed.
2311 
2312   // See if we get an immediate positive hit.  Happens roughly 83% of the
2313   // time.  Test to see if the value loaded just previously from the subklass
2314   // is exactly the superklass.
2315   Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2316   Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2317   IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2318   Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2319   set_control(    _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2320 
2321   // Compile speed common case: Check for being deterministic right now.  If
2322   // chk_off is a constant and not equal to cacheoff then we are NOT a
2323   // subklass.  In this case we need exactly the 1 test above and we can
2324   // return those results immediately.
2325   if (!might_be_cache) {
2326     Node* not_subtype_ctrl = control();
2327     set_control(iftrue1); // We need exactly the 1 test above
2328     return not_subtype_ctrl;
2329   }
2330 
2331   // Gather the various success & failures here
2332   RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2333   record_for_igvn(r_ok_subtype);
2334   RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2335   record_for_igvn(r_not_subtype);
2336 
2337   r_ok_subtype->init_req(1, iftrue1);
2338 
2339   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2340   // is roughly 63% of the remaining cases).  Test to see if the loaded
2341   // check-offset points into the subklass display list or the 1-element
2342   // cache.  If it points to the display (and NOT the cache) and the display
2343   // missed then it's not a subtype.
2344   Node *cacheoff = _gvn.intcon(cacheoff_con);
2345   Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2346   Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2347   IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2348   r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2349   set_control(                _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2350 
2351   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2352   // No performance impact (too rare) but allows sharing of secondary arrays
2353   // which has some footprint reduction.
2354   Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2355   Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2356   IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2357   r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2358   set_control(               _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2359 
2360   // -- Roads not taken here: --
2361   // We could also have chosen to perform the self-check at the beginning
2362   // of this code sequence, as the assembler does.  This would not pay off
2363   // the same way, since the optimizer, unlike the assembler, can perform
2364   // static type analysis to fold away many successful self-checks.
2365   // Non-foldable self checks work better here in second position, because
2366   // the initial primary superclass check subsumes a self-check for most
2367   // types.  An exception would be a secondary type like array-of-interface,
2368   // which does not appear in its own primary supertype display.
2369   // Finally, we could have chosen to move the self-check into the
2370   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2371   // dependent manner.  But it is worthwhile to have the check here,
2372   // where it can be perhaps be optimized.  The cost in code space is
2373   // small (register compare, branch).
2374 
2375   // Now do a linear scan of the secondary super-klass array.  Again, no real
2376   // performance impact (too rare) but it's gotta be done.
2377   // Since the code is rarely used, there is no penalty for moving it
2378   // out of line, and it can only improve I-cache density.
2379   // The decision to inline or out-of-line this final check is platform
2380   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2381   Node* psc = _gvn.transform(
2382     new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2383 
2384   Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2385   Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2386   IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2387   r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2388   r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2389 
2390   // Return false path; set default control to true path.
2391   set_control( _gvn.transform(r_ok_subtype) );
2392   return _gvn.transform(r_not_subtype);
2393 }
2394 
2395 //----------------------------static_subtype_check-----------------------------
2396 // Shortcut important common cases when superklass is exact:
2397 // (0) superklass is java.lang.Object (can occur in reflective code)
2398 // (1) subklass is already limited to a subtype of superklass => always ok
2399 // (2) subklass does not overlap with superklass => always fail
2400 // (3) superklass has NO subtypes and we can check with a simple compare.
2401 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2402   if (StressReflectiveCode) {
2403     return SSC_full_test;       // Let caller generate the general case.
2404   }
2405 
2406   if (superk == env()->Object_klass()) {
2407     return SSC_always_true;     // (0) this test cannot fail
2408   }
2409 
2410   ciType* superelem = superk;
2411   if (superelem->is_array_klass())
2412     superelem = superelem->as_array_klass()->base_element_type();
2413 
2414   if (!subk->is_interface()) {  // cannot trust static interface types yet
2415     if (subk->is_subtype_of(superk)) {
2416       return SSC_always_true;   // (1) false path dead; no dynamic test needed
2417     }
2418     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2419         !superk->is_subtype_of(subk)) {
2420       return SSC_always_false;
2421     }
2422   }
2423 
2424   // If casting to an instance klass, it must have no subtypes
2425   if (superk->is_interface()) {
2426     // Cannot trust interfaces yet.
2427     // %%% S.B. superk->nof_implementors() == 1
2428   } else if (superelem->is_instance_klass()) {
2429     ciInstanceKlass* ik = superelem->as_instance_klass();
2430     if (!ik->has_subklass() && !ik->is_interface()) {
2431       if (!ik->is_final()) {
2432         // Add a dependency if there is a chance of a later subclass.
2433         C->dependencies()->assert_leaf_type(ik);
2434       }
2435       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
2436     }
2437   } else {
2438     // A primitive array type has no subtypes.
2439     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
2440   }
2441 
2442   return SSC_full_test;
2443 }
2444 
2445 // Profile-driven exact type check:
2446 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2447                                     float prob,
2448                                     Node* *casted_receiver) {
2449   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2450   Node* recv_klass = load_object_klass(receiver);
2451   Node* want_klass = makecon(tklass);
2452   Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2453   Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2454   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2455   set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2456   Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2457 
2458   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2459   assert(recv_xtype->klass_is_exact(), "");
2460 
2461   // Subsume downstream occurrences of receiver with a cast to
2462   // recv_xtype, since now we know what the type will be.
2463   Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2464   (*casted_receiver) = _gvn.transform(cast);
2465   // (User must make the replace_in_map call.)
2466 
2467   return fail;
2468 }
2469 
2470 
2471 //------------------------------seems_never_null-------------------------------
2472 // Use null_seen information if it is available from the profile.
2473 // If we see an unexpected null at a type check we record it and force a
2474 // recompile; the offending check will be recompiled to handle NULLs.
2475 // If we see several offending BCIs, then all checks in the
2476 // method will be recompiled.
2477 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2478   if (UncommonNullCast               // Cutout for this technique
2479       && obj != null()               // And not the -Xcomp stupid case?
2480       && !too_many_traps(Deoptimization::Reason_null_check)
2481       ) {
2482     if (data == NULL)
2483       // Edge case:  no mature data.  Be optimistic here.
2484       return true;
2485     // If the profile has not seen a null, assume it won't happen.
2486     assert(java_bc() == Bytecodes::_checkcast ||
2487            java_bc() == Bytecodes::_instanceof ||
2488            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2489     return !data->as_BitData()->null_seen();
2490   }
2491   return false;
2492 }
2493 
2494 //------------------------maybe_cast_profiled_receiver-------------------------
2495 // If the profile has seen exactly one type, narrow to exactly that type.
2496 // Subsequent type checks will always fold up.
2497 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2498                                              ciProfileData* data,
2499                                              ciKlass* require_klass) {
2500   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2501   if (data == NULL)  return NULL;
2502 
2503   // Make sure we haven't already deoptimized from this tactic.
2504   if (too_many_traps(Deoptimization::Reason_class_check))
2505     return NULL;
2506 
2507   // (No, this isn't a call, but it's enough like a virtual call
2508   // to use the same ciMethod accessor to get the profile info...)
2509   ciCallProfile profile = method()->call_profile_at_bci(bci());
2510   if (profile.count() >= 0 &&         // no cast failures here
2511       profile.has_receiver(0) &&
2512       profile.morphism() == 1) {
2513     ciKlass* exact_kls = profile.receiver(0);
2514     if (require_klass == NULL ||
2515         static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2516       // If we narrow the type to match what the type profile sees,
2517       // we can then remove the rest of the cast.
2518       // This is a win, even if the exact_kls is very specific,
2519       // because downstream operations, such as method calls,
2520       // will often benefit from the sharper type.
2521       Node* exact_obj = not_null_obj; // will get updated in place...
2522       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2523                                             &exact_obj);
2524       { PreserveJVMState pjvms(this);
2525         set_control(slow_ctl);
2526         uncommon_trap(Deoptimization::Reason_class_check,
2527                       Deoptimization::Action_maybe_recompile);
2528       }
2529       replace_in_map(not_null_obj, exact_obj);
2530       return exact_obj;
2531     }
2532     // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2533   }
2534 
2535   return NULL;
2536 }
2537 
2538 
2539 //-------------------------------gen_instanceof--------------------------------
2540 // Generate an instance-of idiom.  Used by both the instance-of bytecode
2541 // and the reflective instance-of call.
2542 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) {
2543   kill_dead_locals();           // Benefit all the uncommon traps
2544   assert( !stopped(), "dead parse path should be checked in callers" );
2545   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2546          "must check for not-null not-dead klass in callers");
2547 
2548   // Make the merge point
2549   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2550   RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2551   Node*       phi    = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2552   C->set_has_split_ifs(true); // Has chance for split-if optimization
2553 
2554   ciProfileData* data = NULL;
2555   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
2556     data = method()->method_data()->bci_to_data(bci());
2557   }
2558   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
2559                          && seems_never_null(obj, data));
2560 
2561   // Null check; get casted pointer; set region slot 3
2562   Node* null_ctl = top();
2563   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2564 
2565   // If not_null_obj is dead, only null-path is taken
2566   if (stopped()) {              // Doing instance-of on a NULL?
2567     set_control(null_ctl);
2568     return intcon(0);
2569   }
2570   region->init_req(_null_path, null_ctl);
2571   phi   ->init_req(_null_path, intcon(0)); // Set null path value
2572   if (null_ctl == top()) {
2573     // Do this eagerly, so that pattern matches like is_diamond_phi
2574     // will work even during parsing.
2575     assert(_null_path == PATH_LIMIT-1, "delete last");
2576     region->del_req(_null_path);
2577     phi   ->del_req(_null_path);
2578   }
2579 
2580   if (ProfileDynamicTypes && data != NULL) {
2581     Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL);
2582     if (stopped()) {            // Profile disagrees with this path.
2583       set_control(null_ctl);    // Null is the only remaining possibility.
2584       return intcon(0);
2585     }
2586     if (cast_obj != NULL)
2587       not_null_obj = cast_obj;
2588   }
2589 
2590   // Load the object's klass
2591   Node* obj_klass = load_object_klass(not_null_obj);
2592 
2593   // Generate the subtype check
2594   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2595 
2596   // Plug in the success path to the general merge in slot 1.
2597   region->init_req(_obj_path, control());
2598   phi   ->init_req(_obj_path, intcon(1));
2599 
2600   // Plug in the failing path to the general merge in slot 2.
2601   region->init_req(_fail_path, not_subtype_ctrl);
2602   phi   ->init_req(_fail_path, intcon(0));
2603 
2604   // Return final merged results
2605   set_control( _gvn.transform(region) );
2606   record_for_igvn(region);
2607   return _gvn.transform(phi);
2608 }
2609 
2610 //-------------------------------gen_checkcast---------------------------------
2611 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
2612 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
2613 // uncommon-trap paths work.  Adjust stack after this call.
2614 // If failure_control is supplied and not null, it is filled in with
2615 // the control edge for the cast failure.  Otherwise, an appropriate
2616 // uncommon trap or exception is thrown.
2617 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2618                               Node* *failure_control) {
2619   kill_dead_locals();           // Benefit all the uncommon traps
2620   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2621   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2622 
2623   // Fast cutout:  Check the case that the cast is vacuously true.
2624   // This detects the common cases where the test will short-circuit
2625   // away completely.  We do this before we perform the null check,
2626   // because if the test is going to turn into zero code, we don't
2627   // want a residual null check left around.  (Causes a slowdown,
2628   // for example, in some objArray manipulations, such as a[i]=a[j].)
2629   if (tk->singleton()) {
2630     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2631     if (objtp != NULL && objtp->klass() != NULL) {
2632       switch (static_subtype_check(tk->klass(), objtp->klass())) {
2633       case SSC_always_true:
2634         return obj;
2635       case SSC_always_false:
2636         // It needs a null check because a null will *pass* the cast check.
2637         // A non-null value will always produce an exception.
2638         return do_null_assert(obj, T_OBJECT);
2639       }
2640     }
2641   }
2642 
2643   ciProfileData* data = NULL;
2644   if (failure_control == NULL) {        // use MDO in regular case only
2645     assert(java_bc() == Bytecodes::_aastore ||
2646            java_bc() == Bytecodes::_checkcast,
2647            "interpreter profiles type checks only for these BCs");
2648     data = method()->method_data()->bci_to_data(bci());
2649   }
2650 
2651   // Make the merge point
2652   enum { _obj_path = 1, _null_path, PATH_LIMIT };
2653   RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2654   Node*       phi    = new (C, PATH_LIMIT) PhiNode(region, toop);
2655   C->set_has_split_ifs(true); // Has chance for split-if optimization
2656 
2657   // Use null-cast information if it is available
2658   bool never_see_null = ((failure_control == NULL)  // regular case only
2659                          && seems_never_null(obj, data));
2660 
2661   // Null check; get casted pointer; set region slot 3
2662   Node* null_ctl = top();
2663   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2664 
2665   // If not_null_obj is dead, only null-path is taken
2666   if (stopped()) {              // Doing instance-of on a NULL?
2667     set_control(null_ctl);
2668     return null();
2669   }
2670   region->init_req(_null_path, null_ctl);
2671   phi   ->init_req(_null_path, null());  // Set null path value
2672   if (null_ctl == top()) {
2673     // Do this eagerly, so that pattern matches like is_diamond_phi
2674     // will work even during parsing.
2675     assert(_null_path == PATH_LIMIT-1, "delete last");
2676     region->del_req(_null_path);
2677     phi   ->del_req(_null_path);
2678   }
2679 
2680   Node* cast_obj = NULL;
2681   if (data != NULL &&
2682       // Counter has never been decremented (due to cast failure).
2683       // ...This is a reasonable thing to expect.  It is true of
2684       // all casts inserted by javac to implement generic types.
2685       data->as_CounterData()->count() >= 0) {
2686     cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass());
2687     if (cast_obj != NULL) {
2688       if (failure_control != NULL) // failure is now impossible
2689         (*failure_control) = top();
2690       // adjust the type of the phi to the exact klass:
2691       phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR));
2692     }
2693   }
2694 
2695   if (cast_obj == NULL) {
2696     // Load the object's klass
2697     Node* obj_klass = load_object_klass(not_null_obj);
2698 
2699     // Generate the subtype check
2700     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2701 
2702     // Plug in success path into the merge
2703     cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2704                                                          not_null_obj, toop));
2705     // Failure path ends in uncommon trap (or may be dead - failure impossible)
2706     if (failure_control == NULL) {
2707       if (not_subtype_ctrl != top()) { // If failure is possible
2708         PreserveJVMState pjvms(this);
2709         set_control(not_subtype_ctrl);
2710         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2711       }
2712     } else {
2713       (*failure_control) = not_subtype_ctrl;
2714     }
2715   }
2716 
2717   region->init_req(_obj_path, control());
2718   phi   ->init_req(_obj_path, cast_obj);
2719 
2720   // A merge of NULL or Casted-NotNull obj
2721   Node* res = _gvn.transform(phi);
2722 
2723   // Note I do NOT always 'replace_in_map(obj,result)' here.
2724   //  if( tk->klass()->can_be_primary_super()  )
2725     // This means that if I successfully store an Object into an array-of-String
2726     // I 'forget' that the Object is really now known to be a String.  I have to
2727     // do this because we don't have true union types for interfaces - if I store
2728     // a Baz into an array-of-Interface and then tell the optimizer it's an
2729     // Interface, I forget that it's also a Baz and cannot do Baz-like field
2730     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
2731   //  replace_in_map( obj, res );
2732 
2733   // Return final merged results
2734   set_control( _gvn.transform(region) );
2735   record_for_igvn(region);
2736   return res;
2737 }
2738 
2739 //------------------------------next_monitor-----------------------------------
2740 // What number should be given to the next monitor?
2741 int GraphKit::next_monitor() {
2742   int current = jvms()->monitor_depth()* C->sync_stack_slots();
2743   int next = current + C->sync_stack_slots();
2744   // Keep the toplevel high water mark current:
2745   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
2746   return current;
2747 }
2748 
2749 //------------------------------insert_mem_bar---------------------------------
2750 // Memory barrier to avoid floating things around
2751 // The membar serves as a pinch point between both control and all memory slices.
2752 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2753   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2754   mb->init_req(TypeFunc::Control, control());
2755   mb->init_req(TypeFunc::Memory,  reset_memory());
2756   Node* membar = _gvn.transform(mb);
2757   set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2758   set_all_memory_call(membar);
2759   return membar;
2760 }
2761 
2762 //-------------------------insert_mem_bar_volatile----------------------------
2763 // Memory barrier to avoid floating things around
2764 // The membar serves as a pinch point between both control and memory(alias_idx).
2765 // If you want to make a pinch point on all memory slices, do not use this
2766 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2767 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2768   // When Parse::do_put_xxx updates a volatile field, it appends a series
2769   // of MemBarVolatile nodes, one for *each* volatile field alias category.
2770   // The first membar is on the same memory slice as the field store opcode.
2771   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
2772   // All the other membars (for other volatile slices, including AliasIdxBot,
2773   // which stands for all unknown volatile slices) are control-dependent
2774   // on the first membar.  This prevents later volatile loads or stores
2775   // from sliding up past the just-emitted store.
2776 
2777   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2778   mb->set_req(TypeFunc::Control,control());
2779   if (alias_idx == Compile::AliasIdxBot) {
2780     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2781   } else {
2782     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2783     mb->set_req(TypeFunc::Memory, memory(alias_idx));
2784   }
2785   Node* membar = _gvn.transform(mb);
2786   set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2787   if (alias_idx == Compile::AliasIdxBot) {
2788     merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2789   } else {
2790     set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2791   }
2792   return membar;
2793 }
2794 
2795 //------------------------------shared_lock------------------------------------
2796 // Emit locking code.
2797 FastLockNode* GraphKit::shared_lock(Node* obj) {
2798   // bci is either a monitorenter bc or InvocationEntryBci
2799   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2800   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2801 
2802   if( !GenerateSynchronizationCode )
2803     return NULL;                // Not locking things?
2804   if (stopped())                // Dead monitor?
2805     return NULL;
2806 
2807   assert(dead_locals_are_killed(), "should kill locals before sync. point");
2808 
2809   // Box the stack location
2810   Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2811   Node* mem = reset_memory();
2812 
2813   FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2814   if (PrintPreciseBiasedLockingStatistics) {
2815     // Create the counters for this fast lock.
2816     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2817   }
2818   // Add monitor to debug info for the slow path.  If we block inside the
2819   // slow path and de-opt, we need the monitor hanging around
2820   map()->push_monitor( flock );
2821 
2822   const TypeFunc *tf = LockNode::lock_type();
2823   LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2824 
2825   lock->init_req( TypeFunc::Control, control() );
2826   lock->init_req( TypeFunc::Memory , mem );
2827   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
2828   lock->init_req( TypeFunc::FramePtr, frameptr() );
2829   lock->init_req( TypeFunc::ReturnAdr, top() );
2830 
2831   lock->init_req(TypeFunc::Parms + 0, obj);
2832   lock->init_req(TypeFunc::Parms + 1, box);
2833   lock->init_req(TypeFunc::Parms + 2, flock);
2834   add_safepoint_edges(lock);
2835 
2836   lock = _gvn.transform( lock )->as_Lock();
2837 
2838   // lock has no side-effects, sets few values
2839   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2840 
2841   insert_mem_bar(Op_MemBarAcquire);
2842 
2843   // Add this to the worklist so that the lock can be eliminated
2844   record_for_igvn(lock);
2845 
2846 #ifndef PRODUCT
2847   if (PrintLockStatistics) {
2848     // Update the counter for this lock.  Don't bother using an atomic
2849     // operation since we don't require absolute accuracy.
2850     lock->create_lock_counter(map()->jvms());
2851     increment_counter(lock->counter()->addr());
2852   }
2853 #endif
2854 
2855   return flock;
2856 }
2857 
2858 
2859 //------------------------------shared_unlock----------------------------------
2860 // Emit unlocking code.
2861 void GraphKit::shared_unlock(Node* box, Node* obj) {
2862   // bci is either a monitorenter bc or InvocationEntryBci
2863   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2864   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2865 
2866   if( !GenerateSynchronizationCode )
2867     return;
2868   if (stopped()) {               // Dead monitor?
2869     map()->pop_monitor();        // Kill monitor from debug info
2870     return;
2871   }
2872 
2873   // Memory barrier to avoid floating things down past the locked region
2874   insert_mem_bar(Op_MemBarRelease);
2875 
2876   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2877   UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2878   uint raw_idx = Compile::AliasIdxRaw;
2879   unlock->init_req( TypeFunc::Control, control() );
2880   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2881   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
2882   unlock->init_req( TypeFunc::FramePtr, frameptr() );
2883   unlock->init_req( TypeFunc::ReturnAdr, top() );
2884 
2885   unlock->init_req(TypeFunc::Parms + 0, obj);
2886   unlock->init_req(TypeFunc::Parms + 1, box);
2887   unlock = _gvn.transform(unlock)->as_Unlock();
2888 
2889   Node* mem = reset_memory();
2890 
2891   // unlock has no side-effects, sets few values
2892   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2893 
2894   // Kill monitor from debug info
2895   map()->pop_monitor( );
2896 }
2897 
2898 //-------------------------------get_layout_helper-----------------------------
2899 // If the given klass is a constant or known to be an array,
2900 // fetch the constant layout helper value into constant_value
2901 // and return (Node*)NULL.  Otherwise, load the non-constant
2902 // layout helper value, and return the node which represents it.
2903 // This two-faced routine is useful because allocation sites
2904 // almost always feature constant types.
2905 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2906   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2907   if (!StressReflectiveCode && inst_klass != NULL) {
2908     ciKlass* klass = inst_klass->klass();
2909     bool    xklass = inst_klass->klass_is_exact();
2910     if (xklass || klass->is_array_klass()) {
2911       jint lhelper = klass->layout_helper();
2912       if (lhelper != Klass::_lh_neutral_value) {
2913         constant_value = lhelper;
2914         return (Node*) NULL;
2915       }
2916     }
2917   }
2918   constant_value = Klass::_lh_neutral_value;  // put in a known value
2919   Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
2920   return make_load(NULL, lhp, TypeInt::INT, T_INT);
2921 }
2922 
2923 // We just put in an allocate/initialize with a big raw-memory effect.
2924 // Hook selected additional alias categories on the initialization.
2925 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2926                                 MergeMemNode* init_in_merge,
2927                                 Node* init_out_raw) {
2928   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2929   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2930 
2931   Node* prevmem = kit.memory(alias_idx);
2932   init_in_merge->set_memory_at(alias_idx, prevmem);
2933   kit.set_memory(init_out_raw, alias_idx);
2934 }
2935 
2936 //---------------------------set_output_for_allocation-------------------------
2937 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2938                                           const TypeOopPtr* oop_type,
2939                                           bool raw_mem_only) {
2940   int rawidx = Compile::AliasIdxRaw;
2941   alloc->set_req( TypeFunc::FramePtr, frameptr() );
2942   add_safepoint_edges(alloc);
2943   Node* allocx = _gvn.transform(alloc);
2944   set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2945   // create memory projection for i_o
2946   set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2947   make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2948 
2949   // create a memory projection as for the normal control path
2950   Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2951   set_memory(malloc, rawidx);
2952 
2953   // a normal slow-call doesn't change i_o, but an allocation does
2954   // we create a separate i_o projection for the normal control path
2955   set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2956   Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2957 
2958   // put in an initialization barrier
2959   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2960                                                  rawoop)->as_Initialize();
2961   assert(alloc->initialization() == init,  "2-way macro link must work");
2962   assert(init ->allocation()     == alloc, "2-way macro link must work");
2963   if (ReduceFieldZeroing && !raw_mem_only) {
2964     // Extract memory strands which may participate in the new object's
2965     // initialization, and source them from the new InitializeNode.
2966     // This will allow us to observe initializations when they occur,
2967     // and link them properly (as a group) to the InitializeNode.
2968     assert(init->in(InitializeNode::Memory) == malloc, "");
2969     MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
2970     init->set_req(InitializeNode::Memory, minit_in);
2971     record_for_igvn(minit_in); // fold it up later, if possible
2972     Node* minit_out = memory(rawidx);
2973     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
2974     if (oop_type->isa_aryptr()) {
2975       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
2976       int            elemidx  = C->get_alias_index(telemref);
2977       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
2978     } else if (oop_type->isa_instptr()) {
2979       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
2980       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
2981         ciField* field = ik->nonstatic_field_at(i);
2982         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
2983           continue;  // do not bother to track really large numbers of fields
2984         // Find (or create) the alias category for this field:
2985         int fieldidx = C->alias_type(field)->index();
2986         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
2987       }
2988     }
2989   }
2990 
2991   // Cast raw oop to the real thing...
2992   Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
2993   javaoop = _gvn.transform(javaoop);
2994   C->set_recent_alloc(control(), javaoop);
2995   assert(just_allocated_object(control()) == javaoop, "just allocated");
2996 
2997 #ifdef ASSERT
2998   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
2999     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3000            "Ideal_allocation works");
3001     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3002            "Ideal_allocation works");
3003     if (alloc->is_AllocateArray()) {
3004       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3005              "Ideal_allocation works");
3006       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3007              "Ideal_allocation works");
3008     } else {
3009       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3010     }
3011   }
3012 #endif //ASSERT
3013 
3014   return javaoop;
3015 }
3016 
3017 //---------------------------new_instance--------------------------------------
3018 // This routine takes a klass_node which may be constant (for a static type)
3019 // or may be non-constant (for reflective code).  It will work equally well
3020 // for either, and the graph will fold nicely if the optimizer later reduces
3021 // the type to a constant.
3022 // The optional arguments are for specialized use by intrinsics:
3023 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3024 //  - If 'raw_mem_only', do not cast the result to an oop.
3025 //  - If 'return_size_val', report the the total object size to the caller.
3026 Node* GraphKit::new_instance(Node* klass_node,
3027                              Node* extra_slow_test,
3028                              bool raw_mem_only, // affect only raw memory
3029                              Node* *return_size_val) {
3030   // Compute size in doublewords
3031   // The size is always an integral number of doublewords, represented
3032   // as a positive bytewise size stored in the klass's layout_helper.
3033   // The layout_helper also encodes (in a low bit) the need for a slow path.
3034   jint  layout_con = Klass::_lh_neutral_value;
3035   Node* layout_val = get_layout_helper(klass_node, layout_con);
3036   int   layout_is_con = (layout_val == NULL);
3037 
3038   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3039   // Generate the initial go-slow test.  It's either ALWAYS (return a
3040   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3041   // case) a computed value derived from the layout_helper.
3042   Node* initial_slow_test = NULL;
3043   if (layout_is_con) {
3044     assert(!StressReflectiveCode, "stress mode does not use these paths");
3045     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3046     initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3047 
3048   } else {   // reflective case
3049     // This reflective path is used by Unsafe.allocateInstance.
3050     // (It may be stress-tested by specifying StressReflectiveCode.)
3051     // Basically, we want to get into the VM is there's an illegal argument.
3052     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3053     initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
3054     if (extra_slow_test != intcon(0)) {
3055       initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
3056     }
3057     // (Macro-expander will further convert this to a Bool, if necessary.)
3058   }
3059 
3060   // Find the size in bytes.  This is easy; it's the layout_helper.
3061   // The size value must be valid even if the slow path is taken.
3062   Node* size = NULL;
3063   if (layout_is_con) {
3064     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3065   } else {   // reflective case
3066     // This reflective path is used by clone and Unsafe.allocateInstance.
3067     size = ConvI2X(layout_val);
3068 
3069     // Clear the low bits to extract layout_helper_size_in_bytes:
3070     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3071     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3072     size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
3073   }
3074   if (return_size_val != NULL) {
3075     (*return_size_val) = size;
3076   }
3077 
3078   // This is a precise notnull oop of the klass.
3079   // (Actually, it need not be precise if this is a reflective allocation.)
3080   // It's what we cast the result to.
3081   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3082   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3083   const TypeOopPtr* oop_type = tklass->as_instance_type();
3084 
3085   // Now generate allocation code
3086 
3087   // The entire memory state is needed for slow path of the allocation
3088   // since GC and deoptimization can happened.
3089   Node *mem = reset_memory();
3090   set_all_memory(mem); // Create new memory state
3091 
3092   AllocateNode* alloc
3093     = new (C, AllocateNode::ParmLimit)
3094         AllocateNode(C, AllocateNode::alloc_type(),
3095                      control(), mem, i_o(),
3096                      size, klass_node,
3097                      initial_slow_test);
3098 
3099   return set_output_for_allocation(alloc, oop_type, raw_mem_only);
3100 }
3101 
3102 //-------------------------------new_array-------------------------------------
3103 // helper for both newarray and anewarray
3104 // The 'length' parameter is (obviously) the length of the array.
3105 // See comments on new_instance for the meaning of the other arguments.
3106 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3107                           Node* length,         // number of array elements
3108                           int   nargs,          // number of arguments to push back for uncommon trap
3109                           bool raw_mem_only,    // affect only raw memory
3110                           Node* *return_size_val) {
3111   jint  layout_con = Klass::_lh_neutral_value;
3112   Node* layout_val = get_layout_helper(klass_node, layout_con);
3113   int   layout_is_con = (layout_val == NULL);
3114 
3115   if (!layout_is_con && !StressReflectiveCode &&
3116       !too_many_traps(Deoptimization::Reason_class_check)) {
3117     // This is a reflective array creation site.
3118     // Optimistically assume that it is a subtype of Object[],
3119     // so that we can fold up all the address arithmetic.
3120     layout_con = Klass::array_layout_helper(T_OBJECT);
3121     Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
3122     Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
3123     { BuildCutout unless(this, bol_lh, PROB_MAX);
3124       _sp += nargs;
3125       uncommon_trap(Deoptimization::Reason_class_check,
3126                     Deoptimization::Action_maybe_recompile);
3127     }
3128     layout_val = NULL;
3129     layout_is_con = true;
3130   }
3131 
3132   // Generate the initial go-slow test.  Make sure we do not overflow
3133   // if length is huge (near 2Gig) or negative!  We do not need
3134   // exact double-words here, just a close approximation of needed
3135   // double-words.  We can't add any offset or rounding bits, lest we
3136   // take a size -1 of bytes and make it positive.  Use an unsigned
3137   // compare, so negative sizes look hugely positive.
3138   int fast_size_limit = FastAllocateSizeLimit;
3139   if (layout_is_con) {
3140     assert(!StressReflectiveCode, "stress mode does not use these paths");
3141     // Increase the size limit if we have exact knowledge of array type.
3142     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3143     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3144   }
3145 
3146   Node* initial_slow_cmp  = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3147   Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3148   if (initial_slow_test->is_Bool()) {
3149     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3150     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3151   }
3152 
3153   // --- Size Computation ---
3154   // array_size = round_to_heap(array_header + (length << elem_shift));
3155   // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3156   // and round_to(x, y) == ((x + y-1) & ~(y-1))
3157   // The rounding mask is strength-reduced, if possible.
3158   int round_mask = MinObjAlignmentInBytes - 1;
3159   Node* header_size = NULL;
3160   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3161   // (T_BYTE has the weakest alignment and size restrictions...)
3162   if (layout_is_con) {
3163     int       hsize  = Klass::layout_helper_header_size(layout_con);
3164     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3165     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3166     if ((round_mask & ~right_n_bits(eshift)) == 0)
3167       round_mask = 0;  // strength-reduce it if it goes away completely
3168     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3169     assert(header_size_min <= hsize, "generic minimum is smallest");
3170     header_size_min = hsize;
3171     header_size = intcon(hsize + round_mask);
3172   } else {
3173     Node* hss   = intcon(Klass::_lh_header_size_shift);
3174     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3175     Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3176     hsize       = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3177     Node* mask  = intcon(round_mask);
3178     header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3179   }
3180 
3181   Node* elem_shift = NULL;
3182   if (layout_is_con) {
3183     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3184     if (eshift != 0)
3185       elem_shift = intcon(eshift);
3186   } else {
3187     // There is no need to mask or shift this value.
3188     // The semantics of LShiftINode include an implicit mask to 0x1F.
3189     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3190     elem_shift = layout_val;
3191   }
3192 
3193   // Transition to native address size for all offset calculations:
3194   Node* lengthx = ConvI2X(length);
3195   Node* headerx = ConvI2X(header_size);
3196 #ifdef _LP64
3197   { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3198     if (tllen != NULL && tllen->_lo < 0) {
3199       // Add a manual constraint to a positive range.  Cf. array_element_address.
3200       jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3201       if (size_max > tllen->_hi)  size_max = tllen->_hi;
3202       const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3203       lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3204     }
3205   }
3206 #endif
3207 
3208   // Combine header size (plus rounding) and body size.  Then round down.
3209   // This computation cannot overflow, because it is used only in two
3210   // places, one where the length is sharply limited, and the other
3211   // after a successful allocation.
3212   Node* abody = lengthx;
3213   if (elem_shift != NULL)
3214     abody     = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3215   Node* size  = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3216   if (round_mask != 0) {
3217     Node* mask = MakeConX(~round_mask);
3218     size       = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3219   }
3220   // else if round_mask == 0, the size computation is self-rounding
3221 
3222   if (return_size_val != NULL) {
3223     // This is the size
3224     (*return_size_val) = size;
3225   }
3226 
3227   // Now generate allocation code
3228 
3229   // The entire memory state is needed for slow path of the allocation
3230   // since GC and deoptimization can happened.
3231   Node *mem = reset_memory();
3232   set_all_memory(mem); // Create new memory state
3233 
3234   // Create the AllocateArrayNode and its result projections
3235   AllocateArrayNode* alloc
3236     = new (C, AllocateArrayNode::ParmLimit)
3237         AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3238                           control(), mem, i_o(),
3239                           size, klass_node,
3240                           initial_slow_test,
3241                           length);
3242 
3243   // Cast to correct type.  Note that the klass_node may be constant or not,
3244   // and in the latter case the actual array type will be inexact also.
3245   // (This happens via a non-constant argument to inline_native_newArray.)
3246   // In any case, the value of klass_node provides the desired array type.
3247   const TypeInt* length_type = _gvn.find_int_type(length);
3248   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3249   if (ary_type->isa_aryptr() && length_type != NULL) {
3250     // Try to get a better type than POS for the size
3251     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3252   }
3253 
3254   Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
3255 
3256   // Cast length on remaining path to be as narrow as possible
3257   if (map()->find_edge(length) >= 0) {
3258     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3259     if (ccast != length) {
3260       _gvn.set_type_bottom(ccast);
3261       record_for_igvn(ccast);
3262       replace_in_map(length, ccast);
3263     }
3264   }
3265 
3266   return javaoop;
3267 }
3268 
3269 // The following "Ideal_foo" functions are placed here because they recognize
3270 // the graph shapes created by the functions immediately above.
3271 
3272 //---------------------------Ideal_allocation----------------------------------
3273 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3274 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3275   if (ptr == NULL) {     // reduce dumb test in callers
3276     return NULL;
3277   }
3278   if (ptr->is_CheckCastPP()) {  // strip a raw-to-oop cast
3279     ptr = ptr->in(1);
3280     if (ptr == NULL)  return NULL;
3281   }
3282   if (ptr->is_Proj()) {
3283     Node* allo = ptr->in(0);
3284     if (allo != NULL && allo->is_Allocate()) {
3285       return allo->as_Allocate();
3286     }
3287   }
3288   // Report failure to match.
3289   return NULL;
3290 }
3291 
3292 // Fancy version which also strips off an offset (and reports it to caller).
3293 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3294                                              intptr_t& offset) {
3295   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3296   if (base == NULL)  return NULL;
3297   return Ideal_allocation(base, phase);
3298 }
3299 
3300 // Trace Initialize <- Proj[Parm] <- Allocate
3301 AllocateNode* InitializeNode::allocation() {
3302   Node* rawoop = in(InitializeNode::RawAddress);
3303   if (rawoop->is_Proj()) {
3304     Node* alloc = rawoop->in(0);
3305     if (alloc->is_Allocate()) {
3306       return alloc->as_Allocate();
3307     }
3308   }
3309   return NULL;
3310 }
3311 
3312 // Trace Allocate -> Proj[Parm] -> Initialize
3313 InitializeNode* AllocateNode::initialization() {
3314   ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3315   if (rawoop == NULL)  return NULL;
3316   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3317     Node* init = rawoop->fast_out(i);
3318     if (init->is_Initialize()) {
3319       assert(init->as_Initialize()->allocation() == this, "2-way link");
3320       return init->as_Initialize();
3321     }
3322   }
3323   return NULL;
3324 }
3325 
3326 //----------------------------- store barriers ----------------------------
3327 #define __ ideal.
3328 
3329 void GraphKit::sync_kit(IdealKit& ideal) {
3330   // Final sync IdealKit and graphKit.
3331   __ drain_delay_transform();
3332   set_all_memory(__ merged_memory());
3333   set_control(__ ctrl());
3334 }
3335 
3336 // vanilla/CMS post barrier
3337 // Insert a write-barrier store.  This is to let generational GC work; we have
3338 // to flag all oop-stores before the next GC point.
3339 void GraphKit::write_barrier_post(Node* oop_store,
3340                                   Node* obj,
3341                                   Node* adr,
3342                                   uint  adr_idx,
3343                                   Node* val,
3344                                   bool use_precise) {
3345   // No store check needed if we're storing a NULL or an old object
3346   // (latter case is probably a string constant). The concurrent
3347   // mark sweep garbage collector, however, needs to have all nonNull
3348   // oop updates flagged via card-marks.
3349   if (val != NULL && val->is_Con()) {
3350     // must be either an oop or NULL
3351     const Type* t = val->bottom_type();
3352     if (t == TypePtr::NULL_PTR || t == Type::TOP)
3353       // stores of null never (?) need barriers
3354       return;
3355     ciObject* con = t->is_oopptr()->const_oop();
3356     if (con != NULL
3357         && con->is_perm()
3358         && Universe::heap()->can_elide_permanent_oop_store_barriers())
3359       // no store barrier needed, because no old-to-new ref created
3360       return;
3361   }
3362 
3363   if (use_ReduceInitialCardMarks()
3364       && obj == just_allocated_object(control())) {
3365     // We can skip marks on a freshly-allocated object in Eden.
3366     // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3367     // That routine informs GC to take appropriate compensating steps,
3368     // upon a slow-path allocation, so as to make this card-mark
3369     // elision safe.
3370     return;
3371   }
3372 
3373   if (!use_precise) {
3374     // All card marks for a (non-array) instance are in one place:
3375     adr = obj;
3376   }
3377   // (Else it's an array (or unknown), and we want more precise card marks.)
3378   assert(adr != NULL, "");
3379 
3380   IdealKit ideal(gvn(), control(), merged_memory(), true);
3381 
3382   // Convert the pointer to an int prior to doing math on it
3383   Node* cast = __ CastPX(__ ctrl(), adr);
3384 
3385   // Divide by card size
3386   assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3387          "Only one we handle so far.");
3388   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3389 
3390   // Combine card table base and card offset
3391   Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3392 
3393   // Get the alias_index for raw card-mark memory
3394   int adr_type = Compile::AliasIdxRaw;
3395   // Smash zero into card
3396   Node*   zero = __ ConI(0);
3397   BasicType bt = T_BYTE;
3398   if( !UseConcMarkSweepGC ) {
3399     __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3400   } else {
3401     // Specialized path for CM store barrier
3402     __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3403   }
3404 
3405   // Final sync IdealKit and GraphKit.
3406   sync_kit(ideal);
3407 }
3408 
3409 // G1 pre/post barriers
3410 void GraphKit::g1_write_barrier_pre(Node* obj,
3411                                     Node* adr,
3412                                     uint alias_idx,
3413                                     Node* val,
3414                                     const TypeOopPtr* val_type,
3415                                     BasicType bt) {
3416   IdealKit ideal(gvn(), control(), merged_memory(), true);
3417 
3418   Node* tls = __ thread(); // ThreadLocalStorage
3419 
3420   Node* no_ctrl = NULL;
3421   Node* no_base = __ top();
3422   Node* zero = __ ConI(0);
3423 
3424   float likely  = PROB_LIKELY(0.999);
3425   float unlikely  = PROB_UNLIKELY(0.999);
3426 
3427   BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3428   assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3429 
3430   // Offsets into the thread
3431   const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +  // 648
3432                                           PtrQueue::byte_offset_of_active());
3433   const int index_offset   = in_bytes(JavaThread::satb_mark_queue_offset() +  // 656
3434                                           PtrQueue::byte_offset_of_index());
3435   const int buffer_offset  = in_bytes(JavaThread::satb_mark_queue_offset() +  // 652
3436                                           PtrQueue::byte_offset_of_buf());
3437   // Now the actual pointers into the thread
3438 
3439   // set_control( ctl);
3440 
3441   Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3442   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
3443   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
3444 
3445   // Now some of the values
3446 
3447   Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3448 
3449   // if (!marking)
3450   __ if_then(marking, BoolTest::ne, zero); {
3451     Node* index   = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3452 
3453     const Type* t1 = adr->bottom_type();
3454     const Type* t2 = val->bottom_type();
3455 
3456     Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3457     // if (orig != NULL)
3458     __ if_then(orig, BoolTest::ne, null()); {
3459       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3460 
3461       // load original value
3462       // alias_idx correct??
3463 
3464       // is the queue for this thread full?
3465       __ if_then(index, BoolTest::ne, zero, likely); {
3466 
3467         // decrement the index
3468         Node* next_index = __ SubI(index,  __ ConI(sizeof(intptr_t)));
3469         Node* next_indexX = next_index;
3470 #ifdef _LP64
3471         // We could refine the type for what it's worth
3472         // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3473         next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3474 #endif
3475 
3476         // Now get the buffer location we will log the original value into and store it
3477         Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3478         __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw);
3479 
3480         // update the index
3481         __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3482 
3483       } __ else_(); {
3484 
3485         // logging buffer is full, call the runtime
3486         const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3487         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls);
3488       } __ end_if();  // (!index)
3489     } __ end_if();  // (orig != NULL)
3490   } __ end_if();  // (!marking)
3491 
3492   // Final sync IdealKit and GraphKit.
3493   sync_kit(ideal);
3494 }
3495 
3496 //
3497 // Update the card table and add card address to the queue
3498 //
3499 void GraphKit::g1_mark_card(IdealKit& ideal,
3500                             Node* card_adr,
3501                             Node* oop_store,
3502                             uint oop_alias_idx,
3503                             Node* index,
3504                             Node* index_adr,
3505                             Node* buffer,
3506                             const TypeFunc* tf) {
3507 
3508   Node* zero = __ ConI(0);
3509   Node* no_base = __ top();
3510   BasicType card_bt = T_BYTE;
3511   // Smash zero into card. MUST BE ORDERED WRT TO STORE
3512   __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3513 
3514   //  Now do the queue work
3515   __ if_then(index, BoolTest::ne, zero); {
3516 
3517     Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3518     Node* next_indexX = next_index;
3519 #ifdef _LP64
3520     // We could refine the type for what it's worth
3521     // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3522     next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3523 #endif // _LP64
3524     Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3525 
3526     __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3527     __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3528 
3529   } __ else_(); {
3530     __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3531   } __ end_if();
3532 
3533 }
3534 
3535 void GraphKit::g1_write_barrier_post(Node* oop_store,
3536                                      Node* obj,
3537                                      Node* adr,
3538                                      uint alias_idx,
3539                                      Node* val,
3540                                      BasicType bt,
3541                                      bool use_precise) {
3542   // If we are writing a NULL then we need no post barrier
3543 
3544   if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3545     // Must be NULL
3546     const Type* t = val->bottom_type();
3547     assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3548     // No post barrier if writing NULLx
3549     return;
3550   }
3551 
3552   if (!use_precise) {
3553     // All card marks for a (non-array) instance are in one place:
3554     adr = obj;
3555   }
3556   // (Else it's an array (or unknown), and we want more precise card marks.)
3557   assert(adr != NULL, "");
3558 
3559   IdealKit ideal(gvn(), control(), merged_memory(), true);
3560 
3561   Node* tls = __ thread(); // ThreadLocalStorage
3562 
3563   Node* no_base = __ top();
3564   float likely  = PROB_LIKELY(0.999);
3565   float unlikely  = PROB_UNLIKELY(0.999);
3566   Node* zero = __ ConI(0);
3567   Node* zeroX = __ ConX(0);
3568 
3569   // Get the alias_index for raw card-mark memory
3570   const TypePtr* card_type = TypeRawPtr::BOTTOM;
3571 
3572   const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3573 
3574   // Offsets into the thread
3575   const int index_offset  = in_bytes(JavaThread::dirty_card_queue_offset() +
3576                                      PtrQueue::byte_offset_of_index());
3577   const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3578                                      PtrQueue::byte_offset_of_buf());
3579 
3580   // Pointers into the thread
3581 
3582   Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3583   Node* index_adr =  __ AddP(no_base, tls, __ ConX(index_offset));
3584 
3585   // Now some values
3586   // Use ctrl to avoid hoisting these values past a safepoint, which could
3587   // potentially reset these fields in the JavaThread.
3588   Node* index  = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3589   Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3590 
3591   // Convert the store obj pointer to an int prior to doing math on it
3592   // Must use ctrl to prevent "integerized oop" existing across safepoint
3593   Node* cast =  __ CastPX(__ ctrl(), adr);
3594 
3595   // Divide pointer by card size
3596   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3597 
3598   // Combine card table base and card offset
3599   Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3600 
3601   // If we know the value being stored does it cross regions?
3602 
3603   if (val != NULL) {
3604     // Does the store cause us to cross regions?
3605 
3606     // Should be able to do an unsigned compare of region_size instead of
3607     // and extra shift. Do we have an unsigned compare??
3608     // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3609     Node* xor_res =  __ URShiftX ( __ XorX( cast,  __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3610 
3611     // if (xor_res == 0) same region so skip
3612     __ if_then(xor_res, BoolTest::ne, zeroX); {
3613 
3614       // No barrier if we are storing a NULL
3615       __ if_then(val, BoolTest::ne, null(), unlikely); {
3616 
3617         // Ok must mark the card if not already dirty
3618 
3619         // load the original value of the card
3620         Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3621 
3622         __ if_then(card_val, BoolTest::ne, zero); {
3623           g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3624         } __ end_if();
3625       } __ end_if();
3626     } __ end_if();
3627   } else {
3628     // Object.clone() instrinsic uses this path.
3629     g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3630   }
3631 
3632   // Final sync IdealKit and GraphKit.
3633   sync_kit(ideal);
3634 }
3635 #undef __