src/share/vm/opto/graphKit.cpp

Thu, 12 Jun 2008 13:50:55 -0700

author
ysr
date
Thu, 12 Jun 2008 13:50:55 -0700
changeset 779
6aae2f9d0294
parent 777
37f87013dfd8
parent 599
c436414a719e
child 791
1ee8caae33af
permissions
-rw-r--r--

Merge

     1 /*
     2  * Copyright 2001-2007 Sun Microsystems, Inc.  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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
    20  * CA 95054 USA or visit www.sun.com if you need additional information or
    21  * have any questions.
    22  *
    23  */
    25 #include "incls/_precompiled.incl"
    26 #include "incls/_graphKit.cpp.incl"
    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 }
    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 }
    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 }
    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 }
    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 }
    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 }
   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 }
   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
   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 }
   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 }
   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 }
   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 }
   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 }
   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 }
   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 }
   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
   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 }
   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
   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   }
   235   // No pre-existing exception of the same type.  Chain it on the list.
   236   push_exception_state(ex_map);
   237 }
   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 }
   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 }
   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 }
   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   }
   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();
   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 }
   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   }
   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 }
   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 }
   458 //------------------------------builtin_throw----------------------------------
   459 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
   460   bool must_throw = true;
   462   if (JvmtiExport::can_post_exceptions()) {
   463     // Do not try anything fancy if we're notifying the VM on every throw.
   464     // Cf. case Bytecodes::_athrow in parse2.cpp.
   465     uncommon_trap(reason, Deoptimization::Action_none,
   466                   (ciKlass*)NULL, (char*)NULL, must_throw);
   467     return;
   468   }
   470   // If this particular condition has not yet happened at this
   471   // bytecode, then use the uncommon trap mechanism, and allow for
   472   // a future recompilation if several traps occur here.
   473   // If the throw is hot, try to use a more complicated inline mechanism
   474   // which keeps execution inside the compiled code.
   475   bool treat_throw_as_hot = false;
   476   ciMethodData* md = method()->method_data();
   478   if (ProfileTraps) {
   479     if (too_many_traps(reason)) {
   480       treat_throw_as_hot = true;
   481     }
   482     // (If there is no MDO at all, assume it is early in
   483     // execution, and that any deopts are part of the
   484     // startup transient, and don't need to be remembered.)
   486     // Also, if there is a local exception handler, treat all throws
   487     // as hot if there has been at least one in this method.
   488     if (C->trap_count(reason) != 0
   489         && method()->method_data()->trap_count(reason) != 0
   490         && has_ex_handler()) {
   491         treat_throw_as_hot = true;
   492     }
   493   }
   495   // If this throw happens frequently, an uncommon trap might cause
   496   // a performance pothole.  If there is a local exception handler,
   497   // and if this particular bytecode appears to be deoptimizing often,
   498   // let us handle the throw inline, with a preconstructed instance.
   499   // Note:   If the deopt count has blown up, the uncommon trap
   500   // runtime is going to flush this nmethod, not matter what.
   501   if (treat_throw_as_hot
   502       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
   503     // If the throw is local, we use a pre-existing instance and
   504     // punt on the backtrace.  This would lead to a missing backtrace
   505     // (a repeat of 4292742) if the backtrace object is ever asked
   506     // for its backtrace.
   507     // Fixing this remaining case of 4292742 requires some flavor of
   508     // escape analysis.  Leave that for the future.
   509     ciInstance* ex_obj = NULL;
   510     switch (reason) {
   511     case Deoptimization::Reason_null_check:
   512       ex_obj = env()->NullPointerException_instance();
   513       break;
   514     case Deoptimization::Reason_div0_check:
   515       ex_obj = env()->ArithmeticException_instance();
   516       break;
   517     case Deoptimization::Reason_range_check:
   518       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
   519       break;
   520     case Deoptimization::Reason_class_check:
   521       if (java_bc() == Bytecodes::_aastore) {
   522         ex_obj = env()->ArrayStoreException_instance();
   523       } else {
   524         ex_obj = env()->ClassCastException_instance();
   525       }
   526       break;
   527     }
   528     if (failing()) { stop(); return; }  // exception allocation might fail
   529     if (ex_obj != NULL) {
   530       // Cheat with a preallocated exception object.
   531       if (C->log() != NULL)
   532         C->log()->elem("hot_throw preallocated='1' reason='%s'",
   533                        Deoptimization::trap_reason_name(reason));
   534       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
   535       Node*              ex_node = _gvn.transform( ConNode::make(C, ex_con) );
   537       // Clear the detail message of the preallocated exception object.
   538       // Weblogic sometimes mutates the detail message of exceptions
   539       // using reflection.
   540       int offset = java_lang_Throwable::get_detailMessage_offset();
   541       const TypePtr* adr_typ = ex_con->add_offset(offset);
   543       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
   544       Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT);
   546       add_exception_state(make_exception_state(ex_node));
   547       return;
   548     }
   549   }
   551   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
   552   // It won't be much cheaper than bailing to the interp., since we'll
   553   // have to pass up all the debug-info, and the runtime will have to
   554   // create the stack trace.
   556   // Usual case:  Bail to interpreter.
   557   // Reserve the right to recompile if we haven't seen anything yet.
   559   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
   560   if (treat_throw_as_hot
   561       && (method()->method_data()->trap_recompiled_at(bci())
   562           || C->too_many_traps(reason))) {
   563     // We cannot afford to take more traps here.  Suffer in the interpreter.
   564     if (C->log() != NULL)
   565       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
   566                      Deoptimization::trap_reason_name(reason),
   567                      C->trap_count(reason));
   568     action = Deoptimization::Action_none;
   569   }
   571   // "must_throw" prunes the JVM state to include only the stack, if there
   572   // are no local exception handlers.  This should cut down on register
   573   // allocation time and code size, by drastically reducing the number
   574   // of in-edges on the call to the uncommon trap.
   576   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
   577 }
   580 //----------------------------PreserveJVMState---------------------------------
   581 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
   582   debug_only(kit->verify_map());
   583   _kit    = kit;
   584   _map    = kit->map();   // preserve the map
   585   _sp     = kit->sp();
   586   kit->set_map(clone_map ? kit->clone_map() : NULL);
   587 #ifdef ASSERT
   588   _bci    = kit->bci();
   589   Parse* parser = kit->is_Parse();
   590   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->pre_order();
   591   _block  = block;
   592 #endif
   593 }
   594 PreserveJVMState::~PreserveJVMState() {
   595   GraphKit* kit = _kit;
   596 #ifdef ASSERT
   597   assert(kit->bci() == _bci, "bci must not shift");
   598   Parse* parser = kit->is_Parse();
   599   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->pre_order();
   600   assert(block == _block,    "block must not shift");
   601 #endif
   602   kit->set_map(_map);
   603   kit->set_sp(_sp);
   604 }
   607 //-----------------------------BuildCutout-------------------------------------
   608 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
   609   : PreserveJVMState(kit)
   610 {
   611   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
   612   SafePointNode* outer_map = _map;   // preserved map is caller's
   613   SafePointNode* inner_map = kit->map();
   614   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
   615   outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
   616   inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
   617 }
   618 BuildCutout::~BuildCutout() {
   619   GraphKit* kit = _kit;
   620   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
   621 }
   624 //------------------------------clone_map--------------------------------------
   625 // Implementation of PreserveJVMState
   626 //
   627 // Only clone_map(...) here. If this function is only used in the
   628 // PreserveJVMState class we may want to get rid of this extra
   629 // function eventually and do it all there.
   631 SafePointNode* GraphKit::clone_map() {
   632   if (map() == NULL)  return NULL;
   634   // Clone the memory edge first
   635   Node* mem = MergeMemNode::make(C, map()->memory());
   636   gvn().set_type_bottom(mem);
   638   SafePointNode *clonemap = (SafePointNode*)map()->clone();
   639   JVMState* jvms = this->jvms();
   640   JVMState* clonejvms = jvms->clone_shallow(C);
   641   clonemap->set_memory(mem);
   642   clonemap->set_jvms(clonejvms);
   643   clonejvms->set_map(clonemap);
   644   record_for_igvn(clonemap);
   645   gvn().set_type_bottom(clonemap);
   646   return clonemap;
   647 }
   650 //-----------------------------set_map_clone-----------------------------------
   651 void GraphKit::set_map_clone(SafePointNode* m) {
   652   _map = m;
   653   _map = clone_map();
   654   _map->set_next_exception(NULL);
   655   debug_only(verify_map());
   656 }
   659 //----------------------------kill_dead_locals---------------------------------
   660 // Detect any locals which are known to be dead, and force them to top.
   661 void GraphKit::kill_dead_locals() {
   662   // Consult the liveness information for the locals.  If any
   663   // of them are unused, then they can be replaced by top().  This
   664   // should help register allocation time and cut down on the size
   665   // of the deoptimization information.
   667   // This call is made from many of the bytecode handling
   668   // subroutines called from the Big Switch in do_one_bytecode.
   669   // Every bytecode which might include a slow path is responsible
   670   // for killing its dead locals.  The more consistent we
   671   // are about killing deads, the fewer useless phis will be
   672   // constructed for them at various merge points.
   674   // bci can be -1 (InvocationEntryBci).  We return the entry
   675   // liveness for the method.
   677   if (method() == NULL || method()->code_size() == 0) {
   678     // We are building a graph for a call to a native method.
   679     // All locals are live.
   680     return;
   681   }
   683   ResourceMark rm;
   685   // Consult the liveness information for the locals.  If any
   686   // of them are unused, then they can be replaced by top().  This
   687   // should help register allocation time and cut down on the size
   688   // of the deoptimization information.
   689   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
   691   int len = (int)live_locals.size();
   692   assert(len <= jvms()->loc_size(), "too many live locals");
   693   for (int local = 0; local < len; local++) {
   694     if (!live_locals.at(local)) {
   695       set_local(local, top());
   696     }
   697   }
   698 }
   700 #ifdef ASSERT
   701 //-------------------------dead_locals_are_killed------------------------------
   702 // Return true if all dead locals are set to top in the map.
   703 // Used to assert "clean" debug info at various points.
   704 bool GraphKit::dead_locals_are_killed() {
   705   if (method() == NULL || method()->code_size() == 0) {
   706     // No locals need to be dead, so all is as it should be.
   707     return true;
   708   }
   710   // Make sure somebody called kill_dead_locals upstream.
   711   ResourceMark rm;
   712   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
   713     if (jvms->loc_size() == 0)  continue;  // no locals to consult
   714     SafePointNode* map = jvms->map();
   715     ciMethod* method = jvms->method();
   716     int       bci    = jvms->bci();
   717     if (jvms == this->jvms()) {
   718       bci = this->bci();  // it might not yet be synched
   719     }
   720     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
   721     int len = (int)live_locals.size();
   722     if (!live_locals.is_valid() || len == 0)
   723       // This method is trivial, or is poisoned by a breakpoint.
   724       return true;
   725     assert(len == jvms->loc_size(), "live map consistent with locals map");
   726     for (int local = 0; local < len; local++) {
   727       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
   728         if (PrintMiscellaneous && (Verbose || WizardMode)) {
   729           tty->print_cr("Zombie local %d: ", local);
   730           jvms->dump();
   731         }
   732         return false;
   733       }
   734     }
   735   }
   736   return true;
   737 }
   739 #endif //ASSERT
   741 // Helper function for adding JVMState and debug information to node
   742 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
   743   // Add the safepoint edges to the call (or other safepoint).
   745   // Make sure dead locals are set to top.  This
   746   // should help register allocation time and cut down on the size
   747   // of the deoptimization information.
   748   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
   750   // Walk the inline list to fill in the correct set of JVMState's
   751   // Also fill in the associated edges for each JVMState.
   753   JVMState* youngest_jvms = sync_jvms();
   755   // Do we need debug info here?  If it is a SafePoint and this method
   756   // cannot de-opt, then we do NOT need any debug info.
   757   bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint);
   759   // If we are guaranteed to throw, we can prune everything but the
   760   // input to the current bytecode.
   761   bool can_prune_locals = false;
   762   uint stack_slots_not_pruned = 0;
   763   int inputs = 0, depth = 0;
   764   if (must_throw) {
   765     assert(method() == youngest_jvms->method(), "sanity");
   766     if (compute_stack_effects(inputs, depth)) {
   767       can_prune_locals = true;
   768       stack_slots_not_pruned = inputs;
   769     }
   770   }
   772   if (JvmtiExport::can_examine_or_deopt_anywhere()) {
   773     // At any safepoint, this method can get breakpointed, which would
   774     // then require an immediate deoptimization.
   775     full_info = true;
   776     can_prune_locals = false;  // do not prune locals
   777     stack_slots_not_pruned = 0;
   778   }
   780   // do not scribble on the input jvms
   781   JVMState* out_jvms = youngest_jvms->clone_deep(C);
   782   call->set_jvms(out_jvms); // Start jvms list for call node
   784   // Presize the call:
   785   debug_only(uint non_debug_edges = call->req());
   786   call->add_req_batch(top(), youngest_jvms->debug_depth());
   787   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
   789   // Set up edges so that the call looks like this:
   790   //  Call [state:] ctl io mem fptr retadr
   791   //       [parms:] parm0 ... parmN
   792   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
   793   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
   794   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
   795   // Note that caller debug info precedes callee debug info.
   797   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
   798   uint debug_ptr = call->req();
   800   // Loop over the map input edges associated with jvms, add them
   801   // to the call node, & reset all offsets to match call node array.
   802   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
   803     uint debug_end   = debug_ptr;
   804     uint debug_start = debug_ptr - in_jvms->debug_size();
   805     debug_ptr = debug_start;  // back up the ptr
   807     uint p = debug_start;  // walks forward in [debug_start, debug_end)
   808     uint j, k, l;
   809     SafePointNode* in_map = in_jvms->map();
   810     out_jvms->set_map(call);
   812     if (can_prune_locals) {
   813       assert(in_jvms->method() == out_jvms->method(), "sanity");
   814       // If the current throw can reach an exception handler in this JVMS,
   815       // then we must keep everything live that can reach that handler.
   816       // As a quick and dirty approximation, we look for any handlers at all.
   817       if (in_jvms->method()->has_exception_handlers()) {
   818         can_prune_locals = false;
   819       }
   820     }
   822     // Add the Locals
   823     k = in_jvms->locoff();
   824     l = in_jvms->loc_size();
   825     out_jvms->set_locoff(p);
   826     if (full_info && !can_prune_locals) {
   827       for (j = 0; j < l; j++)
   828         call->set_req(p++, in_map->in(k+j));
   829     } else {
   830       p += l;  // already set to top above by add_req_batch
   831     }
   833     // Add the Expression Stack
   834     k = in_jvms->stkoff();
   835     l = in_jvms->sp();
   836     out_jvms->set_stkoff(p);
   837     if (full_info && !can_prune_locals) {
   838       for (j = 0; j < l; j++)
   839         call->set_req(p++, in_map->in(k+j));
   840     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
   841       // Divide stack into {S0,...,S1}, where S0 is set to top.
   842       uint s1 = stack_slots_not_pruned;
   843       stack_slots_not_pruned = 0;  // for next iteration
   844       if (s1 > l)  s1 = l;
   845       uint s0 = l - s1;
   846       p += s0;  // skip the tops preinstalled by add_req_batch
   847       for (j = s0; j < l; j++)
   848         call->set_req(p++, in_map->in(k+j));
   849     } else {
   850       p += l;  // already set to top above by add_req_batch
   851     }
   853     // Add the Monitors
   854     k = in_jvms->monoff();
   855     l = in_jvms->mon_size();
   856     out_jvms->set_monoff(p);
   857     for (j = 0; j < l; j++)
   858       call->set_req(p++, in_map->in(k+j));
   860     // Copy any scalar object fields.
   861     k = in_jvms->scloff();
   862     l = in_jvms->scl_size();
   863     out_jvms->set_scloff(p);
   864     for (j = 0; j < l; j++)
   865       call->set_req(p++, in_map->in(k+j));
   867     // Finish the new jvms.
   868     out_jvms->set_endoff(p);
   870     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
   871     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
   872     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
   873     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
   874     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
   875     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
   877     // Update the two tail pointers in parallel.
   878     out_jvms = out_jvms->caller();
   879     in_jvms  = in_jvms->caller();
   880   }
   882   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
   884   // Test the correctness of JVMState::debug_xxx accessors:
   885   assert(call->jvms()->debug_start() == non_debug_edges, "");
   886   assert(call->jvms()->debug_end()   == call->req(), "");
   887   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
   888 }
   890 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
   891   Bytecodes::Code code = java_bc();
   892   if (code == Bytecodes::_wide) {
   893     code = method()->java_code_at_bci(bci() + 1);
   894   }
   896   BasicType rtype = T_ILLEGAL;
   897   int       rsize = 0;
   899   if (code != Bytecodes::_illegal) {
   900     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
   901     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
   902     if (rtype < T_CONFLICT)
   903       rsize = type2size[rtype];
   904   }
   906   switch (code) {
   907   case Bytecodes::_illegal:
   908     return false;
   910   case Bytecodes::_ldc:
   911   case Bytecodes::_ldc_w:
   912   case Bytecodes::_ldc2_w:
   913     inputs = 0;
   914     break;
   916   case Bytecodes::_dup:         inputs = 1;  break;
   917   case Bytecodes::_dup_x1:      inputs = 2;  break;
   918   case Bytecodes::_dup_x2:      inputs = 3;  break;
   919   case Bytecodes::_dup2:        inputs = 2;  break;
   920   case Bytecodes::_dup2_x1:     inputs = 3;  break;
   921   case Bytecodes::_dup2_x2:     inputs = 4;  break;
   922   case Bytecodes::_swap:        inputs = 2;  break;
   923   case Bytecodes::_arraylength: inputs = 1;  break;
   925   case Bytecodes::_getstatic:
   926   case Bytecodes::_putstatic:
   927   case Bytecodes::_getfield:
   928   case Bytecodes::_putfield:
   929     {
   930       bool is_get = (depth >= 0), is_static = (depth & 1);
   931       bool ignore;
   932       ciBytecodeStream iter(method());
   933       iter.reset_to_bci(bci());
   934       iter.next();
   935       ciField* field = iter.get_field(ignore);
   936       int      size  = field->type()->size();
   937       inputs  = (is_static ? 0 : 1);
   938       if (is_get) {
   939         depth = size - inputs;
   940       } else {
   941         inputs += size;        // putxxx pops the value from the stack
   942         depth = - inputs;
   943       }
   944     }
   945     break;
   947   case Bytecodes::_invokevirtual:
   948   case Bytecodes::_invokespecial:
   949   case Bytecodes::_invokestatic:
   950   case Bytecodes::_invokeinterface:
   951     {
   952       bool is_static = (depth == 0);
   953       bool ignore;
   954       ciBytecodeStream iter(method());
   955       iter.reset_to_bci(bci());
   956       iter.next();
   957       ciMethod* method = iter.get_method(ignore);
   958       inputs = method->arg_size_no_receiver();
   959       if (!is_static)  inputs += 1;
   960       int size = method->return_type()->size();
   961       depth = size - inputs;
   962     }
   963     break;
   965   case Bytecodes::_multianewarray:
   966     {
   967       ciBytecodeStream iter(method());
   968       iter.reset_to_bci(bci());
   969       iter.next();
   970       inputs = iter.get_dimensions();
   971       assert(rsize == 1, "");
   972       depth = rsize - inputs;
   973     }
   974     break;
   976   case Bytecodes::_ireturn:
   977   case Bytecodes::_lreturn:
   978   case Bytecodes::_freturn:
   979   case Bytecodes::_dreturn:
   980   case Bytecodes::_areturn:
   981     assert(rsize = -depth, "");
   982     inputs = rsize;
   983     break;
   985   case Bytecodes::_jsr:
   986   case Bytecodes::_jsr_w:
   987     inputs = 0;
   988     depth  = 1;                  // S.B. depth=1, not zero
   989     break;
   991   default:
   992     // bytecode produces a typed result
   993     inputs = rsize - depth;
   994     assert(inputs >= 0, "");
   995     break;
   996   }
   998 #ifdef ASSERT
   999   // spot check
  1000   int outputs = depth + inputs;
  1001   assert(outputs >= 0, "sanity");
  1002   switch (code) {
  1003   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
  1004   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
  1005   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
  1006   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
  1007   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
  1009 #endif //ASSERT
  1011   return true;
  1016 //------------------------------basic_plus_adr---------------------------------
  1017 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
  1018   // short-circuit a common case
  1019   if (offset == intcon(0))  return ptr;
  1020   return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
  1023 Node* GraphKit::ConvI2L(Node* offset) {
  1024   // short-circuit a common case
  1025   jint offset_con = find_int_con(offset, Type::OffsetBot);
  1026   if (offset_con != Type::OffsetBot) {
  1027     return longcon((long) offset_con);
  1029   return _gvn.transform( new (C, 2) ConvI2LNode(offset));
  1031 Node* GraphKit::ConvL2I(Node* offset) {
  1032   // short-circuit a common case
  1033   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
  1034   if (offset_con != (jlong)Type::OffsetBot) {
  1035     return intcon((int) offset_con);
  1037   return _gvn.transform( new (C, 2) ConvL2INode(offset));
  1040 //-------------------------load_object_klass-----------------------------------
  1041 Node* GraphKit::load_object_klass(Node* obj) {
  1042   // Special-case a fresh allocation to avoid building nodes:
  1043   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
  1044   if (akls != NULL)  return akls;
  1045   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
  1046   return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
  1049 //-------------------------load_array_length-----------------------------------
  1050 Node* GraphKit::load_array_length(Node* array) {
  1051   // Special-case a fresh allocation to avoid building nodes:
  1052   Node* alen = AllocateArrayNode::Ideal_length(array, &_gvn);
  1053   if (alen != NULL)  return alen;
  1054   Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
  1055   return _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
  1058 //------------------------------do_null_check----------------------------------
  1059 // Helper function to do a NULL pointer check.  Returned value is
  1060 // the incoming address with NULL casted away.  You are allowed to use the
  1061 // not-null value only if you are control dependent on the test.
  1062 extern int explicit_null_checks_inserted,
  1063            explicit_null_checks_elided;
  1064 Node* GraphKit::null_check_common(Node* value, BasicType type,
  1065                                   // optional arguments for variations:
  1066                                   bool assert_null,
  1067                                   Node* *null_control) {
  1068   assert(!assert_null || null_control == NULL, "not both at once");
  1069   if (stopped())  return top();
  1070   if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
  1071     // For some performance testing, we may wish to suppress null checking.
  1072     value = cast_not_null(value);   // Make it appear to be non-null (4962416).
  1073     return value;
  1075   explicit_null_checks_inserted++;
  1077   // Construct NULL check
  1078   Node *chk = NULL;
  1079   switch(type) {
  1080     case T_LONG   : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
  1081     case T_INT    : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
  1082     case T_ARRAY  : // fall through
  1083       type = T_OBJECT;  // simplify further tests
  1084     case T_OBJECT : {
  1085       const Type *t = _gvn.type( value );
  1087       const TypeInstPtr* tp = t->isa_instptr();
  1088       if (tp != NULL && !tp->klass()->is_loaded()
  1089           // Only for do_null_check, not any of its siblings:
  1090           && !assert_null && null_control == NULL) {
  1091         // Usually, any field access or invocation on an unloaded oop type
  1092         // will simply fail to link, since the statically linked class is
  1093         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
  1094         // the static class is loaded but the sharper oop type is not.
  1095         // Rather than checking for this obscure case in lots of places,
  1096         // we simply observe that a null check on an unloaded class
  1097         // will always be followed by a nonsense operation, so we
  1098         // can just issue the uncommon trap here.
  1099         // Our access to the unloaded class will only be correct
  1100         // after it has been loaded and initialized, which requires
  1101         // a trip through the interpreter.
  1102 #ifndef PRODUCT
  1103         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
  1104 #endif
  1105         uncommon_trap(Deoptimization::Reason_unloaded,
  1106                       Deoptimization::Action_reinterpret,
  1107                       tp->klass(), "!loaded");
  1108         return top();
  1111       if (assert_null) {
  1112         // See if the type is contained in NULL_PTR.
  1113         // If so, then the value is already null.
  1114         if (t->higher_equal(TypePtr::NULL_PTR)) {
  1115           explicit_null_checks_elided++;
  1116           return value;           // Elided null assert quickly!
  1118       } else {
  1119         // See if mixing in the NULL pointer changes type.
  1120         // If so, then the NULL pointer was not allowed in the original
  1121         // type.  In other words, "value" was not-null.
  1122         if (t->meet(TypePtr::NULL_PTR) != t) {
  1123           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
  1124           explicit_null_checks_elided++;
  1125           return value;           // Elided null check quickly!
  1128       chk = new (C, 3) CmpPNode( value, null() );
  1129       break;
  1132     default      : ShouldNotReachHere();
  1134   assert(chk != NULL, "sanity check");
  1135   chk = _gvn.transform(chk);
  1137   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
  1138   BoolNode *btst = new (C, 2) BoolNode( chk, btest);
  1139   Node   *tst = _gvn.transform( btst );
  1141   //-----------
  1142   // if peephole optimizations occured, a prior test existed.
  1143   // If a prior test existed, maybe it dominates as we can avoid this test.
  1144   if (tst != btst && type == T_OBJECT) {
  1145     // At this point we want to scan up the CFG to see if we can
  1146     // find an identical test (and so avoid this test altogether).
  1147     Node *cfg = control();
  1148     int depth = 0;
  1149     while( depth < 16 ) {       // Limit search depth for speed
  1150       if( cfg->Opcode() == Op_IfTrue &&
  1151           cfg->in(0)->in(1) == tst ) {
  1152         // Found prior test.  Use "cast_not_null" to construct an identical
  1153         // CastPP (and hence hash to) as already exists for the prior test.
  1154         // Return that casted value.
  1155         if (assert_null) {
  1156           replace_in_map(value, null());
  1157           return null();  // do not issue the redundant test
  1159         Node *oldcontrol = control();
  1160         set_control(cfg);
  1161         Node *res = cast_not_null(value);
  1162         set_control(oldcontrol);
  1163         explicit_null_checks_elided++;
  1164         return res;
  1166       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
  1167       if (cfg == NULL)  break;  // Quit at region nodes
  1168       depth++;
  1172   //-----------
  1173   // Branch to failure if null
  1174   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
  1175   Deoptimization::DeoptReason reason;
  1176   if (assert_null)
  1177     reason = Deoptimization::Reason_null_assert;
  1178   else if (type == T_OBJECT)
  1179     reason = Deoptimization::Reason_null_check;
  1180   else
  1181     reason = Deoptimization::Reason_div0_check;
  1183   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
  1184   // ciMethodData::has_trap_at will return a conservative -1 if any
  1185   // must-be-null assertion has failed.  This could cause performance
  1186   // problems for a method after its first do_null_assert failure.
  1187   // Consider using 'Reason_class_check' instead?
  1189   // To cause an implicit null check, we set the not-null probability
  1190   // to the maximum (PROB_MAX).  For an explicit check the probablity
  1191   // is set to a smaller value.
  1192   if (null_control != NULL || too_many_traps(reason)) {
  1193     // probability is less likely
  1194     ok_prob =  PROB_LIKELY_MAG(3);
  1195   } else if (!assert_null &&
  1196              (ImplicitNullCheckThreshold > 0) &&
  1197              method() != NULL &&
  1198              (method()->method_data()->trap_count(reason)
  1199               >= (uint)ImplicitNullCheckThreshold)) {
  1200     ok_prob =  PROB_LIKELY_MAG(3);
  1203   if (null_control != NULL) {
  1204     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
  1205     Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
  1206     set_control(      _gvn.transform( new (C, 1) IfTrueNode(iff)));
  1207     if (null_true == top())
  1208       explicit_null_checks_elided++;
  1209     (*null_control) = null_true;
  1210   } else {
  1211     BuildCutout unless(this, tst, ok_prob);
  1212     // Check for optimizer eliding test at parse time
  1213     if (stopped()) {
  1214       // Failure not possible; do not bother making uncommon trap.
  1215       explicit_null_checks_elided++;
  1216     } else if (assert_null) {
  1217       uncommon_trap(reason,
  1218                     Deoptimization::Action_make_not_entrant,
  1219                     NULL, "assert_null");
  1220     } else {
  1221       builtin_throw(reason);
  1225   // Must throw exception, fall-thru not possible?
  1226   if (stopped()) {
  1227     return top();               // No result
  1230   if (assert_null) {
  1231     // Cast obj to null on this path.
  1232     replace_in_map(value, zerocon(type));
  1233     return zerocon(type);
  1236   // Cast obj to not-null on this path, if there is no null_control.
  1237   // (If there is a null_control, a non-null value may come back to haunt us.)
  1238   if (type == T_OBJECT) {
  1239     Node* cast = cast_not_null(value, false);
  1240     if (null_control == NULL || (*null_control) == top())
  1241       replace_in_map(value, cast);
  1242     value = cast;
  1245   return value;
  1249 //------------------------------cast_not_null----------------------------------
  1250 // Cast obj to not-null on this path
  1251 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
  1252   const Type *t = _gvn.type(obj);
  1253   const Type *t_not_null = t->join(TypePtr::NOTNULL);
  1254   // Object is already not-null?
  1255   if( t == t_not_null ) return obj;
  1257   Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
  1258   cast->init_req(0, control());
  1259   cast = _gvn.transform( cast );
  1261   // Scan for instances of 'obj' in the current JVM mapping.
  1262   // These instances are known to be not-null after the test.
  1263   if (do_replace_in_map)
  1264     replace_in_map(obj, cast);
  1266   return cast;                  // Return casted value
  1270 //--------------------------replace_in_map-------------------------------------
  1271 void GraphKit::replace_in_map(Node* old, Node* neww) {
  1272   this->map()->replace_edge(old, neww);
  1274   // Note: This operation potentially replaces any edge
  1275   // on the map.  This includes locals, stack, and monitors
  1276   // of the current (innermost) JVM state.
  1278   // We can consider replacing in caller maps.
  1279   // The idea would be that an inlined function's null checks
  1280   // can be shared with the entire inlining tree.
  1281   // The expense of doing this is that the PreserveJVMState class
  1282   // would have to preserve caller states too, with a deep copy.
  1287 //=============================================================================
  1288 //--------------------------------memory---------------------------------------
  1289 Node* GraphKit::memory(uint alias_idx) {
  1290   MergeMemNode* mem = merged_memory();
  1291   Node* p = mem->memory_at(alias_idx);
  1292   _gvn.set_type(p, Type::MEMORY);  // must be mapped
  1293   return p;
  1296 //-----------------------------reset_memory------------------------------------
  1297 Node* GraphKit::reset_memory() {
  1298   Node* mem = map()->memory();
  1299   // do not use this node for any more parsing!
  1300   debug_only( map()->set_memory((Node*)NULL) );
  1301   return _gvn.transform( mem );
  1304 //------------------------------set_all_memory---------------------------------
  1305 void GraphKit::set_all_memory(Node* newmem) {
  1306   Node* mergemem = MergeMemNode::make(C, newmem);
  1307   gvn().set_type_bottom(mergemem);
  1308   map()->set_memory(mergemem);
  1311 //------------------------------set_all_memory_call----------------------------
  1312 void GraphKit::set_all_memory_call(Node* call) {
  1313   Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
  1314   set_all_memory(newmem);
  1317 //=============================================================================
  1318 //
  1319 // parser factory methods for MemNodes
  1320 //
  1321 // These are layered on top of the factory methods in LoadNode and StoreNode,
  1322 // and integrate with the parser's memory state and _gvn engine.
  1323 //
  1325 // factory methods in "int adr_idx"
  1326 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
  1327                           int adr_idx,
  1328                           bool require_atomic_access) {
  1329   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
  1330   const TypePtr* adr_type = NULL; // debug-mode-only argument
  1331   debug_only(adr_type = C->get_adr_type(adr_idx));
  1332   Node* mem = memory(adr_idx);
  1333   Node* ld;
  1334   if (require_atomic_access && bt == T_LONG) {
  1335     ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
  1336   } else {
  1337     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
  1339   return _gvn.transform(ld);
  1342 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
  1343                                 int adr_idx,
  1344                                 bool require_atomic_access) {
  1345   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
  1346   const TypePtr* adr_type = NULL;
  1347   debug_only(adr_type = C->get_adr_type(adr_idx));
  1348   Node *mem = memory(adr_idx);
  1349   Node* st;
  1350   if (require_atomic_access && bt == T_LONG) {
  1351     st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
  1352   } else {
  1353     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
  1355   st = _gvn.transform(st);
  1356   set_memory(st, adr_idx);
  1357   // Back-to-back stores can only remove intermediate store with DU info
  1358   // so push on worklist for optimizer.
  1359   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
  1360     record_for_igvn(st);
  1362   return st;
  1365 void GraphKit::pre_barrier(Node* ctl,
  1366                            Node* obj,
  1367                            Node* adr,
  1368                            uint adr_idx,
  1369                            Node *val,
  1370                            const Type* val_type,
  1371                            BasicType bt) {
  1372   BarrierSet* bs = Universe::heap()->barrier_set();
  1373   set_control(ctl);
  1374   switch (bs->kind()) {
  1375     case BarrierSet::G1SATBCT:
  1376     case BarrierSet::G1SATBCTLogging:
  1377         g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt);
  1378       break;
  1380     case BarrierSet::CardTableModRef:
  1381     case BarrierSet::CardTableExtension:
  1382     case BarrierSet::ModRef:
  1383       break;
  1385     case BarrierSet::Other:
  1386     default      :
  1387       ShouldNotReachHere();
  1392 void GraphKit::post_barrier(Node* ctl,
  1393                             Node* store,
  1394                             Node* obj,
  1395                             Node* adr,
  1396                             uint adr_idx,
  1397                             Node *val,
  1398                             BasicType bt,
  1399                             bool use_precise) {
  1400   BarrierSet* bs = Universe::heap()->barrier_set();
  1401   set_control(ctl);
  1402   switch (bs->kind()) {
  1403     case BarrierSet::G1SATBCT:
  1404     case BarrierSet::G1SATBCTLogging:
  1405         g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
  1406       break;
  1408     case BarrierSet::CardTableModRef:
  1409     case BarrierSet::CardTableExtension:
  1410       write_barrier_post(store, obj, adr, val, use_precise);
  1411       break;
  1413     case BarrierSet::ModRef:
  1414       break;
  1416     case BarrierSet::Other:
  1417     default      :
  1418       ShouldNotReachHere();
  1423 Node* GraphKit::store_oop_to_object(Node* ctl,
  1424                                     Node* obj,
  1425                                     Node* adr,
  1426                                     const TypePtr* adr_type,
  1427                                     Node *val,
  1428                                     const Type* val_type,
  1429                                     BasicType bt) {
  1430   uint adr_idx = C->get_alias_index(adr_type);
  1431   Node* store;
  1432   pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
  1433   store = store_to_memory(control(), adr, val, bt, adr_idx);
  1434   post_barrier(control(), store, obj, adr, adr_idx, val, bt, false);
  1435   return store;
  1438 Node* GraphKit::store_oop_to_array(Node* ctl,
  1439                                    Node* obj,
  1440                                    Node* adr,
  1441                                    const TypePtr* adr_type,
  1442                                    Node *val,
  1443                                    const Type* val_type,
  1444                                    BasicType bt) {
  1445   uint adr_idx = C->get_alias_index(adr_type);
  1446   Node* store;
  1447   pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
  1448   store = store_to_memory(control(), adr, val, bt, adr_idx);
  1449   post_barrier(control(), store, obj, adr, adr_idx, val, bt, true);
  1450   return store;
  1453 Node* GraphKit::store_oop_to_unknown(Node* ctl,
  1454                                      Node* obj,
  1455                                      Node* adr,
  1456                                      const TypePtr* adr_type,
  1457                                      Node *val,
  1458                                      const Type* val_type,
  1459                                      BasicType bt) {
  1460   uint adr_idx = C->get_alias_index(adr_type);
  1461   Node* store;
  1462   pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
  1463   store = store_to_memory(control(), adr, val, bt, adr_idx);
  1464   post_barrier(control(), store, obj, adr, adr_idx, val, bt, true);
  1465   return store;
  1469 //-------------------------array_element_address-------------------------
  1470 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
  1471                                       const TypeInt* sizetype) {
  1472   uint shift  = exact_log2(type2aelembytes(elembt));
  1473   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
  1475   // short-circuit a common case (saves lots of confusing waste motion)
  1476   jint idx_con = find_int_con(idx, -1);
  1477   if (idx_con >= 0) {
  1478     intptr_t offset = header + ((intptr_t)idx_con << shift);
  1479     return basic_plus_adr(ary, offset);
  1482   // must be correct type for alignment purposes
  1483   Node* base  = basic_plus_adr(ary, header);
  1484 #ifdef _LP64
  1485   // The scaled index operand to AddP must be a clean 64-bit value.
  1486   // Java allows a 32-bit int to be incremented to a negative
  1487   // value, which appears in a 64-bit register as a large
  1488   // positive number.  Using that large positive number as an
  1489   // operand in pointer arithmetic has bad consequences.
  1490   // On the other hand, 32-bit overflow is rare, and the possibility
  1491   // can often be excluded, if we annotate the ConvI2L node with
  1492   // a type assertion that its value is known to be a small positive
  1493   // number.  (The prior range check has ensured this.)
  1494   // This assertion is used by ConvI2LNode::Ideal.
  1495   int index_max = max_jint - 1;  // array size is max_jint, index is one less
  1496   if (sizetype != NULL)  index_max = sizetype->_hi - 1;
  1497   const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
  1498   idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
  1499 #endif
  1500   Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
  1501   return basic_plus_adr(ary, base, scale);
  1504 //-------------------------load_array_element-------------------------
  1505 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
  1506   const Type* elemtype = arytype->elem();
  1507   BasicType elembt = elemtype->array_element_basic_type();
  1508   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
  1509   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
  1510   return ld;
  1513 //-------------------------set_arguments_for_java_call-------------------------
  1514 // Arguments (pre-popped from the stack) are taken from the JVMS.
  1515 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
  1516   // Add the call arguments:
  1517   uint nargs = call->method()->arg_size();
  1518   for (uint i = 0; i < nargs; i++) {
  1519     Node* arg = argument(i);
  1520     call->init_req(i + TypeFunc::Parms, arg);
  1524 //---------------------------set_edges_for_java_call---------------------------
  1525 // Connect a newly created call into the current JVMS.
  1526 // A return value node (if any) is returned from set_edges_for_java_call.
  1527 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw) {
  1529   // Add the predefined inputs:
  1530   call->init_req( TypeFunc::Control, control() );
  1531   call->init_req( TypeFunc::I_O    , i_o() );
  1532   call->init_req( TypeFunc::Memory , reset_memory() );
  1533   call->init_req( TypeFunc::FramePtr, frameptr() );
  1534   call->init_req( TypeFunc::ReturnAdr, top() );
  1536   add_safepoint_edges(call, must_throw);
  1538   Node* xcall = _gvn.transform(call);
  1540   if (xcall == top()) {
  1541     set_control(top());
  1542     return;
  1544   assert(xcall == call, "call identity is stable");
  1546   // Re-use the current map to produce the result.
  1548   set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
  1549   set_i_o(    _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O    )));
  1550   set_all_memory_call(xcall);
  1552   //return xcall;   // no need, caller already has it
  1555 Node* GraphKit::set_results_for_java_call(CallJavaNode* call) {
  1556   if (stopped())  return top();  // maybe the call folded up?
  1558   // Capture the return value, if any.
  1559   Node* ret;
  1560   if (call->method() == NULL ||
  1561       call->method()->return_type()->basic_type() == T_VOID)
  1562         ret = top();
  1563   else  ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
  1565   // Note:  Since any out-of-line call can produce an exception,
  1566   // we always insert an I_O projection from the call into the result.
  1568   make_slow_call_ex(call, env()->Throwable_klass(), false);
  1570   return ret;
  1573 //--------------------set_predefined_input_for_runtime_call--------------------
  1574 // Reading and setting the memory state is way conservative here.
  1575 // The real problem is that I am not doing real Type analysis on memory,
  1576 // so I cannot distinguish card mark stores from other stores.  Across a GC
  1577 // point the Store Barrier and the card mark memory has to agree.  I cannot
  1578 // have a card mark store and its barrier split across the GC point from
  1579 // either above or below.  Here I get that to happen by reading ALL of memory.
  1580 // A better answer would be to separate out card marks from other memory.
  1581 // For now, return the input memory state, so that it can be reused
  1582 // after the call, if this call has restricted memory effects.
  1583 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
  1584   // Set fixed predefined input arguments
  1585   Node* memory = reset_memory();
  1586   call->init_req( TypeFunc::Control,   control()  );
  1587   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
  1588   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
  1589   call->init_req( TypeFunc::FramePtr,  frameptr() );
  1590   call->init_req( TypeFunc::ReturnAdr, top()      );
  1591   return memory;
  1594 //-------------------set_predefined_output_for_runtime_call--------------------
  1595 // Set control and memory (not i_o) from the call.
  1596 // If keep_mem is not NULL, use it for the output state,
  1597 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
  1598 // If hook_mem is NULL, this call produces no memory effects at all.
  1599 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
  1600 // then only that memory slice is taken from the call.
  1601 // In the last case, we must put an appropriate memory barrier before
  1602 // the call, so as to create the correct anti-dependencies on loads
  1603 // preceding the call.
  1604 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
  1605                                                       Node* keep_mem,
  1606                                                       const TypePtr* hook_mem) {
  1607   // no i/o
  1608   set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
  1609   if (keep_mem) {
  1610     // First clone the existing memory state
  1611     set_all_memory(keep_mem);
  1612     if (hook_mem != NULL) {
  1613       // Make memory for the call
  1614       Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
  1615       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
  1616       // We also use hook_mem to extract specific effects from arraycopy stubs.
  1617       set_memory(mem, hook_mem);
  1619     // ...else the call has NO memory effects.
  1621     // Make sure the call advertises its memory effects precisely.
  1622     // This lets us build accurate anti-dependences in gcm.cpp.
  1623     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
  1624            "call node must be constructed correctly");
  1625   } else {
  1626     assert(hook_mem == NULL, "");
  1627     // This is not a "slow path" call; all memory comes from the call.
  1628     set_all_memory_call(call);
  1632 //------------------------------increment_counter------------------------------
  1633 // for statistics: increment a VM counter by 1
  1635 void GraphKit::increment_counter(address counter_addr) {
  1636   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
  1637   increment_counter(adr1);
  1640 void GraphKit::increment_counter(Node* counter_addr) {
  1641   int adr_type = Compile::AliasIdxRaw;
  1642   Node* cnt  = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
  1643   Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
  1644   store_to_memory( NULL, counter_addr, incr, T_INT, adr_type );
  1648 //------------------------------uncommon_trap----------------------------------
  1649 // Bail out to the interpreter in mid-method.  Implemented by calling the
  1650 // uncommon_trap blob.  This helper function inserts a runtime call with the
  1651 // right debug info.
  1652 void GraphKit::uncommon_trap(int trap_request,
  1653                              ciKlass* klass, const char* comment,
  1654                              bool must_throw,
  1655                              bool keep_exact_action) {
  1656   if (failing())  stop();
  1657   if (stopped())  return; // trap reachable?
  1659   // Note:  If ProfileTraps is true, and if a deopt. actually
  1660   // occurs here, the runtime will make sure an MDO exists.  There is
  1661   // no need to call method()->build_method_data() at this point.
  1663 #ifdef ASSERT
  1664   if (!must_throw) {
  1665     // Make sure the stack has at least enough depth to execute
  1666     // the current bytecode.
  1667     int inputs, ignore;
  1668     if (compute_stack_effects(inputs, ignore)) {
  1669       assert(sp() >= inputs, "must have enough JVMS stack to execute");
  1670       // It is a frequent error in library_call.cpp to issue an
  1671       // uncommon trap with the _sp value already popped.
  1674 #endif
  1676   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
  1677   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
  1679   switch (action) {
  1680   case Deoptimization::Action_maybe_recompile:
  1681   case Deoptimization::Action_reinterpret:
  1682     // Temporary fix for 6529811 to allow virtual calls to be sure they
  1683     // get the chance to go from mono->bi->mega
  1684     if (!keep_exact_action &&
  1685         Deoptimization::trap_request_index(trap_request) < 0 &&
  1686         too_many_recompiles(reason)) {
  1687       // This BCI is causing too many recompilations.
  1688       action = Deoptimization::Action_none;
  1689       trap_request = Deoptimization::make_trap_request(reason, action);
  1690     } else {
  1691       C->set_trap_can_recompile(true);
  1693     break;
  1694   case Deoptimization::Action_make_not_entrant:
  1695     C->set_trap_can_recompile(true);
  1696     break;
  1697 #ifdef ASSERT
  1698   case Deoptimization::Action_none:
  1699   case Deoptimization::Action_make_not_compilable:
  1700     break;
  1701   default:
  1702     assert(false, "bad action");
  1703 #endif
  1706   if (TraceOptoParse) {
  1707     char buf[100];
  1708     tty->print_cr("Uncommon trap %s at bci:%d",
  1709                   Deoptimization::format_trap_request(buf, sizeof(buf),
  1710                                                       trap_request), bci());
  1713   CompileLog* log = C->log();
  1714   if (log != NULL) {
  1715     int kid = (klass == NULL)? -1: log->identify(klass);
  1716     log->begin_elem("uncommon_trap bci='%d'", bci());
  1717     char buf[100];
  1718     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
  1719                                                           trap_request));
  1720     if (kid >= 0)         log->print(" klass='%d'", kid);
  1721     if (comment != NULL)  log->print(" comment='%s'", comment);
  1722     log->end_elem();
  1725   // Make sure any guarding test views this path as very unlikely
  1726   Node *i0 = control()->in(0);
  1727   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
  1728     IfNode *iff = i0->as_If();
  1729     float f = iff->_prob;   // Get prob
  1730     if (control()->Opcode() == Op_IfTrue) {
  1731       if (f > PROB_UNLIKELY_MAG(4))
  1732         iff->_prob = PROB_MIN;
  1733     } else {
  1734       if (f < PROB_LIKELY_MAG(4))
  1735         iff->_prob = PROB_MAX;
  1739   // Clear out dead values from the debug info.
  1740   kill_dead_locals();
  1742   // Now insert the uncommon trap subroutine call
  1743   address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin();
  1744   const TypePtr* no_memory_effects = NULL;
  1745   // Pass the index of the class to be loaded
  1746   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
  1747                                  (must_throw ? RC_MUST_THROW : 0),
  1748                                  OptoRuntime::uncommon_trap_Type(),
  1749                                  call_addr, "uncommon_trap", no_memory_effects,
  1750                                  intcon(trap_request));
  1751   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
  1752          "must extract request correctly from the graph");
  1753   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
  1755   call->set_req(TypeFunc::ReturnAdr, returnadr());
  1756   // The debug info is the only real input to this call.
  1758   // Halt-and-catch fire here.  The above call should never return!
  1759   HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
  1760   _gvn.set_type_bottom(halt);
  1761   root()->add_req(halt);
  1763   stop_and_kill_map();
  1767 //--------------------------just_allocated_object------------------------------
  1768 // Report the object that was just allocated.
  1769 // It must be the case that there are no intervening safepoints.
  1770 // We use this to determine if an object is so "fresh" that
  1771 // it does not require card marks.
  1772 Node* GraphKit::just_allocated_object(Node* current_control) {
  1773   if (C->recent_alloc_ctl() == current_control)
  1774     return C->recent_alloc_obj();
  1775   return NULL;
  1779 //------------------------------store_barrier----------------------------------
  1780 // Insert a write-barrier store.  This is to let generational GC work; we have
  1781 // to flag all oop-stores before the next GC point.
  1782 void GraphKit::write_barrier_post(Node* oop_store, Node* obj, Node* adr,
  1783                                   Node* val, bool use_precise) {
  1784   // No store check needed if we're storing a NULL or an old object
  1785   // (latter case is probably a string constant). The concurrent
  1786   // mark sweep garbage collector, however, needs to have all nonNull
  1787   // oop updates flagged via card-marks.
  1788   if (val != NULL && val->is_Con()) {
  1789     // must be either an oop or NULL
  1790     const Type* t = val->bottom_type();
  1791     if (t == TypePtr::NULL_PTR || t == Type::TOP)
  1792       // stores of null never (?) need barriers
  1793       return;
  1794     ciObject* con = t->is_oopptr()->const_oop();
  1795     if (con != NULL
  1796         && con->is_perm()
  1797         && Universe::heap()->can_elide_permanent_oop_store_barriers())
  1798       // no store barrier needed, because no old-to-new ref created
  1799       return;
  1802   if (use_ReduceInitialCardMarks()
  1803       && obj == just_allocated_object(control())) {
  1804     // We can skip marks on a freshly-allocated object.
  1805     // Keep this code in sync with do_eager_card_mark in runtime.cpp.
  1806     // That routine eagerly marks the occasional object which is produced
  1807     // by the slow path, so that we don't have to do it here.
  1808     return;
  1811   if (!use_precise) {
  1812     // All card marks for a (non-array) instance are in one place:
  1813     adr = obj;
  1815   // (Else it's an array (or unknown), and we want more precise card marks.)
  1816   assert(adr != NULL, "");
  1818   // Get the alias_index for raw card-mark memory
  1819   int adr_type = Compile::AliasIdxRaw;
  1820   // Convert the pointer to an int prior to doing math on it
  1821   Node* cast = _gvn.transform(new (C, 2) CastP2XNode(control(), adr));
  1822   // Divide by card size
  1823   assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
  1824          "Only one we handle so far.");
  1825   CardTableModRefBS* ct =
  1826     (CardTableModRefBS*)(Universe::heap()->barrier_set());
  1827   Node *b = _gvn.transform(new (C, 3) URShiftXNode( cast, _gvn.intcon(CardTableModRefBS::card_shift) ));
  1828   // We store into a byte array, so do not bother to left-shift by zero
  1829   // Get base of card map
  1830   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte),
  1831          "adjust this code");
  1832   Node *c = makecon(TypeRawPtr::make((address)ct->byte_map_base));
  1833   // Combine
  1834   Node *sb_ctl = control();
  1835   Node *sb_adr = _gvn.transform(new (C, 4) AddPNode( top()/*no base ptr*/, c, b ));
  1836   Node *sb_val = _gvn.intcon(0);
  1837   // Smash zero into card
  1838   if( !UseConcMarkSweepGC ) {
  1839     BasicType bt = T_BYTE;
  1840     store_to_memory(sb_ctl, sb_adr, sb_val, bt, adr_type);
  1841   } else {
  1842     // Specialized path for CM store barrier
  1843     cms_card_mark( sb_ctl, sb_adr, sb_val, oop_store);
  1847 // Specialized path for CMS store barrier
  1848 void GraphKit::cms_card_mark(Node* ctl, Node* adr, Node* val, Node *oop_store) {
  1849   BasicType bt = T_BYTE;
  1850   int adr_idx = Compile::AliasIdxRaw;
  1851   Node* mem = memory(adr_idx);
  1853   // The type input is NULL in PRODUCT builds
  1854   const TypePtr* type = NULL;
  1855   debug_only(type = C->get_adr_type(adr_idx));
  1857   // Add required edge to oop_store, optimizer does not support precedence edges.
  1858   // Convert required edge to precedence edge before allocation.
  1859   Node *store = _gvn.transform( new (C, 5) StoreCMNode(ctl, mem, adr, type, val, oop_store) );
  1860   set_memory(store, adr_idx);
  1862   // For CMS, back-to-back card-marks can only remove the first one
  1863   // and this requires DU info.  Push on worklist for optimizer.
  1864   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
  1865     record_for_igvn(store);
  1869 void GraphKit::round_double_arguments(ciMethod* dest_method) {
  1870   // (Note:  TypeFunc::make has a cache that makes this fast.)
  1871   const TypeFunc* tf    = TypeFunc::make(dest_method);
  1872   int             nargs = tf->_domain->_cnt - TypeFunc::Parms;
  1873   for (int j = 0; j < nargs; j++) {
  1874     const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
  1875     if( targ->basic_type() == T_DOUBLE ) {
  1876       // If any parameters are doubles, they must be rounded before
  1877       // the call, dstore_rounding does gvn.transform
  1878       Node *arg = argument(j);
  1879       arg = dstore_rounding(arg);
  1880       set_argument(j, arg);
  1885 void GraphKit::round_double_result(ciMethod* dest_method) {
  1886   // A non-strict method may return a double value which has an extended
  1887   // exponent, but this must not be visible in a caller which is 'strict'
  1888   // If a strict caller invokes a non-strict callee, round a double result
  1890   BasicType result_type = dest_method->return_type()->basic_type();
  1891   assert( method() != NULL, "must have caller context");
  1892   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
  1893     // Destination method's return value is on top of stack
  1894     // dstore_rounding() does gvn.transform
  1895     Node *result = pop_pair();
  1896     result = dstore_rounding(result);
  1897     push_pair(result);
  1901 // rounding for strict float precision conformance
  1902 Node* GraphKit::precision_rounding(Node* n) {
  1903   return UseStrictFP && _method->flags().is_strict()
  1904     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
  1905     ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
  1906     : n;
  1909 // rounding for strict double precision conformance
  1910 Node* GraphKit::dprecision_rounding(Node *n) {
  1911   return UseStrictFP && _method->flags().is_strict()
  1912     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
  1913     ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
  1914     : n;
  1917 // rounding for non-strict double stores
  1918 Node* GraphKit::dstore_rounding(Node* n) {
  1919   return Matcher::strict_fp_requires_explicit_rounding
  1920     && UseSSE <= 1
  1921     ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
  1922     : n;
  1925 //=============================================================================
  1926 // Generate a fast path/slow path idiom.  Graph looks like:
  1927 // [foo] indicates that 'foo' is a parameter
  1928 //
  1929 //              [in]     NULL
  1930 //                 \    /
  1931 //                  CmpP
  1932 //                  Bool ne
  1933 //                   If
  1934 //                  /  \
  1935 //              True    False-<2>
  1936 //              / |
  1937 //             /  cast_not_null
  1938 //           Load  |    |   ^
  1939 //        [fast_test]   |   |
  1940 // gvn to   opt_test    |   |
  1941 //          /    \      |  <1>
  1942 //      True     False  |
  1943 //        |         \\  |
  1944 //   [slow_call]     \[fast_result]
  1945 //    Ctl   Val       \      \
  1946 //     |               \      \
  1947 //    Catch       <1>   \      \
  1948 //   /    \        ^     \      \
  1949 //  Ex    No_Ex    |      \      \
  1950 //  |       \   \  |       \ <2>  \
  1951 //  ...      \  [slow_res] |  |    \   [null_result]
  1952 //            \         \--+--+---  |  |
  1953 //             \           | /    \ | /
  1954 //              --------Region     Phi
  1955 //
  1956 //=============================================================================
  1957 // Code is structured as a series of driver functions all called 'do_XXX' that
  1958 // call a set of helper functions.  Helper functions first, then drivers.
  1960 //------------------------------null_check_oop---------------------------------
  1961 // Null check oop.  Set null-path control into Region in slot 3.
  1962 // Make a cast-not-nullness use the other not-null control.  Return cast.
  1963 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
  1964                                bool never_see_null) {
  1965   // Initial NULL check taken path
  1966   (*null_control) = top();
  1967   Node* cast = null_check_common(value, T_OBJECT, false, null_control);
  1969   // Generate uncommon_trap:
  1970   if (never_see_null && (*null_control) != top()) {
  1971     // If we see an unexpected null at a check-cast we record it and force a
  1972     // recompile; the offending check-cast will be compiled to handle NULLs.
  1973     // If we see more than one offending BCI, then all checkcasts in the
  1974     // method will be compiled to handle NULLs.
  1975     PreserveJVMState pjvms(this);
  1976     set_control(*null_control);
  1977     uncommon_trap(Deoptimization::Reason_null_check,
  1978                   Deoptimization::Action_make_not_entrant);
  1979     (*null_control) = top();    // NULL path is dead
  1982   // Cast away null-ness on the result
  1983   return cast;
  1986 //------------------------------opt_iff----------------------------------------
  1987 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
  1988 // Return slow-path control.
  1989 Node* GraphKit::opt_iff(Node* region, Node* iff) {
  1990   IfNode *opt_iff = _gvn.transform(iff)->as_If();
  1992   // Fast path taken; set region slot 2
  1993   Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
  1994   region->init_req(2,fast_taken); // Capture fast-control
  1996   // Fast path not-taken, i.e. slow path
  1997   Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
  1998   return slow_taken;
  2001 //-----------------------------make_runtime_call-------------------------------
  2002 Node* GraphKit::make_runtime_call(int flags,
  2003                                   const TypeFunc* call_type, address call_addr,
  2004                                   const char* call_name,
  2005                                   const TypePtr* adr_type,
  2006                                   // The following parms are all optional.
  2007                                   // The first NULL ends the list.
  2008                                   Node* parm0, Node* parm1,
  2009                                   Node* parm2, Node* parm3,
  2010                                   Node* parm4, Node* parm5,
  2011                                   Node* parm6, Node* parm7) {
  2012   // Slow-path call
  2013   int size = call_type->domain()->cnt();
  2014   bool is_leaf = !(flags & RC_NO_LEAF);
  2015   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
  2016   if (call_name == NULL) {
  2017     assert(!is_leaf, "must supply name for leaf");
  2018     call_name = OptoRuntime::stub_name(call_addr);
  2020   CallNode* call;
  2021   if (!is_leaf) {
  2022     call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
  2023                                            bci(), adr_type);
  2024   } else if (flags & RC_NO_FP) {
  2025     call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
  2026   } else {
  2027     call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
  2030   // The following is similar to set_edges_for_java_call,
  2031   // except that the memory effects of the call are restricted to AliasIdxRaw.
  2033   // Slow path call has no side-effects, uses few values
  2034   bool wide_in  = !(flags & RC_NARROW_MEM);
  2035   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
  2037   Node* prev_mem = NULL;
  2038   if (wide_in) {
  2039     prev_mem = set_predefined_input_for_runtime_call(call);
  2040   } else {
  2041     assert(!wide_out, "narrow in => narrow out");
  2042     Node* narrow_mem = memory(adr_type);
  2043     prev_mem = reset_memory();
  2044     map()->set_memory(narrow_mem);
  2045     set_predefined_input_for_runtime_call(call);
  2048   // Hook each parm in order.  Stop looking at the first NULL.
  2049   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
  2050   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
  2051   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
  2052   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
  2053   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
  2054   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
  2055   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
  2056   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
  2057     /* close each nested if ===> */  } } } } } } } }
  2058   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
  2060   if (!is_leaf) {
  2061     // Non-leaves can block and take safepoints:
  2062     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
  2064   // Non-leaves can throw exceptions:
  2065   if (has_io) {
  2066     call->set_req(TypeFunc::I_O, i_o());
  2069   if (flags & RC_UNCOMMON) {
  2070     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
  2071     // (An "if" probability corresponds roughly to an unconditional count.
  2072     // Sort of.)
  2073     call->set_cnt(PROB_UNLIKELY_MAG(4));
  2076   Node* c = _gvn.transform(call);
  2077   assert(c == call, "cannot disappear");
  2079   if (wide_out) {
  2080     // Slow path call has full side-effects.
  2081     set_predefined_output_for_runtime_call(call);
  2082   } else {
  2083     // Slow path call has few side-effects, and/or sets few values.
  2084     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
  2087   if (has_io) {
  2088     set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
  2090   return call;
  2094 //------------------------------merge_memory-----------------------------------
  2095 // Merge memory from one path into the current memory state.
  2096 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
  2097   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
  2098     Node* old_slice = mms.force_memory();
  2099     Node* new_slice = mms.memory2();
  2100     if (old_slice != new_slice) {
  2101       PhiNode* phi;
  2102       if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
  2103         phi = new_slice->as_Phi();
  2104         #ifdef ASSERT
  2105         if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
  2106           old_slice = old_slice->in(new_path);
  2107         // Caller is responsible for ensuring that any pre-existing
  2108         // phis are already aware of old memory.
  2109         int old_path = (new_path > 1) ? 1 : 2;  // choose old_path != new_path
  2110         assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
  2111         #endif
  2112         mms.set_memory(phi);
  2113       } else {
  2114         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
  2115         _gvn.set_type(phi, Type::MEMORY);
  2116         phi->set_req(new_path, new_slice);
  2117         mms.set_memory(_gvn.transform(phi));  // assume it is complete
  2123 //------------------------------make_slow_call_ex------------------------------
  2124 // Make the exception handler hookups for the slow call
  2125 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
  2126   if (stopped())  return;
  2128   // Make a catch node with just two handlers:  fall-through and catch-all
  2129   Node* i_o  = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
  2130   Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
  2131   Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
  2132   Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
  2134   { PreserveJVMState pjvms(this);
  2135     set_control(excp);
  2136     set_i_o(i_o);
  2138     if (excp != top()) {
  2139       // Create an exception state also.
  2140       // Use an exact type if the caller has specified a specific exception.
  2141       const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
  2142       Node*       ex_oop  = new (C, 2) CreateExNode(ex_type, control(), i_o);
  2143       add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
  2147   // Get the no-exception control from the CatchNode.
  2148   set_control(norm);
  2152 //-------------------------------gen_subtype_check-----------------------------
  2153 // Generate a subtyping check.  Takes as input the subtype and supertype.
  2154 // Returns 2 values: sets the default control() to the true path and returns
  2155 // the false path.  Only reads invariant memory; sets no (visible) memory.
  2156 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
  2157 // but that's not exposed to the optimizer.  This call also doesn't take in an
  2158 // Object; if you wish to check an Object you need to load the Object's class
  2159 // prior to coming here.
  2160 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
  2161   // Fast check for identical types, perhaps identical constants.
  2162   // The types can even be identical non-constants, in cases
  2163   // involving Array.newInstance, Object.clone, etc.
  2164   if (subklass == superklass)
  2165     return top();             // false path is dead; no test needed.
  2167   if (_gvn.type(superklass)->singleton()) {
  2168     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
  2169     ciKlass* subk   = _gvn.type(subklass)->is_klassptr()->klass();
  2171     // In the common case of an exact superklass, try to fold up the
  2172     // test before generating code.  You may ask, why not just generate
  2173     // the code and then let it fold up?  The answer is that the generated
  2174     // code will necessarily include null checks, which do not always
  2175     // completely fold away.  If they are also needless, then they turn
  2176     // into a performance loss.  Example:
  2177     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
  2178     // Here, the type of 'fa' is often exact, so the store check
  2179     // of fa[1]=x will fold up, without testing the nullness of x.
  2180     switch (static_subtype_check(superk, subk)) {
  2181     case SSC_always_false:
  2183         Node* always_fail = control();
  2184         set_control(top());
  2185         return always_fail;
  2187     case SSC_always_true:
  2188       return top();
  2189     case SSC_easy_test:
  2191         // Just do a direct pointer compare and be done.
  2192         Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
  2193         Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
  2194         IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
  2195         set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
  2196         return       _gvn.transform( new(C, 1) IfFalseNode(iff) );
  2198     case SSC_full_test:
  2199       break;
  2200     default:
  2201       ShouldNotReachHere();
  2205   // %%% Possible further optimization:  Even if the superklass is not exact,
  2206   // if the subklass is the unique subtype of the superklass, the check
  2207   // will always succeed.  We could leave a dependency behind to ensure this.
  2209   // First load the super-klass's check-offset
  2210   Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
  2211   Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
  2212   int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
  2213   bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
  2215   // Load from the sub-klass's super-class display list, or a 1-word cache of
  2216   // the secondary superclass list, or a failing value with a sentinel offset
  2217   // if the super-klass is an interface or exceptionally deep in the Java
  2218   // hierarchy and we have to scan the secondary superclass list the hard way.
  2219   // Worst-case type is a little odd: NULL is allowed as a result (usually
  2220   // klass loads can never produce a NULL).
  2221   Node *chk_off_X = ConvI2X(chk_off);
  2222   Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
  2223   // For some types like interfaces the following loadKlass is from a 1-word
  2224   // cache which is mutable so can't use immutable memory.  Other
  2225   // types load from the super-class display table which is immutable.
  2226   Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
  2227   Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
  2229   // Compile speed common case: ARE a subtype and we canNOT fail
  2230   if( superklass == nkls )
  2231     return top();             // false path is dead; no test needed.
  2233   // See if we get an immediate positive hit.  Happens roughly 83% of the
  2234   // time.  Test to see if the value loaded just previously from the subklass
  2235   // is exactly the superklass.
  2236   Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
  2237   Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
  2238   IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
  2239   Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
  2240   set_control(    _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
  2242   // Compile speed common case: Check for being deterministic right now.  If
  2243   // chk_off is a constant and not equal to cacheoff then we are NOT a
  2244   // subklass.  In this case we need exactly the 1 test above and we can
  2245   // return those results immediately.
  2246   if (!might_be_cache) {
  2247     Node* not_subtype_ctrl = control();
  2248     set_control(iftrue1); // We need exactly the 1 test above
  2249     return not_subtype_ctrl;
  2252   // Gather the various success & failures here
  2253   RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
  2254   record_for_igvn(r_ok_subtype);
  2255   RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
  2256   record_for_igvn(r_not_subtype);
  2258   r_ok_subtype->init_req(1, iftrue1);
  2260   // Check for immediate negative hit.  Happens roughly 11% of the time (which
  2261   // is roughly 63% of the remaining cases).  Test to see if the loaded
  2262   // check-offset points into the subklass display list or the 1-element
  2263   // cache.  If it points to the display (and NOT the cache) and the display
  2264   // missed then it's not a subtype.
  2265   Node *cacheoff = _gvn.intcon(cacheoff_con);
  2266   Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
  2267   Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
  2268   IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
  2269   r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
  2270   set_control(                _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
  2272   // Check for self.  Very rare to get here, but its taken 1/3 the time.
  2273   // No performance impact (too rare) but allows sharing of secondary arrays
  2274   // which has some footprint reduction.
  2275   Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
  2276   Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
  2277   IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
  2278   r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
  2279   set_control(               _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
  2281   // Now do a linear scan of the secondary super-klass array.  Again, no real
  2282   // performance impact (too rare) but it's gotta be done.
  2283   // (The stub also contains the self-check of subklass == superklass.
  2284   // Since the code is rarely used, there is no penalty for moving it
  2285   // out of line, and it can only improve I-cache density.)
  2286   Node* psc = _gvn.transform(
  2287     new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
  2289   Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
  2290   Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
  2291   IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
  2292   r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
  2293   r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
  2295   // Return false path; set default control to true path.
  2296   set_control( _gvn.transform(r_ok_subtype) );
  2297   return _gvn.transform(r_not_subtype);
  2300 //----------------------------static_subtype_check-----------------------------
  2301 // Shortcut important common cases when superklass is exact:
  2302 // (0) superklass is java.lang.Object (can occur in reflective code)
  2303 // (1) subklass is already limited to a subtype of superklass => always ok
  2304 // (2) subklass does not overlap with superklass => always fail
  2305 // (3) superklass has NO subtypes and we can check with a simple compare.
  2306 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
  2307   if (StressReflectiveCode) {
  2308     return SSC_full_test;       // Let caller generate the general case.
  2311   if (superk == env()->Object_klass()) {
  2312     return SSC_always_true;     // (0) this test cannot fail
  2315   ciType* superelem = superk;
  2316   if (superelem->is_array_klass())
  2317     superelem = superelem->as_array_klass()->base_element_type();
  2319   if (!subk->is_interface()) {  // cannot trust static interface types yet
  2320     if (subk->is_subtype_of(superk)) {
  2321       return SSC_always_true;   // (1) false path dead; no dynamic test needed
  2323     if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
  2324         !superk->is_subtype_of(subk)) {
  2325       return SSC_always_false;
  2329   // If casting to an instance klass, it must have no subtypes
  2330   if (superk->is_interface()) {
  2331     // Cannot trust interfaces yet.
  2332     // %%% S.B. superk->nof_implementors() == 1
  2333   } else if (superelem->is_instance_klass()) {
  2334     ciInstanceKlass* ik = superelem->as_instance_klass();
  2335     if (!ik->has_subklass() && !ik->is_interface()) {
  2336       if (!ik->is_final()) {
  2337         // Add a dependency if there is a chance of a later subclass.
  2338         C->dependencies()->assert_leaf_type(ik);
  2340       return SSC_easy_test;     // (3) caller can do a simple ptr comparison
  2342   } else {
  2343     // A primitive array type has no subtypes.
  2344     return SSC_easy_test;       // (3) caller can do a simple ptr comparison
  2347   return SSC_full_test;
  2350 // Profile-driven exact type check:
  2351 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
  2352                                     float prob,
  2353                                     Node* *casted_receiver) {
  2354   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
  2355   Node* recv_klass = load_object_klass(receiver);
  2356   Node* want_klass = makecon(tklass);
  2357   Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
  2358   Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
  2359   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
  2360   set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
  2361   Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
  2363   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
  2364   assert(recv_xtype->klass_is_exact(), "");
  2366   // Subsume downstream occurrences of receiver with a cast to
  2367   // recv_xtype, since now we know what the type will be.
  2368   Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
  2369   (*casted_receiver) = _gvn.transform(cast);
  2370   // (User must make the replace_in_map call.)
  2372   return fail;
  2376 //-------------------------------gen_instanceof--------------------------------
  2377 // Generate an instance-of idiom.  Used by both the instance-of bytecode
  2378 // and the reflective instance-of call.
  2379 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) {
  2380   C->set_has_split_ifs(true); // Has chance for split-if optimization
  2381   assert( !stopped(), "dead parse path should be checked in callers" );
  2382   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
  2383          "must check for not-null not-dead klass in callers");
  2385   // Make the merge point
  2386   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
  2387   RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
  2388   Node*       phi    = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
  2389   C->set_has_split_ifs(true); // Has chance for split-if optimization
  2391   // Null check; get casted pointer; set region slot 3
  2392   Node* null_ctl = top();
  2393   Node* not_null_obj = null_check_oop(subobj, &null_ctl);
  2395   // If not_null_obj is dead, only null-path is taken
  2396   if (stopped()) {              // Doing instance-of on a NULL?
  2397     set_control(null_ctl);
  2398     return intcon(0);
  2400   region->init_req(_null_path, null_ctl);
  2401   phi   ->init_req(_null_path, intcon(0)); // Set null path value
  2403   // Load the object's klass
  2404   Node* obj_klass = load_object_klass(not_null_obj);
  2406   // Generate the subtype check
  2407   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
  2409   // Plug in the success path to the general merge in slot 1.
  2410   region->init_req(_obj_path, control());
  2411   phi   ->init_req(_obj_path, intcon(1));
  2413   // Plug in the failing path to the general merge in slot 2.
  2414   region->init_req(_fail_path, not_subtype_ctrl);
  2415   phi   ->init_req(_fail_path, intcon(0));
  2417   // Return final merged results
  2418   set_control( _gvn.transform(region) );
  2419   record_for_igvn(region);
  2420   return _gvn.transform(phi);
  2423 //-------------------------------gen_checkcast---------------------------------
  2424 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
  2425 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
  2426 // uncommon-trap paths work.  Adjust stack after this call.
  2427 // If failure_control is supplied and not null, it is filled in with
  2428 // the control edge for the cast failure.  Otherwise, an appropriate
  2429 // uncommon trap or exception is thrown.
  2430 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
  2431                               Node* *failure_control) {
  2432   kill_dead_locals();           // Benefit all the uncommon traps
  2433   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
  2434   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
  2436   // Fast cutout:  Check the case that the cast is vacuously true.
  2437   // This detects the common cases where the test will short-circuit
  2438   // away completely.  We do this before we perform the null check,
  2439   // because if the test is going to turn into zero code, we don't
  2440   // want a residual null check left around.  (Causes a slowdown,
  2441   // for example, in some objArray manipulations, such as a[i]=a[j].)
  2442   if (tk->singleton()) {
  2443     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
  2444     if (objtp != NULL && objtp->klass() != NULL) {
  2445       switch (static_subtype_check(tk->klass(), objtp->klass())) {
  2446       case SSC_always_true:
  2447         return obj;
  2448       case SSC_always_false:
  2449         // It needs a null check because a null will *pass* the cast check.
  2450         // A non-null value will always produce an exception.
  2451         return do_null_assert(obj, T_OBJECT);
  2456   ciProfileData* data = NULL;
  2457   if (failure_control == NULL) {        // use MDO in regular case only
  2458     assert(java_bc() == Bytecodes::_aastore ||
  2459            java_bc() == Bytecodes::_checkcast,
  2460            "interpreter profiles type checks only for these BCs");
  2461     data = method()->method_data()->bci_to_data(bci());
  2464   // Make the merge point
  2465   enum { _obj_path = 1, _null_path, PATH_LIMIT };
  2466   RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
  2467   Node*       phi    = new (C, PATH_LIMIT) PhiNode(region, toop);
  2468   C->set_has_split_ifs(true); // Has chance for split-if optimization
  2470   // Use null-cast information if it is available
  2471   bool never_see_null = false;
  2472   // If we see an unexpected null at a check-cast we record it and force a
  2473   // recompile; the offending check-cast will be compiled to handle NULLs.
  2474   // If we see several offending BCIs, then all checkcasts in the
  2475   // method will be compiled to handle NULLs.
  2476   if (UncommonNullCast            // Cutout for this technique
  2477       && failure_control == NULL  // regular case
  2478       && obj != null()            // And not the -Xcomp stupid case?
  2479       && !too_many_traps(Deoptimization::Reason_null_check)) {
  2480     // Finally, check the "null_seen" bit from the interpreter.
  2481     if (data == NULL || !data->as_BitData()->null_seen()) {
  2482       never_see_null = true;
  2486   // Null check; get casted pointer; set region slot 3
  2487   Node* null_ctl = top();
  2488   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
  2490   // If not_null_obj is dead, only null-path is taken
  2491   if (stopped()) {              // Doing instance-of on a NULL?
  2492     set_control(null_ctl);
  2493     return null();
  2495   region->init_req(_null_path, null_ctl);
  2496   phi   ->init_req(_null_path, null());  // Set null path value
  2498   Node* cast_obj = NULL;        // the casted version of the object
  2500   // If the profile has seen exactly one type, narrow to that type.
  2501   // (The subsequent subtype check will always fold up.)
  2502   if (UseTypeProfile && TypeProfileCasts && data != NULL &&
  2503       // Counter has never been decremented (due to cast failure).
  2504       // ...This is a reasonable thing to expect.  It is true of
  2505       // all casts inserted by javac to implement generic types.
  2506       data->as_CounterData()->count() >= 0 &&
  2507       !too_many_traps(Deoptimization::Reason_class_check)) {
  2508     // (No, this isn't a call, but it's enough like a virtual call
  2509     // to use the same ciMethod accessor to get the profile info...)
  2510     ciCallProfile profile = method()->call_profile_at_bci(bci());
  2511     if (profile.count() >= 0 &&         // no cast failures here
  2512         profile.has_receiver(0) &&
  2513         profile.morphism() == 1) {
  2514       ciKlass* exact_kls = profile.receiver(0);
  2515       int ssc = static_subtype_check(tk->klass(), exact_kls);
  2516       if (ssc == SSC_always_true) {
  2517         // If we narrow the type to match what the type profile sees,
  2518         // we can then remove the rest of the cast.
  2519         // This is a win, even if the exact_kls is very specific,
  2520         // because downstream operations, such as method calls,
  2521         // will often benefit from the sharper type.
  2522         Node* exact_obj = not_null_obj; // will get updated in place...
  2523         Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
  2524                                               &exact_obj);
  2525         { PreserveJVMState pjvms(this);
  2526           set_control(slow_ctl);
  2527           uncommon_trap(Deoptimization::Reason_class_check,
  2528                         Deoptimization::Action_maybe_recompile);
  2530         if (failure_control != NULL) // failure is now impossible
  2531           (*failure_control) = top();
  2532         replace_in_map(not_null_obj, exact_obj);
  2533         // adjust the type of the phi to the exact klass:
  2534         phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR));
  2535         cast_obj = exact_obj;
  2537       // assert(cast_obj != NULL)... except maybe the profile lied to us.
  2541   if (cast_obj == NULL) {
  2542     // Load the object's klass
  2543     Node* obj_klass = load_object_klass(not_null_obj);
  2545     // Generate the subtype check
  2546     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
  2548     // Plug in success path into the merge
  2549     cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
  2550                                                          not_null_obj, toop));
  2551     // Failure path ends in uncommon trap (or may be dead - failure impossible)
  2552     if (failure_control == NULL) {
  2553       if (not_subtype_ctrl != top()) { // If failure is possible
  2554         PreserveJVMState pjvms(this);
  2555         set_control(not_subtype_ctrl);
  2556         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
  2558     } else {
  2559       (*failure_control) = not_subtype_ctrl;
  2563   region->init_req(_obj_path, control());
  2564   phi   ->init_req(_obj_path, cast_obj);
  2566   // A merge of NULL or Casted-NotNull obj
  2567   Node* res = _gvn.transform(phi);
  2569   // Note I do NOT always 'replace_in_map(obj,result)' here.
  2570   //  if( tk->klass()->can_be_primary_super()  )
  2571     // This means that if I successfully store an Object into an array-of-String
  2572     // I 'forget' that the Object is really now known to be a String.  I have to
  2573     // do this because we don't have true union types for interfaces - if I store
  2574     // a Baz into an array-of-Interface and then tell the optimizer it's an
  2575     // Interface, I forget that it's also a Baz and cannot do Baz-like field
  2576     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
  2577   //  replace_in_map( obj, res );
  2579   // Return final merged results
  2580   set_control( _gvn.transform(region) );
  2581   record_for_igvn(region);
  2582   return res;
  2585 //------------------------------next_monitor-----------------------------------
  2586 // What number should be given to the next monitor?
  2587 int GraphKit::next_monitor() {
  2588   int current = jvms()->monitor_depth()* C->sync_stack_slots();
  2589   int next = current + C->sync_stack_slots();
  2590   // Keep the toplevel high water mark current:
  2591   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
  2592   return current;
  2595 //------------------------------insert_mem_bar---------------------------------
  2596 // Memory barrier to avoid floating things around
  2597 // The membar serves as a pinch point between both control and all memory slices.
  2598 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
  2599   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
  2600   mb->init_req(TypeFunc::Control, control());
  2601   mb->init_req(TypeFunc::Memory,  reset_memory());
  2602   Node* membar = _gvn.transform(mb);
  2603   set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
  2604   set_all_memory_call(membar);
  2605   return membar;
  2608 //-------------------------insert_mem_bar_volatile----------------------------
  2609 // Memory barrier to avoid floating things around
  2610 // The membar serves as a pinch point between both control and memory(alias_idx).
  2611 // If you want to make a pinch point on all memory slices, do not use this
  2612 // function (even with AliasIdxBot); use insert_mem_bar() instead.
  2613 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
  2614   // When Parse::do_put_xxx updates a volatile field, it appends a series
  2615   // of MemBarVolatile nodes, one for *each* volatile field alias category.
  2616   // The first membar is on the same memory slice as the field store opcode.
  2617   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
  2618   // All the other membars (for other volatile slices, including AliasIdxBot,
  2619   // which stands for all unknown volatile slices) are control-dependent
  2620   // on the first membar.  This prevents later volatile loads or stores
  2621   // from sliding up past the just-emitted store.
  2623   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
  2624   mb->set_req(TypeFunc::Control,control());
  2625   if (alias_idx == Compile::AliasIdxBot) {
  2626     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
  2627   } else {
  2628     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
  2629     mb->set_req(TypeFunc::Memory, memory(alias_idx));
  2631   Node* membar = _gvn.transform(mb);
  2632   set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
  2633   if (alias_idx == Compile::AliasIdxBot) {
  2634     merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
  2635   } else {
  2636     set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
  2638   return membar;
  2641 //------------------------------shared_lock------------------------------------
  2642 // Emit locking code.
  2643 FastLockNode* GraphKit::shared_lock(Node* obj) {
  2644   // bci is either a monitorenter bc or InvocationEntryBci
  2645   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
  2646   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
  2648   if( !GenerateSynchronizationCode )
  2649     return NULL;                // Not locking things?
  2650   if (stopped())                // Dead monitor?
  2651     return NULL;
  2653   assert(dead_locals_are_killed(), "should kill locals before sync. point");
  2655   // Box the stack location
  2656   Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
  2657   Node* mem = reset_memory();
  2659   FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
  2660   if (PrintPreciseBiasedLockingStatistics) {
  2661     // Create the counters for this fast lock.
  2662     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
  2664   // Add monitor to debug info for the slow path.  If we block inside the
  2665   // slow path and de-opt, we need the monitor hanging around
  2666   map()->push_monitor( flock );
  2668   const TypeFunc *tf = LockNode::lock_type();
  2669   LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
  2671   lock->init_req( TypeFunc::Control, control() );
  2672   lock->init_req( TypeFunc::Memory , mem );
  2673   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
  2674   lock->init_req( TypeFunc::FramePtr, frameptr() );
  2675   lock->init_req( TypeFunc::ReturnAdr, top() );
  2677   lock->init_req(TypeFunc::Parms + 0, obj);
  2678   lock->init_req(TypeFunc::Parms + 1, box);
  2679   lock->init_req(TypeFunc::Parms + 2, flock);
  2680   add_safepoint_edges(lock);
  2682   lock = _gvn.transform( lock )->as_Lock();
  2684   // lock has no side-effects, sets few values
  2685   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
  2687   insert_mem_bar(Op_MemBarAcquire);
  2689   // Add this to the worklist so that the lock can be eliminated
  2690   record_for_igvn(lock);
  2692 #ifndef PRODUCT
  2693   if (PrintLockStatistics) {
  2694     // Update the counter for this lock.  Don't bother using an atomic
  2695     // operation since we don't require absolute accuracy.
  2696     lock->create_lock_counter(map()->jvms());
  2697     int adr_type = Compile::AliasIdxRaw;
  2698     Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr()));
  2699     Node* cnt  = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
  2700     Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
  2701     store_to_memory(control(), counter_addr, incr, T_INT, adr_type);
  2703 #endif
  2705   return flock;
  2709 //------------------------------shared_unlock----------------------------------
  2710 // Emit unlocking code.
  2711 void GraphKit::shared_unlock(Node* box, Node* obj) {
  2712   // bci is either a monitorenter bc or InvocationEntryBci
  2713   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
  2714   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
  2716   if( !GenerateSynchronizationCode )
  2717     return;
  2718   if (stopped()) {               // Dead monitor?
  2719     map()->pop_monitor();        // Kill monitor from debug info
  2720     return;
  2723   // Memory barrier to avoid floating things down past the locked region
  2724   insert_mem_bar(Op_MemBarRelease);
  2726   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
  2727   UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
  2728   uint raw_idx = Compile::AliasIdxRaw;
  2729   unlock->init_req( TypeFunc::Control, control() );
  2730   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
  2731   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
  2732   unlock->init_req( TypeFunc::FramePtr, frameptr() );
  2733   unlock->init_req( TypeFunc::ReturnAdr, top() );
  2735   unlock->init_req(TypeFunc::Parms + 0, obj);
  2736   unlock->init_req(TypeFunc::Parms + 1, box);
  2737   unlock = _gvn.transform(unlock)->as_Unlock();
  2739   Node* mem = reset_memory();
  2741   // unlock has no side-effects, sets few values
  2742   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
  2744   // Kill monitor from debug info
  2745   map()->pop_monitor( );
  2748 //-------------------------------get_layout_helper-----------------------------
  2749 // If the given klass is a constant or known to be an array,
  2750 // fetch the constant layout helper value into constant_value
  2751 // and return (Node*)NULL.  Otherwise, load the non-constant
  2752 // layout helper value, and return the node which represents it.
  2753 // This two-faced routine is useful because allocation sites
  2754 // almost always feature constant types.
  2755 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
  2756   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
  2757   if (!StressReflectiveCode && inst_klass != NULL) {
  2758     ciKlass* klass = inst_klass->klass();
  2759     bool    xklass = inst_klass->klass_is_exact();
  2760     if (xklass || klass->is_array_klass()) {
  2761       jint lhelper = klass->layout_helper();
  2762       if (lhelper != Klass::_lh_neutral_value) {
  2763         constant_value = lhelper;
  2764         return (Node*) NULL;
  2768   constant_value = Klass::_lh_neutral_value;  // put in a known value
  2769   Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
  2770   return make_load(NULL, lhp, TypeInt::INT, T_INT);
  2773 // We just put in an allocate/initialize with a big raw-memory effect.
  2774 // Hook selected additional alias categories on the initialization.
  2775 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
  2776                                 MergeMemNode* init_in_merge,
  2777                                 Node* init_out_raw) {
  2778   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
  2779   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
  2781   Node* prevmem = kit.memory(alias_idx);
  2782   init_in_merge->set_memory_at(alias_idx, prevmem);
  2783   kit.set_memory(init_out_raw, alias_idx);
  2786 //---------------------------set_output_for_allocation-------------------------
  2787 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
  2788                                           const TypeOopPtr* oop_type,
  2789                                           bool raw_mem_only) {
  2790   int rawidx = Compile::AliasIdxRaw;
  2791   alloc->set_req( TypeFunc::FramePtr, frameptr() );
  2792   add_safepoint_edges(alloc);
  2793   Node* allocx = _gvn.transform(alloc);
  2794   set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
  2795   // create memory projection for i_o
  2796   set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
  2797   make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
  2799   // create a memory projection as for the normal control path
  2800   Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
  2801   set_memory(malloc, rawidx);
  2803   // a normal slow-call doesn't change i_o, but an allocation does
  2804   // we create a separate i_o projection for the normal control path
  2805   set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
  2806   Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
  2808   // put in an initialization barrier
  2809   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
  2810                                                  rawoop)->as_Initialize();
  2811   assert(alloc->initialization() == init,  "2-way macro link must work");
  2812   assert(init ->allocation()     == alloc, "2-way macro link must work");
  2813   if (ReduceFieldZeroing && !raw_mem_only) {
  2814     // Extract memory strands which may participate in the new object's
  2815     // initialization, and source them from the new InitializeNode.
  2816     // This will allow us to observe initializations when they occur,
  2817     // and link them properly (as a group) to the InitializeNode.
  2818     assert(init->in(InitializeNode::Memory) == malloc, "");
  2819     MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
  2820     init->set_req(InitializeNode::Memory, minit_in);
  2821     record_for_igvn(minit_in); // fold it up later, if possible
  2822     Node* minit_out = memory(rawidx);
  2823     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
  2824     if (oop_type->isa_aryptr()) {
  2825       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
  2826       int            elemidx  = C->get_alias_index(telemref);
  2827       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
  2828     } else if (oop_type->isa_instptr()) {
  2829       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
  2830       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
  2831         ciField* field = ik->nonstatic_field_at(i);
  2832         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
  2833           continue;  // do not bother to track really large numbers of fields
  2834         // Find (or create) the alias category for this field:
  2835         int fieldidx = C->alias_type(field)->index();
  2836         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
  2841   // Cast raw oop to the real thing...
  2842   Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
  2843   javaoop = _gvn.transform(javaoop);
  2844   C->set_recent_alloc(control(), javaoop);
  2845   assert(just_allocated_object(control()) == javaoop, "just allocated");
  2847 #ifdef ASSERT
  2848   { // Verify that the AllocateNode::Ideal_foo recognizers work:
  2849     Node* kn = alloc->in(AllocateNode::KlassNode);
  2850     Node* ln = alloc->in(AllocateNode::ALength);
  2851     assert(AllocateNode::Ideal_klass(rawoop, &_gvn) == kn,
  2852            "Ideal_klass works");
  2853     assert(AllocateNode::Ideal_klass(javaoop, &_gvn) == kn,
  2854            "Ideal_klass works");
  2855     if (alloc->is_AllocateArray()) {
  2856       assert(AllocateArrayNode::Ideal_length(rawoop, &_gvn) == ln,
  2857              "Ideal_length works");
  2858       assert(AllocateArrayNode::Ideal_length(javaoop, &_gvn) == ln,
  2859              "Ideal_length works");
  2860     } else {
  2861       assert(ln->is_top(), "no length, please");
  2864 #endif //ASSERT
  2866   return javaoop;
  2869 //---------------------------new_instance--------------------------------------
  2870 // This routine takes a klass_node which may be constant (for a static type)
  2871 // or may be non-constant (for reflective code).  It will work equally well
  2872 // for either, and the graph will fold nicely if the optimizer later reduces
  2873 // the type to a constant.
  2874 // The optional arguments are for specialized use by intrinsics:
  2875 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
  2876 //  - If 'raw_mem_only', do not cast the result to an oop.
  2877 //  - If 'return_size_val', report the the total object size to the caller.
  2878 Node* GraphKit::new_instance(Node* klass_node,
  2879                              Node* extra_slow_test,
  2880                              bool raw_mem_only, // affect only raw memory
  2881                              Node* *return_size_val) {
  2882   // Compute size in doublewords
  2883   // The size is always an integral number of doublewords, represented
  2884   // as a positive bytewise size stored in the klass's layout_helper.
  2885   // The layout_helper also encodes (in a low bit) the need for a slow path.
  2886   jint  layout_con = Klass::_lh_neutral_value;
  2887   Node* layout_val = get_layout_helper(klass_node, layout_con);
  2888   int   layout_is_con = (layout_val == NULL);
  2890   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
  2891   // Generate the initial go-slow test.  It's either ALWAYS (return a
  2892   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
  2893   // case) a computed value derived from the layout_helper.
  2894   Node* initial_slow_test = NULL;
  2895   if (layout_is_con) {
  2896     assert(!StressReflectiveCode, "stress mode does not use these paths");
  2897     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
  2898     initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
  2900   } else {   // reflective case
  2901     // This reflective path is used by Unsafe.allocateInstance.
  2902     // (It may be stress-tested by specifying StressReflectiveCode.)
  2903     // Basically, we want to get into the VM is there's an illegal argument.
  2904     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
  2905     initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
  2906     if (extra_slow_test != intcon(0)) {
  2907       initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
  2909     // (Macro-expander will further convert this to a Bool, if necessary.)
  2912   // Find the size in bytes.  This is easy; it's the layout_helper.
  2913   // The size value must be valid even if the slow path is taken.
  2914   Node* size = NULL;
  2915   if (layout_is_con) {
  2916     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
  2917   } else {   // reflective case
  2918     // This reflective path is used by clone and Unsafe.allocateInstance.
  2919     size = ConvI2X(layout_val);
  2921     // Clear the low bits to extract layout_helper_size_in_bytes:
  2922     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
  2923     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
  2924     size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
  2926   if (return_size_val != NULL) {
  2927     (*return_size_val) = size;
  2930   // This is a precise notnull oop of the klass.
  2931   // (Actually, it need not be precise if this is a reflective allocation.)
  2932   // It's what we cast the result to.
  2933   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
  2934   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
  2935   const TypeOopPtr* oop_type = tklass->as_instance_type();
  2937   // Now generate allocation code
  2939   // With escape analysis, the entire memory state is needed to be able to
  2940   // eliminate the allocation.  If the allocations cannot be eliminated, this
  2941   // will be optimized to the raw slice when the allocation is expanded.
  2942   Node *mem;
  2943   if (C->do_escape_analysis()) {
  2944     mem = reset_memory();
  2945     set_all_memory(mem);
  2946   } else {
  2947     mem = memory(Compile::AliasIdxRaw);
  2950   AllocateNode* alloc
  2951     = new (C, AllocateNode::ParmLimit)
  2952         AllocateNode(C, AllocateNode::alloc_type(),
  2953                      control(), mem, i_o(),
  2954                      size, klass_node,
  2955                      initial_slow_test);
  2957   return set_output_for_allocation(alloc, oop_type, raw_mem_only);
  2960 //-------------------------------new_array-------------------------------------
  2961 // helper for both newarray and anewarray
  2962 // The 'length' parameter is (obviously) the length of the array.
  2963 // See comments on new_instance for the meaning of the other arguments.
  2964 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
  2965                           Node* length,         // number of array elements
  2966                           bool raw_mem_only,    // affect only raw memory
  2967                           Node* *return_size_val) {
  2968   jint  layout_con = Klass::_lh_neutral_value;
  2969   Node* layout_val = get_layout_helper(klass_node, layout_con);
  2970   int   layout_is_con = (layout_val == NULL);
  2972   if (!layout_is_con && !StressReflectiveCode &&
  2973       !too_many_traps(Deoptimization::Reason_class_check)) {
  2974     // This is a reflective array creation site.
  2975     // Optimistically assume that it is a subtype of Object[],
  2976     // so that we can fold up all the address arithmetic.
  2977     layout_con = Klass::array_layout_helper(T_OBJECT);
  2978     Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
  2979     Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
  2980     { BuildCutout unless(this, bol_lh, PROB_MAX);
  2981       uncommon_trap(Deoptimization::Reason_class_check,
  2982                     Deoptimization::Action_maybe_recompile);
  2984     layout_val = NULL;
  2985     layout_is_con = true;
  2988   // Generate the initial go-slow test.  Make sure we do not overflow
  2989   // if length is huge (near 2Gig) or negative!  We do not need
  2990   // exact double-words here, just a close approximation of needed
  2991   // double-words.  We can't add any offset or rounding bits, lest we
  2992   // take a size -1 of bytes and make it positive.  Use an unsigned
  2993   // compare, so negative sizes look hugely positive.
  2994   int fast_size_limit = FastAllocateSizeLimit;
  2995   if (layout_is_con) {
  2996     assert(!StressReflectiveCode, "stress mode does not use these paths");
  2997     // Increase the size limit if we have exact knowledge of array type.
  2998     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
  2999     fast_size_limit <<= (LogBytesPerLong - log2_esize);
  3002   Node* initial_slow_cmp  = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
  3003   Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
  3004   if (initial_slow_test->is_Bool()) {
  3005     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
  3006     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
  3009   // --- Size Computation ---
  3010   // array_size = round_to_heap(array_header + (length << elem_shift));
  3011   // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
  3012   // and round_to(x, y) == ((x + y-1) & ~(y-1))
  3013   // The rounding mask is strength-reduced, if possible.
  3014   int round_mask = MinObjAlignmentInBytes - 1;
  3015   Node* header_size = NULL;
  3016   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
  3017   // (T_BYTE has the weakest alignment and size restrictions...)
  3018   if (layout_is_con) {
  3019     int       hsize  = Klass::layout_helper_header_size(layout_con);
  3020     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
  3021     BasicType etype  = Klass::layout_helper_element_type(layout_con);
  3022     if ((round_mask & ~right_n_bits(eshift)) == 0)
  3023       round_mask = 0;  // strength-reduce it if it goes away completely
  3024     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
  3025     assert(header_size_min <= hsize, "generic minimum is smallest");
  3026     header_size_min = hsize;
  3027     header_size = intcon(hsize + round_mask);
  3028   } else {
  3029     Node* hss   = intcon(Klass::_lh_header_size_shift);
  3030     Node* hsm   = intcon(Klass::_lh_header_size_mask);
  3031     Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
  3032     hsize       = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
  3033     Node* mask  = intcon(round_mask);
  3034     header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
  3037   Node* elem_shift = NULL;
  3038   if (layout_is_con) {
  3039     int eshift = Klass::layout_helper_log2_element_size(layout_con);
  3040     if (eshift != 0)
  3041       elem_shift = intcon(eshift);
  3042   } else {
  3043     // There is no need to mask or shift this value.
  3044     // The semantics of LShiftINode include an implicit mask to 0x1F.
  3045     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
  3046     elem_shift = layout_val;
  3049   // Transition to native address size for all offset calculations:
  3050   Node* lengthx = ConvI2X(length);
  3051   Node* headerx = ConvI2X(header_size);
  3052 #ifdef _LP64
  3053   { const TypeLong* tllen = _gvn.find_long_type(lengthx);
  3054     if (tllen != NULL && tllen->_lo < 0) {
  3055       // Add a manual constraint to a positive range.  Cf. array_element_address.
  3056       jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
  3057       if (size_max > tllen->_hi)  size_max = tllen->_hi;
  3058       const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
  3059       lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
  3062 #endif
  3064   // Combine header size (plus rounding) and body size.  Then round down.
  3065   // This computation cannot overflow, because it is used only in two
  3066   // places, one where the length is sharply limited, and the other
  3067   // after a successful allocation.
  3068   Node* abody = lengthx;
  3069   if (elem_shift != NULL)
  3070     abody     = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
  3071   Node* size  = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
  3072   if (round_mask != 0) {
  3073     Node* mask = MakeConX(~round_mask);
  3074     size       = _gvn.transform( new(C, 3) AndXNode(size, mask) );
  3076   // else if round_mask == 0, the size computation is self-rounding
  3078   if (return_size_val != NULL) {
  3079     // This is the size
  3080     (*return_size_val) = size;
  3083   // Now generate allocation code
  3085   // With escape analysis, the entire memory state is needed to be able to
  3086   // eliminate the allocation.  If the allocations cannot be eliminated, this
  3087   // will be optimized to the raw slice when the allocation is expanded.
  3088   Node *mem;
  3089   if (C->do_escape_analysis()) {
  3090     mem = reset_memory();
  3091     set_all_memory(mem);
  3092   } else {
  3093     mem = memory(Compile::AliasIdxRaw);
  3096   // Create the AllocateArrayNode and its result projections
  3097   AllocateArrayNode* alloc
  3098     = new (C, AllocateArrayNode::ParmLimit)
  3099         AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
  3100                           control(), mem, i_o(),
  3101                           size, klass_node,
  3102                           initial_slow_test,
  3103                           length);
  3105   // Cast to correct type.  Note that the klass_node may be constant or not,
  3106   // and in the latter case the actual array type will be inexact also.
  3107   // (This happens via a non-constant argument to inline_native_newArray.)
  3108   // In any case, the value of klass_node provides the desired array type.
  3109   const TypeInt* length_type = _gvn.find_int_type(length);
  3110   const TypeInt* narrow_length_type = NULL;
  3111   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
  3112   if (ary_type->isa_aryptr() && length_type != NULL) {
  3113     // Try to get a better type than POS for the size
  3114     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
  3115     narrow_length_type = ary_type->is_aryptr()->size();
  3116     if (narrow_length_type == length_type)
  3117       narrow_length_type = NULL;
  3120   Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
  3122   // Cast length on remaining path to be positive:
  3123   if (narrow_length_type != NULL) {
  3124     Node* ccast = new (C, 2) CastIINode(length, narrow_length_type);
  3125     ccast->set_req(0, control());
  3126     _gvn.set_type_bottom(ccast);
  3127     record_for_igvn(ccast);
  3128     if (map()->find_edge(length) >= 0) {
  3129       replace_in_map(length, ccast);
  3133   return javaoop;
  3136 // The following "Ideal_foo" functions are placed here because they recognize
  3137 // the graph shapes created by the functions immediately above.
  3139 //---------------------------Ideal_allocation----------------------------------
  3140 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
  3141 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
  3142   if (ptr == NULL) {     // reduce dumb test in callers
  3143     return NULL;
  3145   if (ptr->is_CheckCastPP()) {  // strip a raw-to-oop cast
  3146     ptr = ptr->in(1);
  3147     if (ptr == NULL)  return NULL;
  3149   if (ptr->is_Proj()) {
  3150     Node* allo = ptr->in(0);
  3151     if (allo != NULL && allo->is_Allocate()) {
  3152       return allo->as_Allocate();
  3155   // Report failure to match.
  3156   return NULL;
  3159 // Fancy version which also strips off an offset (and reports it to caller).
  3160 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
  3161                                              intptr_t& offset) {
  3162   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
  3163   if (base == NULL)  return NULL;
  3164   return Ideal_allocation(base, phase);
  3167 // Trace Initialize <- Proj[Parm] <- Allocate
  3168 AllocateNode* InitializeNode::allocation() {
  3169   Node* rawoop = in(InitializeNode::RawAddress);
  3170   if (rawoop->is_Proj()) {
  3171     Node* alloc = rawoop->in(0);
  3172     if (alloc->is_Allocate()) {
  3173       return alloc->as_Allocate();
  3176   return NULL;
  3179 // Trace Allocate -> Proj[Parm] -> Initialize
  3180 InitializeNode* AllocateNode::initialization() {
  3181   ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
  3182   if (rawoop == NULL)  return NULL;
  3183   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
  3184     Node* init = rawoop->fast_out(i);
  3185     if (init->is_Initialize()) {
  3186       assert(init->as_Initialize()->allocation() == this, "2-way link");
  3187       return init->as_Initialize();
  3190   return NULL;
  3193 void GraphKit::g1_write_barrier_pre(Node* obj,
  3194                                     Node* adr,
  3195                                     uint alias_idx,
  3196                                     Node* val,
  3197                                     const Type* val_type,
  3198                                     BasicType bt) {
  3199   IdealKit ideal(gvn(), control(), merged_memory(), true);
  3200 #define __ ideal.
  3201   __ declares_done();
  3203   Node* thread = __ thread();
  3205   Node* no_ctrl = NULL;
  3206   Node* no_base = __ top();
  3207   Node* zero = __ ConI(0);
  3209   float likely  = PROB_LIKELY(0.999);
  3210   float unlikely  = PROB_UNLIKELY(0.999);
  3212   BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
  3213   assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
  3215   // Offsets into the thread
  3216   const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +  // 648
  3217                                           PtrQueue::byte_offset_of_active());
  3218   const int index_offset   = in_bytes(JavaThread::satb_mark_queue_offset() +  // 656
  3219                                           PtrQueue::byte_offset_of_index());
  3220   const int buffer_offset  = in_bytes(JavaThread::satb_mark_queue_offset() +  // 652
  3221                                           PtrQueue::byte_offset_of_buf());
  3222   // Now the actual pointers into the thread
  3224   // set_control( ctl);
  3226   Node* marking_adr = __ AddP(no_base, thread, __ ConX(marking_offset));
  3227   Node* buffer_adr  = __ AddP(no_base, thread, __ ConX(buffer_offset));
  3228   Node* index_adr   = __ AddP(no_base, thread, __ ConX(index_offset));
  3230   // Now some of the values
  3232   Node* marking = __ load(no_ctrl, marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
  3233   Node* index   = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
  3234   Node* buffer  = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
  3236   // if (!marking)
  3237   __ if_then(marking, BoolTest::ne, zero); {
  3239     const Type* t1 = adr->bottom_type();
  3240     const Type* t2 = val->bottom_type();
  3242     Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx);
  3243     // if (orig != NULL)
  3244     __ if_then(orig, BoolTest::ne, null()); {
  3246       // load original value
  3247       // alias_idx correct??
  3249       // is the queue for this thread full?
  3250       __ if_then(index, BoolTest::ne, zero, likely); {
  3252         // decrement the index
  3253         Node* next_index = __ SubI(index,  __ ConI(sizeof(intptr_t)));
  3254         Node* next_indexX = next_index;
  3255 #ifdef _LP64
  3256           // We could refine the type for what it's worth
  3257           // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
  3258           next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
  3259 #endif // _LP64
  3261         // Now get the buffer location we will log the original value into and store it
  3263         Node *log_addr = __ AddP(no_base, buffer, next_indexX);
  3264         // __ store(__ ctrl(), log_addr, orig, T_OBJECT, C->get_alias_index(TypeOopPtr::BOTTOM));
  3265         __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw);
  3268         // update the index
  3269         // __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
  3270         // This is a hack to force this store to occur before the oop store that is coming up
  3271         __ store(__ ctrl(), index_adr, next_index, T_INT, C->get_alias_index(TypeOopPtr::BOTTOM));
  3273       } __ else_(); {
  3275         // logging buffer is full, call the runtime
  3276         const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
  3277         // __ make_leaf_call(tf, OptoRuntime::g1_wb_pre_Java(), "g1_wb_pre", orig, thread);
  3278         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, thread);
  3279       } __ end_if();
  3280     } __ end_if();
  3281   } __ end_if();
  3283   __ drain_delay_transform();
  3284   set_control( __ ctrl());
  3285   set_all_memory( __ merged_memory());
  3287 #undef __
  3290 //
  3291 // Update the card table and add card address to the queue
  3292 //
  3293 void GraphKit::g1_mark_card(IdealKit* ideal, Node* card_adr, Node* store,  Node* index, Node* index_adr, Node* buffer, const TypeFunc* tf) {
  3294 #define __ ideal->
  3295   Node* zero = __ ConI(0);
  3296   Node* no_base = __ top();
  3297   BasicType card_bt = T_BYTE;
  3298   // Smash zero into card. MUST BE ORDERED WRT TO STORE
  3299   __ storeCM(__ ctrl(), card_adr, zero, store, card_bt, Compile::AliasIdxRaw);
  3301   //  Now do the queue work
  3302   __ if_then(index, BoolTest::ne, zero); {
  3304     Node* next_index = __ SubI(index,  __ ConI(sizeof(intptr_t)));
  3305     Node* next_indexX = next_index;
  3306 #ifdef _LP64
  3307     // We could refine the type for what it's worth
  3308     // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
  3309     next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
  3310 #endif // _LP64
  3311     Node* log_addr = __ AddP(no_base, buffer, next_indexX);
  3313     __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
  3314     __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
  3316   } __ else_(); {
  3317     __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
  3318   } __ end_if();
  3319 #undef __
  3322 void GraphKit::g1_write_barrier_post(Node* store,
  3323                                      Node* obj,
  3324                                      Node* adr,
  3325                                      uint alias_idx,
  3326                                      Node* val,
  3327                                      BasicType bt,
  3328                                      bool use_precise) {
  3329   // If we are writing a NULL then we need no post barrier
  3331   if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
  3332     // Must be NULL
  3333     const Type* t = val->bottom_type();
  3334     assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
  3335     // No post barrier if writing NULLx
  3336     return;
  3339   if (!use_precise) {
  3340     // All card marks for a (non-array) instance are in one place:
  3341     adr = obj;
  3343   // (Else it's an array (or unknown), and we want more precise card marks.)
  3344   assert(adr != NULL, "");
  3346   IdealKit ideal(gvn(), control(), merged_memory(), true);
  3347 #define __ ideal.
  3348   __ declares_done();
  3350   Node* thread = __ thread();
  3352   Node* no_ctrl = NULL;
  3353   Node* no_base = __ top();
  3354   float likely  = PROB_LIKELY(0.999);
  3355   float unlikely  = PROB_UNLIKELY(0.999);
  3356   Node* zero = __ ConI(0);
  3357   Node* zeroX = __ ConX(0);
  3359   // Get the alias_index for raw card-mark memory
  3360   const TypePtr* card_type = TypeRawPtr::BOTTOM;
  3362   const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
  3364   // Get the address of the card table
  3365   CardTableModRefBS* ct =
  3366     (CardTableModRefBS*)(Universe::heap()->barrier_set());
  3367   Node *card_table = __ makecon(TypeRawPtr::make((address)ct->byte_map_base));
  3368   // Get base of card map
  3369   assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
  3372   // Offsets into the thread
  3373   const int index_offset  = in_bytes(JavaThread::dirty_card_queue_offset() +
  3374                                      PtrQueue::byte_offset_of_index());
  3375   const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
  3376                                      PtrQueue::byte_offset_of_buf());
  3378   // Pointers into the thread
  3380   Node* buffer_adr = __ AddP(no_base, thread, __ ConX(buffer_offset));
  3381   Node* index_adr =  __ AddP(no_base, thread, __ ConX(index_offset));
  3383   // Now some values
  3385   Node* index  = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
  3386   Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
  3389   // Convert the store obj pointer to an int prior to doing math on it
  3390   // Use addr not obj gets accurate card marks
  3392   // Node* cast = __ CastPX(no_ctrl, adr /* obj */);
  3394   // Must use ctrl to prevent "integerized oop" existing across safepoint
  3395   Node* cast =  __ CastPX(__ ctrl(), ( use_precise ? adr : obj ));
  3397   // Divide pointer by card size
  3398   Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
  3400   // Combine card table base and card offset
  3401   Node *card_adr = __ AddP(no_base, card_table, card_offset );
  3403   // If we know the value being stored does it cross regions?
  3405   if (val != NULL) {
  3406     // Does the store cause us to cross regions?
  3408     // Should be able to do an unsigned compare of region_size instead of
  3409     // and extra shift. Do we have an unsigned compare??
  3410     // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
  3411     Node* xor_res =  __ URShiftX ( __ XorX( cast,  __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
  3413     // if (xor_res == 0) same region so skip
  3414     __ if_then(xor_res, BoolTest::ne, zeroX); {
  3416       // No barrier if we are storing a NULL
  3417       __ if_then(val, BoolTest::ne, null(), unlikely); {
  3419         // Ok must mark the card if not already dirty
  3421         // load the original value of the card
  3422         Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
  3424         __ if_then(card_val, BoolTest::ne, zero); {
  3425           g1_mark_card(&ideal, card_adr, store, index, index_adr, buffer, tf);
  3426         } __ end_if();
  3427       } __ end_if();
  3428     } __ end_if();
  3429   } else {
  3430     g1_mark_card(&ideal, card_adr, store, index, index_adr, buffer, tf);
  3434   __ drain_delay_transform();
  3435   set_control( __ ctrl());
  3436   set_all_memory( __ merged_memory());
  3437 #undef __

mercurial