G1BarrierSet_merge

0 /*
1  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
2  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
3  *
4  * This code is free software; you can redistribute it and/or modify it
5  * under the terms of the GNU General Public License version 2 only, as
6  * published by the Free Software Foundation.
7  *
8  * This code is distributed in the hope that it will be useful, but WITHOUT
9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
11  * version 2 for more details (a copy is included in the LICENSE file that
12  * accompanied this code).
13  *
14  * You should have received a copy of the GNU General Public License version
15  * 2 along with this work; if not, write to the Free Software Foundation,
16  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17  *
18  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19  * or visit www.oracle.com if you need additional information or have any
20  * questions.
21  *
22  */
23 
24 #include "precompiled.hpp"
25 #include "ci/ciUtilities.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/g1/g1BarrierSet.hpp"
28 #include "gc/g1/g1CardTable.hpp"

29 #include "gc/g1/heapRegion.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/cardTable.hpp"
32 #include "gc/shared/cardTableModRefBS.hpp"
33 #include "gc/shared/collectedHeap.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/intrinsicnode.hpp"
41 #include "opto/locknode.hpp"
42 #include "opto/machnode.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/parse.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 
50 //----------------------------GraphKit-----------------------------------------
51 // Main utility constructor.
52 GraphKit::GraphKit(JVMState* jvms)
53   : Phase(Phase::Parser),
54     _env(C->env()),
55     _gvn(*C->initial_gvn())
56 {
57   _exceptions = jvms->map()->next_exception();
58   if (_exceptions != NULL)  jvms->map()->set_next_exception(NULL);
59   set_jvms(jvms);
60 }
61 
62 // Private constructor for parser.
63 GraphKit::GraphKit()
64   : Phase(Phase::Parser),
65     _env(C->env()),
66     _gvn(*C->initial_gvn())
67 {
68   _exceptions = NULL;
69   set_map(NULL);
70   debug_only(_sp = -99);
71   debug_only(set_bci(-99));
72 }
73 
74 
75 
76 //---------------------------clean_stack---------------------------------------
77 // Clear away rubbish from the stack area of the JVM state.
78 // This destroys any arguments that may be waiting on the stack.
79 void GraphKit::clean_stack(int from_sp) {
80   SafePointNode* map      = this->map();
81   JVMState*      jvms     = this->jvms();
82   int            stk_size = jvms->stk_size();
83   int            stkoff   = jvms->stkoff();
84   Node*          top      = this->top();
85   for (int i = from_sp; i < stk_size; i++) {
86     if (map->in(stkoff + i) != top) {
87       map->set_req(stkoff + i, top);
88     }
89   }
90 }
91 
92 
93 //--------------------------------sync_jvms-----------------------------------
94 // Make sure our current jvms agrees with our parse state.
95 JVMState* GraphKit::sync_jvms() const {
96   JVMState* jvms = this->jvms();
97   jvms->set_bci(bci());       // Record the new bci in the JVMState
98   jvms->set_sp(sp());         // Record the new sp in the JVMState
99   assert(jvms_in_sync(), "jvms is now in sync");
100   return jvms;
101 }
102 
103 //--------------------------------sync_jvms_for_reexecute---------------------
104 // Make sure our current jvms agrees with our parse state.  This version
105 // uses the reexecute_sp for reexecuting bytecodes.
106 JVMState* GraphKit::sync_jvms_for_reexecute() {
107   JVMState* jvms = this->jvms();
108   jvms->set_bci(bci());          // Record the new bci in the JVMState
109   jvms->set_sp(reexecute_sp());  // Record the new sp in the JVMState
110   return jvms;
111 }
112 
113 #ifdef ASSERT
114 bool GraphKit::jvms_in_sync() const {
115   Parse* parse = is_Parse();
116   if (parse == NULL) {
117     if (bci() !=      jvms()->bci())          return false;
118     if (sp()  != (int)jvms()->sp())           return false;
119     return true;
120   }
121   if (jvms()->method() != parse->method())    return false;
122   if (jvms()->bci()    != parse->bci())       return false;
123   int jvms_sp = jvms()->sp();
124   if (jvms_sp          != parse->sp())        return false;
125   int jvms_depth = jvms()->depth();
126   if (jvms_depth       != parse->depth())     return false;
127   return true;
128 }
129 
130 // Local helper checks for special internal merge points
131 // used to accumulate and merge exception states.
132 // They are marked by the region's in(0) edge being the map itself.
133 // Such merge points must never "escape" into the parser at large,
134 // until they have been handed to gvn.transform.
135 static bool is_hidden_merge(Node* reg) {
136   if (reg == NULL)  return false;
137   if (reg->is_Phi()) {
138     reg = reg->in(0);
139     if (reg == NULL)  return false;
140   }
141   return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
142 }
143 
144 void GraphKit::verify_map() const {
145   if (map() == NULL)  return;  // null map is OK
146   assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
147   assert(!map()->has_exceptions(),    "call add_exception_states_from 1st");
148   assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
149 }
150 
151 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
152   assert(ex_map->next_exception() == NULL, "not already part of a chain");
153   assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
154 }
155 #endif
156 
157 //---------------------------stop_and_kill_map---------------------------------
158 // Set _map to NULL, signalling a stop to further bytecode execution.
159 // First smash the current map's control to a constant, to mark it dead.
160 void GraphKit::stop_and_kill_map() {
161   SafePointNode* dead_map = stop();
162   if (dead_map != NULL) {
163     dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
164     assert(dead_map->is_killed(), "must be so marked");
165   }
166 }
167 
168 
169 //--------------------------------stopped--------------------------------------
170 // Tell if _map is NULL, or control is top.
171 bool GraphKit::stopped() {
172   if (map() == NULL)           return true;
173   else if (control() == top()) return true;
174   else                         return false;
175 }
176 
177 
178 //-----------------------------has_ex_handler----------------------------------
179 // Tell if this method or any caller method has exception handlers.
180 bool GraphKit::has_ex_handler() {
181   for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
182     if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
183       return true;
184     }
185   }
186   return false;
187 }
188 
189 //------------------------------save_ex_oop------------------------------------
190 // Save an exception without blowing stack contents or other JVM state.
191 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
192   assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
193   ex_map->add_req(ex_oop);
194   debug_only(verify_exception_state(ex_map));
195 }
196 
197 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
198   assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
199   Node* ex_oop = ex_map->in(ex_map->req()-1);
200   if (clear_it)  ex_map->del_req(ex_map->req()-1);
201   return ex_oop;
202 }
203 
204 //-----------------------------saved_ex_oop------------------------------------
205 // Recover a saved exception from its map.
206 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
207   return common_saved_ex_oop(ex_map, false);
208 }
209 
210 //--------------------------clear_saved_ex_oop---------------------------------
211 // Erase a previously saved exception from its map.
212 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
213   return common_saved_ex_oop(ex_map, true);
214 }
215 
216 #ifdef ASSERT
217 //---------------------------has_saved_ex_oop----------------------------------
218 // Erase a previously saved exception from its map.
219 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
220   return ex_map->req() == ex_map->jvms()->endoff()+1;
221 }
222 #endif
223 
224 //-------------------------make_exception_state--------------------------------
225 // Turn the current JVM state into an exception state, appending the ex_oop.
226 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
227   sync_jvms();
228   SafePointNode* ex_map = stop();  // do not manipulate this map any more
229   set_saved_ex_oop(ex_map, ex_oop);
230   return ex_map;
231 }
232 
233 
234 //--------------------------add_exception_state--------------------------------
235 // Add an exception to my list of exceptions.
236 void GraphKit::add_exception_state(SafePointNode* ex_map) {
237   if (ex_map == NULL || ex_map->control() == top()) {
238     return;
239   }
240 #ifdef ASSERT
241   verify_exception_state(ex_map);
242   if (has_exceptions()) {
243     assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
244   }
245 #endif
246 
247   // If there is already an exception of exactly this type, merge with it.
248   // In particular, null-checks and other low-level exceptions common up here.
249   Node*       ex_oop  = saved_ex_oop(ex_map);
250   const Type* ex_type = _gvn.type(ex_oop);
251   if (ex_oop == top()) {
252     // No action needed.
253     return;
254   }
255   assert(ex_type->isa_instptr(), "exception must be an instance");
256   for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
257     const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
258     // We check sp also because call bytecodes can generate exceptions
259     // both before and after arguments are popped!
260     if (ex_type2 == ex_type
261         && e2->_jvms->sp() == ex_map->_jvms->sp()) {
262       combine_exception_states(ex_map, e2);
263       return;
264     }
265   }
266 
267   // No pre-existing exception of the same type.  Chain it on the list.
268   push_exception_state(ex_map);
269 }
270 
271 //-----------------------add_exception_states_from-----------------------------
272 void GraphKit::add_exception_states_from(JVMState* jvms) {
273   SafePointNode* ex_map = jvms->map()->next_exception();
274   if (ex_map != NULL) {
275     jvms->map()->set_next_exception(NULL);
276     for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
277       next_map = ex_map->next_exception();
278       ex_map->set_next_exception(NULL);
279       add_exception_state(ex_map);
280     }
281   }
282 }
283 
284 //-----------------------transfer_exceptions_into_jvms-------------------------
285 JVMState* GraphKit::transfer_exceptions_into_jvms() {
286   if (map() == NULL) {
287     // We need a JVMS to carry the exceptions, but the map has gone away.
288     // Create a scratch JVMS, cloned from any of the exception states...
289     if (has_exceptions()) {
290       _map = _exceptions;
291       _map = clone_map();
292       _map->set_next_exception(NULL);
293       clear_saved_ex_oop(_map);
294       debug_only(verify_map());
295     } else {
296       // ...or created from scratch
297       JVMState* jvms = new (C) JVMState(_method, NULL);
298       jvms->set_bci(_bci);
299       jvms->set_sp(_sp);
300       jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
301       set_jvms(jvms);
302       for (uint i = 0; i < map()->req(); i++)  map()->init_req(i, top());
303       set_all_memory(top());
304       while (map()->req() < jvms->endoff())  map()->add_req(top());
305     }
306     // (This is a kludge, in case you didn't notice.)
307     set_control(top());
308   }
309   JVMState* jvms = sync_jvms();
310   assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
311   jvms->map()->set_next_exception(_exceptions);
312   _exceptions = NULL;   // done with this set of exceptions
313   return jvms;
314 }
315 
316 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
317   assert(is_hidden_merge(dstphi), "must be a special merge node");
318   assert(is_hidden_merge(srcphi), "must be a special merge node");
319   uint limit = srcphi->req();
320   for (uint i = PhiNode::Input; i < limit; i++) {
321     dstphi->add_req(srcphi->in(i));
322   }
323 }
324 static inline void add_one_req(Node* dstphi, Node* src) {
325   assert(is_hidden_merge(dstphi), "must be a special merge node");
326   assert(!is_hidden_merge(src), "must not be a special merge node");
327   dstphi->add_req(src);
328 }
329 
330 //-----------------------combine_exception_states------------------------------
331 // This helper function combines exception states by building phis on a
332 // specially marked state-merging region.  These regions and phis are
333 // untransformed, and can build up gradually.  The region is marked by
334 // having a control input of its exception map, rather than NULL.  Such
335 // regions do not appear except in this function, and in use_exception_state.
336 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
337   if (failing())  return;  // dying anyway...
338   JVMState* ex_jvms = ex_map->_jvms;
339   assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
340   assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
341   assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
342   assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
343   assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
344   assert(ex_map->req() == phi_map->req(), "matching maps");
345   uint tos = ex_jvms->stkoff() + ex_jvms->sp();
346   Node*         hidden_merge_mark = root();
347   Node*         region  = phi_map->control();
348   MergeMemNode* phi_mem = phi_map->merged_memory();
349   MergeMemNode* ex_mem  = ex_map->merged_memory();
350   if (region->in(0) != hidden_merge_mark) {
351     // The control input is not (yet) a specially-marked region in phi_map.
352     // Make it so, and build some phis.
353     region = new RegionNode(2);
354     _gvn.set_type(region, Type::CONTROL);
355     region->set_req(0, hidden_merge_mark);  // marks an internal ex-state
356     region->init_req(1, phi_map->control());
357     phi_map->set_control(region);
358     Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
359     record_for_igvn(io_phi);
360     _gvn.set_type(io_phi, Type::ABIO);
361     phi_map->set_i_o(io_phi);
362     for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
363       Node* m = mms.memory();
364       Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
365       record_for_igvn(m_phi);
366       _gvn.set_type(m_phi, Type::MEMORY);
367       mms.set_memory(m_phi);
368     }
369   }
370 
371   // Either or both of phi_map and ex_map might already be converted into phis.
372   Node* ex_control = ex_map->control();
373   // if there is special marking on ex_map also, we add multiple edges from src
374   bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
375   // how wide was the destination phi_map, originally?
376   uint orig_width = region->req();
377 
378   if (add_multiple) {
379     add_n_reqs(region, ex_control);
380     add_n_reqs(phi_map->i_o(), ex_map->i_o());
381   } else {
382     // ex_map has no merges, so we just add single edges everywhere
383     add_one_req(region, ex_control);
384     add_one_req(phi_map->i_o(), ex_map->i_o());
385   }
386   for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
387     if (mms.is_empty()) {
388       // get a copy of the base memory, and patch some inputs into it
389       const TypePtr* adr_type = mms.adr_type(C);
390       Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
391       assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
392       mms.set_memory(phi);
393       // Prepare to append interesting stuff onto the newly sliced phi:
394       while (phi->req() > orig_width)  phi->del_req(phi->req()-1);
395     }
396     // Append stuff from ex_map:
397     if (add_multiple) {
398       add_n_reqs(mms.memory(), mms.memory2());
399     } else {
400       add_one_req(mms.memory(), mms.memory2());
401     }
402   }
403   uint limit = ex_map->req();
404   for (uint i = TypeFunc::Parms; i < limit; i++) {
405     // Skip everything in the JVMS after tos.  (The ex_oop follows.)
406     if (i == tos)  i = ex_jvms->monoff();
407     Node* src = ex_map->in(i);
408     Node* dst = phi_map->in(i);
409     if (src != dst) {
410       PhiNode* phi;
411       if (dst->in(0) != region) {
412         dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
413         record_for_igvn(phi);
414         _gvn.set_type(phi, phi->type());
415         phi_map->set_req(i, dst);
416         // Prepare to append interesting stuff onto the new phi:
417         while (dst->req() > orig_width)  dst->del_req(dst->req()-1);
418       } else {
419         assert(dst->is_Phi(), "nobody else uses a hidden region");
420         phi = dst->as_Phi();
421       }
422       if (add_multiple && src->in(0) == ex_control) {
423         // Both are phis.
424         add_n_reqs(dst, src);
425       } else {
426         while (dst->req() < region->req())  add_one_req(dst, src);
427       }
428       const Type* srctype = _gvn.type(src);
429       if (phi->type() != srctype) {
430         const Type* dsttype = phi->type()->meet_speculative(srctype);
431         if (phi->type() != dsttype) {
432           phi->set_type(dsttype);
433           _gvn.set_type(phi, dsttype);
434         }
435       }
436     }
437   }
438   phi_map->merge_replaced_nodes_with(ex_map);
439 }
440 
441 //--------------------------use_exception_state--------------------------------
442 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
443   if (failing()) { stop(); return top(); }
444   Node* region = phi_map->control();
445   Node* hidden_merge_mark = root();
446   assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
447   Node* ex_oop = clear_saved_ex_oop(phi_map);
448   if (region->in(0) == hidden_merge_mark) {
449     // Special marking for internal ex-states.  Process the phis now.
450     region->set_req(0, region);  // now it's an ordinary region
451     set_jvms(phi_map->jvms());   // ...so now we can use it as a map
452     // Note: Setting the jvms also sets the bci and sp.
453     set_control(_gvn.transform(region));
454     uint tos = jvms()->stkoff() + sp();
455     for (uint i = 1; i < tos; i++) {
456       Node* x = phi_map->in(i);
457       if (x->in(0) == region) {
458         assert(x->is_Phi(), "expected a special phi");
459         phi_map->set_req(i, _gvn.transform(x));
460       }
461     }
462     for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
463       Node* x = mms.memory();
464       if (x->in(0) == region) {
465         assert(x->is_Phi(), "nobody else uses a hidden region");
466         mms.set_memory(_gvn.transform(x));
467       }
468     }
469     if (ex_oop->in(0) == region) {
470       assert(ex_oop->is_Phi(), "expected a special phi");
471       ex_oop = _gvn.transform(ex_oop);
472     }
473   } else {
474     set_jvms(phi_map->jvms());
475   }
476 
477   assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
478   assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
479   return ex_oop;
480 }
481 
482 //---------------------------------java_bc-------------------------------------
483 Bytecodes::Code GraphKit::java_bc() const {
484   ciMethod* method = this->method();
485   int       bci    = this->bci();
486   if (method != NULL && bci != InvocationEntryBci)
487     return method->java_code_at_bci(bci);
488   else
489     return Bytecodes::_illegal;
490 }
491 
492 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
493                                                           bool must_throw) {
494     // if the exception capability is set, then we will generate code
495     // to check the JavaThread.should_post_on_exceptions flag to see
496     // if we actually need to report exception events (for this
497     // thread).  If we don't need to report exception events, we will
498     // take the normal fast path provided by add_exception_events.  If
499     // exception event reporting is enabled for this thread, we will
500     // take the uncommon_trap in the BuildCutout below.
501 
502     // first must access the should_post_on_exceptions_flag in this thread's JavaThread
503     Node* jthread = _gvn.transform(new ThreadLocalNode());
504     Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
505     Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
506 
507     // Test the should_post_on_exceptions_flag vs. 0
508     Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
509     Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
510 
511     // Branch to slow_path if should_post_on_exceptions_flag was true
512     { BuildCutout unless(this, tst, PROB_MAX);
513       // Do not try anything fancy if we're notifying the VM on every throw.
514       // Cf. case Bytecodes::_athrow in parse2.cpp.
515       uncommon_trap(reason, Deoptimization::Action_none,
516                     (ciKlass*)NULL, (char*)NULL, must_throw);
517     }
518 
519 }
520 
521 //------------------------------builtin_throw----------------------------------
522 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
523   bool must_throw = true;
524 
525   if (env()->jvmti_can_post_on_exceptions()) {
526     // check if we must post exception events, take uncommon trap if so
527     uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
528     // here if should_post_on_exceptions is false
529     // continue on with the normal codegen
530   }
531 
532   // If this particular condition has not yet happened at this
533   // bytecode, then use the uncommon trap mechanism, and allow for
534   // a future recompilation if several traps occur here.
535   // If the throw is hot, try to use a more complicated inline mechanism
536   // which keeps execution inside the compiled code.
537   bool treat_throw_as_hot = false;
538   ciMethodData* md = method()->method_data();
539 
540   if (ProfileTraps) {
541     if (too_many_traps(reason)) {
542       treat_throw_as_hot = true;
543     }
544     // (If there is no MDO at all, assume it is early in
545     // execution, and that any deopts are part of the
546     // startup transient, and don't need to be remembered.)
547 
548     // Also, if there is a local exception handler, treat all throws
549     // as hot if there has been at least one in this method.
550     if (C->trap_count(reason) != 0
551         && method()->method_data()->trap_count(reason) != 0
552         && has_ex_handler()) {
553         treat_throw_as_hot = true;
554     }
555   }
556 
557   // If this throw happens frequently, an uncommon trap might cause
558   // a performance pothole.  If there is a local exception handler,
559   // and if this particular bytecode appears to be deoptimizing often,
560   // let us handle the throw inline, with a preconstructed instance.
561   // Note:   If the deopt count has blown up, the uncommon trap
562   // runtime is going to flush this nmethod, not matter what.
563   if (treat_throw_as_hot
564       && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
565     // If the throw is local, we use a pre-existing instance and
566     // punt on the backtrace.  This would lead to a missing backtrace
567     // (a repeat of 4292742) if the backtrace object is ever asked
568     // for its backtrace.
569     // Fixing this remaining case of 4292742 requires some flavor of
570     // escape analysis.  Leave that for the future.
571     ciInstance* ex_obj = NULL;
572     switch (reason) {
573     case Deoptimization::Reason_null_check:
574       ex_obj = env()->NullPointerException_instance();
575       break;
576     case Deoptimization::Reason_div0_check:
577       ex_obj = env()->ArithmeticException_instance();
578       break;
579     case Deoptimization::Reason_range_check:
580       ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
581       break;
582     case Deoptimization::Reason_class_check:
583       if (java_bc() == Bytecodes::_aastore) {
584         ex_obj = env()->ArrayStoreException_instance();
585       } else {
586         ex_obj = env()->ClassCastException_instance();
587       }
588       break;
589     default:
590       break;
591     }
592     if (failing()) { stop(); return; }  // exception allocation might fail
593     if (ex_obj != NULL) {
594       // Cheat with a preallocated exception object.
595       if (C->log() != NULL)
596         C->log()->elem("hot_throw preallocated='1' reason='%s'",
597                        Deoptimization::trap_reason_name(reason));
598       const TypeInstPtr* ex_con  = TypeInstPtr::make(ex_obj);
599       Node*              ex_node = _gvn.transform(ConNode::make(ex_con));
600 
601       // Clear the detail message of the preallocated exception object.
602       // Weblogic sometimes mutates the detail message of exceptions
603       // using reflection.
604       int offset = java_lang_Throwable::get_detailMessage_offset();
605       const TypePtr* adr_typ = ex_con->add_offset(offset);
606 
607       Node *adr = basic_plus_adr(ex_node, ex_node, offset);
608       const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
609       // Conservatively release stores of object references.
610       Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
611 
612       add_exception_state(make_exception_state(ex_node));
613       return;
614     }
615   }
616 
617   // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
618   // It won't be much cheaper than bailing to the interp., since we'll
619   // have to pass up all the debug-info, and the runtime will have to
620   // create the stack trace.
621 
622   // Usual case:  Bail to interpreter.
623   // Reserve the right to recompile if we haven't seen anything yet.
624 
625   ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
626   Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
627   if (treat_throw_as_hot
628       && (method()->method_data()->trap_recompiled_at(bci(), m)
629           || C->too_many_traps(reason))) {
630     // We cannot afford to take more traps here.  Suffer in the interpreter.
631     if (C->log() != NULL)
632       C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
633                      Deoptimization::trap_reason_name(reason),
634                      C->trap_count(reason));
635     action = Deoptimization::Action_none;
636   }
637 
638   // "must_throw" prunes the JVM state to include only the stack, if there
639   // are no local exception handlers.  This should cut down on register
640   // allocation time and code size, by drastically reducing the number
641   // of in-edges on the call to the uncommon trap.
642 
643   uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
644 }
645 
646 
647 //----------------------------PreserveJVMState---------------------------------
648 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
649   debug_only(kit->verify_map());
650   _kit    = kit;
651   _map    = kit->map();   // preserve the map
652   _sp     = kit->sp();
653   kit->set_map(clone_map ? kit->clone_map() : NULL);
654 #ifdef ASSERT
655   _bci    = kit->bci();
656   Parse* parser = kit->is_Parse();
657   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
658   _block  = block;
659 #endif
660 }
661 PreserveJVMState::~PreserveJVMState() {
662   GraphKit* kit = _kit;
663 #ifdef ASSERT
664   assert(kit->bci() == _bci, "bci must not shift");
665   Parse* parser = kit->is_Parse();
666   int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
667   assert(block == _block,    "block must not shift");
668 #endif
669   kit->set_map(_map);
670   kit->set_sp(_sp);
671 }
672 
673 
674 //-----------------------------BuildCutout-------------------------------------
675 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
676   : PreserveJVMState(kit)
677 {
678   assert(p->is_Con() || p->is_Bool(), "test must be a bool");
679   SafePointNode* outer_map = _map;   // preserved map is caller's
680   SafePointNode* inner_map = kit->map();
681   IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
682   outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
683   inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
684 }
685 BuildCutout::~BuildCutout() {
686   GraphKit* kit = _kit;
687   assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
688 }
689 
690 //---------------------------PreserveReexecuteState----------------------------
691 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
692   assert(!kit->stopped(), "must call stopped() before");
693   _kit    =    kit;
694   _sp     =    kit->sp();
695   _reexecute = kit->jvms()->_reexecute;
696 }
697 PreserveReexecuteState::~PreserveReexecuteState() {
698   if (_kit->stopped()) return;
699   _kit->jvms()->_reexecute = _reexecute;
700   _kit->set_sp(_sp);
701 }
702 
703 //------------------------------clone_map--------------------------------------
704 // Implementation of PreserveJVMState
705 //
706 // Only clone_map(...) here. If this function is only used in the
707 // PreserveJVMState class we may want to get rid of this extra
708 // function eventually and do it all there.
709 
710 SafePointNode* GraphKit::clone_map() {
711   if (map() == NULL)  return NULL;
712 
713   // Clone the memory edge first
714   Node* mem = MergeMemNode::make(map()->memory());
715   gvn().set_type_bottom(mem);
716 
717   SafePointNode *clonemap = (SafePointNode*)map()->clone();
718   JVMState* jvms = this->jvms();
719   JVMState* clonejvms = jvms->clone_shallow(C);
720   clonemap->set_memory(mem);
721   clonemap->set_jvms(clonejvms);
722   clonejvms->set_map(clonemap);
723   record_for_igvn(clonemap);
724   gvn().set_type_bottom(clonemap);
725   return clonemap;
726 }
727 
728 
729 //-----------------------------set_map_clone-----------------------------------
730 void GraphKit::set_map_clone(SafePointNode* m) {
731   _map = m;
732   _map = clone_map();
733   _map->set_next_exception(NULL);
734   debug_only(verify_map());
735 }
736 
737 
738 //----------------------------kill_dead_locals---------------------------------
739 // Detect any locals which are known to be dead, and force them to top.
740 void GraphKit::kill_dead_locals() {
741   // Consult the liveness information for the locals.  If any
742   // of them are unused, then they can be replaced by top().  This
743   // should help register allocation time and cut down on the size
744   // of the deoptimization information.
745 
746   // This call is made from many of the bytecode handling
747   // subroutines called from the Big Switch in do_one_bytecode.
748   // Every bytecode which might include a slow path is responsible
749   // for killing its dead locals.  The more consistent we
750   // are about killing deads, the fewer useless phis will be
751   // constructed for them at various merge points.
752 
753   // bci can be -1 (InvocationEntryBci).  We return the entry
754   // liveness for the method.
755 
756   if (method() == NULL || method()->code_size() == 0) {
757     // We are building a graph for a call to a native method.
758     // All locals are live.
759     return;
760   }
761 
762   ResourceMark rm;
763 
764   // Consult the liveness information for the locals.  If any
765   // of them are unused, then they can be replaced by top().  This
766   // should help register allocation time and cut down on the size
767   // of the deoptimization information.
768   MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
769 
770   int len = (int)live_locals.size();
771   assert(len <= jvms()->loc_size(), "too many live locals");
772   for (int local = 0; local < len; local++) {
773     if (!live_locals.at(local)) {
774       set_local(local, top());
775     }
776   }
777 }
778 
779 #ifdef ASSERT
780 //-------------------------dead_locals_are_killed------------------------------
781 // Return true if all dead locals are set to top in the map.
782 // Used to assert "clean" debug info at various points.
783 bool GraphKit::dead_locals_are_killed() {
784   if (method() == NULL || method()->code_size() == 0) {
785     // No locals need to be dead, so all is as it should be.
786     return true;
787   }
788 
789   // Make sure somebody called kill_dead_locals upstream.
790   ResourceMark rm;
791   for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
792     if (jvms->loc_size() == 0)  continue;  // no locals to consult
793     SafePointNode* map = jvms->map();
794     ciMethod* method = jvms->method();
795     int       bci    = jvms->bci();
796     if (jvms == this->jvms()) {
797       bci = this->bci();  // it might not yet be synched
798     }
799     MethodLivenessResult live_locals = method->liveness_at_bci(bci);
800     int len = (int)live_locals.size();
801     if (!live_locals.is_valid() || len == 0)
802       // This method is trivial, or is poisoned by a breakpoint.
803       return true;
804     assert(len == jvms->loc_size(), "live map consistent with locals map");
805     for (int local = 0; local < len; local++) {
806       if (!live_locals.at(local) && map->local(jvms, local) != top()) {
807         if (PrintMiscellaneous && (Verbose || WizardMode)) {
808           tty->print_cr("Zombie local %d: ", local);
809           jvms->dump();
810         }
811         return false;
812       }
813     }
814   }
815   return true;
816 }
817 
818 #endif //ASSERT
819 
820 // Helper function for enforcing certain bytecodes to reexecute if
821 // deoptimization happens
822 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
823   ciMethod* cur_method = jvms->method();
824   int       cur_bci   = jvms->bci();
825   if (cur_method != NULL && cur_bci != InvocationEntryBci) {
826     Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
827     return Interpreter::bytecode_should_reexecute(code) ||
828            (is_anewarray && code == Bytecodes::_multianewarray);
829     // Reexecute _multianewarray bytecode which was replaced with
830     // sequence of [a]newarray. See Parse::do_multianewarray().
831     //
832     // Note: interpreter should not have it set since this optimization
833     // is limited by dimensions and guarded by flag so in some cases
834     // multianewarray() runtime calls will be generated and
835     // the bytecode should not be reexecutes (stack will not be reset).
836   } else
837     return false;
838 }
839 
840 // Helper function for adding JVMState and debug information to node
841 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
842   // Add the safepoint edges to the call (or other safepoint).
843 
844   // Make sure dead locals are set to top.  This
845   // should help register allocation time and cut down on the size
846   // of the deoptimization information.
847   assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
848 
849   // Walk the inline list to fill in the correct set of JVMState's
850   // Also fill in the associated edges for each JVMState.
851 
852   // If the bytecode needs to be reexecuted we need to put
853   // the arguments back on the stack.
854   const bool should_reexecute = jvms()->should_reexecute();
855   JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
856 
857   // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
858   // undefined if the bci is different.  This is normal for Parse but it
859   // should not happen for LibraryCallKit because only one bci is processed.
860   assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
861          "in LibraryCallKit the reexecute bit should not change");
862 
863   // If we are guaranteed to throw, we can prune everything but the
864   // input to the current bytecode.
865   bool can_prune_locals = false;
866   uint stack_slots_not_pruned = 0;
867   int inputs = 0, depth = 0;
868   if (must_throw) {
869     assert(method() == youngest_jvms->method(), "sanity");
870     if (compute_stack_effects(inputs, depth)) {
871       can_prune_locals = true;
872       stack_slots_not_pruned = inputs;
873     }
874   }
875 
876   if (env()->should_retain_local_variables()) {
877     // At any safepoint, this method can get breakpointed, which would
878     // then require an immediate deoptimization.
879     can_prune_locals = false;  // do not prune locals
880     stack_slots_not_pruned = 0;
881   }
882 
883   // do not scribble on the input jvms
884   JVMState* out_jvms = youngest_jvms->clone_deep(C);
885   call->set_jvms(out_jvms); // Start jvms list for call node
886 
887   // For a known set of bytecodes, the interpreter should reexecute them if
888   // deoptimization happens. We set the reexecute state for them here
889   if (out_jvms->is_reexecute_undefined() && //don't change if already specified
890       should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
891     out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
892   }
893 
894   // Presize the call:
895   DEBUG_ONLY(uint non_debug_edges = call->req());
896   call->add_req_batch(top(), youngest_jvms->debug_depth());
897   assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
898 
899   // Set up edges so that the call looks like this:
900   //  Call [state:] ctl io mem fptr retadr
901   //       [parms:] parm0 ... parmN
902   //       [root:]  loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
903   //    [...mid:]   loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
904   //       [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
905   // Note that caller debug info precedes callee debug info.
906 
907   // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
908   uint debug_ptr = call->req();
909 
910   // Loop over the map input edges associated with jvms, add them
911   // to the call node, & reset all offsets to match call node array.
912   for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
913     uint debug_end   = debug_ptr;
914     uint debug_start = debug_ptr - in_jvms->debug_size();
915     debug_ptr = debug_start;  // back up the ptr
916 
917     uint p = debug_start;  // walks forward in [debug_start, debug_end)
918     uint j, k, l;
919     SafePointNode* in_map = in_jvms->map();
920     out_jvms->set_map(call);
921 
922     if (can_prune_locals) {
923       assert(in_jvms->method() == out_jvms->method(), "sanity");
924       // If the current throw can reach an exception handler in this JVMS,
925       // then we must keep everything live that can reach that handler.
926       // As a quick and dirty approximation, we look for any handlers at all.
927       if (in_jvms->method()->has_exception_handlers()) {
928         can_prune_locals = false;
929       }
930     }
931 
932     // Add the Locals
933     k = in_jvms->locoff();
934     l = in_jvms->loc_size();
935     out_jvms->set_locoff(p);
936     if (!can_prune_locals) {
937       for (j = 0; j < l; j++)
938         call->set_req(p++, in_map->in(k+j));
939     } else {
940       p += l;  // already set to top above by add_req_batch
941     }
942 
943     // Add the Expression Stack
944     k = in_jvms->stkoff();
945     l = in_jvms->sp();
946     out_jvms->set_stkoff(p);
947     if (!can_prune_locals) {
948       for (j = 0; j < l; j++)
949         call->set_req(p++, in_map->in(k+j));
950     } else if (can_prune_locals && stack_slots_not_pruned != 0) {
951       // Divide stack into {S0,...,S1}, where S0 is set to top.
952       uint s1 = stack_slots_not_pruned;
953       stack_slots_not_pruned = 0;  // for next iteration
954       if (s1 > l)  s1 = l;
955       uint s0 = l - s1;
956       p += s0;  // skip the tops preinstalled by add_req_batch
957       for (j = s0; j < l; j++)
958         call->set_req(p++, in_map->in(k+j));
959     } else {
960       p += l;  // already set to top above by add_req_batch
961     }
962 
963     // Add the Monitors
964     k = in_jvms->monoff();
965     l = in_jvms->mon_size();
966     out_jvms->set_monoff(p);
967     for (j = 0; j < l; j++)
968       call->set_req(p++, in_map->in(k+j));
969 
970     // Copy any scalar object fields.
971     k = in_jvms->scloff();
972     l = in_jvms->scl_size();
973     out_jvms->set_scloff(p);
974     for (j = 0; j < l; j++)
975       call->set_req(p++, in_map->in(k+j));
976 
977     // Finish the new jvms.
978     out_jvms->set_endoff(p);
979 
980     assert(out_jvms->endoff()     == debug_end,             "fill ptr must match");
981     assert(out_jvms->depth()      == in_jvms->depth(),      "depth must match");
982     assert(out_jvms->loc_size()   == in_jvms->loc_size(),   "size must match");
983     assert(out_jvms->mon_size()   == in_jvms->mon_size(),   "size must match");
984     assert(out_jvms->scl_size()   == in_jvms->scl_size(),   "size must match");
985     assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
986 
987     // Update the two tail pointers in parallel.
988     out_jvms = out_jvms->caller();
989     in_jvms  = in_jvms->caller();
990   }
991 
992   assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
993 
994   // Test the correctness of JVMState::debug_xxx accessors:
995   assert(call->jvms()->debug_start() == non_debug_edges, "");
996   assert(call->jvms()->debug_end()   == call->req(), "");
997   assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
998 }
999 
1000 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1001   Bytecodes::Code code = java_bc();
1002   if (code == Bytecodes::_wide) {
1003     code = method()->java_code_at_bci(bci() + 1);
1004   }
1005 
1006   BasicType rtype = T_ILLEGAL;
1007   int       rsize = 0;
1008 
1009   if (code != Bytecodes::_illegal) {
1010     depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1011     rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1012     if (rtype < T_CONFLICT)
1013       rsize = type2size[rtype];
1014   }
1015 
1016   switch (code) {
1017   case Bytecodes::_illegal:
1018     return false;
1019 
1020   case Bytecodes::_ldc:
1021   case Bytecodes::_ldc_w:
1022   case Bytecodes::_ldc2_w:
1023     inputs = 0;
1024     break;
1025 
1026   case Bytecodes::_dup:         inputs = 1;  break;
1027   case Bytecodes::_dup_x1:      inputs = 2;  break;
1028   case Bytecodes::_dup_x2:      inputs = 3;  break;
1029   case Bytecodes::_dup2:        inputs = 2;  break;
1030   case Bytecodes::_dup2_x1:     inputs = 3;  break;
1031   case Bytecodes::_dup2_x2:     inputs = 4;  break;
1032   case Bytecodes::_swap:        inputs = 2;  break;
1033   case Bytecodes::_arraylength: inputs = 1;  break;
1034 
1035   case Bytecodes::_getstatic:
1036   case Bytecodes::_putstatic:
1037   case Bytecodes::_getfield:
1038   case Bytecodes::_putfield:
1039     {
1040       bool ignored_will_link;
1041       ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1042       int      size  = field->type()->size();
1043       bool is_get = (depth >= 0), is_static = (depth & 1);
1044       inputs = (is_static ? 0 : 1);
1045       if (is_get) {
1046         depth = size - inputs;
1047       } else {
1048         inputs += size;        // putxxx pops the value from the stack
1049         depth = - inputs;
1050       }
1051     }
1052     break;
1053 
1054   case Bytecodes::_invokevirtual:
1055   case Bytecodes::_invokespecial:
1056   case Bytecodes::_invokestatic:
1057   case Bytecodes::_invokedynamic:
1058   case Bytecodes::_invokeinterface:
1059     {
1060       bool ignored_will_link;
1061       ciSignature* declared_signature = NULL;
1062       ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1063       assert(declared_signature != NULL, "cannot be null");
1064       inputs   = declared_signature->arg_size_for_bc(code);
1065       int size = declared_signature->return_type()->size();
1066       depth = size - inputs;
1067     }
1068     break;
1069 
1070   case Bytecodes::_multianewarray:
1071     {
1072       ciBytecodeStream iter(method());
1073       iter.reset_to_bci(bci());
1074       iter.next();
1075       inputs = iter.get_dimensions();
1076       assert(rsize == 1, "");
1077       depth = rsize - inputs;
1078     }
1079     break;
1080 
1081   case Bytecodes::_ireturn:
1082   case Bytecodes::_lreturn:
1083   case Bytecodes::_freturn:
1084   case Bytecodes::_dreturn:
1085   case Bytecodes::_areturn:
1086     assert(rsize == -depth, "");
1087     inputs = rsize;
1088     break;
1089 
1090   case Bytecodes::_jsr:
1091   case Bytecodes::_jsr_w:
1092     inputs = 0;
1093     depth  = 1;                  // S.B. depth=1, not zero
1094     break;
1095 
1096   default:
1097     // bytecode produces a typed result
1098     inputs = rsize - depth;
1099     assert(inputs >= 0, "");
1100     break;
1101   }
1102 
1103 #ifdef ASSERT
1104   // spot check
1105   int outputs = depth + inputs;
1106   assert(outputs >= 0, "sanity");
1107   switch (code) {
1108   case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1109   case Bytecodes::_athrow:    assert(inputs == 1 && outputs == 0, ""); break;
1110   case Bytecodes::_aload_0:   assert(inputs == 0 && outputs == 1, ""); break;
1111   case Bytecodes::_return:    assert(inputs == 0 && outputs == 0, ""); break;
1112   case Bytecodes::_drem:      assert(inputs == 4 && outputs == 2, ""); break;
1113   default:                    break;
1114   }
1115 #endif //ASSERT
1116 
1117   return true;
1118 }
1119 
1120 
1121 
1122 //------------------------------basic_plus_adr---------------------------------
1123 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1124   // short-circuit a common case
1125   if (offset == intcon(0))  return ptr;
1126   return _gvn.transform( new AddPNode(base, ptr, offset) );
1127 }
1128 
1129 Node* GraphKit::ConvI2L(Node* offset) {
1130   // short-circuit a common case
1131   jint offset_con = find_int_con(offset, Type::OffsetBot);
1132   if (offset_con != Type::OffsetBot) {
1133     return longcon((jlong) offset_con);
1134   }
1135   return _gvn.transform( new ConvI2LNode(offset));
1136 }
1137 
1138 Node* GraphKit::ConvI2UL(Node* offset) {
1139   juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1140   if (offset_con != (juint) Type::OffsetBot) {
1141     return longcon((julong) offset_con);
1142   }
1143   Node* conv = _gvn.transform( new ConvI2LNode(offset));
1144   Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1145   return _gvn.transform( new AndLNode(conv, mask) );
1146 }
1147 
1148 Node* GraphKit::ConvL2I(Node* offset) {
1149   // short-circuit a common case
1150   jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1151   if (offset_con != (jlong)Type::OffsetBot) {
1152     return intcon((int) offset_con);
1153   }
1154   return _gvn.transform( new ConvL2INode(offset));
1155 }
1156 
1157 //-------------------------load_object_klass-----------------------------------
1158 Node* GraphKit::load_object_klass(Node* obj) {
1159   // Special-case a fresh allocation to avoid building nodes:
1160   Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1161   if (akls != NULL)  return akls;
1162   Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1163   return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1164 }
1165 
1166 //-------------------------load_array_length-----------------------------------
1167 Node* GraphKit::load_array_length(Node* array) {
1168   // Special-case a fresh allocation to avoid building nodes:
1169   AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1170   Node *alen;
1171   if (alloc == NULL) {
1172     Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1173     alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1174   } else {
1175     alen = alloc->Ideal_length();
1176     Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1177     if (ccast != alen) {
1178       alen = _gvn.transform(ccast);
1179     }
1180   }
1181   return alen;
1182 }
1183 
1184 //------------------------------do_null_check----------------------------------
1185 // Helper function to do a NULL pointer check.  Returned value is
1186 // the incoming address with NULL casted away.  You are allowed to use the
1187 // not-null value only if you are control dependent on the test.
1188 #ifndef PRODUCT
1189 extern int explicit_null_checks_inserted,
1190            explicit_null_checks_elided;
1191 #endif
1192 Node* GraphKit::null_check_common(Node* value, BasicType type,
1193                                   // optional arguments for variations:
1194                                   bool assert_null,
1195                                   Node* *null_control,
1196                                   bool speculative) {
1197   assert(!assert_null || null_control == NULL, "not both at once");
1198   if (stopped())  return top();
1199   NOT_PRODUCT(explicit_null_checks_inserted++);
1200 
1201   // Construct NULL check
1202   Node *chk = NULL;
1203   switch(type) {
1204     case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1205     case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;
1206     case T_ARRAY  : // fall through
1207       type = T_OBJECT;  // simplify further tests
1208     case T_OBJECT : {
1209       const Type *t = _gvn.type( value );
1210 
1211       const TypeOopPtr* tp = t->isa_oopptr();
1212       if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1213           // Only for do_null_check, not any of its siblings:
1214           && !assert_null && null_control == NULL) {
1215         // Usually, any field access or invocation on an unloaded oop type
1216         // will simply fail to link, since the statically linked class is
1217         // likely also to be unloaded.  However, in -Xcomp mode, sometimes
1218         // the static class is loaded but the sharper oop type is not.
1219         // Rather than checking for this obscure case in lots of places,
1220         // we simply observe that a null check on an unloaded class
1221         // will always be followed by a nonsense operation, so we
1222         // can just issue the uncommon trap here.
1223         // Our access to the unloaded class will only be correct
1224         // after it has been loaded and initialized, which requires
1225         // a trip through the interpreter.
1226 #ifndef PRODUCT
1227         if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1228 #endif
1229         uncommon_trap(Deoptimization::Reason_unloaded,
1230                       Deoptimization::Action_reinterpret,
1231                       tp->klass(), "!loaded");
1232         return top();
1233       }
1234 
1235       if (assert_null) {
1236         // See if the type is contained in NULL_PTR.
1237         // If so, then the value is already null.
1238         if (t->higher_equal(TypePtr::NULL_PTR)) {
1239           NOT_PRODUCT(explicit_null_checks_elided++);
1240           return value;           // Elided null assert quickly!
1241         }
1242       } else {
1243         // See if mixing in the NULL pointer changes type.
1244         // If so, then the NULL pointer was not allowed in the original
1245         // type.  In other words, "value" was not-null.
1246         if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1247           // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1248           NOT_PRODUCT(explicit_null_checks_elided++);
1249           return value;           // Elided null check quickly!
1250         }
1251       }
1252       chk = new CmpPNode( value, null() );
1253       break;
1254     }
1255 
1256     default:
1257       fatal("unexpected type: %s", type2name(type));
1258   }
1259   assert(chk != NULL, "sanity check");
1260   chk = _gvn.transform(chk);
1261 
1262   BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1263   BoolNode *btst = new BoolNode( chk, btest);
1264   Node   *tst = _gvn.transform( btst );
1265 
1266   //-----------
1267   // if peephole optimizations occurred, a prior test existed.
1268   // If a prior test existed, maybe it dominates as we can avoid this test.
1269   if (tst != btst && type == T_OBJECT) {
1270     // At this point we want to scan up the CFG to see if we can
1271     // find an identical test (and so avoid this test altogether).
1272     Node *cfg = control();
1273     int depth = 0;
1274     while( depth < 16 ) {       // Limit search depth for speed
1275       if( cfg->Opcode() == Op_IfTrue &&
1276           cfg->in(0)->in(1) == tst ) {
1277         // Found prior test.  Use "cast_not_null" to construct an identical
1278         // CastPP (and hence hash to) as already exists for the prior test.
1279         // Return that casted value.
1280         if (assert_null) {
1281           replace_in_map(value, null());
1282           return null();  // do not issue the redundant test
1283         }
1284         Node *oldcontrol = control();
1285         set_control(cfg);
1286         Node *res = cast_not_null(value);
1287         set_control(oldcontrol);
1288         NOT_PRODUCT(explicit_null_checks_elided++);
1289         return res;
1290       }
1291       cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1292       if (cfg == NULL)  break;  // Quit at region nodes
1293       depth++;
1294     }
1295   }
1296 
1297   //-----------
1298   // Branch to failure if null
1299   float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
1300   Deoptimization::DeoptReason reason;
1301   if (assert_null) {
1302     reason = Deoptimization::reason_null_assert(speculative);
1303   } else if (type == T_OBJECT) {
1304     reason = Deoptimization::reason_null_check(speculative);
1305   } else {
1306     reason = Deoptimization::Reason_div0_check;
1307   }
1308   // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1309   // ciMethodData::has_trap_at will return a conservative -1 if any
1310   // must-be-null assertion has failed.  This could cause performance
1311   // problems for a method after its first do_null_assert failure.
1312   // Consider using 'Reason_class_check' instead?
1313 
1314   // To cause an implicit null check, we set the not-null probability
1315   // to the maximum (PROB_MAX).  For an explicit check the probability
1316   // is set to a smaller value.
1317   if (null_control != NULL || too_many_traps(reason)) {
1318     // probability is less likely
1319     ok_prob =  PROB_LIKELY_MAG(3);
1320   } else if (!assert_null &&
1321              (ImplicitNullCheckThreshold > 0) &&
1322              method() != NULL &&
1323              (method()->method_data()->trap_count(reason)
1324               >= (uint)ImplicitNullCheckThreshold)) {
1325     ok_prob =  PROB_LIKELY_MAG(3);
1326   }
1327 
1328   if (null_control != NULL) {
1329     IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1330     Node* null_true = _gvn.transform( new IfFalseNode(iff));
1331     set_control(      _gvn.transform( new IfTrueNode(iff)));
1332 #ifndef PRODUCT
1333     if (null_true == top()) {
1334       explicit_null_checks_elided++;
1335     }
1336 #endif
1337     (*null_control) = null_true;
1338   } else {
1339     BuildCutout unless(this, tst, ok_prob);
1340     // Check for optimizer eliding test at parse time
1341     if (stopped()) {
1342       // Failure not possible; do not bother making uncommon trap.
1343       NOT_PRODUCT(explicit_null_checks_elided++);
1344     } else if (assert_null) {
1345       uncommon_trap(reason,
1346                     Deoptimization::Action_make_not_entrant,
1347                     NULL, "assert_null");
1348     } else {
1349       replace_in_map(value, zerocon(type));
1350       builtin_throw(reason);
1351     }
1352   }
1353 
1354   // Must throw exception, fall-thru not possible?
1355   if (stopped()) {
1356     return top();               // No result
1357   }
1358 
1359   if (assert_null) {
1360     // Cast obj to null on this path.
1361     replace_in_map(value, zerocon(type));
1362     return zerocon(type);
1363   }
1364 
1365   // Cast obj to not-null on this path, if there is no null_control.
1366   // (If there is a null_control, a non-null value may come back to haunt us.)
1367   if (type == T_OBJECT) {
1368     Node* cast = cast_not_null(value, false);
1369     if (null_control == NULL || (*null_control) == top())
1370       replace_in_map(value, cast);
1371     value = cast;
1372   }
1373 
1374   return value;
1375 }
1376 
1377 
1378 //------------------------------cast_not_null----------------------------------
1379 // Cast obj to not-null on this path
1380 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1381   const Type *t = _gvn.type(obj);
1382   const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1383   // Object is already not-null?
1384   if( t == t_not_null ) return obj;
1385 
1386   Node *cast = new CastPPNode(obj,t_not_null);
1387   cast->init_req(0, control());
1388   cast = _gvn.transform( cast );
1389 
1390   // Scan for instances of 'obj' in the current JVM mapping.
1391   // These instances are known to be not-null after the test.
1392   if (do_replace_in_map)
1393     replace_in_map(obj, cast);
1394 
1395   return cast;                  // Return casted value
1396 }
1397 
1398 // Sometimes in intrinsics, we implicitly know an object is not null
1399 // (there's no actual null check) so we can cast it to not null. In
1400 // the course of optimizations, the input to the cast can become null.
1401 // In that case that data path will die and we need the control path
1402 // to become dead as well to keep the graph consistent. So we have to
1403 // add a check for null for which one branch can't be taken. It uses
1404 // an Opaque4 node that will cause the check to be removed after loop
1405 // opts so the test goes away and the compiled code doesn't execute a
1406 // useless check.
1407 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1408   Node* chk = _gvn.transform(new CmpPNode(value, null()));
1409   Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1410   Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1411   IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1412   _gvn.set_type(iff, iff->Value(&_gvn));
1413   Node *if_f = _gvn.transform(new IfFalseNode(iff));
1414   Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1415   Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1416   C->root()->add_req(halt);
1417   Node *if_t = _gvn.transform(new IfTrueNode(iff));
1418   set_control(if_t);
1419   return cast_not_null(value, do_replace_in_map);
1420 }
1421 
1422 
1423 //--------------------------replace_in_map-------------------------------------
1424 void GraphKit::replace_in_map(Node* old, Node* neww) {
1425   if (old == neww) {
1426     return;
1427   }
1428 
1429   map()->replace_edge(old, neww);
1430 
1431   // Note: This operation potentially replaces any edge
1432   // on the map.  This includes locals, stack, and monitors
1433   // of the current (innermost) JVM state.
1434 
1435   // don't let inconsistent types from profiling escape this
1436   // method
1437 
1438   const Type* told = _gvn.type(old);
1439   const Type* tnew = _gvn.type(neww);
1440 
1441   if (!tnew->higher_equal(told)) {
1442     return;
1443   }
1444 
1445   map()->record_replaced_node(old, neww);
1446 }
1447 
1448 
1449 //=============================================================================
1450 //--------------------------------memory---------------------------------------
1451 Node* GraphKit::memory(uint alias_idx) {
1452   MergeMemNode* mem = merged_memory();
1453   Node* p = mem->memory_at(alias_idx);
1454   _gvn.set_type(p, Type::MEMORY);  // must be mapped
1455   return p;
1456 }
1457 
1458 //-----------------------------reset_memory------------------------------------
1459 Node* GraphKit::reset_memory() {
1460   Node* mem = map()->memory();
1461   // do not use this node for any more parsing!
1462   debug_only( map()->set_memory((Node*)NULL) );
1463   return _gvn.transform( mem );
1464 }
1465 
1466 //------------------------------set_all_memory---------------------------------
1467 void GraphKit::set_all_memory(Node* newmem) {
1468   Node* mergemem = MergeMemNode::make(newmem);
1469   gvn().set_type_bottom(mergemem);
1470   map()->set_memory(mergemem);
1471 }
1472 
1473 //------------------------------set_all_memory_call----------------------------
1474 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1475   Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1476   set_all_memory(newmem);
1477 }
1478 
1479 //=============================================================================
1480 //
1481 // parser factory methods for MemNodes
1482 //
1483 // These are layered on top of the factory methods in LoadNode and StoreNode,
1484 // and integrate with the parser's memory state and _gvn engine.
1485 //
1486 
1487 // factory methods in "int adr_idx"
1488 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1489                           int adr_idx,
1490                           MemNode::MemOrd mo,
1491                           LoadNode::ControlDependency control_dependency,
1492                           bool require_atomic_access,
1493                           bool unaligned,
1494                           bool mismatched) {
1495   assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1496   const TypePtr* adr_type = NULL; // debug-mode-only argument
1497   debug_only(adr_type = C->get_adr_type(adr_idx));
1498   Node* mem = memory(adr_idx);
1499   Node* ld;
1500   if (require_atomic_access && bt == T_LONG) {
1501     ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1502   } else if (require_atomic_access && bt == T_DOUBLE) {
1503     ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1504   } else {
1505     ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1506   }
1507   ld = _gvn.transform(ld);
1508   if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1509     // Improve graph before escape analysis and boxing elimination.
1510     record_for_igvn(ld);
1511   }
1512   return ld;
1513 }
1514 
1515 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1516                                 int adr_idx,
1517                                 MemNode::MemOrd mo,
1518                                 bool require_atomic_access,
1519                                 bool unaligned,
1520                                 bool mismatched) {
1521   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1522   const TypePtr* adr_type = NULL;
1523   debug_only(adr_type = C->get_adr_type(adr_idx));
1524   Node *mem = memory(adr_idx);
1525   Node* st;
1526   if (require_atomic_access && bt == T_LONG) {
1527     st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1528   } else if (require_atomic_access && bt == T_DOUBLE) {
1529     st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1530   } else {
1531     st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1532   }
1533   if (unaligned) {
1534     st->as_Store()->set_unaligned_access();
1535   }
1536   if (mismatched) {
1537     st->as_Store()->set_mismatched_access();
1538   }
1539   st = _gvn.transform(st);
1540   set_memory(st, adr_idx);
1541   // Back-to-back stores can only remove intermediate store with DU info
1542   // so push on worklist for optimizer.
1543   if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1544     record_for_igvn(st);
1545 
1546   return st;
1547 }
1548 
1549 
1550 void GraphKit::pre_barrier(bool do_load,
1551                            Node* ctl,
1552                            Node* obj,
1553                            Node* adr,
1554                            uint  adr_idx,
1555                            Node* val,
1556                            const TypeOopPtr* val_type,
1557                            Node* pre_val,
1558                            BasicType bt) {
1559 
1560   BarrierSet* bs = Universe::heap()->barrier_set();
1561   set_control(ctl);
1562   switch (bs->kind()) {
1563     case BarrierSet::G1BarrierSet:
1564       g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1565       break;
1566 
1567     case BarrierSet::CardTableModRef:
1568       break;
1569 
1570     default      :
1571       ShouldNotReachHere();
1572 
1573   }
1574 }
1575 
1576 bool GraphKit::can_move_pre_barrier() const {
1577   BarrierSet* bs = Universe::heap()->barrier_set();
1578   switch (bs->kind()) {
1579     case BarrierSet::G1BarrierSet:
1580       return true; // Can move it if no safepoint
1581 
1582     case BarrierSet::CardTableModRef:
1583       return true; // There is no pre-barrier
1584 
1585     default      :
1586       ShouldNotReachHere();
1587   }
1588   return false;
1589 }
1590 
1591 void GraphKit::post_barrier(Node* ctl,
1592                             Node* store,
1593                             Node* obj,
1594                             Node* adr,
1595                             uint  adr_idx,
1596                             Node* val,
1597                             BasicType bt,
1598                             bool use_precise) {
1599   BarrierSet* bs = Universe::heap()->barrier_set();
1600   set_control(ctl);
1601   switch (bs->kind()) {
1602     case BarrierSet::G1BarrierSet:
1603       g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1604       break;
1605 
1606     case BarrierSet::CardTableModRef:
1607       write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1608       break;
1609 
1610     default      :
1611       ShouldNotReachHere();
1612 
1613   }
1614 }
1615 
1616 Node* GraphKit::store_oop(Node* ctl,
1617                           Node* obj,
1618                           Node* adr,
1619                           const TypePtr* adr_type,
1620                           Node* val,
1621                           const TypeOopPtr* val_type,
1622                           BasicType bt,
1623                           bool use_precise,
1624                           MemNode::MemOrd mo,
1625                           bool mismatched) {
1626   // Transformation of a value which could be NULL pointer (CastPP #NULL)
1627   // could be delayed during Parse (for example, in adjust_map_after_if()).
1628   // Execute transformation here to avoid barrier generation in such case.
1629   if (_gvn.type(val) == TypePtr::NULL_PTR)
1630     val = _gvn.makecon(TypePtr::NULL_PTR);
1631 
1632   set_control(ctl);
1633   if (stopped()) return top(); // Dead path ?
1634 
1635   assert(bt == T_OBJECT, "sanity");
1636   assert(val != NULL, "not dead path");
1637   uint adr_idx = C->get_alias_index(adr_type);
1638   assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1639 
1640   pre_barrier(true /* do_load */,
1641               control(), obj, adr, adr_idx, val, val_type,
1642               NULL /* pre_val */,
1643               bt);
1644 
1645   Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1646   post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1647   return store;
1648 }
1649 
1650 // Could be an array or object we don't know at compile time (unsafe ref.)
1651 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1652                              Node* obj,   // containing obj
1653                              Node* adr,  // actual adress to store val at
1654                              const TypePtr* adr_type,
1655                              Node* val,
1656                              BasicType bt,
1657                              MemNode::MemOrd mo,
1658                              bool mismatched) {
1659   Compile::AliasType* at = C->alias_type(adr_type);
1660   const TypeOopPtr* val_type = NULL;
1661   if (adr_type->isa_instptr()) {
1662     if (at->field() != NULL) {
1663       // known field.  This code is a copy of the do_put_xxx logic.
1664       ciField* field = at->field();
1665       if (!field->type()->is_loaded()) {
1666         val_type = TypeInstPtr::BOTTOM;
1667       } else {
1668         val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1669       }
1670     }
1671   } else if (adr_type->isa_aryptr()) {
1672     val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1673   }
1674   if (val_type == NULL) {
1675     val_type = TypeInstPtr::BOTTOM;
1676   }
1677   return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1678 }
1679 
1680 
1681 //-------------------------array_element_address-------------------------
1682 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1683                                       const TypeInt* sizetype, Node* ctrl) {
1684   uint shift  = exact_log2(type2aelembytes(elembt));
1685   uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1686 
1687   // short-circuit a common case (saves lots of confusing waste motion)
1688   jint idx_con = find_int_con(idx, -1);
1689   if (idx_con >= 0) {
1690     intptr_t offset = header + ((intptr_t)idx_con << shift);
1691     return basic_plus_adr(ary, offset);
1692   }
1693 
1694   // must be correct type for alignment purposes
1695   Node* base  = basic_plus_adr(ary, header);
1696   idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1697   Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1698   return basic_plus_adr(ary, base, scale);
1699 }
1700 
1701 //-------------------------load_array_element-------------------------
1702 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1703   const Type* elemtype = arytype->elem();
1704   BasicType elembt = elemtype->array_element_basic_type();
1705   Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1706   if (elembt == T_NARROWOOP) {
1707     elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1708   }
1709   Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1710   return ld;
1711 }
1712 
1713 //-------------------------set_arguments_for_java_call-------------------------
1714 // Arguments (pre-popped from the stack) are taken from the JVMS.
1715 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1716   // Add the call arguments:
1717   uint nargs = call->method()->arg_size();
1718   for (uint i = 0; i < nargs; i++) {
1719     Node* arg = argument(i);
1720     call->init_req(i + TypeFunc::Parms, arg);
1721   }
1722 }
1723 
1724 //---------------------------set_edges_for_java_call---------------------------
1725 // Connect a newly created call into the current JVMS.
1726 // A return value node (if any) is returned from set_edges_for_java_call.
1727 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1728 
1729   // Add the predefined inputs:
1730   call->init_req( TypeFunc::Control, control() );
1731   call->init_req( TypeFunc::I_O    , i_o() );
1732   call->init_req( TypeFunc::Memory , reset_memory() );
1733   call->init_req( TypeFunc::FramePtr, frameptr() );
1734   call->init_req( TypeFunc::ReturnAdr, top() );
1735 
1736   add_safepoint_edges(call, must_throw);
1737 
1738   Node* xcall = _gvn.transform(call);
1739 
1740   if (xcall == top()) {
1741     set_control(top());
1742     return;
1743   }
1744   assert(xcall == call, "call identity is stable");
1745 
1746   // Re-use the current map to produce the result.
1747 
1748   set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1749   set_i_o(    _gvn.transform(new ProjNode(call, TypeFunc::I_O    , separate_io_proj)));
1750   set_all_memory_call(xcall, separate_io_proj);
1751 
1752   //return xcall;   // no need, caller already has it
1753 }
1754 
1755 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1756   if (stopped())  return top();  // maybe the call folded up?
1757 
1758   // Capture the return value, if any.
1759   Node* ret;
1760   if (call->method() == NULL ||
1761       call->method()->return_type()->basic_type() == T_VOID)
1762         ret = top();
1763   else  ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1764 
1765   // Note:  Since any out-of-line call can produce an exception,
1766   // we always insert an I_O projection from the call into the result.
1767 
1768   make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1769 
1770   if (separate_io_proj) {
1771     // The caller requested separate projections be used by the fall
1772     // through and exceptional paths, so replace the projections for
1773     // the fall through path.
1774     set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1775     set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1776   }
1777   return ret;
1778 }
1779 
1780 //--------------------set_predefined_input_for_runtime_call--------------------
1781 // Reading and setting the memory state is way conservative here.
1782 // The real problem is that I am not doing real Type analysis on memory,
1783 // so I cannot distinguish card mark stores from other stores.  Across a GC
1784 // point the Store Barrier and the card mark memory has to agree.  I cannot
1785 // have a card mark store and its barrier split across the GC point from
1786 // either above or below.  Here I get that to happen by reading ALL of memory.
1787 // A better answer would be to separate out card marks from other memory.
1788 // For now, return the input memory state, so that it can be reused
1789 // after the call, if this call has restricted memory effects.
1790 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1791   // Set fixed predefined input arguments
1792   Node* memory = reset_memory();
1793   call->init_req( TypeFunc::Control,   control()  );
1794   call->init_req( TypeFunc::I_O,       top()      ); // does no i/o
1795   call->init_req( TypeFunc::Memory,    memory     ); // may gc ptrs
1796   call->init_req( TypeFunc::FramePtr,  frameptr() );
1797   call->init_req( TypeFunc::ReturnAdr, top()      );
1798   return memory;
1799 }
1800 
1801 //-------------------set_predefined_output_for_runtime_call--------------------
1802 // Set control and memory (not i_o) from the call.
1803 // If keep_mem is not NULL, use it for the output state,
1804 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1805 // If hook_mem is NULL, this call produces no memory effects at all.
1806 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1807 // then only that memory slice is taken from the call.
1808 // In the last case, we must put an appropriate memory barrier before
1809 // the call, so as to create the correct anti-dependencies on loads
1810 // preceding the call.
1811 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1812                                                       Node* keep_mem,
1813                                                       const TypePtr* hook_mem) {
1814   // no i/o
1815   set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1816   if (keep_mem) {
1817     // First clone the existing memory state
1818     set_all_memory(keep_mem);
1819     if (hook_mem != NULL) {
1820       // Make memory for the call
1821       Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1822       // Set the RawPtr memory state only.  This covers all the heap top/GC stuff
1823       // We also use hook_mem to extract specific effects from arraycopy stubs.
1824       set_memory(mem, hook_mem);
1825     }
1826     // ...else the call has NO memory effects.
1827 
1828     // Make sure the call advertises its memory effects precisely.
1829     // This lets us build accurate anti-dependences in gcm.cpp.
1830     assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1831            "call node must be constructed correctly");
1832   } else {
1833     assert(hook_mem == NULL, "");
1834     // This is not a "slow path" call; all memory comes from the call.
1835     set_all_memory_call(call);
1836   }
1837 }
1838 
1839 
1840 // Replace the call with the current state of the kit.
1841 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1842   JVMState* ejvms = NULL;
1843   if (has_exceptions()) {
1844     ejvms = transfer_exceptions_into_jvms();
1845   }
1846 
1847   ReplacedNodes replaced_nodes = map()->replaced_nodes();
1848   ReplacedNodes replaced_nodes_exception;
1849   Node* ex_ctl = top();
1850 
1851   SafePointNode* final_state = stop();
1852 
1853   // Find all the needed outputs of this call
1854   CallProjections callprojs;
1855   call->extract_projections(&callprojs, true);
1856 
1857   Node* init_mem = call->in(TypeFunc::Memory);
1858   Node* final_mem = final_state->in(TypeFunc::Memory);
1859   Node* final_ctl = final_state->in(TypeFunc::Control);
1860   Node* final_io = final_state->in(TypeFunc::I_O);
1861 
1862   // Replace all the old call edges with the edges from the inlining result
1863   if (callprojs.fallthrough_catchproj != NULL) {
1864     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1865   }
1866   if (callprojs.fallthrough_memproj != NULL) {
1867     if (final_mem->is_MergeMem()) {
1868       // Parser's exits MergeMem was not transformed but may be optimized
1869       final_mem = _gvn.transform(final_mem);
1870     }
1871     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
1872   }
1873   if (callprojs.fallthrough_ioproj != NULL) {
1874     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
1875   }
1876 
1877   // Replace the result with the new result if it exists and is used
1878   if (callprojs.resproj != NULL && result != NULL) {
1879     C->gvn_replace_by(callprojs.resproj, result);
1880   }
1881 
1882   if (ejvms == NULL) {
1883     // No exception edges to simply kill off those paths
1884     if (callprojs.catchall_catchproj != NULL) {
1885       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1886     }
1887     if (callprojs.catchall_memproj != NULL) {
1888       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
1889     }
1890     if (callprojs.catchall_ioproj != NULL) {
1891       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
1892     }
1893     // Replace the old exception object with top
1894     if (callprojs.exobj != NULL) {
1895       C->gvn_replace_by(callprojs.exobj, C->top());
1896     }
1897   } else {
1898     GraphKit ekit(ejvms);
1899 
1900     // Load my combined exception state into the kit, with all phis transformed:
1901     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1902     replaced_nodes_exception = ex_map->replaced_nodes();
1903 
1904     Node* ex_oop = ekit.use_exception_state(ex_map);
1905 
1906     if (callprojs.catchall_catchproj != NULL) {
1907       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1908       ex_ctl = ekit.control();
1909     }
1910     if (callprojs.catchall_memproj != NULL) {
1911       C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
1912     }
1913     if (callprojs.catchall_ioproj != NULL) {
1914       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
1915     }
1916 
1917     // Replace the old exception object with the newly created one
1918     if (callprojs.exobj != NULL) {
1919       C->gvn_replace_by(callprojs.exobj, ex_oop);
1920     }
1921   }
1922 
1923   // Disconnect the call from the graph
1924   call->disconnect_inputs(NULL, C);
1925   C->gvn_replace_by(call, C->top());
1926 
1927   // Clean up any MergeMems that feed other MergeMems since the
1928   // optimizer doesn't like that.
1929   if (final_mem->is_MergeMem()) {
1930     Node_List wl;
1931     for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1932       Node* m = i.get();
1933       if (m->is_MergeMem() && !wl.contains(m)) {
1934         wl.push(m);
1935       }
1936     }
1937     while (wl.size()  > 0) {
1938       _gvn.transform(wl.pop());
1939     }
1940   }
1941 
1942   if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1943     replaced_nodes.apply(C, final_ctl);
1944   }
1945   if (!ex_ctl->is_top() && do_replaced_nodes) {
1946     replaced_nodes_exception.apply(C, ex_ctl);
1947   }
1948 }
1949 
1950 
1951 //------------------------------increment_counter------------------------------
1952 // for statistics: increment a VM counter by 1
1953 
1954 void GraphKit::increment_counter(address counter_addr) {
1955   Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1956   increment_counter(adr1);
1957 }
1958 
1959 void GraphKit::increment_counter(Node* counter_addr) {
1960   int adr_type = Compile::AliasIdxRaw;
1961   Node* ctrl = control();
1962   Node* cnt  = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1963   Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
1964   store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1965 }
1966 
1967 
1968 //------------------------------uncommon_trap----------------------------------
1969 // Bail out to the interpreter in mid-method.  Implemented by calling the
1970 // uncommon_trap blob.  This helper function inserts a runtime call with the
1971 // right debug info.
1972 void GraphKit::uncommon_trap(int trap_request,
1973                              ciKlass* klass, const char* comment,
1974                              bool must_throw,
1975                              bool keep_exact_action) {
1976   if (failing())  stop();
1977   if (stopped())  return; // trap reachable?
1978 
1979   // Note:  If ProfileTraps is true, and if a deopt. actually
1980   // occurs here, the runtime will make sure an MDO exists.  There is
1981   // no need to call method()->ensure_method_data() at this point.
1982 
1983   // Set the stack pointer to the right value for reexecution:
1984   set_sp(reexecute_sp());
1985 
1986 #ifdef ASSERT
1987   if (!must_throw) {
1988     // Make sure the stack has at least enough depth to execute
1989     // the current bytecode.
1990     int inputs, ignored_depth;
1991     if (compute_stack_effects(inputs, ignored_depth)) {
1992       assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
1993              Bytecodes::name(java_bc()), sp(), inputs);
1994     }
1995   }
1996 #endif
1997 
1998   Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1999   Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2000 
2001   switch (action) {
2002   case Deoptimization::Action_maybe_recompile:
2003   case Deoptimization::Action_reinterpret:
2004     // Temporary fix for 6529811 to allow virtual calls to be sure they
2005     // get the chance to go from mono->bi->mega
2006     if (!keep_exact_action &&
2007         Deoptimization::trap_request_index(trap_request) < 0 &&
2008         too_many_recompiles(reason)) {
2009       // This BCI is causing too many recompilations.
2010       if (C->log() != NULL) {
2011         C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2012                 Deoptimization::trap_reason_name(reason),
2013                 Deoptimization::trap_action_name(action));
2014       }
2015       action = Deoptimization::Action_none;
2016       trap_request = Deoptimization::make_trap_request(reason, action);
2017     } else {
2018       C->set_trap_can_recompile(true);
2019     }
2020     break;
2021   case Deoptimization::Action_make_not_entrant:
2022     C->set_trap_can_recompile(true);
2023     break;
2024   case Deoptimization::Action_none:
2025   case Deoptimization::Action_make_not_compilable:
2026     break;
2027   default:
2028 #ifdef ASSERT
2029     fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2030 #endif
2031     break;
2032   }
2033 
2034   if (TraceOptoParse) {
2035     char buf[100];
2036     tty->print_cr("Uncommon trap %s at bci:%d",
2037                   Deoptimization::format_trap_request(buf, sizeof(buf),
2038                                                       trap_request), bci());
2039   }
2040 
2041   CompileLog* log = C->log();
2042   if (log != NULL) {
2043     int kid = (klass == NULL)? -1: log->identify(klass);
2044     log->begin_elem("uncommon_trap bci='%d'", bci());
2045     char buf[100];
2046     log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2047                                                           trap_request));
2048     if (kid >= 0)         log->print(" klass='%d'", kid);
2049     if (comment != NULL)  log->print(" comment='%s'", comment);
2050     log->end_elem();
2051   }
2052 
2053   // Make sure any guarding test views this path as very unlikely
2054   Node *i0 = control()->in(0);
2055   if (i0 != NULL && i0->is_If()) {        // Found a guarding if test?
2056     IfNode *iff = i0->as_If();
2057     float f = iff->_prob;   // Get prob
2058     if (control()->Opcode() == Op_IfTrue) {
2059       if (f > PROB_UNLIKELY_MAG(4))
2060         iff->_prob = PROB_MIN;
2061     } else {
2062       if (f < PROB_LIKELY_MAG(4))
2063         iff->_prob = PROB_MAX;
2064     }
2065   }
2066 
2067   // Clear out dead values from the debug info.
2068   kill_dead_locals();
2069 
2070   // Now insert the uncommon trap subroutine call
2071   address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2072   const TypePtr* no_memory_effects = NULL;
2073   // Pass the index of the class to be loaded
2074   Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2075                                  (must_throw ? RC_MUST_THROW : 0),
2076                                  OptoRuntime::uncommon_trap_Type(),
2077                                  call_addr, "uncommon_trap", no_memory_effects,
2078                                  intcon(trap_request));
2079   assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2080          "must extract request correctly from the graph");
2081   assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2082 
2083   call->set_req(TypeFunc::ReturnAdr, returnadr());
2084   // The debug info is the only real input to this call.
2085 
2086   // Halt-and-catch fire here.  The above call should never return!
2087   HaltNode* halt = new HaltNode(control(), frameptr());
2088   _gvn.set_type_bottom(halt);
2089   root()->add_req(halt);
2090 
2091   stop_and_kill_map();
2092 }
2093 
2094 
2095 //--------------------------just_allocated_object------------------------------
2096 // Report the object that was just allocated.
2097 // It must be the case that there are no intervening safepoints.
2098 // We use this to determine if an object is so "fresh" that
2099 // it does not require card marks.
2100 Node* GraphKit::just_allocated_object(Node* current_control) {
2101   if (C->recent_alloc_ctl() == current_control)
2102     return C->recent_alloc_obj();
2103   return NULL;
2104 }
2105 
2106 
2107 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2108   // (Note:  TypeFunc::make has a cache that makes this fast.)
2109   const TypeFunc* tf    = TypeFunc::make(dest_method);
2110   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2111   for (int j = 0; j < nargs; j++) {
2112     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2113     if( targ->basic_type() == T_DOUBLE ) {
2114       // If any parameters are doubles, they must be rounded before
2115       // the call, dstore_rounding does gvn.transform
2116       Node *arg = argument(j);
2117       arg = dstore_rounding(arg);
2118       set_argument(j, arg);
2119     }
2120   }
2121 }
2122 
2123 /**
2124  * Record profiling data exact_kls for Node n with the type system so
2125  * that it can propagate it (speculation)
2126  *
2127  * @param n          node that the type applies to
2128  * @param exact_kls  type from profiling
2129  * @param maybe_null did profiling see null?
2130  *
2131  * @return           node with improved type
2132  */
2133 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2134   const Type* current_type = _gvn.type(n);
2135   assert(UseTypeSpeculation, "type speculation must be on");
2136 
2137   const TypePtr* speculative = current_type->speculative();
2138 
2139   // Should the klass from the profile be recorded in the speculative type?
2140   if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2141     const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2142     const TypeOopPtr* xtype = tklass->as_instance_type();
2143     assert(xtype->klass_is_exact(), "Should be exact");
2144     // Any reason to believe n is not null (from this profiling or a previous one)?
2145     assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2146     const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNUL
2147     // record the new speculative type's depth
2148     speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2149     speculative = speculative->with_inline_depth(jvms()->depth());
2150   } else if (current_type->would_improve_ptr(ptr_kind)) {
2151     // Profiling report that null was never seen so we can change the
2152     // speculative type to non null ptr.
2153     if (ptr_kind == ProfileAlwaysNull) {
2154       speculative = TypePtr::NULL_PTR;
2155     } else {
2156       assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2157       const TypePtr* ptr = TypePtr::NOTNULL;
2158       if (speculative != NULL) {
2159         speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2160       } else {
2161         speculative = ptr;
2162       }
2163     }
2164   }
2165 
2166   if (speculative != current_type->speculative()) {
2167     // Build a type with a speculative type (what we think we know
2168     // about the type but will need a guard when we use it)
2169     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2170     // We're changing the type, we need a new CheckCast node to carry
2171     // the new type. The new type depends on the control: what
2172     // profiling tells us is only valid from here as far as we can
2173     // tell.
2174     Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2175     cast = _gvn.transform(cast);
2176     replace_in_map(n, cast);
2177     n = cast;
2178   }
2179 
2180   return n;
2181 }
2182 
2183 /**
2184  * Record profiling data from receiver profiling at an invoke with the
2185  * type system so that it can propagate it (speculation)
2186  *
2187  * @param n  receiver node
2188  *
2189  * @return   node with improved type
2190  */
2191 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2192   if (!UseTypeSpeculation) {
2193     return n;
2194   }
2195   ciKlass* exact_kls = profile_has_unique_klass();
2196   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2197   if ((java_bc() == Bytecodes::_checkcast ||
2198        java_bc() == Bytecodes::_instanceof ||
2199        java_bc() == Bytecodes::_aastore) &&
2200       method()->method_data()->is_mature()) {
2201     ciProfileData* data = method()->method_data()->bci_to_data(bci());
2202     if (data != NULL) {
2203       if (!data->as_BitData()->null_seen()) {
2204         ptr_kind = ProfileNeverNull;
2205       } else {
2206         assert(data->is_ReceiverTypeData(), "bad profile data type");
2207         ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2208         uint i = 0;
2209         for (; i < call->row_limit(); i++) {
2210           ciKlass* receiver = call->receiver(i);
2211           if (receiver != NULL) {
2212             break;
2213           }
2214         }
2215         ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2216       }
2217     }
2218   }
2219   return record_profile_for_speculation(n, exact_kls, ptr_kind);
2220 }
2221 
2222 /**
2223  * Record profiling data from argument profiling at an invoke with the
2224  * type system so that it can propagate it (speculation)
2225  *
2226  * @param dest_method  target method for the call
2227  * @param bc           what invoke bytecode is this?
2228  */
2229 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2230   if (!UseTypeSpeculation) {
2231     return;
2232   }
2233   const TypeFunc* tf    = TypeFunc::make(dest_method);
2234   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
2235   int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2236   for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2237     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2238     if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2239       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2240       ciKlass* better_type = NULL;
2241       if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2242         record_profile_for_speculation(argument(j), better_type, ptr_kind);
2243       }
2244       i++;
2245     }
2246   }
2247 }
2248 
2249 /**
2250  * Record profiling data from parameter profiling at an invoke with
2251  * the type system so that it can propagate it (speculation)
2252  */
2253 void GraphKit::record_profiled_parameters_for_speculation() {
2254   if (!UseTypeSpeculation) {
2255     return;
2256   }
2257   for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2258     if (_gvn.type(local(i))->isa_oopptr()) {
2259       ProfilePtrKind ptr_kind = ProfileMaybeNull;
2260       ciKlass* better_type = NULL;
2261       if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2262         record_profile_for_speculation(local(i), better_type, ptr_kind);
2263       }
2264       j++;
2265     }
2266   }
2267 }
2268 
2269 /**
2270  * Record profiling data from return value profiling at an invoke with
2271  * the type system so that it can propagate it (speculation)
2272  */
2273 void GraphKit::record_profiled_return_for_speculation() {
2274   if (!UseTypeSpeculation) {
2275     return;
2276   }
2277   ProfilePtrKind ptr_kind = ProfileMaybeNull;
2278   ciKlass* better_type = NULL;
2279   if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2280     // If profiling reports a single type for the return value,
2281     // feed it to the type system so it can propagate it as a
2282     // speculative type
2283     record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2284   }
2285 }
2286 
2287 void GraphKit::round_double_result(ciMethod* dest_method) {
2288   // A non-strict method may return a double value which has an extended
2289   // exponent, but this must not be visible in a caller which is 'strict'
2290   // If a strict caller invokes a non-strict callee, round a double result
2291 
2292   BasicType result_type = dest_method->return_type()->basic_type();
2293   assert( method() != NULL, "must have caller context");
2294   if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2295     // Destination method's return value is on top of stack
2296     // dstore_rounding() does gvn.transform
2297     Node *result = pop_pair();
2298     result = dstore_rounding(result);
2299     push_pair(result);
2300   }
2301 }
2302 
2303 // rounding for strict float precision conformance
2304 Node* GraphKit::precision_rounding(Node* n) {
2305   return UseStrictFP && _method->flags().is_strict()
2306     && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2307     ? _gvn.transform( new RoundFloatNode(0, n) )
2308     : n;
2309 }
2310 
2311 // rounding for strict double precision conformance
2312 Node* GraphKit::dprecision_rounding(Node *n) {
2313   return UseStrictFP && _method->flags().is_strict()
2314     && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2315     ? _gvn.transform( new RoundDoubleNode(0, n) )
2316     : n;
2317 }
2318 
2319 // rounding for non-strict double stores
2320 Node* GraphKit::dstore_rounding(Node* n) {
2321   return Matcher::strict_fp_requires_explicit_rounding
2322     && UseSSE <= 1
2323     ? _gvn.transform( new RoundDoubleNode(0, n) )
2324     : n;
2325 }
2326 
2327 //=============================================================================
2328 // Generate a fast path/slow path idiom.  Graph looks like:
2329 // [foo] indicates that 'foo' is a parameter
2330 //
2331 //              [in]     NULL
2332 //                 \    /
2333 //                  CmpP
2334 //                  Bool ne
2335 //                   If
2336 //                  /  \
2337 //              True    False-<2>
2338 //              / |
2339 //             /  cast_not_null
2340 //           Load  |    |   ^
2341 //        [fast_test]   |   |
2342 // gvn to   opt_test    |   |
2343 //          /    \      |  <1>
2344 //      True     False  |
2345 //        |         \\  |
2346 //   [slow_call]     \[fast_result]
2347 //    Ctl   Val       \      \
2348 //     |               \      \
2349 //    Catch       <1>   \      \
2350 //   /    \        ^     \      \
2351 //  Ex    No_Ex    |      \      \
2352 //  |       \   \  |       \ <2>  \
2353 //  ...      \  [slow_res] |  |    \   [null_result]
2354 //            \         \--+--+---  |  |
2355 //             \           | /    \ | /
2356 //              --------Region     Phi
2357 //
2358 //=============================================================================
2359 // Code is structured as a series of driver functions all called 'do_XXX' that
2360 // call a set of helper functions.  Helper functions first, then drivers.
2361 
2362 //------------------------------null_check_oop---------------------------------
2363 // Null check oop.  Set null-path control into Region in slot 3.
2364 // Make a cast-not-nullness use the other not-null control.  Return cast.
2365 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2366                                bool never_see_null,
2367                                bool safe_for_replace,
2368                                bool speculative) {
2369   // Initial NULL check taken path
2370   (*null_control) = top();
2371   Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2372 
2373   // Generate uncommon_trap:
2374   if (never_see_null && (*null_control) != top()) {
2375     // If we see an unexpected null at a check-cast we record it and force a
2376     // recompile; the offending check-cast will be compiled to handle NULLs.
2377     // If we see more than one offending BCI, then all checkcasts in the
2378     // method will be compiled to handle NULLs.
2379     PreserveJVMState pjvms(this);
2380     set_control(*null_control);
2381     replace_in_map(value, null());
2382     Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2383     uncommon_trap(reason,
2384                   Deoptimization::Action_make_not_entrant);
2385     (*null_control) = top();    // NULL path is dead
2386   }
2387   if ((*null_control) == top() && safe_for_replace) {
2388     replace_in_map(value, cast);
2389   }
2390 
2391   // Cast away null-ness on the result
2392   return cast;
2393 }
2394 
2395 //------------------------------opt_iff----------------------------------------
2396 // Optimize the fast-check IfNode.  Set the fast-path region slot 2.
2397 // Return slow-path control.
2398 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2399   IfNode *opt_iff = _gvn.transform(iff)->as_If();
2400 
2401   // Fast path taken; set region slot 2
2402   Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2403   region->init_req(2,fast_taken); // Capture fast-control
2404 
2405   // Fast path not-taken, i.e. slow path
2406   Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2407   return slow_taken;
2408 }
2409 
2410 //-----------------------------make_runtime_call-------------------------------
2411 Node* GraphKit::make_runtime_call(int flags,
2412                                   const TypeFunc* call_type, address call_addr,
2413                                   const char* call_name,
2414                                   const TypePtr* adr_type,
2415                                   // The following parms are all optional.
2416                                   // The first NULL ends the list.
2417                                   Node* parm0, Node* parm1,
2418                                   Node* parm2, Node* parm3,
2419                                   Node* parm4, Node* parm5,
2420                                   Node* parm6, Node* parm7) {
2421   // Slow-path call
2422   bool is_leaf = !(flags & RC_NO_LEAF);
2423   bool has_io  = (!is_leaf && !(flags & RC_NO_IO));
2424   if (call_name == NULL) {
2425     assert(!is_leaf, "must supply name for leaf");
2426     call_name = OptoRuntime::stub_name(call_addr);
2427   }
2428   CallNode* call;
2429   if (!is_leaf) {
2430     call = new CallStaticJavaNode(call_type, call_addr, call_name,
2431                                            bci(), adr_type);
2432   } else if (flags & RC_NO_FP) {
2433     call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2434   } else {
2435     call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2436   }
2437 
2438   // The following is similar to set_edges_for_java_call,
2439   // except that the memory effects of the call are restricted to AliasIdxRaw.
2440 
2441   // Slow path call has no side-effects, uses few values
2442   bool wide_in  = !(flags & RC_NARROW_MEM);
2443   bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2444 
2445   Node* prev_mem = NULL;
2446   if (wide_in) {
2447     prev_mem = set_predefined_input_for_runtime_call(call);
2448   } else {
2449     assert(!wide_out, "narrow in => narrow out");
2450     Node* narrow_mem = memory(adr_type);
2451     prev_mem = reset_memory();
2452     map()->set_memory(narrow_mem);
2453     set_predefined_input_for_runtime_call(call);
2454   }
2455 
2456   // Hook each parm in order.  Stop looking at the first NULL.
2457   if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2458   if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2459   if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2460   if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2461   if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2462   if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2463   if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2464   if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2465     /* close each nested if ===> */  } } } } } } } }
2466   assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2467 
2468   if (!is_leaf) {
2469     // Non-leaves can block and take safepoints:
2470     add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2471   }
2472   // Non-leaves can throw exceptions:
2473   if (has_io) {
2474     call->set_req(TypeFunc::I_O, i_o());
2475   }
2476 
2477   if (flags & RC_UNCOMMON) {
2478     // Set the count to a tiny probability.  Cf. Estimate_Block_Frequency.
2479     // (An "if" probability corresponds roughly to an unconditional count.
2480     // Sort of.)
2481     call->set_cnt(PROB_UNLIKELY_MAG(4));
2482   }
2483 
2484   Node* c = _gvn.transform(call);
2485   assert(c == call, "cannot disappear");
2486 
2487   if (wide_out) {
2488     // Slow path call has full side-effects.
2489     set_predefined_output_for_runtime_call(call);
2490   } else {
2491     // Slow path call has few side-effects, and/or sets few values.
2492     set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2493   }
2494 
2495   if (has_io) {
2496     set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2497   }
2498   return call;
2499 
2500 }
2501 
2502 //------------------------------merge_memory-----------------------------------
2503 // Merge memory from one path into the current memory state.
2504 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2505   for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2506     Node* old_slice = mms.force_memory();
2507     Node* new_slice = mms.memory2();
2508     if (old_slice != new_slice) {
2509       PhiNode* phi;
2510       if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2511         if (mms.is_empty()) {
2512           // clone base memory Phi's inputs for this memory slice
2513           assert(old_slice == mms.base_memory(), "sanity");
2514           phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2515           _gvn.set_type(phi, Type::MEMORY);
2516           for (uint i = 1; i < phi->req(); i++) {
2517             phi->init_req(i, old_slice->in(i));
2518           }
2519         } else {
2520           phi = old_slice->as_Phi(); // Phi was generated already
2521         }
2522       } else {
2523         phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2524         _gvn.set_type(phi, Type::MEMORY);
2525       }
2526       phi->set_req(new_path, new_slice);
2527       mms.set_memory(phi);
2528     }
2529   }
2530 }
2531 
2532 //------------------------------make_slow_call_ex------------------------------
2533 // Make the exception handler hookups for the slow call
2534 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2535   if (stopped())  return;
2536 
2537   // Make a catch node with just two handlers:  fall-through and catch-all
2538   Node* i_o  = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2539   Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2540   Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2541   Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index,    CatchProjNode::no_handler_bci) );
2542 
2543   { PreserveJVMState pjvms(this);
2544     set_control(excp);
2545     set_i_o(i_o);
2546 
2547     if (excp != top()) {
2548       if (deoptimize) {
2549         // Deoptimize if an exception is caught. Don't construct exception state in this case.
2550         uncommon_trap(Deoptimization::Reason_unhandled,
2551                       Deoptimization::Action_none);
2552       } else {
2553         // Create an exception state also.
2554         // Use an exact type if the caller has specified a specific exception.
2555         const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2556         Node*       ex_oop  = new CreateExNode(ex_type, control(), i_o);
2557         add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2558       }
2559     }
2560   }
2561 
2562   // Get the no-exception control from the CatchNode.
2563   set_control(norm);
2564 }
2565 
2566 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt)
2567   Node* cmp = NULL;
2568   switch(bt) {
2569   case T_INT: cmp = new CmpINode(in1, in2); break;
2570   case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2571   default: fatal("unexpected comparison type %s", type2name(bt));
2572   }
2573   gvn->transform(cmp);
2574   Node* bol = gvn->transform(new BoolNode(cmp, test));
2575   IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2576   gvn->transform(iff);
2577   if (!bol->is_Con()) gvn->record_for_igvn(iff);
2578   return iff;
2579 }
2580 
2581 
2582 //-------------------------------gen_subtype_check-----------------------------
2583 // Generate a subtyping check.  Takes as input the subtype and supertype.
2584 // Returns 2 values: sets the default control() to the true path and returns
2585 // the false path.  Only reads invariant memory; sets no (visible) memory.
2586 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2587 // but that's not exposed to the optimizer.  This call also doesn't take in an
2588 // Object; if you wish to check an Object you need to load the Object's class
2589 // prior to coming here.
2590 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2591   Compile* C = gvn->C;
2592 
2593   if ((*ctrl)->is_top()) {
2594     return C->top();
2595   }
2596 
2597   // Fast check for identical types, perhaps identical constants.
2598   // The types can even be identical non-constants, in cases
2599   // involving Array.newInstance, Object.clone, etc.
2600   if (subklass == superklass)
2601     return C->top();             // false path is dead; no test needed.
2602 
2603   if (gvn->type(superklass)->singleton()) {
2604     ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2605     ciKlass* subk   = gvn->type(subklass)->is_klassptr()->klass();
2606 
2607     // In the common case of an exact superklass, try to fold up the
2608     // test before generating code.  You may ask, why not just generate
2609     // the code and then let it fold up?  The answer is that the generated
2610     // code will necessarily include null checks, which do not always
2611     // completely fold away.  If they are also needless, then they turn
2612     // into a performance loss.  Example:
2613     //    Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2614     // Here, the type of 'fa' is often exact, so the store check
2615     // of fa[1]=x will fold up, without testing the nullness of x.
2616     switch (C->static_subtype_check(superk, subk)) {
2617     case Compile::SSC_always_false:
2618       {
2619         Node* always_fail = *ctrl;
2620         *ctrl = gvn->C->top();
2621         return always_fail;
2622       }
2623     case Compile::SSC_always_true:
2624       return C->top();
2625     case Compile::SSC_easy_test:
2626       {
2627         // Just do a direct pointer compare and be done.
2628         IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2629         *ctrl = gvn->transform(new IfTrueNode(iff));
2630         return gvn->transform(new IfFalseNode(iff));
2631       }
2632     case Compile::SSC_full_test:
2633       break;
2634     default:
2635       ShouldNotReachHere();
2636     }
2637   }
2638 
2639   // %%% Possible further optimization:  Even if the superklass is not exact,
2640   // if the subklass is the unique subtype of the superklass, the check
2641   // will always succeed.  We could leave a dependency behind to ensure this.
2642 
2643   // First load the super-klass's check-offset
2644   Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2645   Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2646   Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2647   int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2648   bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2649 
2650   // Load from the sub-klass's super-class display list, or a 1-word cache of
2651   // the secondary superclass list, or a failing value with a sentinel offset
2652   // if the super-klass is an interface or exceptionally deep in the Java
2653   // hierarchy and we have to scan the secondary superclass list the hard way.
2654   // Worst-case type is a little odd: NULL is allowed as a result (usually
2655   // klass loads can never produce a NULL).
2656   Node *chk_off_X = chk_off;
2657 #ifdef _LP64
2658   chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2659 #endif
2660   Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2661   // For some types like interfaces the following loadKlass is from a 1-word
2662   // cache which is mutable so can't use immutable memory.  Other
2663   // types load from the super-class display table which is immutable.
2664   m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2665   Node *kmem = might_be_cache ? m : C->immutable_memory();
2666   Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2667 
2668   // Compile speed common case: ARE a subtype and we canNOT fail
2669   if( superklass == nkls )
2670     return C->top();             // false path is dead; no test needed.
2671 
2672   // See if we get an immediate positive hit.  Happens roughly 83% of the
2673   // time.  Test to see if the value loaded just previously from the subklass
2674   // is exactly the superklass.
2675   IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2676   Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2677   *ctrl = gvn->transform(new IfFalseNode(iff1));
2678 
2679   // Compile speed common case: Check for being deterministic right now.  If
2680   // chk_off is a constant and not equal to cacheoff then we are NOT a
2681   // subklass.  In this case we need exactly the 1 test above and we can
2682   // return those results immediately.
2683   if (!might_be_cache) {
2684     Node* not_subtype_ctrl = *ctrl;
2685     *ctrl = iftrue1; // We need exactly the 1 test above
2686     return not_subtype_ctrl;
2687   }
2688 
2689   // Gather the various success & failures here
2690   RegionNode *r_ok_subtype = new RegionNode(4);
2691   gvn->record_for_igvn(r_ok_subtype);
2692   RegionNode *r_not_subtype = new RegionNode(3);
2693   gvn->record_for_igvn(r_not_subtype);
2694 
2695   r_ok_subtype->init_req(1, iftrue1);
2696 
2697   // Check for immediate negative hit.  Happens roughly 11% of the time (which
2698   // is roughly 63% of the remaining cases).  Test to see if the loaded
2699   // check-offset points into the subklass display list or the 1-element
2700   // cache.  If it points to the display (and NOT the cache) and the display
2701   // missed then it's not a subtype.
2702   Node *cacheoff = gvn->intcon(cacheoff_con);
2703   IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2704   r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2705   *ctrl = gvn->transform(new IfFalseNode(iff2));
2706 
2707   // Check for self.  Very rare to get here, but it is taken 1/3 the time.
2708   // No performance impact (too rare) but allows sharing of secondary arrays
2709   // which has some footprint reduction.
2710   IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2711   r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2712   *ctrl = gvn->transform(new IfFalseNode(iff3));
2713 
2714   // -- Roads not taken here: --
2715   // We could also have chosen to perform the self-check at the beginning
2716   // of this code sequence, as the assembler does.  This would not pay off
2717   // the same way, since the optimizer, unlike the assembler, can perform
2718   // static type analysis to fold away many successful self-checks.
2719   // Non-foldable self checks work better here in second position, because
2720   // the initial primary superclass check subsumes a self-check for most
2721   // types.  An exception would be a secondary type like array-of-interface,
2722   // which does not appear in its own primary supertype display.
2723   // Finally, we could have chosen to move the self-check into the
2724   // PartialSubtypeCheckNode, and from there out-of-line in a platform
2725   // dependent manner.  But it is worthwhile to have the check here,
2726   // where it can be perhaps be optimized.  The cost in code space is
2727   // small (register compare, branch).
2728 
2729   // Now do a linear scan of the secondary super-klass array.  Again, no real
2730   // performance impact (too rare) but it's gotta be done.
2731   // Since the code is rarely used, there is no penalty for moving it
2732   // out of line, and it can only improve I-cache density.
2733   // The decision to inline or out-of-line this final check is platform
2734   // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2735   Node* psc = gvn->transform(
2736     new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2737 
2738   IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2739   r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2740   r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2741 
2742   // Return false path; set default control to true path.
2743   *ctrl = gvn->transform(r_ok_subtype);
2744   return gvn->transform(r_not_subtype);
2745 }
2746 
2747 // Profile-driven exact type check:
2748 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2749                                     float prob,
2750                                     Node* *casted_receiver) {
2751   const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2752   Node* recv_klass = load_object_klass(receiver);
2753   Node* want_klass = makecon(tklass);
2754   Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2755   Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2756   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2757   set_control( _gvn.transform( new IfTrueNode (iff) ));
2758   Node* fail = _gvn.transform( new IfFalseNode(iff) );
2759 
2760   const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2761   assert(recv_xtype->klass_is_exact(), "");
2762 
2763   // Subsume downstream occurrences of receiver with a cast to
2764   // recv_xtype, since now we know what the type will be.
2765   Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2766   (*casted_receiver) = _gvn.transform(cast);
2767   // (User must make the replace_in_map call.)
2768 
2769   return fail;
2770 }
2771 
2772 
2773 //------------------------------seems_never_null-------------------------------
2774 // Use null_seen information if it is available from the profile.
2775 // If we see an unexpected null at a type check we record it and force a
2776 // recompile; the offending check will be recompiled to handle NULLs.
2777 // If we see several offending BCIs, then all checks in the
2778 // method will be recompiled.
2779 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2780   speculating = !_gvn.type(obj)->speculative_maybe_null();
2781   Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2782   if (UncommonNullCast               // Cutout for this technique
2783       && obj != null()               // And not the -Xcomp stupid case?
2784       && !too_many_traps(reason)
2785       ) {
2786     if (speculating) {
2787       return true;
2788     }
2789     if (data == NULL)
2790       // Edge case:  no mature data.  Be optimistic here.
2791       return true;
2792     // If the profile has not seen a null, assume it won't happen.
2793     assert(java_bc() == Bytecodes::_checkcast ||
2794            java_bc() == Bytecodes::_instanceof ||
2795            java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2796     return !data->as_BitData()->null_seen();
2797   }
2798   speculating = false;
2799   return false;
2800 }
2801 
2802 //------------------------maybe_cast_profiled_receiver-------------------------
2803 // If the profile has seen exactly one type, narrow to exactly that type.
2804 // Subsequent type checks will always fold up.
2805 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2806                                              ciKlass* require_klass,
2807                                              ciKlass* spec_klass,
2808                                              bool safe_for_replace) {
2809   if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2810 
2811   Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2812 
2813   // Make sure we haven't already deoptimized from this tactic.
2814   if (too_many_traps(reason) || too_many_recompiles(reason))
2815     return NULL;
2816 
2817   // (No, this isn't a call, but it's enough like a virtual call
2818   // to use the same ciMethod accessor to get the profile info...)
2819   // If we have a speculative type use it instead of profiling (which
2820   // may not help us)
2821   ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2822   if (exact_kls != NULL) {// no cast failures here
2823     if (require_klass == NULL ||
2824         C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2825       // If we narrow the type to match what the type profile sees or
2826       // the speculative type, we can then remove the rest of the
2827       // cast.
2828       // This is a win, even if the exact_kls is very specific,
2829       // because downstream operations, such as method calls,
2830       // will often benefit from the sharper type.
2831       Node* exact_obj = not_null_obj; // will get updated in place...
2832       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2833                                             &exact_obj);
2834       { PreserveJVMState pjvms(this);
2835         set_control(slow_ctl);
2836         uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2837       }
2838       if (safe_for_replace) {
2839         replace_in_map(not_null_obj, exact_obj);
2840       }
2841       return exact_obj;
2842     }
2843     // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2844   }
2845 
2846   return NULL;
2847 }
2848 
2849 /**
2850  * Cast obj to type and emit guard unless we had too many traps here
2851  * already
2852  *
2853  * @param obj       node being casted
2854  * @param type      type to cast the node to
2855  * @param not_null  true if we know node cannot be null
2856  */
2857 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2858                                         ciKlass* type,
2859                                         bool not_null) {
2860   if (stopped()) {
2861     return obj;
2862   }
2863 
2864   // type == NULL if profiling tells us this object is always null
2865   if (type != NULL) {
2866     Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2867     Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2868 
2869     if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2870         !too_many_traps(class_reason) &&
2871         !too_many_recompiles(class_reason)) {
2872       Node* not_null_obj = NULL;
2873       // not_null is true if we know the object is not null and
2874       // there's no need for a null check
2875       if (!not_null) {
2876         Node* null_ctl = top();
2877         not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2878         assert(null_ctl->is_top(), "no null control here");
2879       } else {
2880         not_null_obj = obj;
2881       }
2882 
2883       Node* exact_obj = not_null_obj;
2884       ciKlass* exact_kls = type;
2885       Node* slow_ctl  = type_check_receiver(exact_obj, exact_kls, 1.0,
2886                                             &exact_obj);
2887       {
2888         PreserveJVMState pjvms(this);
2889         set_control(slow_ctl);
2890         uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2891       }
2892       replace_in_map(not_null_obj, exact_obj);
2893       obj = exact_obj;
2894     }
2895   } else {
2896     if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2897         !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2898       Node* exact_obj = null_assert(obj);
2899       replace_in_map(obj, exact_obj);
2900       obj = exact_obj;
2901     }
2902   }
2903   return obj;
2904 }
2905 
2906 //-------------------------------gen_instanceof--------------------------------
2907 // Generate an instance-of idiom.  Used by both the instance-of bytecode
2908 // and the reflective instance-of call.
2909 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2910   kill_dead_locals();           // Benefit all the uncommon traps
2911   assert( !stopped(), "dead parse path should be checked in callers" );
2912   assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2913          "must check for not-null not-dead klass in callers");
2914 
2915   // Make the merge point
2916   enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2917   RegionNode* region = new RegionNode(PATH_LIMIT);
2918   Node*       phi    = new PhiNode(region, TypeInt::BOOL);
2919   C->set_has_split_ifs(true); // Has chance for split-if optimization
2920 
2921   ciProfileData* data = NULL;
2922   if (java_bc() == Bytecodes::_instanceof) {  // Only for the bytecode
2923     data = method()->method_data()->bci_to_data(bci());
2924   }
2925   bool speculative_not_null = false;
2926   bool never_see_null = (ProfileDynamicTypes  // aggressive use of profile
2927                          && seems_never_null(obj, data, speculative_not_null));
2928 
2929   // Null check; get casted pointer; set region slot 3
2930   Node* null_ctl = top();
2931   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2932 
2933   // If not_null_obj is dead, only null-path is taken
2934   if (stopped()) {              // Doing instance-of on a NULL?
2935     set_control(null_ctl);
2936     return intcon(0);
2937   }
2938   region->init_req(_null_path, null_ctl);
2939   phi   ->init_req(_null_path, intcon(0)); // Set null path value
2940   if (null_ctl == top()) {
2941     // Do this eagerly, so that pattern matches like is_diamond_phi
2942     // will work even during parsing.
2943     assert(_null_path == PATH_LIMIT-1, "delete last");
2944     region->del_req(_null_path);
2945     phi   ->del_req(_null_path);
2946   }
2947 
2948   // Do we know the type check always succeed?
2949   bool known_statically = false;
2950   if (_gvn.type(superklass)->singleton()) {
2951     ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2952     ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2953     if (subk != NULL && subk->is_loaded()) {
2954       int static_res = C->static_subtype_check(superk, subk);
2955       known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
2956     }
2957   }
2958 
2959   if (!known_statically) {
2960     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2961     // We may not have profiling here or it may not help us. If we
2962     // have a speculative type use it to perform an exact cast.
2963     ciKlass* spec_obj_type = obj_type->speculative_type();
2964     if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2965       Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2966       if (stopped()) {            // Profile disagrees with this path.
2967         set_control(null_ctl);    // Null is the only remaining possibility.
2968         return intcon(0);
2969       }
2970       if (cast_obj != NULL) {
2971         not_null_obj = cast_obj;
2972       }
2973     }
2974   }
2975 
2976   // Load the object's klass
2977   Node* obj_klass = load_object_klass(not_null_obj);
2978 
2979   // Generate the subtype check
2980   Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2981 
2982   // Plug in the success path to the general merge in slot 1.
2983   region->init_req(_obj_path, control());
2984   phi   ->init_req(_obj_path, intcon(1));
2985 
2986   // Plug in the failing path to the general merge in slot 2.
2987   region->init_req(_fail_path, not_subtype_ctrl);
2988   phi   ->init_req(_fail_path, intcon(0));
2989 
2990   // Return final merged results
2991   set_control( _gvn.transform(region) );
2992   record_for_igvn(region);
2993 
2994   // If we know the type check always succeeds then we don't use the
2995   // profiling data at this bytecode. Don't lose it, feed it to the
2996   // type system as a speculative type.
2997   if (safe_for_replace) {
2998     Node* casted_obj = record_profiled_receiver_for_speculation(obj);
2999     replace_in_map(obj, casted_obj);
3000   }
3001 
3002   return _gvn.transform(phi);
3003 }
3004 
3005 //-------------------------------gen_checkcast---------------------------------
3006 // Generate a checkcast idiom.  Used by both the checkcast bytecode and the
3007 // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
3008 // uncommon-trap paths work.  Adjust stack after this call.
3009 // If failure_control is supplied and not null, it is filled in with
3010 // the control edge for the cast failure.  Otherwise, an appropriate
3011 // uncommon trap or exception is thrown.
3012 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3013                               Node* *failure_control) {
3014   kill_dead_locals();           // Benefit all the uncommon traps
3015   const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3016   const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3017 
3018   // Fast cutout:  Check the case that the cast is vacuously true.
3019   // This detects the common cases where the test will short-circuit
3020   // away completely.  We do this before we perform the null check,
3021   // because if the test is going to turn into zero code, we don't
3022   // want a residual null check left around.  (Causes a slowdown,
3023   // for example, in some objArray manipulations, such as a[i]=a[j].)
3024   if (tk->singleton()) {
3025     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3026     if (objtp != NULL && objtp->klass() != NULL) {
3027       switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3028       case Compile::SSC_always_true:
3029         // If we know the type check always succeed then we don't use
3030         // the profiling data at this bytecode. Don't lose it, feed it
3031         // to the type system as a speculative type.
3032         return record_profiled_receiver_for_speculation(obj);
3033       case Compile::SSC_always_false:
3034         // It needs a null check because a null will *pass* the cast check.
3035         // A non-null value will always produce an exception.
3036         return null_assert(obj);
3037       }
3038     }
3039   }
3040 
3041   ciProfileData* data = NULL;
3042   bool safe_for_replace = false;
3043   if (failure_control == NULL) {        // use MDO in regular case only
3044     assert(java_bc() == Bytecodes::_aastore ||
3045            java_bc() == Bytecodes::_checkcast,
3046            "interpreter profiles type checks only for these BCs");
3047     data = method()->method_data()->bci_to_data(bci());
3048     safe_for_replace = true;
3049   }
3050 
3051   // Make the merge point
3052   enum { _obj_path = 1, _null_path, PATH_LIMIT };
3053   RegionNode* region = new RegionNode(PATH_LIMIT);
3054   Node*       phi    = new PhiNode(region, toop);
3055   C->set_has_split_ifs(true); // Has chance for split-if optimization
3056 
3057   // Use null-cast information if it is available
3058   bool speculative_not_null = false;
3059   bool never_see_null = ((failure_control == NULL)  // regular case only
3060                          && seems_never_null(obj, data, speculative_not_null));
3061 
3062   // Null check; get casted pointer; set region slot 3
3063   Node* null_ctl = top();
3064   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3065 
3066   // If not_null_obj is dead, only null-path is taken
3067   if (stopped()) {              // Doing instance-of on a NULL?
3068     set_control(null_ctl);
3069     return null();
3070   }
3071   region->init_req(_null_path, null_ctl);
3072   phi   ->init_req(_null_path, null());  // Set null path value
3073   if (null_ctl == top()) {
3074     // Do this eagerly, so that pattern matches like is_diamond_phi
3075     // will work even during parsing.
3076     assert(_null_path == PATH_LIMIT-1, "delete last");
3077     region->del_req(_null_path);
3078     phi   ->del_req(_null_path);
3079   }
3080 
3081   Node* cast_obj = NULL;
3082   if (tk->klass_is_exact()) {
3083     // The following optimization tries to statically cast the speculative type of the object
3084     // (for example obtained during profiling) to the type of the superklass and then do a
3085     // dynamic check that the type of the object is what we expect. To work correctly
3086     // for checkcast and aastore the type of superklass should be exact.
3087     const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3088     // We may not have profiling here or it may not help us. If we have
3089     // a speculative type use it to perform an exact cast.
3090     ciKlass* spec_obj_type = obj_type->speculative_type();
3091     if (spec_obj_type != NULL || data != NULL) {
3092       cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3093       if (cast_obj != NULL) {
3094         if (failure_control != NULL) // failure is now impossible
3095           (*failure_control) = top();
3096         // adjust the type of the phi to the exact klass:
3097         phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3098       }
3099     }
3100   }
3101 
3102   if (cast_obj == NULL) {
3103     // Load the object's klass
3104     Node* obj_klass = load_object_klass(not_null_obj);
3105 
3106     // Generate the subtype check
3107     Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3108 
3109     // Plug in success path into the merge
3110     cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3111     // Failure path ends in uncommon trap (or may be dead - failure impossible)
3112     if (failure_control == NULL) {
3113       if (not_subtype_ctrl != top()) { // If failure is possible
3114         PreserveJVMState pjvms(this);
3115         set_control(not_subtype_ctrl);
3116         builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3117       }
3118     } else {
3119       (*failure_control) = not_subtype_ctrl;
3120     }
3121   }
3122 
3123   region->init_req(_obj_path, control());
3124   phi   ->init_req(_obj_path, cast_obj);
3125 
3126   // A merge of NULL or Casted-NotNull obj
3127   Node* res = _gvn.transform(phi);
3128 
3129   // Note I do NOT always 'replace_in_map(obj,result)' here.
3130   //  if( tk->klass()->can_be_primary_super()  )
3131     // This means that if I successfully store an Object into an array-of-String
3132     // I 'forget' that the Object is really now known to be a String.  I have to
3133     // do this because we don't have true union types for interfaces - if I store
3134     // a Baz into an array-of-Interface and then tell the optimizer it's an
3135     // Interface, I forget that it's also a Baz and cannot do Baz-like field
3136     // references to it.  FIX THIS WHEN UNION TYPES APPEAR!
3137   //  replace_in_map( obj, res );
3138 
3139   // Return final merged results
3140   set_control( _gvn.transform(region) );
3141   record_for_igvn(region);
3142 
3143   return record_profiled_receiver_for_speculation(res);
3144 }
3145 
3146 //------------------------------next_monitor-----------------------------------
3147 // What number should be given to the next monitor?
3148 int GraphKit::next_monitor() {
3149   int current = jvms()->monitor_depth()* C->sync_stack_slots();
3150   int next = current + C->sync_stack_slots();
3151   // Keep the toplevel high water mark current:
3152   if (C->fixed_slots() < next)  C->set_fixed_slots(next);
3153   return current;
3154 }
3155 
3156 //------------------------------insert_mem_bar---------------------------------
3157 // Memory barrier to avoid floating things around
3158 // The membar serves as a pinch point between both control and all memory slices.
3159 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3160   MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3161   mb->init_req(TypeFunc::Control, control());
3162   mb->init_req(TypeFunc::Memory,  reset_memory());
3163   Node* membar = _gvn.transform(mb);
3164   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3165   set_all_memory_call(membar);
3166   return membar;
3167 }
3168 
3169 //-------------------------insert_mem_bar_volatile----------------------------
3170 // Memory barrier to avoid floating things around
3171 // The membar serves as a pinch point between both control and memory(alias_idx).
3172 // If you want to make a pinch point on all memory slices, do not use this
3173 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3174 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3175   // When Parse::do_put_xxx updates a volatile field, it appends a series
3176   // of MemBarVolatile nodes, one for *each* volatile field alias category.
3177   // The first membar is on the same memory slice as the field store opcode.
3178   // This forces the membar to follow the store.  (Bug 6500685 broke this.)
3179   // All the other membars (for other volatile slices, including AliasIdxBot,
3180   // which stands for all unknown volatile slices) are control-dependent
3181   // on the first membar.  This prevents later volatile loads or stores
3182   // from sliding up past the just-emitted store.
3183 
3184   MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3185   mb->set_req(TypeFunc::Control,control());
3186   if (alias_idx == Compile::AliasIdxBot) {
3187     mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3188   } else {
3189     assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3190     mb->set_req(TypeFunc::Memory, memory(alias_idx));
3191   }
3192   Node* membar = _gvn.transform(mb);
3193   set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3194   if (alias_idx == Compile::AliasIdxBot) {
3195     merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3196   } else {
3197     set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3198   }
3199   return membar;
3200 }
3201 
3202 //------------------------------shared_lock------------------------------------
3203 // Emit locking code.
3204 FastLockNode* GraphKit::shared_lock(Node* obj) {
3205   // bci is either a monitorenter bc or InvocationEntryBci
3206   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3207   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3208 
3209   if( !GenerateSynchronizationCode )
3210     return NULL;                // Not locking things?
3211   if (stopped())                // Dead monitor?
3212     return NULL;
3213 
3214   assert(dead_locals_are_killed(), "should kill locals before sync. point");
3215 
3216   // Box the stack location
3217   Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3218   Node* mem = reset_memory();
3219 
3220   FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3221   if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3222     // Create the counters for this fast lock.
3223     flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3224   }
3225 
3226   // Create the rtm counters for this fast lock if needed.
3227   flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3228 
3229   // Add monitor to debug info for the slow path.  If we block inside the
3230   // slow path and de-opt, we need the monitor hanging around
3231   map()->push_monitor( flock );
3232 
3233   const TypeFunc *tf = LockNode::lock_type();
3234   LockNode *lock = new LockNode(C, tf);
3235 
3236   lock->init_req( TypeFunc::Control, control() );
3237   lock->init_req( TypeFunc::Memory , mem );
3238   lock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3239   lock->init_req( TypeFunc::FramePtr, frameptr() );
3240   lock->init_req( TypeFunc::ReturnAdr, top() );
3241 
3242   lock->init_req(TypeFunc::Parms + 0, obj);
3243   lock->init_req(TypeFunc::Parms + 1, box);
3244   lock->init_req(TypeFunc::Parms + 2, flock);
3245   add_safepoint_edges(lock);
3246 
3247   lock = _gvn.transform( lock )->as_Lock();
3248 
3249   // lock has no side-effects, sets few values
3250   set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3251 
3252   insert_mem_bar(Op_MemBarAcquireLock);
3253 
3254   // Add this to the worklist so that the lock can be eliminated
3255   record_for_igvn(lock);
3256 
3257 #ifndef PRODUCT
3258   if (PrintLockStatistics) {
3259     // Update the counter for this lock.  Don't bother using an atomic
3260     // operation since we don't require absolute accuracy.
3261     lock->create_lock_counter(map()->jvms());
3262     increment_counter(lock->counter()->addr());
3263   }
3264 #endif
3265 
3266   return flock;
3267 }
3268 
3269 
3270 //------------------------------shared_unlock----------------------------------
3271 // Emit unlocking code.
3272 void GraphKit::shared_unlock(Node* box, Node* obj) {
3273   // bci is either a monitorenter bc or InvocationEntryBci
3274   // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3275   assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3276 
3277   if( !GenerateSynchronizationCode )
3278     return;
3279   if (stopped()) {               // Dead monitor?
3280     map()->pop_monitor();        // Kill monitor from debug info
3281     return;
3282   }
3283 
3284   // Memory barrier to avoid floating things down past the locked region
3285   insert_mem_bar(Op_MemBarReleaseLock);
3286 
3287   const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3288   UnlockNode *unlock = new UnlockNode(C, tf);
3289 #ifdef ASSERT
3290   unlock->set_dbg_jvms(sync_jvms());
3291 #endif
3292   uint raw_idx = Compile::AliasIdxRaw;
3293   unlock->init_req( TypeFunc::Control, control() );
3294   unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3295   unlock->init_req( TypeFunc::I_O    , top() )     ;   // does no i/o
3296   unlock->init_req( TypeFunc::FramePtr, frameptr() );
3297   unlock->init_req( TypeFunc::ReturnAdr, top() );
3298 
3299   unlock->init_req(TypeFunc::Parms + 0, obj);
3300   unlock->init_req(TypeFunc::Parms + 1, box);
3301   unlock = _gvn.transform(unlock)->as_Unlock();
3302 
3303   Node* mem = reset_memory();
3304 
3305   // unlock has no side-effects, sets few values
3306   set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3307 
3308   // Kill monitor from debug info
3309   map()->pop_monitor( );
3310 }
3311 
3312 //-------------------------------get_layout_helper-----------------------------
3313 // If the given klass is a constant or known to be an array,
3314 // fetch the constant layout helper value into constant_value
3315 // and return (Node*)NULL.  Otherwise, load the non-constant
3316 // layout helper value, and return the node which represents it.
3317 // This two-faced routine is useful because allocation sites
3318 // almost always feature constant types.
3319 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3320   const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3321   if (!StressReflectiveCode && inst_klass != NULL) {
3322     ciKlass* klass = inst_klass->klass();
3323     bool    xklass = inst_klass->klass_is_exact();
3324     if (xklass || klass->is_array_klass()) {
3325       jint lhelper = klass->layout_helper();
3326       if (lhelper != Klass::_lh_neutral_value) {
3327         constant_value = lhelper;
3328         return (Node*) NULL;
3329       }
3330     }
3331   }
3332   constant_value = Klass::_lh_neutral_value;  // put in a known value
3333   Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3334   return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3335 }
3336 
3337 // We just put in an allocate/initialize with a big raw-memory effect.
3338 // Hook selected additional alias categories on the initialization.
3339 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3340                                 MergeMemNode* init_in_merge,
3341                                 Node* init_out_raw) {
3342   DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3343   assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3344 
3345   Node* prevmem = kit.memory(alias_idx);
3346   init_in_merge->set_memory_at(alias_idx, prevmem);
3347   kit.set_memory(init_out_raw, alias_idx);
3348 }
3349 
3350 //---------------------------set_output_for_allocation-------------------------
3351 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3352                                           const TypeOopPtr* oop_type,
3353                                           bool deoptimize_on_exception) {
3354   int rawidx = Compile::AliasIdxRaw;
3355   alloc->set_req( TypeFunc::FramePtr, frameptr() );
3356   add_safepoint_edges(alloc);
3357   Node* allocx = _gvn.transform(alloc);
3358   set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3359   // create memory projection for i_o
3360   set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3361   make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3362 
3363   // create a memory projection as for the normal control path
3364   Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3365   set_memory(malloc, rawidx);
3366 
3367   // a normal slow-call doesn't change i_o, but an allocation does
3368   // we create a separate i_o projection for the normal control path
3369   set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3370   Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3371 
3372   // put in an initialization barrier
3373   InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3374                                                  rawoop)->as_Initialize();
3375   assert(alloc->initialization() == init,  "2-way macro link must work");
3376   assert(init ->allocation()     == alloc, "2-way macro link must work");
3377   {
3378     // Extract memory strands which may participate in the new object's
3379     // initialization, and source them from the new InitializeNode.
3380     // This will allow us to observe initializations when they occur,
3381     // and link them properly (as a group) to the InitializeNode.
3382     assert(init->in(InitializeNode::Memory) == malloc, "");
3383     MergeMemNode* minit_in = MergeMemNode::make(malloc);
3384     init->set_req(InitializeNode::Memory, minit_in);
3385     record_for_igvn(minit_in); // fold it up later, if possible
3386     Node* minit_out = memory(rawidx);
3387     assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3388     if (oop_type->isa_aryptr()) {
3389       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3390       int            elemidx  = C->get_alias_index(telemref);
3391       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3392     } else if (oop_type->isa_instptr()) {
3393       ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3394       for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3395         ciField* field = ik->nonstatic_field_at(i);
3396         if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3397           continue;  // do not bother to track really large numbers of fields
3398         // Find (or create) the alias category for this field:
3399         int fieldidx = C->alias_type(field)->index();
3400         hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3401       }
3402     }
3403   }
3404 
3405   // Cast raw oop to the real thing...
3406   Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3407   javaoop = _gvn.transform(javaoop);
3408   C->set_recent_alloc(control(), javaoop);
3409   assert(just_allocated_object(control()) == javaoop, "just allocated");
3410 
3411 #ifdef ASSERT
3412   { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3413     assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3414            "Ideal_allocation works");
3415     assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3416            "Ideal_allocation works");
3417     if (alloc->is_AllocateArray()) {
3418       assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3419              "Ideal_allocation works");
3420       assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3421              "Ideal_allocation works");
3422     } else {
3423       assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3424     }
3425   }
3426 #endif //ASSERT
3427 
3428   return javaoop;
3429 }
3430 
3431 //---------------------------new_instance--------------------------------------
3432 // This routine takes a klass_node which may be constant (for a static type)
3433 // or may be non-constant (for reflective code).  It will work equally well
3434 // for either, and the graph will fold nicely if the optimizer later reduces
3435 // the type to a constant.
3436 // The optional arguments are for specialized use by intrinsics:
3437 //  - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3438 //  - If 'return_size_val', report the the total object size to the caller.
3439 //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3440 Node* GraphKit::new_instance(Node* klass_node,
3441                              Node* extra_slow_test,
3442                              Node* *return_size_val,
3443                              bool deoptimize_on_exception) {
3444   // Compute size in doublewords
3445   // The size is always an integral number of doublewords, represented
3446   // as a positive bytewise size stored in the klass's layout_helper.
3447   // The layout_helper also encodes (in a low bit) the need for a slow path.
3448   jint  layout_con = Klass::_lh_neutral_value;
3449   Node* layout_val = get_layout_helper(klass_node, layout_con);
3450   int   layout_is_con = (layout_val == NULL);
3451 
3452   if (extra_slow_test == NULL)  extra_slow_test = intcon(0);
3453   // Generate the initial go-slow test.  It's either ALWAYS (return a
3454   // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3455   // case) a computed value derived from the layout_helper.
3456   Node* initial_slow_test = NULL;
3457   if (layout_is_con) {
3458     assert(!StressReflectiveCode, "stress mode does not use these paths");
3459     bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3460     initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3461   } else {   // reflective case
3462     // This reflective path is used by Unsafe.allocateInstance.
3463     // (It may be stress-tested by specifying StressReflectiveCode.)
3464     // Basically, we want to get into the VM is there's an illegal argument.
3465     Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3466     initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3467     if (extra_slow_test != intcon(0)) {
3468       initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3469     }
3470     // (Macro-expander will further convert this to a Bool, if necessary.)
3471   }
3472 
3473   // Find the size in bytes.  This is easy; it's the layout_helper.
3474   // The size value must be valid even if the slow path is taken.
3475   Node* size = NULL;
3476   if (layout_is_con) {
3477     size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3478   } else {   // reflective case
3479     // This reflective path is used by clone and Unsafe.allocateInstance.
3480     size = ConvI2X(layout_val);
3481 
3482     // Clear the low bits to extract layout_helper_size_in_bytes:
3483     assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3484     Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3485     size = _gvn.transform( new AndXNode(size, mask) );
3486   }
3487   if (return_size_val != NULL) {
3488     (*return_size_val) = size;
3489   }
3490 
3491   // This is a precise notnull oop of the klass.
3492   // (Actually, it need not be precise if this is a reflective allocation.)
3493   // It's what we cast the result to.
3494   const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3495   if (!tklass)  tklass = TypeKlassPtr::OBJECT;
3496   const TypeOopPtr* oop_type = tklass->as_instance_type();
3497 
3498   // Now generate allocation code
3499 
3500   // The entire memory state is needed for slow path of the allocation
3501   // since GC and deoptimization can happened.
3502   Node *mem = reset_memory();
3503   set_all_memory(mem); // Create new memory state
3504 
3505   AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3506                                          control(), mem, i_o(),
3507                                          size, klass_node,
3508                                          initial_slow_test);
3509 
3510   return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3511 }
3512 
3513 //-------------------------------new_array-------------------------------------
3514 // helper for both newarray and anewarray
3515 // The 'length' parameter is (obviously) the length of the array.
3516 // See comments on new_instance for the meaning of the other arguments.
3517 Node* GraphKit::new_array(Node* klass_node,     // array klass (maybe variable)
3518                           Node* length,         // number of array elements
3519                           int   nargs,          // number of arguments to push back for uncommon trap
3520                           Node* *return_size_val,
3521                           bool deoptimize_on_exception) {
3522   jint  layout_con = Klass::_lh_neutral_value;
3523   Node* layout_val = get_layout_helper(klass_node, layout_con);
3524   int   layout_is_con = (layout_val == NULL);
3525 
3526   if (!layout_is_con && !StressReflectiveCode &&
3527       !too_many_traps(Deoptimization::Reason_class_check)) {
3528     // This is a reflective array creation site.
3529     // Optimistically assume that it is a subtype of Object[],
3530     // so that we can fold up all the address arithmetic.
3531     layout_con = Klass::array_layout_helper(T_OBJECT);
3532     Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3533     Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3534     { BuildCutout unless(this, bol_lh, PROB_MAX);
3535       inc_sp(nargs);
3536       uncommon_trap(Deoptimization::Reason_class_check,
3537                     Deoptimization::Action_maybe_recompile);
3538     }
3539     layout_val = NULL;
3540     layout_is_con = true;
3541   }
3542 
3543   // Generate the initial go-slow test.  Make sure we do not overflow
3544   // if length is huge (near 2Gig) or negative!  We do not need
3545   // exact double-words here, just a close approximation of needed
3546   // double-words.  We can't add any offset or rounding bits, lest we
3547   // take a size -1 of bytes and make it positive.  Use an unsigned
3548   // compare, so negative sizes look hugely positive.
3549   int fast_size_limit = FastAllocateSizeLimit;
3550   if (layout_is_con) {
3551     assert(!StressReflectiveCode, "stress mode does not use these paths");
3552     // Increase the size limit if we have exact knowledge of array type.
3553     int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3554     fast_size_limit <<= (LogBytesPerLong - log2_esize);
3555   }
3556 
3557   Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3558   Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3559 
3560   // --- Size Computation ---
3561   // array_size = round_to_heap(array_header + (length << elem_shift));
3562   // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3563   // and align_to(x, y) == ((x + y-1) & ~(y-1))
3564   // The rounding mask is strength-reduced, if possible.
3565   int round_mask = MinObjAlignmentInBytes - 1;
3566   Node* header_size = NULL;
3567   int   header_size_min  = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3568   // (T_BYTE has the weakest alignment and size restrictions...)
3569   if (layout_is_con) {
3570     int       hsize  = Klass::layout_helper_header_size(layout_con);
3571     int       eshift = Klass::layout_helper_log2_element_size(layout_con);
3572     BasicType etype  = Klass::layout_helper_element_type(layout_con);
3573     if ((round_mask & ~right_n_bits(eshift)) == 0)
3574       round_mask = 0;  // strength-reduce it if it goes away completely
3575     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3576     assert(header_size_min <= hsize, "generic minimum is smallest");
3577     header_size_min = hsize;
3578     header_size = intcon(hsize + round_mask);
3579   } else {
3580     Node* hss   = intcon(Klass::_lh_header_size_shift);
3581     Node* hsm   = intcon(Klass::_lh_header_size_mask);
3582     Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3583     hsize       = _gvn.transform( new AndINode(hsize, hsm) );
3584     Node* mask  = intcon(round_mask);
3585     header_size = _gvn.transform( new AddINode(hsize, mask) );
3586   }
3587 
3588   Node* elem_shift = NULL;
3589   if (layout_is_con) {
3590     int eshift = Klass::layout_helper_log2_element_size(layout_con);
3591     if (eshift != 0)
3592       elem_shift = intcon(eshift);
3593   } else {
3594     // There is no need to mask or shift this value.
3595     // The semantics of LShiftINode include an implicit mask to 0x1F.
3596     assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3597     elem_shift = layout_val;
3598   }
3599 
3600   // Transition to native address size for all offset calculations:
3601   Node* lengthx = ConvI2X(length);
3602   Node* headerx = ConvI2X(header_size);
3603 #ifdef _LP64
3604   { const TypeInt* tilen = _gvn.find_int_type(length);
3605     if (tilen != NULL && tilen->_lo < 0) {
3606       // Add a manual constraint to a positive range.  Cf. array_element_address.
3607       jint size_max = fast_size_limit;
3608       if (size_max > tilen->_hi)  size_max = tilen->_hi;
3609       const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3610 
3611       // Only do a narrow I2L conversion if the range check passed.
3612       IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3613       _gvn.transform(iff);
3614       RegionNode* region = new RegionNode(3);
3615       _gvn.set_type(region, Type::CONTROL);
3616       lengthx = new PhiNode(region, TypeLong::LONG);
3617       _gvn.set_type(lengthx, TypeLong::LONG);
3618 
3619       // Range check passed. Use ConvI2L node with narrow type.
3620       Node* passed = IfFalse(iff);
3621       region->init_req(1, passed);
3622       // Make I2L conversion control dependent to prevent it from
3623       // floating above the range check during loop optimizations.
3624       lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3625 
3626       // Range check failed. Use ConvI2L with wide type because length may be invalid.
3627       region->init_req(2, IfTrue(iff));
3628       lengthx->init_req(2, ConvI2X(length));
3629 
3630       set_control(region);
3631       record_for_igvn(region);
3632       record_for_igvn(lengthx);
3633     }
3634   }
3635 #endif
3636 
3637   // Combine header size (plus rounding) and body size.  Then round down.
3638   // This computation cannot overflow, because it is used only in two
3639   // places, one where the length is sharply limited, and the other
3640   // after a successful allocation.
3641   Node* abody = lengthx;
3642   if (elem_shift != NULL)
3643     abody     = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3644   Node* size  = _gvn.transform( new AddXNode(headerx, abody) );
3645   if (round_mask != 0) {
3646     Node* mask = MakeConX(~round_mask);
3647     size       = _gvn.transform( new AndXNode(size, mask) );
3648   }
3649   // else if round_mask == 0, the size computation is self-rounding
3650 
3651   if (return_size_val != NULL) {
3652     // This is the size
3653     (*return_size_val) = size;
3654   }
3655 
3656   // Now generate allocation code
3657 
3658   // The entire memory state is needed for slow path of the allocation
3659   // since GC and deoptimization can happened.
3660   Node *mem = reset_memory();
3661   set_all_memory(mem); // Create new memory state
3662 
3663   if (initial_slow_test->is_Bool()) {
3664     // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3665     initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3666   }
3667 
3668   // Create the AllocateArrayNode and its result projections
3669   AllocateArrayNode* alloc
3670     = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3671                             control(), mem, i_o(),
3672                             size, klass_node,
3673                             initial_slow_test,
3674                             length);
3675 
3676   // Cast to correct type.  Note that the klass_node may be constant or not,
3677   // and in the latter case the actual array type will be inexact also.
3678   // (This happens via a non-constant argument to inline_native_newArray.)
3679   // In any case, the value of klass_node provides the desired array type.
3680   const TypeInt* length_type = _gvn.find_int_type(length);
3681   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3682   if (ary_type->isa_aryptr() && length_type != NULL) {
3683     // Try to get a better type than POS for the size
3684     ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3685   }
3686 
3687   Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3688 
3689   // Cast length on remaining path to be as narrow as possible
3690   if (map()->find_edge(length) >= 0) {
3691     Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3692     if (ccast != length) {
3693       _gvn.set_type_bottom(ccast);
3694       record_for_igvn(ccast);
3695       replace_in_map(length, ccast);
3696     }
3697   }
3698 
3699   return javaoop;
3700 }
3701 
3702 // The following "Ideal_foo" functions are placed here because they recognize
3703 // the graph shapes created by the functions immediately above.
3704 
3705 //---------------------------Ideal_allocation----------------------------------
3706 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3707 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3708   if (ptr == NULL) {     // reduce dumb test in callers
3709     return NULL;
3710   }
3711   if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3712     ptr = ptr->in(1);
3713     if (ptr == NULL) return NULL;
3714   }
3715   // Return NULL for allocations with several casts:
3716   //   j.l.reflect.Array.newInstance(jobject, jint)
3717   //   Object.clone()
3718   // to keep more precise type from last cast.
3719   if (ptr->is_Proj()) {
3720     Node* allo = ptr->in(0);
3721     if (allo != NULL && allo->is_Allocate()) {
3722       return allo->as_Allocate();
3723     }
3724   }
3725   // Report failure to match.
3726   return NULL;
3727 }
3728 
3729 // Fancy version which also strips off an offset (and reports it to caller).
3730 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3731                                              intptr_t& offset) {
3732   Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3733   if (base == NULL)  return NULL;
3734   return Ideal_allocation(base, phase);
3735 }
3736 
3737 // Trace Initialize <- Proj[Parm] <- Allocate
3738 AllocateNode* InitializeNode::allocation() {
3739   Node* rawoop = in(InitializeNode::RawAddress);
3740   if (rawoop->is_Proj()) {
3741     Node* alloc = rawoop->in(0);
3742     if (alloc->is_Allocate()) {
3743       return alloc->as_Allocate();
3744     }
3745   }
3746   return NULL;
3747 }
3748 
3749 // Trace Allocate -> Proj[Parm] -> Initialize
3750 InitializeNode* AllocateNode::initialization() {
3751   ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3752   if (rawoop == NULL)  return NULL;
3753   for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3754     Node* init = rawoop->fast_out(i);
3755     if (init->is_Initialize()) {
3756       assert(init->as_Initialize()->allocation() == this, "2-way link");
3757       return init->as_Initialize();
3758     }
3759   }
3760   return NULL;
3761 }
3762 
3763 //----------------------------- loop predicates ---------------------------
3764 
3765 //------------------------------add_predicate_impl----------------------------
3766 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3767   // Too many traps seen?
3768   if (too_many_traps(reason)) {
3769 #ifdef ASSERT
3770     if (TraceLoopPredicate) {
3771       int tc = C->trap_count(reason);
3772       tty->print("too many traps=%s tcount=%d in ",
3773                     Deoptimization::trap_reason_name(reason), tc);
3774       method()->print(); // which method has too many predicate traps
3775       tty->cr();
3776     }
3777 #endif
3778     // We cannot afford to take more traps here,
3779     // do not generate predicate.
3780     return;
3781   }
3782 
3783   Node *cont    = _gvn.intcon(1);
3784   Node* opq     = _gvn.transform(new Opaque1Node(C, cont));
3785   Node *bol     = _gvn.transform(new Conv2BNode(opq));
3786   IfNode* iff   = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3787   Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3788   C->add_predicate_opaq(opq);
3789   {
3790     PreserveJVMState pjvms(this);
3791     set_control(iffalse);
3792     inc_sp(nargs);
3793     uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3794   }
3795   Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3796   set_control(iftrue);
3797 }
3798 
3799 //------------------------------add_predicate---------------------------------
3800 void GraphKit::add_predicate(int nargs) {
3801   if (UseLoopPredicate) {
3802     add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3803   }
3804   // loop's limit check predicate should be near the loop.
3805   add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3806 }
3807 
3808 //----------------------------- store barriers ----------------------------
3809 #define __ ideal.
3810 
3811 bool GraphKit::use_ReduceInitialCardMarks() {
3812   BarrierSet *bs = Universe::heap()->barrier_set();
3813   return bs->is_a(BarrierSet::CardTableModRef)
3814          && barrier_set_cast<CardTableModRefBS>(bs)->can_elide_tlab_store_barriers()
3815          && ReduceInitialCardMarks;
3816 }
3817 
3818 void GraphKit::sync_kit(IdealKit& ideal) {
3819   set_all_memory(__ merged_memory());
3820   set_i_o(__ i_o());
3821   set_control(__ ctrl());
3822 }
3823 
3824 void GraphKit::final_sync(IdealKit& ideal) {
3825   // Final sync IdealKit and graphKit.
3826   sync_kit(ideal);
3827 }
3828 
3829 Node* GraphKit::byte_map_base_node() {
3830   // Get base of card map
3831   jbyte* card_table_base = ci_card_table_address();
3832   if (card_table_base != NULL) {
3833     return makecon(TypeRawPtr::make((address)card_table_base));
3834   } else {
3835     return null();
3836   }
3837 }
3838 
3839 // vanilla/CMS post barrier
3840 // Insert a write-barrier store.  This is to let generational GC work; we have
3841 // to flag all oop-stores before the next GC point.
3842 void GraphKit::write_barrier_post(Node* oop_store,
3843                                   Node* obj,
3844                                   Node* adr,
3845                                   uint  adr_idx,
3846                                   Node* val,
3847                                   bool use_precise) {
3848   // No store check needed if we're storing a NULL or an old object
3849   // (latter case is probably a string constant). The concurrent
3850   // mark sweep garbage collector, however, needs to have all nonNull
3851   // oop updates flagged via card-marks.
3852   if (val != NULL && val->is_Con()) {
3853     // must be either an oop or NULL
3854     const Type* t = val->bottom_type();
3855     if (t == TypePtr::NULL_PTR || t == Type::TOP)
3856       // stores of null never (?) need barriers
3857       return;
3858   }
3859 
3860   if (use_ReduceInitialCardMarks()
3861       && obj == just_allocated_object(control())) {
3862     // We can skip marks on a freshly-allocated object in Eden.
3863     // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
3864     // That routine informs GC to take appropriate compensating steps,
3865     // upon a slow-path allocation, so as to make this card-mark
3866     // elision safe.
3867     return;
3868   }
3869 
3870   if (!use_precise) {
3871     // All card marks for a (non-array) instance are in one place:
3872     adr = obj;
3873   }
3874   // (Else it's an array (or unknown), and we want more precise card marks.)
3875   assert(adr != NULL, "");
3876 
3877   IdealKit ideal(this, true);
3878 
3879   // Convert the pointer to an int prior to doing math on it
3880   Node* cast = __ CastPX(__ ctrl(), adr);
3881 
3882   // Divide by card size
3883   assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3884          "Only one we handle so far.");
3885   Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
3886 
3887   // Combine card table base and card offset
3888   Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3889 
3890   // Get the alias_index for raw card-mark memory
3891   int adr_type = Compile::AliasIdxRaw;
3892   Node*   zero = __ ConI(0); // Dirty card value
3893   BasicType bt = T_BYTE;
3894 
3895   if (UseConcMarkSweepGC && UseCondCardMark) {
3896     insert_mem_bar(Op_MemBarVolatile);   // StoreLoad barrier
3897     __ sync_kit(this);
3898   }
3899 
3900   if (UseCondCardMark) {
3901     // The classic GC reference write barrier is typically implemented
3902     // as a store into the global card mark table.  Unfortunately
3903     // unconditional stores can result in false sharing and excessive
3904     // coherence traffic as well as false transactional aborts.
3905     // UseCondCardMark enables MP "polite" conditional card mark
3906     // stores.  In theory we could relax the load from ctrl() to
3907     // no_ctrl, but that doesn't buy much latitude.
3908     Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3909     __ if_then(card_val, BoolTest::ne, zero);
3910   }
3911 
3912   // Smash zero into card
3913   if( !UseConcMarkSweepGC ) {
3914     __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3915   } else {
3916     // Specialized path for CM store barrier
3917     __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3918   }
3919 
3920   if (UseCondCardMark) {
3921     __ end_if();
3922   }
3923 
3924   // Final sync IdealKit and GraphKit.
3925   final_sync(ideal);
3926 }
3927 /*
3928  * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3929  * required by SATB to make sure all objects live at the start of the
3930  * marking are kept alive, all reference updates need to any previous
3931  * reference stored before writing.
3932  *
3933  * If the previous value is NULL there is no need to save the old value.
3934  * References that are NULL are filtered during runtime by the barrier
3935  * code to avoid unnecessary queuing.
3936  *
3937  * However in the case of newly allocated objects it might be possible to
3938  * prove that the reference about to be overwritten is NULL during compile
3939  * time and avoid adding the barrier code completely.
3940  *
3941  * The compiler needs to determine that the object in which a field is about
3942  * to be written is newly allocated, and that no prior store to the same field
3943  * has happened since the allocation.
3944  *
3945  * Returns true if the pre-barrier can be removed
3946  */
3947 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
3948                                          BasicType bt, uint adr_idx) {
3949   intptr_t offset = 0;
3950   Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
3951   AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
3952 
3953   if (offset == Type::OffsetBot) {
3954     return false; // cannot unalias unless there are precise offsets
3955   }
3956 
3957   if (alloc == NULL) {
3958     return false; // No allocation found
3959   }
3960 
3961   intptr_t size_in_bytes = type2aelembytes(bt);
3962 
3963   Node* mem = memory(adr_idx); // start searching here...
3964 
3965   for (int cnt = 0; cnt < 50; cnt++) {
3966 
3967     if (mem->is_Store()) {
3968 
3969       Node* st_adr = mem->in(MemNode::Address);
3970       intptr_t st_offset = 0;
3971       Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
3972 
3973       if (st_base == NULL) {
3974         break; // inscrutable pointer
3975       }
3976 
3977       // Break we have found a store with same base and offset as ours so break
3978       if (st_base == base && st_offset == offset) {
3979         break;
3980       }
3981 
3982       if (st_offset != offset && st_offset != Type::OffsetBot) {
3983         const int MAX_STORE = BytesPerLong;
3984         if (st_offset >= offset + size_in_bytes ||
3985             st_offset <= offset - MAX_STORE ||
3986             st_offset <= offset - mem->as_Store()->memory_size()) {
3987           // Success:  The offsets are provably independent.
3988           // (You may ask, why not just test st_offset != offset and be done?
3989           // The answer is that stores of different sizes can co-exist
3990           // in the same sequence of RawMem effects.  We sometimes initialize
3991           // a whole 'tile' of array elements with a single jint or jlong.)
3992           mem = mem->in(MemNode::Memory);
3993           continue; // advance through independent store memory
3994         }
3995       }
3996 
3997       if (st_base != base
3998           && MemNode::detect_ptr_independence(base, alloc, st_base,
3999                                               AllocateNode::Ideal_allocation(st_base, phase),
4000                                               phase)) {
4001         // Success:  The bases are provably independent.
4002         mem = mem->in(MemNode::Memory);
4003         continue; // advance through independent store memory
4004       }
4005     } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4006 
4007       InitializeNode* st_init = mem->in(0)->as_Initialize();
4008       AllocateNode* st_alloc = st_init->allocation();
4009 
4010       // Make sure that we are looking at the same allocation site.
4011       // The alloc variable is guaranteed to not be null here from earlier check.
4012       if (alloc == st_alloc) {
4013         // Check that the initialization is storing NULL so that no previous store
4014         // has been moved up and directly write a reference
4015         Node* captured_store = st_init->find_captured_store(offset,
4016                                                             type2aelembytes(T_OBJECT),
4017                                                             phase);
4018         if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4019           return true;
4020         }
4021       }
4022     }
4023 
4024     // Unless there is an explicit 'continue', we must bail out here,
4025     // because 'mem' is an inscrutable memory state (e.g., a call).
4026     break;
4027   }
4028 
4029   return false;
4030 }
4031 
4032 // G1 pre/post barriers
4033 void GraphKit::g1_write_barrier_pre(bool do_load,
4034                                     Node* obj,
4035                                     Node* adr,
4036                                     uint alias_idx,
4037                                     Node* val,
4038                                     const TypeOopPtr* val_type,
4039                                     Node* pre_val,
4040                                     BasicType bt) {
4041 
4042   // Some sanity checks
4043   // Note: val is unused in this routine.
4044 
4045   if (do_load) {
4046     // We need to generate the load of the previous value
4047     assert(obj != NULL, "must have a base");
4048     assert(adr != NULL, "where are loading from?");
4049     assert(pre_val == NULL, "loaded already?");
4050     assert(val_type != NULL, "need a type");
4051 
4052     if (use_ReduceInitialCardMarks()
4053         && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4054       return;
4055     }
4056 
4057   } else {
4058     // In this case both val_type and alias_idx are unused.
4059     assert(pre_val != NULL, "must be loaded already");
4060     // Nothing to be done if pre_val is null.
4061     if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4062     assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4063   }
4064   assert(bt == T_OBJECT, "or we shouldn't be here");
4065 
4066   IdealKit ideal(this, true);
4067 
4068   Node* tls = __ thread(); // ThreadLocalStorage
4069 
4070   Node* no_ctrl = NULL;
4071   Node* no_base = __ top();
4072   Node* zero  = __ ConI(0);
4073   Node* zeroX = __ ConX(0);
4074 
4075   float likely  = PROB_LIKELY(0.999);
4076   float unlikely  = PROB_UNLIKELY(0.999);
4077 
4078   BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4079   assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4080 
4081   // Offsets into the thread
4082   const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +  // 648
4083                                           SATBMarkQueue::byte_offset_of_active());
4084   const int index_offset   = in_bytes(JavaThread::satb_mark_queue_offset() +  // 656
4085                                           SATBMarkQueue::byte_offset_of_index());
4086   const int buffer_offset  = in_bytes(JavaThread::satb_mark_queue_offset() +  // 652
4087                                           SATBMarkQueue::byte_offset_of_buf());
4088 
4089   // Now the actual pointers into the thread
4090   Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4091   Node* buffer_adr  = __ AddP(no_base, tls, __ ConX(buffer_offset));
4092   Node* index_adr   = __ AddP(no_base, tls, __ ConX(index_offset));
4093 
4094   // Now some of the values
4095   Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4096 
4097   // if (!marking)
4098   __ if_then(marking, BoolTest::ne, zero, unlikely); {
4099     BasicType index_bt = TypeX_X->basic_type();
4100     assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4101     Node* index   = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4102 
4103     if (do_load) {
4104       // load original value
4105       // alias_idx correct??
4106       pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4107     }
4108 
4109     // if (pre_val != NULL)
4110     __ if_then(pre_val, BoolTest::ne, null()); {
4111       Node* buffer  = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4112 
4113       // is the queue for this thread full?
4114       __ if_then(index, BoolTest::ne, zeroX, likely); {
4115 
4116         // decrement the index
4117         Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4118 
4119         // Now get the buffer location we will log the previous value into and store it
4120         Node *log_addr = __ AddP(no_base, buffer, next_index);
4121         __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4122         // update the index
4123         __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4124 
4125       } __ else_(); {
4126 
4127         // logging buffer is full, call the runtime
4128         const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4129         __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4130       } __ end_if();  // (!index)
4131     } __ end_if();  // (pre_val != NULL)
4132   } __ end_if();  // (!marking)
4133 
4134   // Final sync IdealKit and GraphKit.
4135   final_sync(ideal);
4136 }
4137 
4138 /*
4139  * G1 similar to any GC with a Young Generation requires a way to keep track of
4140  * references from Old Generation to Young Generation to make sure all live
4141  * objects are found. G1 also requires to keep track of object references
4142  * between different regions to enable evacuation of old regions, which is done
4143  * as part of mixed collections. References are tracked in remembered sets and
4144  * is continuously updated as reference are written to with the help of the
4145  * post-barrier.
4146  *
4147  * To reduce the number of updates to the remembered set the post-barrier
4148  * filters updates to fields in objects located in the Young Generation,
4149  * the same region as the reference, when the NULL is being written or
4150  * if the card is already marked as dirty by an earlier write.
4151  *
4152  * Under certain circumstances it is possible to avoid generating the
4153  * post-barrier completely if it is possible during compile time to prove
4154  * the object is newly allocated and that no safepoint exists between the
4155  * allocation and the store.
4156  *
4157  * In the case of slow allocation the allocation code must handle the barrier
4158  * as part of the allocation in the case the allocated object is not located
4159  * in the nursery, this would happen for humongous objects. This is similar to
4160  * how CMS is required to handle this case, see the comments for the method
4161  * CardTableModRefBS::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier.
4162  * A deferred card mark is required for these objects and handled in the above
4163  * mentioned methods.
4164  *
4165  * Returns true if the post barrier can be removed
4166  */
4167 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4168                                           Node* adr) {
4169   intptr_t      offset = 0;
4170   Node*         base   = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4171   AllocateNode* alloc  = AllocateNode::Ideal_allocation(base, phase);
4172 
4173   if (offset == Type::OffsetBot) {
4174     return false; // cannot unalias unless there are precise offsets
4175   }
4176 
4177   if (alloc == NULL) {
4178      return false; // No allocation found
4179   }
4180 
4181   // Start search from Store node
4182   Node* mem = store->in(MemNode::Control);
4183   if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4184 
4185     InitializeNode* st_init = mem->in(0)->as_Initialize();
4186     AllocateNode*  st_alloc = st_init->allocation();
4187 
4188     // Make sure we are looking at the same allocation
4189     if (alloc == st_alloc) {
4190       return true;
4191     }
4192   }
4193 
4194   return false;
4195 }
4196 
4197 //
4198 // Update the card table and add card address to the queue
4199 //
4200 void GraphKit::g1_mark_card(IdealKit& ideal,
4201                             Node* card_adr,
4202                             Node* oop_store,
4203                             uint oop_alias_idx,
4204                             Node* index,
4205                             Node* index_adr,
4206                             Node* buffer,
4207                             const TypeFunc* tf) {
4208 
4209   Node* zero  = __ ConI(0);
4210   Node* zeroX = __ ConX(0);
4211   Node* no_base = __ top();
4212   BasicType card_bt = T_BYTE;
4213   // Smash zero into card. MUST BE ORDERED WRT TO STORE
4214   __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4215 
4216   //  Now do the queue work
4217   __ if_then(index, BoolTest::ne, zeroX); {
4218 
4219     Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4220     Node* log_addr = __ AddP(no_base, buffer, next_index);
4221 
4222     // Order, see storeCM.
4223     __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4224     __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4225 
4226   } __ else_(); {
4227     __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4228   } __ end_if();
4229 
4230 }
4231 
4232 void GraphKit::g1_write_barrier_post(Node* oop_store,
4233                                      Node* obj,
4234                                      Node* adr,
4235                                      uint alias_idx,
4236                                      Node* val,
4237                                      BasicType bt,
4238                                      bool use_precise) {
4239   // If we are writing a NULL then we need no post barrier
4240 
4241   if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4242     // Must be NULL
4243     const Type* t = val->bottom_type();
4244     assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4245     // No post barrier if writing NULLx
4246     return;
4247   }
4248 
4249   if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4250     // We can skip marks on a freshly-allocated object in Eden.
4251     // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
4252     // That routine informs GC to take appropriate compensating steps,
4253     // upon a slow-path allocation, so as to make this card-mark
4254     // elision safe.
4255     return;
4256   }
4257 
4258   if (use_ReduceInitialCardMarks()
4259       && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4260     return;
4261   }
4262 
4263   if (!use_precise) {
4264     // All card marks for a (non-array) instance are in one place:
4265     adr = obj;
4266   }
4267   // (Else it's an array (or unknown), and we want more precise card marks.)
4268   assert(adr != NULL, "");
4269 
4270   IdealKit ideal(this, true);
4271 
4272   Node* tls = __ thread(); // ThreadLocalStorage
4273 
4274   Node* no_base = __ top();
4275   float likely  = PROB_LIKELY(0.999);
4276   float unlikely  = PROB_UNLIKELY(0.999);
4277   Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
4278   Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val());
4279   Node* zeroX = __ ConX(0);
4280 
4281   // Get the alias_index for raw card-mark memory
4282   const TypePtr* card_type = TypeRawPtr::BOTTOM;
4283 
4284   const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4285 
4286   // Offsets into the thread
4287   const int index_offset  = in_bytes(JavaThread::dirty_card_queue_offset() +
4288                                      DirtyCardQueue::byte_offset_of_index());
4289   const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4290                                      DirtyCardQueue::byte_offset_of_buf());
4291 
4292   // Pointers into the thread
4293 
4294   Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4295   Node* index_adr =  __ AddP(no_base, tls, __ ConX(index_offset));
4296 
4297   // Now some values
4298   // Use ctrl to avoid hoisting these values past a safepoint, which could
4299   // potentially reset these fields in the JavaThread.
4300   Node* index  = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4301   Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4302 
4303   // Convert the store obj pointer to an int prior to doing math on it
4304   // Must use ctrl to prevent "integerized oop" existing across safepoint
4305   Node* cast =  __ CastPX(__ ctrl(), adr);
4306 
4307   // Divide pointer by card size
4308   Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
4309 
4310   // Combine card table base and card offset
4311   Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4312 
4313   // If we know the value being stored does it cross regions?
4314 
4315   if (val != NULL) {
4316     // Does the store cause us to cross regions?
4317 
4318     // Should be able to do an unsigned compare of region_size instead of
4319     // and extra shift. Do we have an unsigned compare??
4320     // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4321     Node* xor_res =  __ URShiftX ( __ XorX( cast,  __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4322 
4323     // if (xor_res == 0) same region so skip
4324     __ if_then(xor_res, BoolTest::ne, zeroX); {
4325 
4326       // No barrier if we are storing a NULL
4327       __ if_then(val, BoolTest::ne, null(), unlikely); {
4328 
4329         // Ok must mark the card if not already dirty
4330 
4331         // load the original value of the card
4332         Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4333 
4334         __ if_then(card_val, BoolTest::ne, young_card); {
4335           sync_kit(ideal);
4336           // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4337           insert_mem_bar(Op_MemBarVolatile, oop_store);
4338           __ sync_kit(this);
4339 
4340           Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4341           __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4342             g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4343           } __ end_if();
4344         } __ end_if();
4345       } __ end_if();
4346     } __ end_if();
4347   } else {
4348     // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4349     // We don't need a barrier here if the destination is a newly allocated object
4350     // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4351     // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
4352     assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4353     Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4354     __ if_then(card_val, BoolTest::ne, young_card); {
4355       g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4356     } __ end_if();
4357   }
4358 
4359   // Final sync IdealKit and GraphKit.
4360   final_sync(ideal);
4361 }
4362 #undef __
4363 
4364 
4365 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4366   Node* len = load_array_length(load_String_value(ctrl, str));
4367   Node* coder = load_String_coder(ctrl, str);
4368   // Divide length by 2 if coder is UTF16
4369   return _gvn.transform(new RShiftINode(len, coder));
4370 }
4371 
4372 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4373   int value_offset = java_lang_String::value_offset_in_bytes();
4374   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4375                                                      false, NULL, 0);
4376   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4377   const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4378                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4379                                                   ciTypeArrayKlass::make(T_BYTE), true, 0);
4380   int value_field_idx = C->get_alias_index(value_field_type);
4381   Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4382                          value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4383   // String.value field is known to be @Stable.
4384   if (UseImplicitStableValues) {
4385     load = cast_array_to_stable(load, value_type);
4386   }
4387   return load;
4388 }
4389 
4390 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4391   if (!CompactStrings) {
4392     return intcon(java_lang_String::CODER_UTF16);
4393   }
4394   int coder_offset = java_lang_String::coder_offset_in_bytes();
4395   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4396                                                      false, NULL, 0);
4397   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4398   int coder_field_idx = C->get_alias_index(coder_field_type);
4399   return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4400                    TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4401 }
4402 
4403 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4404   int value_offset = java_lang_String::value_offset_in_bytes();
4405   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4406                                                      false, NULL, 0);
4407   const TypePtr* value_field_type = string_type->add_offset(value_offset);
4408   store_oop_to_object(ctrl, str,  basic_plus_adr(str, value_offset), value_field_type,
4409       value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4410 }
4411 
4412 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4413   int coder_offset = java_lang_String::coder_offset_in_bytes();
4414   const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4415                                                      false, NULL, 0);
4416   const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4417   int coder_field_idx = C->get_alias_index(coder_field_type);
4418   store_to_memory(ctrl, basic_plus_adr(str, coder_offset),
4419                   value, T_BYTE, coder_field_idx, MemNode::unordered);
4420 }
4421 
4422 // Capture src and dst memory state with a MergeMemNode
4423 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4424   if (src_type == dst_type) {
4425     // Types are equal, we don't need a MergeMemNode
4426     return memory(src_type);
4427   }
4428   MergeMemNode* merge = MergeMemNode::make(map()->memory());
4429   record_for_igvn(merge); // fold it up later, if possible
4430   int src_idx = C->get_alias_index(src_type);
4431   int dst_idx = C->get_alias_index(dst_type);
4432   merge->set_memory_at(src_idx, memory(src_idx));
4433   merge->set_memory_at(dst_idx, memory(dst_idx));
4434   return merge;
4435 }
4436 
4437 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4438   assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4439   assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4440   // If input and output memory types differ, capture both states to preserve
4441   // the dependency between preceding and subsequent loads/stores.
4442   // For example, the following program:
4443   //  StoreB
4444   //  compress_string
4445   //  LoadB
4446   // has this memory graph (use->def):
4447   //  LoadB -> compress_string -> CharMem
4448   //             ... -> StoreB -> ByteMem
4449   // The intrinsic hides the dependency between LoadB and StoreB, causing
4450   // the load to read from memory not containing the result of the StoreB.
4451   // The correct memory graph should look like this:
4452   //  LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4453   Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4454   StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4455   Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4456   set_memory(res_mem, TypeAryPtr::BYTES);
4457   return str;
4458 }
4459 
4460 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4461   assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4462   assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4463   // Capture src and dst memory (see comment in 'compress_string').
4464   Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4465   StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4466   set_memory(_gvn.transform(str), dst_type);
4467 }
4468 
4469 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4470   /**
4471    * int i_char = start;
4472    * for (int i_byte = 0; i_byte < count; i_byte++) {
4473    *   dst[i_char++] = (char)(src[i_byte] & 0xff);
4474    * }
4475    */
4476   add_predicate();
4477   RegionNode* head = new RegionNode(3);
4478   head->init_req(1, control());
4479   gvn().set_type(head, Type::CONTROL);
4480   record_for_igvn(head);
4481 
4482   Node* i_byte = new PhiNode(head, TypeInt::INT);
4483   i_byte->init_req(1, intcon(0));
4484   gvn().set_type(i_byte, TypeInt::INT);
4485   record_for_igvn(i_byte);
4486 
4487   Node* i_char = new PhiNode(head, TypeInt::INT);
4488   i_char->init_req(1, start);
4489   gvn().set_type(i_char, TypeInt::INT);
4490   record_for_igvn(i_char);
4491 
4492   Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4493   gvn().set_type(mem, Type::MEMORY);
4494   record_for_igvn(mem);
4495   set_control(head);
4496   set_memory(mem, TypeAryPtr::BYTES);
4497   Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4498   Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4499                              AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4500                              false, false, true /* mismatched */);
4501 
4502   IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4503   head->init_req(2, IfTrue(iff));
4504   mem->init_req(2, st);
4505   i_byte->init_req(2, AddI(i_byte, intcon(1)));
4506   i_char->init_req(2, AddI(i_char, intcon(2)));
4507 
4508   set_control(IfFalse(iff));
4509   set_memory(st, TypeAryPtr::BYTES);
4510 }
4511 
4512 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4513   if (!field->is_constant()) {
4514     return NULL; // Field not marked as constant.
4515   }
4516   ciInstance* holder = NULL;
4517   if (!field->is_static()) {
4518     ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4519     if (const_oop != NULL && const_oop->is_instance()) {
4520       holder = const_oop->as_instance();
4521     }
4522   }
4523   const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4524                                                         /*is_unsigned_load=*/false);
4525   if (con_type != NULL) {
4526     return makecon(con_type);
4527   }
4528   return NULL;
4529 }
4530 
4531 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4532   // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4533   // assumption of CCP analysis.
4534   return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4535 }
--- EOF ---