1 /*
2 * Copyright 2001-2008 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 #include "incls/_precompiled.incl"
26 #include "incls/_graphKit.cpp.incl"
27
28 //----------------------------GraphKit-----------------------------------------
29 // Main utility constructor.
30 GraphKit::GraphKit(JVMState* jvms)
31 : Phase(Phase::Parser),
32 _env(C->env()),
33 _gvn(*C->initial_gvn())
34 {
35 _exceptions = jvms->map()->next_exception();
36 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
37 set_jvms(jvms);
38 }
39
40 // Private constructor for parser.
41 GraphKit::GraphKit()
42 : Phase(Phase::Parser),
43 _env(C->env()),
44 _gvn(*C->initial_gvn())
45 {
46 _exceptions = NULL;
47 set_map(NULL);
48 debug_only(_sp = -99);
49 debug_only(set_bci(-99));
50 }
51
52
53
54 //---------------------------clean_stack---------------------------------------
55 // Clear away rubbish from the stack area of the JVM state.
56 // This destroys any arguments that may be waiting on the stack.
57 void GraphKit::clean_stack(int from_sp) {
58 SafePointNode* map = this->map();
59 JVMState* jvms = this->jvms();
60 int stk_size = jvms->stk_size();
61 int stkoff = jvms->stkoff();
62 Node* top = this->top();
63 for (int i = from_sp; i < stk_size; i++) {
64 if (map->in(stkoff + i) != top) {
65 map->set_req(stkoff + i, top);
66 }
67 }
68 }
69
70
71 //--------------------------------sync_jvms-----------------------------------
72 // Make sure our current jvms agrees with our parse state.
73 JVMState* GraphKit::sync_jvms() const {
74 JVMState* jvms = this->jvms();
75 jvms->set_bci(bci()); // Record the new bci in the JVMState
76 jvms->set_sp(sp()); // Record the new sp in the JVMState
77 assert(jvms_in_sync(), "jvms is now in sync");
78 return jvms;
79 }
80
81 #ifdef ASSERT
82 bool GraphKit::jvms_in_sync() const {
83 Parse* parse = is_Parse();
84 if (parse == NULL) {
85 if (bci() != jvms()->bci()) return false;
86 if (sp() != (int)jvms()->sp()) return false;
87 return true;
88 }
89 if (jvms()->method() != parse->method()) return false;
90 if (jvms()->bci() != parse->bci()) return false;
91 int jvms_sp = jvms()->sp();
92 if (jvms_sp != parse->sp()) return false;
93 int jvms_depth = jvms()->depth();
94 if (jvms_depth != parse->depth()) return false;
95 return true;
96 }
97
98 // Local helper checks for special internal merge points
99 // used to accumulate and merge exception states.
100 // They are marked by the region's in(0) edge being the map itself.
101 // Such merge points must never "escape" into the parser at large,
102 // until they have been handed to gvn.transform.
103 static bool is_hidden_merge(Node* reg) {
104 if (reg == NULL) return false;
105 if (reg->is_Phi()) {
106 reg = reg->in(0);
107 if (reg == NULL) return false;
108 }
109 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
110 }
111
112 void GraphKit::verify_map() const {
113 if (map() == NULL) return; // null map is OK
114 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
115 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
116 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
117 }
118
119 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
120 assert(ex_map->next_exception() == NULL, "not already part of a chain");
121 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
122 }
123 #endif
124
125 //---------------------------stop_and_kill_map---------------------------------
126 // Set _map to NULL, signalling a stop to further bytecode execution.
127 // First smash the current map's control to a constant, to mark it dead.
128 void GraphKit::stop_and_kill_map() {
129 SafePointNode* dead_map = stop();
130 if (dead_map != NULL) {
131 dead_map->disconnect_inputs(NULL); // Mark the map as killed.
132 assert(dead_map->is_killed(), "must be so marked");
133 }
134 }
135
136
137 //--------------------------------stopped--------------------------------------
138 // Tell if _map is NULL, or control is top.
139 bool GraphKit::stopped() {
140 if (map() == NULL) return true;
141 else if (control() == top()) return true;
142 else return false;
143 }
144
145
146 //-----------------------------has_ex_handler----------------------------------
147 // Tell if this method or any caller method has exception handlers.
148 bool GraphKit::has_ex_handler() {
149 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
150 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
151 return true;
152 }
153 }
154 return false;
155 }
156
157 //------------------------------save_ex_oop------------------------------------
158 // Save an exception without blowing stack contents or other JVM state.
159 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
160 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
161 ex_map->add_req(ex_oop);
162 debug_only(verify_exception_state(ex_map));
163 }
164
165 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
166 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
167 Node* ex_oop = ex_map->in(ex_map->req()-1);
168 if (clear_it) ex_map->del_req(ex_map->req()-1);
169 return ex_oop;
170 }
171
172 //-----------------------------saved_ex_oop------------------------------------
173 // Recover a saved exception from its map.
174 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
175 return common_saved_ex_oop(ex_map, false);
176 }
177
178 //--------------------------clear_saved_ex_oop---------------------------------
179 // Erase a previously saved exception from its map.
180 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
181 return common_saved_ex_oop(ex_map, true);
182 }
183
184 #ifdef ASSERT
185 //---------------------------has_saved_ex_oop----------------------------------
186 // Erase a previously saved exception from its map.
187 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
188 return ex_map->req() == ex_map->jvms()->endoff()+1;
189 }
190 #endif
191
192 //-------------------------make_exception_state--------------------------------
193 // Turn the current JVM state into an exception state, appending the ex_oop.
194 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
195 sync_jvms();
196 SafePointNode* ex_map = stop(); // do not manipulate this map any more
197 set_saved_ex_oop(ex_map, ex_oop);
198 return ex_map;
199 }
200
201
202 //--------------------------add_exception_state--------------------------------
203 // Add an exception to my list of exceptions.
204 void GraphKit::add_exception_state(SafePointNode* ex_map) {
205 if (ex_map == NULL || ex_map->control() == top()) {
206 return;
207 }
208 #ifdef ASSERT
209 verify_exception_state(ex_map);
210 if (has_exceptions()) {
211 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
212 }
213 #endif
214
215 // If there is already an exception of exactly this type, merge with it.
216 // In particular, null-checks and other low-level exceptions common up here.
217 Node* ex_oop = saved_ex_oop(ex_map);
218 const Type* ex_type = _gvn.type(ex_oop);
219 if (ex_oop == top()) {
220 // No action needed.
221 return;
222 }
223 assert(ex_type->isa_instptr(), "exception must be an instance");
224 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
225 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
226 // We check sp also because call bytecodes can generate exceptions
227 // both before and after arguments are popped!
228 if (ex_type2 == ex_type
229 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
230 combine_exception_states(ex_map, e2);
231 return;
232 }
233 }
234
235 // No pre-existing exception of the same type. Chain it on the list.
236 push_exception_state(ex_map);
237 }
238
239 //-----------------------add_exception_states_from-----------------------------
240 void GraphKit::add_exception_states_from(JVMState* jvms) {
241 SafePointNode* ex_map = jvms->map()->next_exception();
242 if (ex_map != NULL) {
243 jvms->map()->set_next_exception(NULL);
244 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
245 next_map = ex_map->next_exception();
246 ex_map->set_next_exception(NULL);
247 add_exception_state(ex_map);
248 }
249 }
250 }
251
252 //-----------------------transfer_exceptions_into_jvms-------------------------
253 JVMState* GraphKit::transfer_exceptions_into_jvms() {
254 if (map() == NULL) {
255 // We need a JVMS to carry the exceptions, but the map has gone away.
256 // Create a scratch JVMS, cloned from any of the exception states...
257 if (has_exceptions()) {
258 _map = _exceptions;
259 _map = clone_map();
260 _map->set_next_exception(NULL);
261 clear_saved_ex_oop(_map);
262 debug_only(verify_map());
263 } else {
264 // ...or created from scratch
265 JVMState* jvms = new (C) JVMState(_method, NULL);
266 jvms->set_bci(_bci);
267 jvms->set_sp(_sp);
268 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms));
269 set_jvms(jvms);
270 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
271 set_all_memory(top());
272 while (map()->req() < jvms->endoff()) map()->add_req(top());
273 }
274 // (This is a kludge, in case you didn't notice.)
275 set_control(top());
276 }
277 JVMState* jvms = sync_jvms();
278 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
279 jvms->map()->set_next_exception(_exceptions);
280 _exceptions = NULL; // done with this set of exceptions
281 return jvms;
282 }
283
284 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
285 assert(is_hidden_merge(dstphi), "must be a special merge node");
286 assert(is_hidden_merge(srcphi), "must be a special merge node");
287 uint limit = srcphi->req();
288 for (uint i = PhiNode::Input; i < limit; i++) {
289 dstphi->add_req(srcphi->in(i));
290 }
291 }
292 static inline void add_one_req(Node* dstphi, Node* src) {
293 assert(is_hidden_merge(dstphi), "must be a special merge node");
294 assert(!is_hidden_merge(src), "must not be a special merge node");
295 dstphi->add_req(src);
296 }
297
298 //-----------------------combine_exception_states------------------------------
299 // This helper function combines exception states by building phis on a
300 // specially marked state-merging region. These regions and phis are
301 // untransformed, and can build up gradually. The region is marked by
302 // having a control input of its exception map, rather than NULL. Such
303 // regions do not appear except in this function, and in use_exception_state.
304 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
305 if (failing()) return; // dying anyway...
306 JVMState* ex_jvms = ex_map->_jvms;
307 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
308 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
309 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
310 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
311 assert(ex_map->req() == phi_map->req(), "matching maps");
312 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
313 Node* hidden_merge_mark = root();
314 Node* region = phi_map->control();
315 MergeMemNode* phi_mem = phi_map->merged_memory();
316 MergeMemNode* ex_mem = ex_map->merged_memory();
317 if (region->in(0) != hidden_merge_mark) {
318 // The control input is not (yet) a specially-marked region in phi_map.
319 // Make it so, and build some phis.
320 region = new (C, 2) RegionNode(2);
321 _gvn.set_type(region, Type::CONTROL);
322 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
323 region->init_req(1, phi_map->control());
324 phi_map->set_control(region);
325 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
326 record_for_igvn(io_phi);
327 _gvn.set_type(io_phi, Type::ABIO);
328 phi_map->set_i_o(io_phi);
329 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
330 Node* m = mms.memory();
331 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
332 record_for_igvn(m_phi);
333 _gvn.set_type(m_phi, Type::MEMORY);
334 mms.set_memory(m_phi);
335 }
336 }
337
338 // Either or both of phi_map and ex_map might already be converted into phis.
339 Node* ex_control = ex_map->control();
340 // if there is special marking on ex_map also, we add multiple edges from src
341 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
342 // how wide was the destination phi_map, originally?
343 uint orig_width = region->req();
344
345 if (add_multiple) {
346 add_n_reqs(region, ex_control);
347 add_n_reqs(phi_map->i_o(), ex_map->i_o());
348 } else {
349 // ex_map has no merges, so we just add single edges everywhere
350 add_one_req(region, ex_control);
351 add_one_req(phi_map->i_o(), ex_map->i_o());
352 }
353 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
354 if (mms.is_empty()) {
355 // get a copy of the base memory, and patch some inputs into it
356 const TypePtr* adr_type = mms.adr_type(C);
357 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
358 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
359 mms.set_memory(phi);
360 // Prepare to append interesting stuff onto the newly sliced phi:
361 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
362 }
363 // Append stuff from ex_map:
364 if (add_multiple) {
365 add_n_reqs(mms.memory(), mms.memory2());
366 } else {
367 add_one_req(mms.memory(), mms.memory2());
368 }
369 }
370 uint limit = ex_map->req();
371 for (uint i = TypeFunc::Parms; i < limit; i++) {
372 // Skip everything in the JVMS after tos. (The ex_oop follows.)
373 if (i == tos) i = ex_jvms->monoff();
374 Node* src = ex_map->in(i);
375 Node* dst = phi_map->in(i);
376 if (src != dst) {
377 PhiNode* phi;
378 if (dst->in(0) != region) {
379 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
380 record_for_igvn(phi);
381 _gvn.set_type(phi, phi->type());
382 phi_map->set_req(i, dst);
383 // Prepare to append interesting stuff onto the new phi:
384 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
385 } else {
386 assert(dst->is_Phi(), "nobody else uses a hidden region");
387 phi = (PhiNode*)dst;
388 }
389 if (add_multiple && src->in(0) == ex_control) {
390 // Both are phis.
391 add_n_reqs(dst, src);
392 } else {
393 while (dst->req() < region->req()) add_one_req(dst, src);
394 }
395 const Type* srctype = _gvn.type(src);
396 if (phi->type() != srctype) {
397 const Type* dsttype = phi->type()->meet(srctype);
398 if (phi->type() != dsttype) {
399 phi->set_type(dsttype);
400 _gvn.set_type(phi, dsttype);
401 }
402 }
403 }
404 }
405 }
406
407 //--------------------------use_exception_state--------------------------------
408 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
409 if (failing()) { stop(); return top(); }
410 Node* region = phi_map->control();
411 Node* hidden_merge_mark = root();
412 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
413 Node* ex_oop = clear_saved_ex_oop(phi_map);
414 if (region->in(0) == hidden_merge_mark) {
415 // Special marking for internal ex-states. Process the phis now.
416 region->set_req(0, region); // now it's an ordinary region
417 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
418 // Note: Setting the jvms also sets the bci and sp.
419 set_control(_gvn.transform(region));
420 uint tos = jvms()->stkoff() + sp();
421 for (uint i = 1; i < tos; i++) {
422 Node* x = phi_map->in(i);
423 if (x->in(0) == region) {
424 assert(x->is_Phi(), "expected a special phi");
425 phi_map->set_req(i, _gvn.transform(x));
426 }
427 }
428 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
429 Node* x = mms.memory();
430 if (x->in(0) == region) {
431 assert(x->is_Phi(), "nobody else uses a hidden region");
432 mms.set_memory(_gvn.transform(x));
433 }
434 }
435 if (ex_oop->in(0) == region) {
436 assert(ex_oop->is_Phi(), "expected a special phi");
437 ex_oop = _gvn.transform(ex_oop);
438 }
439 } else {
440 set_jvms(phi_map->jvms());
441 }
442
443 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
444 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
445 return ex_oop;
446 }
447
448 //---------------------------------java_bc-------------------------------------
449 Bytecodes::Code GraphKit::java_bc() const {
450 ciMethod* method = this->method();
451 int bci = this->bci();
452 if (method != NULL && bci != InvocationEntryBci)
453 return method->java_code_at_bci(bci);
454 else
455 return Bytecodes::_illegal;
456 }
457
458 //------------------------------builtin_throw----------------------------------
459 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
460 bool must_throw = true;
461
462 if (JvmtiExport::can_post_exceptions()) {
463 // Do not try anything fancy if we're notifying the VM on every throw.
464 // Cf. case Bytecodes::_athrow in parse2.cpp.
465 uncommon_trap(reason, Deoptimization::Action_none,
466 (ciKlass*)NULL, (char*)NULL, must_throw);
467 return;
468 }
469
470 // If this particular condition has not yet happened at this
471 // bytecode, then use the uncommon trap mechanism, and allow for
472 // a future recompilation if several traps occur here.
473 // If the throw is hot, try to use a more complicated inline mechanism
474 // which keeps execution inside the compiled code.
475 bool treat_throw_as_hot = false;
476 ciMethodData* md = method()->method_data();
477
478 if (ProfileTraps) {
479 if (too_many_traps(reason)) {
480 treat_throw_as_hot = true;
481 }
482 // (If there is no MDO at all, assume it is early in
483 // execution, and that any deopts are part of the
484 // startup transient, and don't need to be remembered.)
485
486 // Also, if there is a local exception handler, treat all throws
487 // as hot if there has been at least one in this method.
488 if (C->trap_count(reason) != 0
489 && method()->method_data()->trap_count(reason) != 0
490 && has_ex_handler()) {
491 treat_throw_as_hot = true;
492 }
493 }
494
495 // If this throw happens frequently, an uncommon trap might cause
496 // a performance pothole. If there is a local exception handler,
497 // and if this particular bytecode appears to be deoptimizing often,
498 // let us handle the throw inline, with a preconstructed instance.
499 // Note: If the deopt count has blown up, the uncommon trap
500 // runtime is going to flush this nmethod, not matter what.
501 if (treat_throw_as_hot
502 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
503 // If the throw is local, we use a pre-existing instance and
504 // punt on the backtrace. This would lead to a missing backtrace
505 // (a repeat of 4292742) if the backtrace object is ever asked
506 // for its backtrace.
507 // Fixing this remaining case of 4292742 requires some flavor of
508 // escape analysis. Leave that for the future.
509 ciInstance* ex_obj = NULL;
510 switch (reason) {
511 case Deoptimization::Reason_null_check:
512 ex_obj = env()->NullPointerException_instance();
513 break;
514 case Deoptimization::Reason_div0_check:
515 ex_obj = env()->ArithmeticException_instance();
516 break;
517 case Deoptimization::Reason_range_check:
518 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
519 break;
520 case Deoptimization::Reason_class_check:
521 if (java_bc() == Bytecodes::_aastore) {
522 ex_obj = env()->ArrayStoreException_instance();
523 } else {
524 ex_obj = env()->ClassCastException_instance();
525 }
526 break;
527 }
528 if (failing()) { stop(); return; } // exception allocation might fail
529 if (ex_obj != NULL) {
530 // Cheat with a preallocated exception object.
531 if (C->log() != NULL)
532 C->log()->elem("hot_throw preallocated='1' reason='%s'",
533 Deoptimization::trap_reason_name(reason));
534 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
535 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
536
537 // Clear the detail message of the preallocated exception object.
538 // Weblogic sometimes mutates the detail message of exceptions
539 // using reflection.
540 int offset = java_lang_Throwable::get_detailMessage_offset();
541 const TypePtr* adr_typ = ex_con->add_offset(offset);
542
543 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
544 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT);
545
546 add_exception_state(make_exception_state(ex_node));
547 return;
548 }
549 }
550
551 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
552 // It won't be much cheaper than bailing to the interp., since we'll
553 // have to pass up all the debug-info, and the runtime will have to
554 // create the stack trace.
555
556 // Usual case: Bail to interpreter.
557 // Reserve the right to recompile if we haven't seen anything yet.
558
559 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
560 if (treat_throw_as_hot
561 && (method()->method_data()->trap_recompiled_at(bci())
562 || C->too_many_traps(reason))) {
563 // We cannot afford to take more traps here. Suffer in the interpreter.
564 if (C->log() != NULL)
565 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
566 Deoptimization::trap_reason_name(reason),
567 C->trap_count(reason));
568 action = Deoptimization::Action_none;
569 }
570
571 // "must_throw" prunes the JVM state to include only the stack, if there
572 // are no local exception handlers. This should cut down on register
573 // allocation time and code size, by drastically reducing the number
574 // of in-edges on the call to the uncommon trap.
575
576 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
577 }
578
579
580 //----------------------------PreserveJVMState---------------------------------
581 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
582 debug_only(kit->verify_map());
583 _kit = kit;
584 _map = kit->map(); // preserve the map
585 _sp = kit->sp();
586 kit->set_map(clone_map ? kit->clone_map() : NULL);
587 #ifdef ASSERT
588 _bci = kit->bci();
589 Parse* parser = kit->is_Parse();
590 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
591 _block = block;
592 #endif
593 }
594 PreserveJVMState::~PreserveJVMState() {
595 GraphKit* kit = _kit;
596 #ifdef ASSERT
597 assert(kit->bci() == _bci, "bci must not shift");
598 Parse* parser = kit->is_Parse();
599 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
600 assert(block == _block, "block must not shift");
601 #endif
602 kit->set_map(_map);
603 kit->set_sp(_sp);
604 }
605
606
607 //-----------------------------BuildCutout-------------------------------------
608 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
609 : PreserveJVMState(kit)
610 {
611 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
612 SafePointNode* outer_map = _map; // preserved map is caller's
613 SafePointNode* inner_map = kit->map();
614 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
615 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
616 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
617 }
618 BuildCutout::~BuildCutout() {
619 GraphKit* kit = _kit;
620 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
621 }
622
623
624 //------------------------------clone_map--------------------------------------
625 // Implementation of PreserveJVMState
626 //
627 // Only clone_map(...) here. If this function is only used in the
628 // PreserveJVMState class we may want to get rid of this extra
629 // function eventually and do it all there.
630
631 SafePointNode* GraphKit::clone_map() {
632 if (map() == NULL) return NULL;
633
634 // Clone the memory edge first
635 Node* mem = MergeMemNode::make(C, map()->memory());
636 gvn().set_type_bottom(mem);
637
638 SafePointNode *clonemap = (SafePointNode*)map()->clone();
639 JVMState* jvms = this->jvms();
640 JVMState* clonejvms = jvms->clone_shallow(C);
641 clonemap->set_memory(mem);
642 clonemap->set_jvms(clonejvms);
643 clonejvms->set_map(clonemap);
644 record_for_igvn(clonemap);
645 gvn().set_type_bottom(clonemap);
646 return clonemap;
647 }
648
649
650 //-----------------------------set_map_clone-----------------------------------
651 void GraphKit::set_map_clone(SafePointNode* m) {
652 _map = m;
653 _map = clone_map();
654 _map->set_next_exception(NULL);
655 debug_only(verify_map());
656 }
657
658
659 //----------------------------kill_dead_locals---------------------------------
660 // Detect any locals which are known to be dead, and force them to top.
661 void GraphKit::kill_dead_locals() {
662 // Consult the liveness information for the locals. If any
663 // of them are unused, then they can be replaced by top(). This
664 // should help register allocation time and cut down on the size
665 // of the deoptimization information.
666
667 // This call is made from many of the bytecode handling
668 // subroutines called from the Big Switch in do_one_bytecode.
669 // Every bytecode which might include a slow path is responsible
670 // for killing its dead locals. The more consistent we
671 // are about killing deads, the fewer useless phis will be
672 // constructed for them at various merge points.
673
674 // bci can be -1 (InvocationEntryBci). We return the entry
675 // liveness for the method.
676
677 if (method() == NULL || method()->code_size() == 0) {
678 // We are building a graph for a call to a native method.
679 // All locals are live.
680 return;
681 }
682
683 ResourceMark rm;
684
685 // Consult the liveness information for the locals. If any
686 // of them are unused, then they can be replaced by top(). This
687 // should help register allocation time and cut down on the size
688 // of the deoptimization information.
689 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
690
691 int len = (int)live_locals.size();
692 assert(len <= jvms()->loc_size(), "too many live locals");
693 for (int local = 0; local < len; local++) {
694 if (!live_locals.at(local)) {
695 set_local(local, top());
696 }
697 }
698 }
699
700 #ifdef ASSERT
701 //-------------------------dead_locals_are_killed------------------------------
702 // Return true if all dead locals are set to top in the map.
703 // Used to assert "clean" debug info at various points.
704 bool GraphKit::dead_locals_are_killed() {
705 if (method() == NULL || method()->code_size() == 0) {
706 // No locals need to be dead, so all is as it should be.
707 return true;
708 }
709
710 // Make sure somebody called kill_dead_locals upstream.
711 ResourceMark rm;
712 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
713 if (jvms->loc_size() == 0) continue; // no locals to consult
714 SafePointNode* map = jvms->map();
715 ciMethod* method = jvms->method();
716 int bci = jvms->bci();
717 if (jvms == this->jvms()) {
718 bci = this->bci(); // it might not yet be synched
719 }
720 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
721 int len = (int)live_locals.size();
722 if (!live_locals.is_valid() || len == 0)
723 // This method is trivial, or is poisoned by a breakpoint.
724 return true;
725 assert(len == jvms->loc_size(), "live map consistent with locals map");
726 for (int local = 0; local < len; local++) {
727 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
728 if (PrintMiscellaneous && (Verbose || WizardMode)) {
729 tty->print_cr("Zombie local %d: ", local);
730 jvms->dump();
731 }
732 return false;
733 }
734 }
735 }
736 return true;
737 }
738
739 #endif //ASSERT
740
741 // Helper function for adding JVMState and debug information to node
742 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
743 // Add the safepoint edges to the call (or other safepoint).
744
745 // Make sure dead locals are set to top. This
746 // should help register allocation time and cut down on the size
747 // of the deoptimization information.
748 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
749
750 // Walk the inline list to fill in the correct set of JVMState's
751 // Also fill in the associated edges for each JVMState.
752
753 JVMState* youngest_jvms = sync_jvms();
754
755 // Do we need debug info here? If it is a SafePoint and this method
756 // cannot de-opt, then we do NOT need any debug info.
757 bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint);
758
759 // If we are guaranteed to throw, we can prune everything but the
760 // input to the current bytecode.
761 bool can_prune_locals = false;
762 uint stack_slots_not_pruned = 0;
763 int inputs = 0, depth = 0;
764 if (must_throw) {
765 assert(method() == youngest_jvms->method(), "sanity");
766 if (compute_stack_effects(inputs, depth)) {
767 can_prune_locals = true;
768 stack_slots_not_pruned = inputs;
769 }
770 }
771
772 if (JvmtiExport::can_examine_or_deopt_anywhere()) {
773 // At any safepoint, this method can get breakpointed, which would
774 // then require an immediate deoptimization.
775 full_info = true;
776 can_prune_locals = false; // do not prune locals
777 stack_slots_not_pruned = 0;
778 }
779
780 // do not scribble on the input jvms
781 JVMState* out_jvms = youngest_jvms->clone_deep(C);
782 call->set_jvms(out_jvms); // Start jvms list for call node
783
784 // Presize the call:
785 debug_only(uint non_debug_edges = call->req());
786 call->add_req_batch(top(), youngest_jvms->debug_depth());
787 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
788
789 // Set up edges so that the call looks like this:
790 // Call [state:] ctl io mem fptr retadr
791 // [parms:] parm0 ... parmN
792 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
793 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
794 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
795 // Note that caller debug info precedes callee debug info.
796
797 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
798 uint debug_ptr = call->req();
799
800 // Loop over the map input edges associated with jvms, add them
801 // to the call node, & reset all offsets to match call node array.
802 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
803 uint debug_end = debug_ptr;
804 uint debug_start = debug_ptr - in_jvms->debug_size();
805 debug_ptr = debug_start; // back up the ptr
806
807 uint p = debug_start; // walks forward in [debug_start, debug_end)
808 uint j, k, l;
809 SafePointNode* in_map = in_jvms->map();
810 out_jvms->set_map(call);
811
812 if (can_prune_locals) {
813 assert(in_jvms->method() == out_jvms->method(), "sanity");
814 // If the current throw can reach an exception handler in this JVMS,
815 // then we must keep everything live that can reach that handler.
816 // As a quick and dirty approximation, we look for any handlers at all.
817 if (in_jvms->method()->has_exception_handlers()) {
818 can_prune_locals = false;
819 }
820 }
821
822 // Add the Locals
823 k = in_jvms->locoff();
824 l = in_jvms->loc_size();
825 out_jvms->set_locoff(p);
826 if (full_info && !can_prune_locals) {
827 for (j = 0; j < l; j++)
828 call->set_req(p++, in_map->in(k+j));
829 } else {
830 p += l; // already set to top above by add_req_batch
831 }
832
833 // Add the Expression Stack
834 k = in_jvms->stkoff();
835 l = in_jvms->sp();
836 out_jvms->set_stkoff(p);
837 if (full_info && !can_prune_locals) {
838 for (j = 0; j < l; j++)
839 call->set_req(p++, in_map->in(k+j));
840 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
841 // Divide stack into {S0,...,S1}, where S0 is set to top.
842 uint s1 = stack_slots_not_pruned;
843 stack_slots_not_pruned = 0; // for next iteration
844 if (s1 > l) s1 = l;
845 uint s0 = l - s1;
846 p += s0; // skip the tops preinstalled by add_req_batch
847 for (j = s0; j < l; j++)
848 call->set_req(p++, in_map->in(k+j));
849 } else {
850 p += l; // already set to top above by add_req_batch
851 }
852
853 // Add the Monitors
854 k = in_jvms->monoff();
855 l = in_jvms->mon_size();
856 out_jvms->set_monoff(p);
857 for (j = 0; j < l; j++)
858 call->set_req(p++, in_map->in(k+j));
859
860 // Copy any scalar object fields.
861 k = in_jvms->scloff();
862 l = in_jvms->scl_size();
863 out_jvms->set_scloff(p);
864 for (j = 0; j < l; j++)
865 call->set_req(p++, in_map->in(k+j));
866
867 // Finish the new jvms.
868 out_jvms->set_endoff(p);
869
870 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
871 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
872 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
873 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
874 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
875 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
876
877 // Update the two tail pointers in parallel.
878 out_jvms = out_jvms->caller();
879 in_jvms = in_jvms->caller();
880 }
881
882 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
883
884 // Test the correctness of JVMState::debug_xxx accessors:
885 assert(call->jvms()->debug_start() == non_debug_edges, "");
886 assert(call->jvms()->debug_end() == call->req(), "");
887 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
888 }
889
890 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
891 Bytecodes::Code code = java_bc();
892 if (code == Bytecodes::_wide) {
893 code = method()->java_code_at_bci(bci() + 1);
894 }
895
896 BasicType rtype = T_ILLEGAL;
897 int rsize = 0;
898
899 if (code != Bytecodes::_illegal) {
900 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
901 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
902 if (rtype < T_CONFLICT)
903 rsize = type2size[rtype];
904 }
905
906 switch (code) {
907 case Bytecodes::_illegal:
908 return false;
909
910 case Bytecodes::_ldc:
911 case Bytecodes::_ldc_w:
912 case Bytecodes::_ldc2_w:
913 inputs = 0;
914 break;
915
916 case Bytecodes::_dup: inputs = 1; break;
917 case Bytecodes::_dup_x1: inputs = 2; break;
918 case Bytecodes::_dup_x2: inputs = 3; break;
919 case Bytecodes::_dup2: inputs = 2; break;
920 case Bytecodes::_dup2_x1: inputs = 3; break;
921 case Bytecodes::_dup2_x2: inputs = 4; break;
922 case Bytecodes::_swap: inputs = 2; break;
923 case Bytecodes::_arraylength: inputs = 1; break;
924
925 case Bytecodes::_getstatic:
926 case Bytecodes::_putstatic:
927 case Bytecodes::_getfield:
928 case Bytecodes::_putfield:
929 {
930 bool is_get = (depth >= 0), is_static = (depth & 1);
931 bool ignore;
932 ciBytecodeStream iter(method());
933 iter.reset_to_bci(bci());
934 iter.next();
935 ciField* field = iter.get_field(ignore);
936 int size = field->type()->size();
937 inputs = (is_static ? 0 : 1);
938 if (is_get) {
939 depth = size - inputs;
940 } else {
941 inputs += size; // putxxx pops the value from the stack
942 depth = - inputs;
943 }
944 }
945 break;
946
947 case Bytecodes::_invokevirtual:
948 case Bytecodes::_invokespecial:
949 case Bytecodes::_invokestatic:
950 case Bytecodes::_invokeinterface:
951 {
952 bool is_static = (depth == 0);
953 bool ignore;
954 ciBytecodeStream iter(method());
955 iter.reset_to_bci(bci());
956 iter.next();
957 ciMethod* method = iter.get_method(ignore);
958 inputs = method->arg_size_no_receiver();
959 if (!is_static) inputs += 1;
960 int size = method->return_type()->size();
961 depth = size - inputs;
962 }
963 break;
964
965 case Bytecodes::_multianewarray:
966 {
967 ciBytecodeStream iter(method());
968 iter.reset_to_bci(bci());
969 iter.next();
970 inputs = iter.get_dimensions();
971 assert(rsize == 1, "");
972 depth = rsize - inputs;
973 }
974 break;
975
976 case Bytecodes::_ireturn:
977 case Bytecodes::_lreturn:
978 case Bytecodes::_freturn:
979 case Bytecodes::_dreturn:
980 case Bytecodes::_areturn:
981 assert(rsize = -depth, "");
982 inputs = rsize;
983 break;
984
985 case Bytecodes::_jsr:
986 case Bytecodes::_jsr_w:
987 inputs = 0;
988 depth = 1; // S.B. depth=1, not zero
989 break;
990
991 default:
992 // bytecode produces a typed result
993 inputs = rsize - depth;
994 assert(inputs >= 0, "");
995 break;
996 }
997
998 #ifdef ASSERT
999 // spot check
1000 int outputs = depth + inputs;
1001 assert(outputs >= 0, "sanity");
1002 switch (code) {
1003 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1004 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1005 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1006 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1007 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1008 }
1009 #endif //ASSERT
1010
1011 return true;
1012 }
1013
1014
1015
1016 //------------------------------basic_plus_adr---------------------------------
1017 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1018 // short-circuit a common case
1019 if (offset == intcon(0)) return ptr;
1020 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1021 }
1022
1023 Node* GraphKit::ConvI2L(Node* offset) {
1024 // short-circuit a common case
1025 jint offset_con = find_int_con(offset, Type::OffsetBot);
1026 if (offset_con != Type::OffsetBot) {
1027 return longcon((long) offset_con);
1028 }
1029 return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1030 }
1031 Node* GraphKit::ConvL2I(Node* offset) {
1032 // short-circuit a common case
1033 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1034 if (offset_con != (jlong)Type::OffsetBot) {
1035 return intcon((int) offset_con);
1036 }
1037 return _gvn.transform( new (C, 2) ConvL2INode(offset));
1038 }
1039
1040 //-------------------------load_object_klass-----------------------------------
1041 Node* GraphKit::load_object_klass(Node* obj) {
1042 // Special-case a fresh allocation to avoid building nodes:
1043 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1044 if (akls != NULL) return akls;
1045 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1046 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1047 }
1048
1049 //-------------------------load_array_length-----------------------------------
1050 Node* GraphKit::load_array_length(Node* array) {
1051 // Special-case a fresh allocation to avoid building nodes:
1052 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1053 Node *alen;
1054 if (alloc == NULL) {
1055 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1056 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1057 } else {
1058 alen = alloc->Ideal_length();
1059 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_aryptr(), &_gvn);
1060 if (ccast != alen) {
1061 alen = _gvn.transform(ccast);
1062 }
1063 }
1064 return alen;
1065 }
1066
1067 //------------------------------do_null_check----------------------------------
1068 // Helper function to do a NULL pointer check. Returned value is
1069 // the incoming address with NULL casted away. You are allowed to use the
1070 // not-null value only if you are control dependent on the test.
1071 extern int explicit_null_checks_inserted,
1072 explicit_null_checks_elided;
1073 Node* GraphKit::null_check_common(Node* value, BasicType type,
1074 // optional arguments for variations:
1075 bool assert_null,
1076 Node* *null_control) {
1077 assert(!assert_null || null_control == NULL, "not both at once");
1078 if (stopped()) return top();
1079 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1080 // For some performance testing, we may wish to suppress null checking.
1081 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1082 return value;
1083 }
1084 explicit_null_checks_inserted++;
1085
1086 // Construct NULL check
1087 Node *chk = NULL;
1088 switch(type) {
1089 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1090 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1091 case T_ARRAY : // fall through
1092 type = T_OBJECT; // simplify further tests
1093 case T_OBJECT : {
1094 const Type *t = _gvn.type( value );
1095
1096 const TypeInstPtr* tp = t->isa_instptr();
1097 if (tp != NULL && !tp->klass()->is_loaded()
1098 // Only for do_null_check, not any of its siblings:
1099 && !assert_null && null_control == NULL) {
1100 // Usually, any field access or invocation on an unloaded oop type
1101 // will simply fail to link, since the statically linked class is
1102 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1103 // the static class is loaded but the sharper oop type is not.
1104 // Rather than checking for this obscure case in lots of places,
1105 // we simply observe that a null check on an unloaded class
1106 // will always be followed by a nonsense operation, so we
1107 // can just issue the uncommon trap here.
1108 // Our access to the unloaded class will only be correct
1109 // after it has been loaded and initialized, which requires
1110 // a trip through the interpreter.
1111 #ifndef PRODUCT
1112 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1113 #endif
1114 uncommon_trap(Deoptimization::Reason_unloaded,
1115 Deoptimization::Action_reinterpret,
1116 tp->klass(), "!loaded");
1117 return top();
1118 }
1119
1120 if (assert_null) {
1121 // See if the type is contained in NULL_PTR.
1122 // If so, then the value is already null.
1123 if (t->higher_equal(TypePtr::NULL_PTR)) {
1124 explicit_null_checks_elided++;
1125 return value; // Elided null assert quickly!
1126 }
1127 } else {
1128 // See if mixing in the NULL pointer changes type.
1129 // If so, then the NULL pointer was not allowed in the original
1130 // type. In other words, "value" was not-null.
1131 if (t->meet(TypePtr::NULL_PTR) != t) {
1132 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1133 explicit_null_checks_elided++;
1134 return value; // Elided null check quickly!
1135 }
1136 }
1137 chk = new (C, 3) CmpPNode( value, null() );
1138 break;
1139 }
1140
1141 default : ShouldNotReachHere();
1142 }
1143 assert(chk != NULL, "sanity check");
1144 chk = _gvn.transform(chk);
1145
1146 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1147 BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1148 Node *tst = _gvn.transform( btst );
1149
1150 //-----------
1151 // if peephole optimizations occured, a prior test existed.
1152 // If a prior test existed, maybe it dominates as we can avoid this test.
1153 if (tst != btst && type == T_OBJECT) {
1154 // At this point we want to scan up the CFG to see if we can
1155 // find an identical test (and so avoid this test altogether).
1156 Node *cfg = control();
1157 int depth = 0;
1158 while( depth < 16 ) { // Limit search depth for speed
1159 if( cfg->Opcode() == Op_IfTrue &&
1160 cfg->in(0)->in(1) == tst ) {
1161 // Found prior test. Use "cast_not_null" to construct an identical
1162 // CastPP (and hence hash to) as already exists for the prior test.
1163 // Return that casted value.
1164 if (assert_null) {
1165 replace_in_map(value, null());
1166 return null(); // do not issue the redundant test
1167 }
1168 Node *oldcontrol = control();
1169 set_control(cfg);
1170 Node *res = cast_not_null(value);
1171 set_control(oldcontrol);
1172 explicit_null_checks_elided++;
1173 return res;
1174 }
1175 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1176 if (cfg == NULL) break; // Quit at region nodes
1177 depth++;
1178 }
1179 }
1180
1181 //-----------
1182 // Branch to failure if null
1183 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1184 Deoptimization::DeoptReason reason;
1185 if (assert_null)
1186 reason = Deoptimization::Reason_null_assert;
1187 else if (type == T_OBJECT)
1188 reason = Deoptimization::Reason_null_check;
1189 else
1190 reason = Deoptimization::Reason_div0_check;
1191
1192 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1193 // ciMethodData::has_trap_at will return a conservative -1 if any
1194 // must-be-null assertion has failed. This could cause performance
1195 // problems for a method after its first do_null_assert failure.
1196 // Consider using 'Reason_class_check' instead?
1197
1198 // To cause an implicit null check, we set the not-null probability
1199 // to the maximum (PROB_MAX). For an explicit check the probablity
1200 // is set to a smaller value.
1201 if (null_control != NULL || too_many_traps(reason)) {
1202 // probability is less likely
1203 ok_prob = PROB_LIKELY_MAG(3);
1204 } else if (!assert_null &&
1205 (ImplicitNullCheckThreshold > 0) &&
1206 method() != NULL &&
1207 (method()->method_data()->trap_count(reason)
1208 >= (uint)ImplicitNullCheckThreshold)) {
1209 ok_prob = PROB_LIKELY_MAG(3);
1210 }
1211
1212 if (null_control != NULL) {
1213 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1214 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1215 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
1216 if (null_true == top())
1217 explicit_null_checks_elided++;
1218 (*null_control) = null_true;
1219 } else {
1220 BuildCutout unless(this, tst, ok_prob);
1221 // Check for optimizer eliding test at parse time
1222 if (stopped()) {
1223 // Failure not possible; do not bother making uncommon trap.
1224 explicit_null_checks_elided++;
1225 } else if (assert_null) {
1226 uncommon_trap(reason,
1227 Deoptimization::Action_make_not_entrant,
1228 NULL, "assert_null");
1229 } else {
1230 replace_in_map(value, zerocon(type));
1231 builtin_throw(reason);
1232 }
1233 }
1234
1235 // Must throw exception, fall-thru not possible?
1236 if (stopped()) {
1237 return top(); // No result
1238 }
1239
1240 if (assert_null) {
1241 // Cast obj to null on this path.
1242 replace_in_map(value, zerocon(type));
1243 return zerocon(type);
1244 }
1245
1246 // Cast obj to not-null on this path, if there is no null_control.
1247 // (If there is a null_control, a non-null value may come back to haunt us.)
1248 if (type == T_OBJECT) {
1249 Node* cast = cast_not_null(value, false);
1250 if (null_control == NULL || (*null_control) == top())
1251 replace_in_map(value, cast);
1252 value = cast;
1253 }
1254
1255 return value;
1256 }
1257
1258
1259 //------------------------------cast_not_null----------------------------------
1260 // Cast obj to not-null on this path
1261 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1262 const Type *t = _gvn.type(obj);
1263 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1264 // Object is already not-null?
1265 if( t == t_not_null ) return obj;
1266
1267 Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1268 cast->init_req(0, control());
1269 cast = _gvn.transform( cast );
1270
1271 // Scan for instances of 'obj' in the current JVM mapping.
1272 // These instances are known to be not-null after the test.
1273 if (do_replace_in_map)
1274 replace_in_map(obj, cast);
1275
1276 return cast; // Return casted value
1277 }
1278
1279
1280 //--------------------------replace_in_map-------------------------------------
1281 void GraphKit::replace_in_map(Node* old, Node* neww) {
1282 this->map()->replace_edge(old, neww);
1283
1284 // Note: This operation potentially replaces any edge
1285 // on the map. This includes locals, stack, and monitors
1286 // of the current (innermost) JVM state.
1287
1288 // We can consider replacing in caller maps.
1289 // The idea would be that an inlined function's null checks
1290 // can be shared with the entire inlining tree.
1291 // The expense of doing this is that the PreserveJVMState class
1292 // would have to preserve caller states too, with a deep copy.
1293 }
1294
1295
1296
1297 //=============================================================================
1298 //--------------------------------memory---------------------------------------
1299 Node* GraphKit::memory(uint alias_idx) {
1300 MergeMemNode* mem = merged_memory();
1301 Node* p = mem->memory_at(alias_idx);
1302 _gvn.set_type(p, Type::MEMORY); // must be mapped
1303 return p;
1304 }
1305
1306 //-----------------------------reset_memory------------------------------------
1307 Node* GraphKit::reset_memory() {
1308 Node* mem = map()->memory();
1309 // do not use this node for any more parsing!
1310 debug_only( map()->set_memory((Node*)NULL) );
1311 return _gvn.transform( mem );
1312 }
1313
1314 //------------------------------set_all_memory---------------------------------
1315 void GraphKit::set_all_memory(Node* newmem) {
1316 Node* mergemem = MergeMemNode::make(C, newmem);
1317 gvn().set_type_bottom(mergemem);
1318 map()->set_memory(mergemem);
1319 }
1320
1321 //------------------------------set_all_memory_call----------------------------
1322 void GraphKit::set_all_memory_call(Node* call) {
1323 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1324 set_all_memory(newmem);
1325 }
1326
1327 //=============================================================================
1328 //
1329 // parser factory methods for MemNodes
1330 //
1331 // These are layered on top of the factory methods in LoadNode and StoreNode,
1332 // and integrate with the parser's memory state and _gvn engine.
1333 //
1334
1335 // factory methods in "int adr_idx"
1336 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1337 int adr_idx,
1338 bool require_atomic_access) {
1339 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1340 const TypePtr* adr_type = NULL; // debug-mode-only argument
1341 debug_only(adr_type = C->get_adr_type(adr_idx));
1342 Node* mem = memory(adr_idx);
1343 Node* ld;
1344 if (require_atomic_access && bt == T_LONG) {
1345 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1346 } else {
1347 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1348 }
1349 return _gvn.transform(ld);
1350 }
1351
1352 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1353 int adr_idx,
1354 bool require_atomic_access) {
1355 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1356 const TypePtr* adr_type = NULL;
1357 debug_only(adr_type = C->get_adr_type(adr_idx));
1358 Node *mem = memory(adr_idx);
1359 Node* st;
1360 if (require_atomic_access && bt == T_LONG) {
1361 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1362 } else {
1363 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1364 }
1365 st = _gvn.transform(st);
1366 set_memory(st, adr_idx);
1367 // Back-to-back stores can only remove intermediate store with DU info
1368 // so push on worklist for optimizer.
1369 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1370 record_for_igvn(st);
1371
1372 return st;
1373 }
1374
1375 void GraphKit::pre_barrier(Node* ctl,
1376 Node* obj,
1377 Node* adr,
1378 uint adr_idx,
1379 Node *val,
1380 const Type* val_type,
1381 BasicType bt) {
1382 BarrierSet* bs = Universe::heap()->barrier_set();
1383 set_control(ctl);
1384 switch (bs->kind()) {
1385 case BarrierSet::G1SATBCT:
1386 case BarrierSet::G1SATBCTLogging:
1387 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt);
1388 break;
1389
1390 case BarrierSet::CardTableModRef:
1391 case BarrierSet::CardTableExtension:
1392 case BarrierSet::ModRef:
1393 break;
1394
1395 case BarrierSet::Other:
1396 default :
1397 ShouldNotReachHere();
1398
1399 }
1400 }
1401
1402 void GraphKit::post_barrier(Node* ctl,
1403 Node* store,
1404 Node* obj,
1405 Node* adr,
1406 uint adr_idx,
1407 Node *val,
1408 BasicType bt,
1409 bool use_precise) {
1410 BarrierSet* bs = Universe::heap()->barrier_set();
1411 set_control(ctl);
1412 switch (bs->kind()) {
1413 case BarrierSet::G1SATBCT:
1414 case BarrierSet::G1SATBCTLogging:
1415 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1416 break;
1417
1418 case BarrierSet::CardTableModRef:
1419 case BarrierSet::CardTableExtension:
1420 write_barrier_post(store, obj, adr, val, use_precise);
1421 break;
1422
1423 case BarrierSet::ModRef:
1424 break;
1425
1426 case BarrierSet::Other:
1427 default :
1428 ShouldNotReachHere();
1429
1430 }
1431 }
1432
1433 Node* GraphKit::store_oop_to_object(Node* ctl,
1434 Node* obj,
1435 Node* adr,
1436 const TypePtr* adr_type,
1437 Node *val,
1438 const Type* val_type,
1439 BasicType bt) {
1440 uint adr_idx = C->get_alias_index(adr_type);
1441 Node* store;
1442 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
1443 store = store_to_memory(control(), adr, val, bt, adr_idx);
1444 post_barrier(control(), store, obj, adr, adr_idx, val, bt, false);
1445 return store;
1446 }
1447
1448 Node* GraphKit::store_oop_to_array(Node* ctl,
1449 Node* obj,
1450 Node* adr,
1451 const TypePtr* adr_type,
1452 Node *val,
1453 const Type* val_type,
1454 BasicType bt) {
1455 uint adr_idx = C->get_alias_index(adr_type);
1456 Node* store;
1457 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
1458 store = store_to_memory(control(), adr, val, bt, adr_idx);
1459 post_barrier(control(), store, obj, adr, adr_idx, val, bt, true);
1460 return store;
1461 }
1462
1463 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1464 Node* obj,
1465 Node* adr,
1466 const TypePtr* adr_type,
1467 Node *val,
1468 const Type* val_type,
1469 BasicType bt) {
1470 uint adr_idx = C->get_alias_index(adr_type);
1471 Node* store;
1472 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
1473 store = store_to_memory(control(), adr, val, bt, adr_idx);
1474 post_barrier(control(), store, obj, adr, adr_idx, val, bt, true);
1475 return store;
1476 }
1477
1478
1479 //-------------------------array_element_address-------------------------
1480 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1481 const TypeInt* sizetype) {
1482 uint shift = exact_log2(type2aelembytes(elembt));
1483 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1484
1485 // short-circuit a common case (saves lots of confusing waste motion)
1486 jint idx_con = find_int_con(idx, -1);
1487 if (idx_con >= 0) {
1488 intptr_t offset = header + ((intptr_t)idx_con << shift);
1489 return basic_plus_adr(ary, offset);
1490 }
1491
1492 // must be correct type for alignment purposes
1493 Node* base = basic_plus_adr(ary, header);
1494 #ifdef _LP64
1495 // The scaled index operand to AddP must be a clean 64-bit value.
1496 // Java allows a 32-bit int to be incremented to a negative
1497 // value, which appears in a 64-bit register as a large
1498 // positive number. Using that large positive number as an
1499 // operand in pointer arithmetic has bad consequences.
1500 // On the other hand, 32-bit overflow is rare, and the possibility
1501 // can often be excluded, if we annotate the ConvI2L node with
1502 // a type assertion that its value is known to be a small positive
1503 // number. (The prior range check has ensured this.)
1504 // This assertion is used by ConvI2LNode::Ideal.
1505 int index_max = max_jint - 1; // array size is max_jint, index is one less
1506 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1507 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1508 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1509 #endif
1510 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1511 return basic_plus_adr(ary, base, scale);
1512 }
1513
1514 //-------------------------load_array_element-------------------------
1515 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1516 const Type* elemtype = arytype->elem();
1517 BasicType elembt = elemtype->array_element_basic_type();
1518 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1519 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1520 return ld;
1521 }
1522
1523 //-------------------------set_arguments_for_java_call-------------------------
1524 // Arguments (pre-popped from the stack) are taken from the JVMS.
1525 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1526 // Add the call arguments:
1527 uint nargs = call->method()->arg_size();
1528 for (uint i = 0; i < nargs; i++) {
1529 Node* arg = argument(i);
1530 call->init_req(i + TypeFunc::Parms, arg);
1531 }
1532 }
1533
1534 //---------------------------set_edges_for_java_call---------------------------
1535 // Connect a newly created call into the current JVMS.
1536 // A return value node (if any) is returned from set_edges_for_java_call.
1537 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw) {
1538
1539 // Add the predefined inputs:
1540 call->init_req( TypeFunc::Control, control() );
1541 call->init_req( TypeFunc::I_O , i_o() );
1542 call->init_req( TypeFunc::Memory , reset_memory() );
1543 call->init_req( TypeFunc::FramePtr, frameptr() );
1544 call->init_req( TypeFunc::ReturnAdr, top() );
1545
1546 add_safepoint_edges(call, must_throw);
1547
1548 Node* xcall = _gvn.transform(call);
1549
1550 if (xcall == top()) {
1551 set_control(top());
1552 return;
1553 }
1554 assert(xcall == call, "call identity is stable");
1555
1556 // Re-use the current map to produce the result.
1557
1558 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1559 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O )));
1560 set_all_memory_call(xcall);
1561
1562 //return xcall; // no need, caller already has it
1563 }
1564
1565 Node* GraphKit::set_results_for_java_call(CallJavaNode* call) {
1566 if (stopped()) return top(); // maybe the call folded up?
1567
1568 // Capture the return value, if any.
1569 Node* ret;
1570 if (call->method() == NULL ||
1571 call->method()->return_type()->basic_type() == T_VOID)
1572 ret = top();
1573 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1574
1575 // Note: Since any out-of-line call can produce an exception,
1576 // we always insert an I_O projection from the call into the result.
1577
1578 make_slow_call_ex(call, env()->Throwable_klass(), false);
1579
1580 return ret;
1581 }
1582
1583 //--------------------set_predefined_input_for_runtime_call--------------------
1584 // Reading and setting the memory state is way conservative here.
1585 // The real problem is that I am not doing real Type analysis on memory,
1586 // so I cannot distinguish card mark stores from other stores. Across a GC
1587 // point the Store Barrier and the card mark memory has to agree. I cannot
1588 // have a card mark store and its barrier split across the GC point from
1589 // either above or below. Here I get that to happen by reading ALL of memory.
1590 // A better answer would be to separate out card marks from other memory.
1591 // For now, return the input memory state, so that it can be reused
1592 // after the call, if this call has restricted memory effects.
1593 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1594 // Set fixed predefined input arguments
1595 Node* memory = reset_memory();
1596 call->init_req( TypeFunc::Control, control() );
1597 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1598 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1599 call->init_req( TypeFunc::FramePtr, frameptr() );
1600 call->init_req( TypeFunc::ReturnAdr, top() );
1601 return memory;
1602 }
1603
1604 //-------------------set_predefined_output_for_runtime_call--------------------
1605 // Set control and memory (not i_o) from the call.
1606 // If keep_mem is not NULL, use it for the output state,
1607 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1608 // If hook_mem is NULL, this call produces no memory effects at all.
1609 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1610 // then only that memory slice is taken from the call.
1611 // In the last case, we must put an appropriate memory barrier before
1612 // the call, so as to create the correct anti-dependencies on loads
1613 // preceding the call.
1614 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1615 Node* keep_mem,
1616 const TypePtr* hook_mem) {
1617 // no i/o
1618 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1619 if (keep_mem) {
1620 // First clone the existing memory state
1621 set_all_memory(keep_mem);
1622 if (hook_mem != NULL) {
1623 // Make memory for the call
1624 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1625 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1626 // We also use hook_mem to extract specific effects from arraycopy stubs.
1627 set_memory(mem, hook_mem);
1628 }
1629 // ...else the call has NO memory effects.
1630
1631 // Make sure the call advertises its memory effects precisely.
1632 // This lets us build accurate anti-dependences in gcm.cpp.
1633 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1634 "call node must be constructed correctly");
1635 } else {
1636 assert(hook_mem == NULL, "");
1637 // This is not a "slow path" call; all memory comes from the call.
1638 set_all_memory_call(call);
1639 }
1640 }
1641
1642 //------------------------------increment_counter------------------------------
1643 // for statistics: increment a VM counter by 1
1644
1645 void GraphKit::increment_counter(address counter_addr) {
1646 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1647 increment_counter(adr1);
1648 }
1649
1650 void GraphKit::increment_counter(Node* counter_addr) {
1651 int adr_type = Compile::AliasIdxRaw;
1652 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
1653 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1654 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type );
1655 }
1656
1657
1658 //------------------------------uncommon_trap----------------------------------
1659 // Bail out to the interpreter in mid-method. Implemented by calling the
1660 // uncommon_trap blob. This helper function inserts a runtime call with the
1661 // right debug info.
1662 void GraphKit::uncommon_trap(int trap_request,
1663 ciKlass* klass, const char* comment,
1664 bool must_throw,
1665 bool keep_exact_action) {
1666 if (failing()) stop();
1667 if (stopped()) return; // trap reachable?
1668
1669 // Note: If ProfileTraps is true, and if a deopt. actually
1670 // occurs here, the runtime will make sure an MDO exists. There is
1671 // no need to call method()->build_method_data() at this point.
1672
1673 #ifdef ASSERT
1674 if (!must_throw) {
1675 // Make sure the stack has at least enough depth to execute
1676 // the current bytecode.
1677 int inputs, ignore;
1678 if (compute_stack_effects(inputs, ignore)) {
1679 assert(sp() >= inputs, "must have enough JVMS stack to execute");
1680 // It is a frequent error in library_call.cpp to issue an
1681 // uncommon trap with the _sp value already popped.
1682 }
1683 }
1684 #endif
1685
1686 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1687 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1688
1689 switch (action) {
1690 case Deoptimization::Action_maybe_recompile:
1691 case Deoptimization::Action_reinterpret:
1692 // Temporary fix for 6529811 to allow virtual calls to be sure they
1693 // get the chance to go from mono->bi->mega
1694 if (!keep_exact_action &&
1695 Deoptimization::trap_request_index(trap_request) < 0 &&
1696 too_many_recompiles(reason)) {
1697 // This BCI is causing too many recompilations.
1698 action = Deoptimization::Action_none;
1699 trap_request = Deoptimization::make_trap_request(reason, action);
1700 } else {
1701 C->set_trap_can_recompile(true);
1702 }
1703 break;
1704 case Deoptimization::Action_make_not_entrant:
1705 C->set_trap_can_recompile(true);
1706 break;
1707 #ifdef ASSERT
1708 case Deoptimization::Action_none:
1709 case Deoptimization::Action_make_not_compilable:
1710 break;
1711 default:
1712 assert(false, "bad action");
1713 #endif
1714 }
1715
1716 if (TraceOptoParse) {
1717 char buf[100];
1718 tty->print_cr("Uncommon trap %s at bci:%d",
1719 Deoptimization::format_trap_request(buf, sizeof(buf),
1720 trap_request), bci());
1721 }
1722
1723 CompileLog* log = C->log();
1724 if (log != NULL) {
1725 int kid = (klass == NULL)? -1: log->identify(klass);
1726 log->begin_elem("uncommon_trap bci='%d'", bci());
1727 char buf[100];
1728 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1729 trap_request));
1730 if (kid >= 0) log->print(" klass='%d'", kid);
1731 if (comment != NULL) log->print(" comment='%s'", comment);
1732 log->end_elem();
1733 }
1734
1735 // Make sure any guarding test views this path as very unlikely
1736 Node *i0 = control()->in(0);
1737 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1738 IfNode *iff = i0->as_If();
1739 float f = iff->_prob; // Get prob
1740 if (control()->Opcode() == Op_IfTrue) {
1741 if (f > PROB_UNLIKELY_MAG(4))
1742 iff->_prob = PROB_MIN;
1743 } else {
1744 if (f < PROB_LIKELY_MAG(4))
1745 iff->_prob = PROB_MAX;
1746 }
1747 }
1748
1749 // Clear out dead values from the debug info.
1750 kill_dead_locals();
1751
1752 // Now insert the uncommon trap subroutine call
1753 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin();
1754 const TypePtr* no_memory_effects = NULL;
1755 // Pass the index of the class to be loaded
1756 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1757 (must_throw ? RC_MUST_THROW : 0),
1758 OptoRuntime::uncommon_trap_Type(),
1759 call_addr, "uncommon_trap", no_memory_effects,
1760 intcon(trap_request));
1761 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1762 "must extract request correctly from the graph");
1763 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1764
1765 call->set_req(TypeFunc::ReturnAdr, returnadr());
1766 // The debug info is the only real input to this call.
1767
1768 // Halt-and-catch fire here. The above call should never return!
1769 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1770 _gvn.set_type_bottom(halt);
1771 root()->add_req(halt);
1772
1773 stop_and_kill_map();
1774 }
1775
1776
1777 //--------------------------just_allocated_object------------------------------
1778 // Report the object that was just allocated.
1779 // It must be the case that there are no intervening safepoints.
1780 // We use this to determine if an object is so "fresh" that
1781 // it does not require card marks.
1782 Node* GraphKit::just_allocated_object(Node* current_control) {
1783 if (C->recent_alloc_ctl() == current_control)
1784 return C->recent_alloc_obj();
1785 return NULL;
1786 }
1787
1788
1789 //------------------------------store_barrier----------------------------------
1790 // Insert a write-barrier store. This is to let generational GC work; we have
1791 // to flag all oop-stores before the next GC point.
1792 void GraphKit::write_barrier_post(Node* oop_store, Node* obj, Node* adr,
1793 Node* val, bool use_precise) {
1794 // No store check needed if we're storing a NULL or an old object
1795 // (latter case is probably a string constant). The concurrent
1796 // mark sweep garbage collector, however, needs to have all nonNull
1797 // oop updates flagged via card-marks.
1798 if (val != NULL && val->is_Con()) {
1799 // must be either an oop or NULL
1800 const Type* t = val->bottom_type();
1801 if (t == TypePtr::NULL_PTR || t == Type::TOP)
1802 // stores of null never (?) need barriers
1803 return;
1804 ciObject* con = t->is_oopptr()->const_oop();
1805 if (con != NULL
1806 && con->is_perm()
1807 && Universe::heap()->can_elide_permanent_oop_store_barriers())
1808 // no store barrier needed, because no old-to-new ref created
1809 return;
1810 }
1811
1812 if (use_ReduceInitialCardMarks()
1813 && obj == just_allocated_object(control())) {
1814 // We can skip marks on a freshly-allocated object.
1815 // Keep this code in sync with do_eager_card_mark in runtime.cpp.
1816 // That routine eagerly marks the occasional object which is produced
1817 // by the slow path, so that we don't have to do it here.
1818 return;
1819 }
1820
1821 if (!use_precise) {
1822 // All card marks for a (non-array) instance are in one place:
1823 adr = obj;
1824 }
1825 // (Else it's an array (or unknown), and we want more precise card marks.)
1826 assert(adr != NULL, "");
1827
1828 // Get the alias_index for raw card-mark memory
1829 int adr_type = Compile::AliasIdxRaw;
1830 // Convert the pointer to an int prior to doing math on it
1831 Node* cast = _gvn.transform(new (C, 2) CastP2XNode(control(), adr));
1832 // Divide by card size
1833 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
1834 "Only one we handle so far.");
1835 CardTableModRefBS* ct =
1836 (CardTableModRefBS*)(Universe::heap()->barrier_set());
1837 Node *b = _gvn.transform(new (C, 3) URShiftXNode( cast, _gvn.intcon(CardTableModRefBS::card_shift) ));
1838 // We store into a byte array, so do not bother to left-shift by zero
1839 Node *c = byte_map_base_node();
1840 // Combine
1841 Node *sb_ctl = control();
1842 Node *sb_adr = _gvn.transform(new (C, 4) AddPNode( top()/*no base ptr*/, c, b ));
1843 Node *sb_val = _gvn.intcon(0);
1844 // Smash zero into card
1845 if( !UseConcMarkSweepGC ) {
1846 BasicType bt = T_BYTE;
1847 store_to_memory(sb_ctl, sb_adr, sb_val, bt, adr_type);
1848 } else {
1849 // Specialized path for CM store barrier
1850 cms_card_mark( sb_ctl, sb_adr, sb_val, oop_store);
1851 }
1852 }
1853
1854 // Specialized path for CMS store barrier
1855 void GraphKit::cms_card_mark(Node* ctl, Node* adr, Node* val, Node *oop_store) {
1856 BasicType bt = T_BYTE;
1857 int adr_idx = Compile::AliasIdxRaw;
1858 Node* mem = memory(adr_idx);
1859
1860 // The type input is NULL in PRODUCT builds
1861 const TypePtr* type = NULL;
1862 debug_only(type = C->get_adr_type(adr_idx));
1863
1864 // Add required edge to oop_store, optimizer does not support precedence edges.
1865 // Convert required edge to precedence edge before allocation.
1866 Node *store = _gvn.transform( new (C, 5) StoreCMNode(ctl, mem, adr, type, val, oop_store) );
1867 set_memory(store, adr_idx);
1868
1869 // For CMS, back-to-back card-marks can only remove the first one
1870 // and this requires DU info. Push on worklist for optimizer.
1871 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1872 record_for_igvn(store);
1873 }
1874
1875
1876 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1877 // (Note: TypeFunc::make has a cache that makes this fast.)
1878 const TypeFunc* tf = TypeFunc::make(dest_method);
1879 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
1880 for (int j = 0; j < nargs; j++) {
1881 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1882 if( targ->basic_type() == T_DOUBLE ) {
1883 // If any parameters are doubles, they must be rounded before
1884 // the call, dstore_rounding does gvn.transform
1885 Node *arg = argument(j);
1886 arg = dstore_rounding(arg);
1887 set_argument(j, arg);
1888 }
1889 }
1890 }
1891
1892 void GraphKit::round_double_result(ciMethod* dest_method) {
1893 // A non-strict method may return a double value which has an extended
1894 // exponent, but this must not be visible in a caller which is 'strict'
1895 // If a strict caller invokes a non-strict callee, round a double result
1896
1897 BasicType result_type = dest_method->return_type()->basic_type();
1898 assert( method() != NULL, "must have caller context");
1899 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
1900 // Destination method's return value is on top of stack
1901 // dstore_rounding() does gvn.transform
1902 Node *result = pop_pair();
1903 result = dstore_rounding(result);
1904 push_pair(result);
1905 }
1906 }
1907
1908 // rounding for strict float precision conformance
1909 Node* GraphKit::precision_rounding(Node* n) {
1910 return UseStrictFP && _method->flags().is_strict()
1911 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
1912 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
1913 : n;
1914 }
1915
1916 // rounding for strict double precision conformance
1917 Node* GraphKit::dprecision_rounding(Node *n) {
1918 return UseStrictFP && _method->flags().is_strict()
1919 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
1920 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1921 : n;
1922 }
1923
1924 // rounding for non-strict double stores
1925 Node* GraphKit::dstore_rounding(Node* n) {
1926 return Matcher::strict_fp_requires_explicit_rounding
1927 && UseSSE <= 1
1928 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1929 : n;
1930 }
1931
1932 //=============================================================================
1933 // Generate a fast path/slow path idiom. Graph looks like:
1934 // [foo] indicates that 'foo' is a parameter
1935 //
1936 // [in] NULL
1937 // \ /
1938 // CmpP
1939 // Bool ne
1940 // If
1941 // / \
1942 // True False-<2>
1943 // / |
1944 // / cast_not_null
1945 // Load | | ^
1946 // [fast_test] | |
1947 // gvn to opt_test | |
1948 // / \ | <1>
1949 // True False |
1950 // | \\ |
1951 // [slow_call] \[fast_result]
1952 // Ctl Val \ \
1953 // | \ \
1954 // Catch <1> \ \
1955 // / \ ^ \ \
1956 // Ex No_Ex | \ \
1957 // | \ \ | \ <2> \
1958 // ... \ [slow_res] | | \ [null_result]
1959 // \ \--+--+--- | |
1960 // \ | / \ | /
1961 // --------Region Phi
1962 //
1963 //=============================================================================
1964 // Code is structured as a series of driver functions all called 'do_XXX' that
1965 // call a set of helper functions. Helper functions first, then drivers.
1966
1967 //------------------------------null_check_oop---------------------------------
1968 // Null check oop. Set null-path control into Region in slot 3.
1969 // Make a cast-not-nullness use the other not-null control. Return cast.
1970 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
1971 bool never_see_null) {
1972 // Initial NULL check taken path
1973 (*null_control) = top();
1974 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
1975
1976 // Generate uncommon_trap:
1977 if (never_see_null && (*null_control) != top()) {
1978 // If we see an unexpected null at a check-cast we record it and force a
1979 // recompile; the offending check-cast will be compiled to handle NULLs.
1980 // If we see more than one offending BCI, then all checkcasts in the
1981 // method will be compiled to handle NULLs.
1982 PreserveJVMState pjvms(this);
1983 set_control(*null_control);
1984 replace_in_map(value, null());
1985 uncommon_trap(Deoptimization::Reason_null_check,
1986 Deoptimization::Action_make_not_entrant);
1987 (*null_control) = top(); // NULL path is dead
1988 }
1989
1990 // Cast away null-ness on the result
1991 return cast;
1992 }
1993
1994 //------------------------------opt_iff----------------------------------------
1995 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
1996 // Return slow-path control.
1997 Node* GraphKit::opt_iff(Node* region, Node* iff) {
1998 IfNode *opt_iff = _gvn.transform(iff)->as_If();
1999
2000 // Fast path taken; set region slot 2
2001 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2002 region->init_req(2,fast_taken); // Capture fast-control
2003
2004 // Fast path not-taken, i.e. slow path
2005 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2006 return slow_taken;
2007 }
2008
2009 //-----------------------------make_runtime_call-------------------------------
2010 Node* GraphKit::make_runtime_call(int flags,
2011 const TypeFunc* call_type, address call_addr,
2012 const char* call_name,
2013 const TypePtr* adr_type,
2014 // The following parms are all optional.
2015 // The first NULL ends the list.
2016 Node* parm0, Node* parm1,
2017 Node* parm2, Node* parm3,
2018 Node* parm4, Node* parm5,
2019 Node* parm6, Node* parm7) {
2020 // Slow-path call
2021 int size = call_type->domain()->cnt();
2022 bool is_leaf = !(flags & RC_NO_LEAF);
2023 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2024 if (call_name == NULL) {
2025 assert(!is_leaf, "must supply name for leaf");
2026 call_name = OptoRuntime::stub_name(call_addr);
2027 }
2028 CallNode* call;
2029 if (!is_leaf) {
2030 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2031 bci(), adr_type);
2032 } else if (flags & RC_NO_FP) {
2033 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2034 } else {
2035 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2036 }
2037
2038 // The following is similar to set_edges_for_java_call,
2039 // except that the memory effects of the call are restricted to AliasIdxRaw.
2040
2041 // Slow path call has no side-effects, uses few values
2042 bool wide_in = !(flags & RC_NARROW_MEM);
2043 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2044
2045 Node* prev_mem = NULL;
2046 if (wide_in) {
2047 prev_mem = set_predefined_input_for_runtime_call(call);
2048 } else {
2049 assert(!wide_out, "narrow in => narrow out");
2050 Node* narrow_mem = memory(adr_type);
2051 prev_mem = reset_memory();
2052 map()->set_memory(narrow_mem);
2053 set_predefined_input_for_runtime_call(call);
2054 }
2055
2056 // Hook each parm in order. Stop looking at the first NULL.
2057 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2058 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2059 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2060 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2061 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2062 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2063 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2064 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2065 /* close each nested if ===> */ } } } } } } } }
2066 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2067
2068 if (!is_leaf) {
2069 // Non-leaves can block and take safepoints:
2070 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2071 }
2072 // Non-leaves can throw exceptions:
2073 if (has_io) {
2074 call->set_req(TypeFunc::I_O, i_o());
2075 }
2076
2077 if (flags & RC_UNCOMMON) {
2078 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2079 // (An "if" probability corresponds roughly to an unconditional count.
2080 // Sort of.)
2081 call->set_cnt(PROB_UNLIKELY_MAG(4));
2082 }
2083
2084 Node* c = _gvn.transform(call);
2085 assert(c == call, "cannot disappear");
2086
2087 if (wide_out) {
2088 // Slow path call has full side-effects.
2089 set_predefined_output_for_runtime_call(call);
2090 } else {
2091 // Slow path call has few side-effects, and/or sets few values.
2092 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2093 }
2094
2095 if (has_io) {
2096 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2097 }
2098 return call;
2099
2100 }
2101
2102 //------------------------------merge_memory-----------------------------------
2103 // Merge memory from one path into the current memory state.
2104 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2105 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2106 Node* old_slice = mms.force_memory();
2107 Node* new_slice = mms.memory2();
2108 if (old_slice != new_slice) {
2109 PhiNode* phi;
2110 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2111 phi = new_slice->as_Phi();
2112 #ifdef ASSERT
2113 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2114 old_slice = old_slice->in(new_path);
2115 // Caller is responsible for ensuring that any pre-existing
2116 // phis are already aware of old memory.
2117 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2118 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2119 #endif
2120 mms.set_memory(phi);
2121 } else {
2122 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2123 _gvn.set_type(phi, Type::MEMORY);
2124 phi->set_req(new_path, new_slice);
2125 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2126 }
2127 }
2128 }
2129 }
2130
2131 //------------------------------make_slow_call_ex------------------------------
2132 // Make the exception handler hookups for the slow call
2133 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2134 if (stopped()) return;
2135
2136 // Make a catch node with just two handlers: fall-through and catch-all
2137 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2138 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2139 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2140 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2141
2142 { PreserveJVMState pjvms(this);
2143 set_control(excp);
2144 set_i_o(i_o);
2145
2146 if (excp != top()) {
2147 // Create an exception state also.
2148 // Use an exact type if the caller has specified a specific exception.
2149 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2150 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
2151 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2152 }
2153 }
2154
2155 // Get the no-exception control from the CatchNode.
2156 set_control(norm);
2157 }
2158
2159
2160 //-------------------------------gen_subtype_check-----------------------------
2161 // Generate a subtyping check. Takes as input the subtype and supertype.
2162 // Returns 2 values: sets the default control() to the true path and returns
2163 // the false path. Only reads invariant memory; sets no (visible) memory.
2164 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2165 // but that's not exposed to the optimizer. This call also doesn't take in an
2166 // Object; if you wish to check an Object you need to load the Object's class
2167 // prior to coming here.
2168 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2169 // Fast check for identical types, perhaps identical constants.
2170 // The types can even be identical non-constants, in cases
2171 // involving Array.newInstance, Object.clone, etc.
2172 if (subklass == superklass)
2173 return top(); // false path is dead; no test needed.
2174
2175 if (_gvn.type(superklass)->singleton()) {
2176 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2177 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2178
2179 // In the common case of an exact superklass, try to fold up the
2180 // test before generating code. You may ask, why not just generate
2181 // the code and then let it fold up? The answer is that the generated
2182 // code will necessarily include null checks, which do not always
2183 // completely fold away. If they are also needless, then they turn
2184 // into a performance loss. Example:
2185 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2186 // Here, the type of 'fa' is often exact, so the store check
2187 // of fa[1]=x will fold up, without testing the nullness of x.
2188 switch (static_subtype_check(superk, subk)) {
2189 case SSC_always_false:
2190 {
2191 Node* always_fail = control();
2192 set_control(top());
2193 return always_fail;
2194 }
2195 case SSC_always_true:
2196 return top();
2197 case SSC_easy_test:
2198 {
2199 // Just do a direct pointer compare and be done.
2200 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2201 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2202 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2203 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2204 return _gvn.transform( new(C, 1) IfFalseNode(iff) );
2205 }
2206 case SSC_full_test:
2207 break;
2208 default:
2209 ShouldNotReachHere();
2210 }
2211 }
2212
2213 // %%% Possible further optimization: Even if the superklass is not exact,
2214 // if the subklass is the unique subtype of the superklass, the check
2215 // will always succeed. We could leave a dependency behind to ensure this.
2216
2217 // First load the super-klass's check-offset
2218 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
2219 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2220 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
2221 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2222
2223 // Load from the sub-klass's super-class display list, or a 1-word cache of
2224 // the secondary superclass list, or a failing value with a sentinel offset
2225 // if the super-klass is an interface or exceptionally deep in the Java
2226 // hierarchy and we have to scan the secondary superclass list the hard way.
2227 // Worst-case type is a little odd: NULL is allowed as a result (usually
2228 // klass loads can never produce a NULL).
2229 Node *chk_off_X = ConvI2X(chk_off);
2230 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2231 // For some types like interfaces the following loadKlass is from a 1-word
2232 // cache which is mutable so can't use immutable memory. Other
2233 // types load from the super-class display table which is immutable.
2234 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2235 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2236
2237 // Compile speed common case: ARE a subtype and we canNOT fail
2238 if( superklass == nkls )
2239 return top(); // false path is dead; no test needed.
2240
2241 // See if we get an immediate positive hit. Happens roughly 83% of the
2242 // time. Test to see if the value loaded just previously from the subklass
2243 // is exactly the superklass.
2244 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2245 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2246 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2247 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2248 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2249
2250 // Compile speed common case: Check for being deterministic right now. If
2251 // chk_off is a constant and not equal to cacheoff then we are NOT a
2252 // subklass. In this case we need exactly the 1 test above and we can
2253 // return those results immediately.
2254 if (!might_be_cache) {
2255 Node* not_subtype_ctrl = control();
2256 set_control(iftrue1); // We need exactly the 1 test above
2257 return not_subtype_ctrl;
2258 }
2259
2260 // Gather the various success & failures here
2261 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2262 record_for_igvn(r_ok_subtype);
2263 RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2264 record_for_igvn(r_not_subtype);
2265
2266 r_ok_subtype->init_req(1, iftrue1);
2267
2268 // Check for immediate negative hit. Happens roughly 11% of the time (which
2269 // is roughly 63% of the remaining cases). Test to see if the loaded
2270 // check-offset points into the subklass display list or the 1-element
2271 // cache. If it points to the display (and NOT the cache) and the display
2272 // missed then it's not a subtype.
2273 Node *cacheoff = _gvn.intcon(cacheoff_con);
2274 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2275 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2276 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2277 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2278 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2279
2280 // Check for self. Very rare to get here, but its taken 1/3 the time.
2281 // No performance impact (too rare) but allows sharing of secondary arrays
2282 // which has some footprint reduction.
2283 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2284 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2285 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2286 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2287 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2288
2289 // Now do a linear scan of the secondary super-klass array. Again, no real
2290 // performance impact (too rare) but it's gotta be done.
2291 // (The stub also contains the self-check of subklass == superklass.
2292 // Since the code is rarely used, there is no penalty for moving it
2293 // out of line, and it can only improve I-cache density.)
2294 Node* psc = _gvn.transform(
2295 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2296
2297 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2298 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2299 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2300 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2301 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2302
2303 // Return false path; set default control to true path.
2304 set_control( _gvn.transform(r_ok_subtype) );
2305 return _gvn.transform(r_not_subtype);
2306 }
2307
2308 //----------------------------static_subtype_check-----------------------------
2309 // Shortcut important common cases when superklass is exact:
2310 // (0) superklass is java.lang.Object (can occur in reflective code)
2311 // (1) subklass is already limited to a subtype of superklass => always ok
2312 // (2) subklass does not overlap with superklass => always fail
2313 // (3) superklass has NO subtypes and we can check with a simple compare.
2314 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2315 if (StressReflectiveCode) {
2316 return SSC_full_test; // Let caller generate the general case.
2317 }
2318
2319 if (superk == env()->Object_klass()) {
2320 return SSC_always_true; // (0) this test cannot fail
2321 }
2322
2323 ciType* superelem = superk;
2324 if (superelem->is_array_klass())
2325 superelem = superelem->as_array_klass()->base_element_type();
2326
2327 if (!subk->is_interface()) { // cannot trust static interface types yet
2328 if (subk->is_subtype_of(superk)) {
2329 return SSC_always_true; // (1) false path dead; no dynamic test needed
2330 }
2331 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2332 !superk->is_subtype_of(subk)) {
2333 return SSC_always_false;
2334 }
2335 }
2336
2337 // If casting to an instance klass, it must have no subtypes
2338 if (superk->is_interface()) {
2339 // Cannot trust interfaces yet.
2340 // %%% S.B. superk->nof_implementors() == 1
2341 } else if (superelem->is_instance_klass()) {
2342 ciInstanceKlass* ik = superelem->as_instance_klass();
2343 if (!ik->has_subklass() && !ik->is_interface()) {
2344 if (!ik->is_final()) {
2345 // Add a dependency if there is a chance of a later subclass.
2346 C->dependencies()->assert_leaf_type(ik);
2347 }
2348 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2349 }
2350 } else {
2351 // A primitive array type has no subtypes.
2352 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2353 }
2354
2355 return SSC_full_test;
2356 }
2357
2358 // Profile-driven exact type check:
2359 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2360 float prob,
2361 Node* *casted_receiver) {
2362 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2363 Node* recv_klass = load_object_klass(receiver);
2364 Node* want_klass = makecon(tklass);
2365 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2366 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2367 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2368 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2369 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2370
2371 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2372 assert(recv_xtype->klass_is_exact(), "");
2373
2374 // Subsume downstream occurrences of receiver with a cast to
2375 // recv_xtype, since now we know what the type will be.
2376 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2377 (*casted_receiver) = _gvn.transform(cast);
2378 // (User must make the replace_in_map call.)
2379
2380 return fail;
2381 }
2382
2383
2384 //-------------------------------gen_instanceof--------------------------------
2385 // Generate an instance-of idiom. Used by both the instance-of bytecode
2386 // and the reflective instance-of call.
2387 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) {
2388 C->set_has_split_ifs(true); // Has chance for split-if optimization
2389 assert( !stopped(), "dead parse path should be checked in callers" );
2390 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2391 "must check for not-null not-dead klass in callers");
2392
2393 // Make the merge point
2394 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2395 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2396 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2397 C->set_has_split_ifs(true); // Has chance for split-if optimization
2398
2399 // Null check; get casted pointer; set region slot 3
2400 Node* null_ctl = top();
2401 Node* not_null_obj = null_check_oop(subobj, &null_ctl);
2402
2403 // If not_null_obj is dead, only null-path is taken
2404 if (stopped()) { // Doing instance-of on a NULL?
2405 set_control(null_ctl);
2406 return intcon(0);
2407 }
2408 region->init_req(_null_path, null_ctl);
2409 phi ->init_req(_null_path, intcon(0)); // Set null path value
2410
2411 // Load the object's klass
2412 Node* obj_klass = load_object_klass(not_null_obj);
2413
2414 // Generate the subtype check
2415 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2416
2417 // Plug in the success path to the general merge in slot 1.
2418 region->init_req(_obj_path, control());
2419 phi ->init_req(_obj_path, intcon(1));
2420
2421 // Plug in the failing path to the general merge in slot 2.
2422 region->init_req(_fail_path, not_subtype_ctrl);
2423 phi ->init_req(_fail_path, intcon(0));
2424
2425 // Return final merged results
2426 set_control( _gvn.transform(region) );
2427 record_for_igvn(region);
2428 return _gvn.transform(phi);
2429 }
2430
2431 //-------------------------------gen_checkcast---------------------------------
2432 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2433 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2434 // uncommon-trap paths work. Adjust stack after this call.
2435 // If failure_control is supplied and not null, it is filled in with
2436 // the control edge for the cast failure. Otherwise, an appropriate
2437 // uncommon trap or exception is thrown.
2438 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2439 Node* *failure_control) {
2440 kill_dead_locals(); // Benefit all the uncommon traps
2441 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2442 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2443
2444 // Fast cutout: Check the case that the cast is vacuously true.
2445 // This detects the common cases where the test will short-circuit
2446 // away completely. We do this before we perform the null check,
2447 // because if the test is going to turn into zero code, we don't
2448 // want a residual null check left around. (Causes a slowdown,
2449 // for example, in some objArray manipulations, such as a[i]=a[j].)
2450 if (tk->singleton()) {
2451 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2452 if (objtp != NULL && objtp->klass() != NULL) {
2453 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2454 case SSC_always_true:
2455 return obj;
2456 case SSC_always_false:
2457 // It needs a null check because a null will *pass* the cast check.
2458 // A non-null value will always produce an exception.
2459 return do_null_assert(obj, T_OBJECT);
2460 }
2461 }
2462 }
2463
2464 ciProfileData* data = NULL;
2465 if (failure_control == NULL) { // use MDO in regular case only
2466 assert(java_bc() == Bytecodes::_aastore ||
2467 java_bc() == Bytecodes::_checkcast,
2468 "interpreter profiles type checks only for these BCs");
2469 data = method()->method_data()->bci_to_data(bci());
2470 }
2471
2472 // Make the merge point
2473 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2474 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2475 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
2476 C->set_has_split_ifs(true); // Has chance for split-if optimization
2477
2478 // Use null-cast information if it is available
2479 bool never_see_null = false;
2480 // If we see an unexpected null at a check-cast we record it and force a
2481 // recompile; the offending check-cast will be compiled to handle NULLs.
2482 // If we see several offending BCIs, then all checkcasts in the
2483 // method will be compiled to handle NULLs.
2484 if (UncommonNullCast // Cutout for this technique
2485 && failure_control == NULL // regular case
2486 && obj != null() // And not the -Xcomp stupid case?
2487 && !too_many_traps(Deoptimization::Reason_null_check)) {
2488 // Finally, check the "null_seen" bit from the interpreter.
2489 if (data == NULL || !data->as_BitData()->null_seen()) {
2490 never_see_null = true;
2491 }
2492 }
2493
2494 // Null check; get casted pointer; set region slot 3
2495 Node* null_ctl = top();
2496 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2497
2498 // If not_null_obj is dead, only null-path is taken
2499 if (stopped()) { // Doing instance-of on a NULL?
2500 set_control(null_ctl);
2501 return null();
2502 }
2503 region->init_req(_null_path, null_ctl);
2504 phi ->init_req(_null_path, null()); // Set null path value
2505
2506 Node* cast_obj = NULL; // the casted version of the object
2507
2508 // If the profile has seen exactly one type, narrow to that type.
2509 // (The subsequent subtype check will always fold up.)
2510 if (UseTypeProfile && TypeProfileCasts && data != NULL &&
2511 // Counter has never been decremented (due to cast failure).
2512 // ...This is a reasonable thing to expect. It is true of
2513 // all casts inserted by javac to implement generic types.
2514 data->as_CounterData()->count() >= 0 &&
2515 !too_many_traps(Deoptimization::Reason_class_check)) {
2516 // (No, this isn't a call, but it's enough like a virtual call
2517 // to use the same ciMethod accessor to get the profile info...)
2518 ciCallProfile profile = method()->call_profile_at_bci(bci());
2519 if (profile.count() >= 0 && // no cast failures here
2520 profile.has_receiver(0) &&
2521 profile.morphism() == 1) {
2522 ciKlass* exact_kls = profile.receiver(0);
2523 int ssc = static_subtype_check(tk->klass(), exact_kls);
2524 if (ssc == SSC_always_true) {
2525 // If we narrow the type to match what the type profile sees,
2526 // we can then remove the rest of the cast.
2527 // This is a win, even if the exact_kls is very specific,
2528 // because downstream operations, such as method calls,
2529 // will often benefit from the sharper type.
2530 Node* exact_obj = not_null_obj; // will get updated in place...
2531 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2532 &exact_obj);
2533 { PreserveJVMState pjvms(this);
2534 set_control(slow_ctl);
2535 uncommon_trap(Deoptimization::Reason_class_check,
2536 Deoptimization::Action_maybe_recompile);
2537 }
2538 if (failure_control != NULL) // failure is now impossible
2539 (*failure_control) = top();
2540 replace_in_map(not_null_obj, exact_obj);
2541 // adjust the type of the phi to the exact klass:
2542 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR));
2543 cast_obj = exact_obj;
2544 }
2545 // assert(cast_obj != NULL)... except maybe the profile lied to us.
2546 }
2547 }
2548
2549 if (cast_obj == NULL) {
2550 // Load the object's klass
2551 Node* obj_klass = load_object_klass(not_null_obj);
2552
2553 // Generate the subtype check
2554 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2555
2556 // Plug in success path into the merge
2557 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2558 not_null_obj, toop));
2559 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2560 if (failure_control == NULL) {
2561 if (not_subtype_ctrl != top()) { // If failure is possible
2562 PreserveJVMState pjvms(this);
2563 set_control(not_subtype_ctrl);
2564 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2565 }
2566 } else {
2567 (*failure_control) = not_subtype_ctrl;
2568 }
2569 }
2570
2571 region->init_req(_obj_path, control());
2572 phi ->init_req(_obj_path, cast_obj);
2573
2574 // A merge of NULL or Casted-NotNull obj
2575 Node* res = _gvn.transform(phi);
2576
2577 // Note I do NOT always 'replace_in_map(obj,result)' here.
2578 // if( tk->klass()->can_be_primary_super() )
2579 // This means that if I successfully store an Object into an array-of-String
2580 // I 'forget' that the Object is really now known to be a String. I have to
2581 // do this because we don't have true union types for interfaces - if I store
2582 // a Baz into an array-of-Interface and then tell the optimizer it's an
2583 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2584 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2585 // replace_in_map( obj, res );
2586
2587 // Return final merged results
2588 set_control( _gvn.transform(region) );
2589 record_for_igvn(region);
2590 return res;
2591 }
2592
2593 //------------------------------next_monitor-----------------------------------
2594 // What number should be given to the next monitor?
2595 int GraphKit::next_monitor() {
2596 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2597 int next = current + C->sync_stack_slots();
2598 // Keep the toplevel high water mark current:
2599 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2600 return current;
2601 }
2602
2603 //------------------------------insert_mem_bar---------------------------------
2604 // Memory barrier to avoid floating things around
2605 // The membar serves as a pinch point between both control and all memory slices.
2606 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2607 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2608 mb->init_req(TypeFunc::Control, control());
2609 mb->init_req(TypeFunc::Memory, reset_memory());
2610 Node* membar = _gvn.transform(mb);
2611 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2612 set_all_memory_call(membar);
2613 return membar;
2614 }
2615
2616 //-------------------------insert_mem_bar_volatile----------------------------
2617 // Memory barrier to avoid floating things around
2618 // The membar serves as a pinch point between both control and memory(alias_idx).
2619 // If you want to make a pinch point on all memory slices, do not use this
2620 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2621 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2622 // When Parse::do_put_xxx updates a volatile field, it appends a series
2623 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2624 // The first membar is on the same memory slice as the field store opcode.
2625 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2626 // All the other membars (for other volatile slices, including AliasIdxBot,
2627 // which stands for all unknown volatile slices) are control-dependent
2628 // on the first membar. This prevents later volatile loads or stores
2629 // from sliding up past the just-emitted store.
2630
2631 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2632 mb->set_req(TypeFunc::Control,control());
2633 if (alias_idx == Compile::AliasIdxBot) {
2634 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2635 } else {
2636 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2637 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2638 }
2639 Node* membar = _gvn.transform(mb);
2640 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2641 if (alias_idx == Compile::AliasIdxBot) {
2642 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2643 } else {
2644 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2645 }
2646 return membar;
2647 }
2648
2649 //------------------------------shared_lock------------------------------------
2650 // Emit locking code.
2651 FastLockNode* GraphKit::shared_lock(Node* obj) {
2652 // bci is either a monitorenter bc or InvocationEntryBci
2653 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2654 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2655
2656 if( !GenerateSynchronizationCode )
2657 return NULL; // Not locking things?
2658 if (stopped()) // Dead monitor?
2659 return NULL;
2660
2661 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2662
2663 // Box the stack location
2664 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2665 Node* mem = reset_memory();
2666
2667 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2668 if (PrintPreciseBiasedLockingStatistics) {
2669 // Create the counters for this fast lock.
2670 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2671 }
2672 // Add monitor to debug info for the slow path. If we block inside the
2673 // slow path and de-opt, we need the monitor hanging around
2674 map()->push_monitor( flock );
2675
2676 const TypeFunc *tf = LockNode::lock_type();
2677 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2678
2679 lock->init_req( TypeFunc::Control, control() );
2680 lock->init_req( TypeFunc::Memory , mem );
2681 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2682 lock->init_req( TypeFunc::FramePtr, frameptr() );
2683 lock->init_req( TypeFunc::ReturnAdr, top() );
2684
2685 lock->init_req(TypeFunc::Parms + 0, obj);
2686 lock->init_req(TypeFunc::Parms + 1, box);
2687 lock->init_req(TypeFunc::Parms + 2, flock);
2688 add_safepoint_edges(lock);
2689
2690 lock = _gvn.transform( lock )->as_Lock();
2691
2692 // lock has no side-effects, sets few values
2693 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2694
2695 insert_mem_bar(Op_MemBarAcquire);
2696
2697 // Add this to the worklist so that the lock can be eliminated
2698 record_for_igvn(lock);
2699
2700 #ifndef PRODUCT
2701 if (PrintLockStatistics) {
2702 // Update the counter for this lock. Don't bother using an atomic
2703 // operation since we don't require absolute accuracy.
2704 lock->create_lock_counter(map()->jvms());
2705 int adr_type = Compile::AliasIdxRaw;
2706 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr()));
2707 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
2708 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
2709 store_to_memory(control(), counter_addr, incr, T_INT, adr_type);
2710 }
2711 #endif
2712
2713 return flock;
2714 }
2715
2716
2717 //------------------------------shared_unlock----------------------------------
2718 // Emit unlocking code.
2719 void GraphKit::shared_unlock(Node* box, Node* obj) {
2720 // bci is either a monitorenter bc or InvocationEntryBci
2721 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2722 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2723
2724 if( !GenerateSynchronizationCode )
2725 return;
2726 if (stopped()) { // Dead monitor?
2727 map()->pop_monitor(); // Kill monitor from debug info
2728 return;
2729 }
2730
2731 // Memory barrier to avoid floating things down past the locked region
2732 insert_mem_bar(Op_MemBarRelease);
2733
2734 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2735 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2736 uint raw_idx = Compile::AliasIdxRaw;
2737 unlock->init_req( TypeFunc::Control, control() );
2738 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2739 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2740 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2741 unlock->init_req( TypeFunc::ReturnAdr, top() );
2742
2743 unlock->init_req(TypeFunc::Parms + 0, obj);
2744 unlock->init_req(TypeFunc::Parms + 1, box);
2745 unlock = _gvn.transform(unlock)->as_Unlock();
2746
2747 Node* mem = reset_memory();
2748
2749 // unlock has no side-effects, sets few values
2750 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2751
2752 // Kill monitor from debug info
2753 map()->pop_monitor( );
2754 }
2755
2756 //-------------------------------get_layout_helper-----------------------------
2757 // If the given klass is a constant or known to be an array,
2758 // fetch the constant layout helper value into constant_value
2759 // and return (Node*)NULL. Otherwise, load the non-constant
2760 // layout helper value, and return the node which represents it.
2761 // This two-faced routine is useful because allocation sites
2762 // almost always feature constant types.
2763 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2764 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2765 if (!StressReflectiveCode && inst_klass != NULL) {
2766 ciKlass* klass = inst_klass->klass();
2767 bool xklass = inst_klass->klass_is_exact();
2768 if (xklass || klass->is_array_klass()) {
2769 jint lhelper = klass->layout_helper();
2770 if (lhelper != Klass::_lh_neutral_value) {
2771 constant_value = lhelper;
2772 return (Node*) NULL;
2773 }
2774 }
2775 }
2776 constant_value = Klass::_lh_neutral_value; // put in a known value
2777 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
2778 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2779 }
2780
2781 // We just put in an allocate/initialize with a big raw-memory effect.
2782 // Hook selected additional alias categories on the initialization.
2783 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2784 MergeMemNode* init_in_merge,
2785 Node* init_out_raw) {
2786 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2787 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2788
2789 Node* prevmem = kit.memory(alias_idx);
2790 init_in_merge->set_memory_at(alias_idx, prevmem);
2791 kit.set_memory(init_out_raw, alias_idx);
2792 }
2793
2794 //---------------------------set_output_for_allocation-------------------------
2795 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2796 const TypeOopPtr* oop_type,
2797 bool raw_mem_only) {
2798 int rawidx = Compile::AliasIdxRaw;
2799 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2800 add_safepoint_edges(alloc);
2801 Node* allocx = _gvn.transform(alloc);
2802 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2803 // create memory projection for i_o
2804 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2805 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2806
2807 // create a memory projection as for the normal control path
2808 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2809 set_memory(malloc, rawidx);
2810
2811 // a normal slow-call doesn't change i_o, but an allocation does
2812 // we create a separate i_o projection for the normal control path
2813 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2814 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2815
2816 // put in an initialization barrier
2817 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2818 rawoop)->as_Initialize();
2819 assert(alloc->initialization() == init, "2-way macro link must work");
2820 assert(init ->allocation() == alloc, "2-way macro link must work");
2821 if (ReduceFieldZeroing && !raw_mem_only) {
2822 // Extract memory strands which may participate in the new object's
2823 // initialization, and source them from the new InitializeNode.
2824 // This will allow us to observe initializations when they occur,
2825 // and link them properly (as a group) to the InitializeNode.
2826 assert(init->in(InitializeNode::Memory) == malloc, "");
2827 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
2828 init->set_req(InitializeNode::Memory, minit_in);
2829 record_for_igvn(minit_in); // fold it up later, if possible
2830 Node* minit_out = memory(rawidx);
2831 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
2832 if (oop_type->isa_aryptr()) {
2833 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
2834 int elemidx = C->get_alias_index(telemref);
2835 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
2836 } else if (oop_type->isa_instptr()) {
2837 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
2838 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
2839 ciField* field = ik->nonstatic_field_at(i);
2840 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
2841 continue; // do not bother to track really large numbers of fields
2842 // Find (or create) the alias category for this field:
2843 int fieldidx = C->alias_type(field)->index();
2844 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
2845 }
2846 }
2847 }
2848
2849 // Cast raw oop to the real thing...
2850 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
2851 javaoop = _gvn.transform(javaoop);
2852 C->set_recent_alloc(control(), javaoop);
2853 assert(just_allocated_object(control()) == javaoop, "just allocated");
2854
2855 #ifdef ASSERT
2856 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
2857 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
2858 "Ideal_allocation works");
2859 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
2860 "Ideal_allocation works");
2861 if (alloc->is_AllocateArray()) {
2862 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
2863 "Ideal_allocation works");
2864 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
2865 "Ideal_allocation works");
2866 } else {
2867 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
2868 }
2869 }
2870 #endif //ASSERT
2871
2872 return javaoop;
2873 }
2874
2875 //---------------------------new_instance--------------------------------------
2876 // This routine takes a klass_node which may be constant (for a static type)
2877 // or may be non-constant (for reflective code). It will work equally well
2878 // for either, and the graph will fold nicely if the optimizer later reduces
2879 // the type to a constant.
2880 // The optional arguments are for specialized use by intrinsics:
2881 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
2882 // - If 'raw_mem_only', do not cast the result to an oop.
2883 // - If 'return_size_val', report the the total object size to the caller.
2884 Node* GraphKit::new_instance(Node* klass_node,
2885 Node* extra_slow_test,
2886 bool raw_mem_only, // affect only raw memory
2887 Node* *return_size_val) {
2888 // Compute size in doublewords
2889 // The size is always an integral number of doublewords, represented
2890 // as a positive bytewise size stored in the klass's layout_helper.
2891 // The layout_helper also encodes (in a low bit) the need for a slow path.
2892 jint layout_con = Klass::_lh_neutral_value;
2893 Node* layout_val = get_layout_helper(klass_node, layout_con);
2894 int layout_is_con = (layout_val == NULL);
2895
2896 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
2897 // Generate the initial go-slow test. It's either ALWAYS (return a
2898 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
2899 // case) a computed value derived from the layout_helper.
2900 Node* initial_slow_test = NULL;
2901 if (layout_is_con) {
2902 assert(!StressReflectiveCode, "stress mode does not use these paths");
2903 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
2904 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
2905
2906 } else { // reflective case
2907 // This reflective path is used by Unsafe.allocateInstance.
2908 // (It may be stress-tested by specifying StressReflectiveCode.)
2909 // Basically, we want to get into the VM is there's an illegal argument.
2910 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
2911 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
2912 if (extra_slow_test != intcon(0)) {
2913 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
2914 }
2915 // (Macro-expander will further convert this to a Bool, if necessary.)
2916 }
2917
2918 // Find the size in bytes. This is easy; it's the layout_helper.
2919 // The size value must be valid even if the slow path is taken.
2920 Node* size = NULL;
2921 if (layout_is_con) {
2922 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
2923 } else { // reflective case
2924 // This reflective path is used by clone and Unsafe.allocateInstance.
2925 size = ConvI2X(layout_val);
2926
2927 // Clear the low bits to extract layout_helper_size_in_bytes:
2928 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
2929 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
2930 size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
2931 }
2932 if (return_size_val != NULL) {
2933 (*return_size_val) = size;
2934 }
2935
2936 // This is a precise notnull oop of the klass.
2937 // (Actually, it need not be precise if this is a reflective allocation.)
2938 // It's what we cast the result to.
2939 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
2940 if (!tklass) tklass = TypeKlassPtr::OBJECT;
2941 const TypeOopPtr* oop_type = tklass->as_instance_type();
2942
2943 // Now generate allocation code
2944
2945 // With escape analysis, the entire memory state is needed to be able to
2946 // eliminate the allocation. If the allocations cannot be eliminated, this
2947 // will be optimized to the raw slice when the allocation is expanded.
2948 Node *mem;
2949 if (C->do_escape_analysis()) {
2950 mem = reset_memory();
2951 set_all_memory(mem);
2952 } else {
2953 mem = memory(Compile::AliasIdxRaw);
2954 }
2955
2956 AllocateNode* alloc
2957 = new (C, AllocateNode::ParmLimit)
2958 AllocateNode(C, AllocateNode::alloc_type(),
2959 control(), mem, i_o(),
2960 size, klass_node,
2961 initial_slow_test);
2962
2963 return set_output_for_allocation(alloc, oop_type, raw_mem_only);
2964 }
2965
2966 //-------------------------------new_array-------------------------------------
2967 // helper for both newarray and anewarray
2968 // The 'length' parameter is (obviously) the length of the array.
2969 // See comments on new_instance for the meaning of the other arguments.
2970 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
2971 Node* length, // number of array elements
2972 bool raw_mem_only, // affect only raw memory
2973 Node* *return_size_val) {
2974 jint layout_con = Klass::_lh_neutral_value;
2975 Node* layout_val = get_layout_helper(klass_node, layout_con);
2976 int layout_is_con = (layout_val == NULL);
2977
2978 if (!layout_is_con && !StressReflectiveCode &&
2979 !too_many_traps(Deoptimization::Reason_class_check)) {
2980 // This is a reflective array creation site.
2981 // Optimistically assume that it is a subtype of Object[],
2982 // so that we can fold up all the address arithmetic.
2983 layout_con = Klass::array_layout_helper(T_OBJECT);
2984 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
2985 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
2986 { BuildCutout unless(this, bol_lh, PROB_MAX);
2987 uncommon_trap(Deoptimization::Reason_class_check,
2988 Deoptimization::Action_maybe_recompile);
2989 }
2990 layout_val = NULL;
2991 layout_is_con = true;
2992 }
2993
2994 // Generate the initial go-slow test. Make sure we do not overflow
2995 // if length is huge (near 2Gig) or negative! We do not need
2996 // exact double-words here, just a close approximation of needed
2997 // double-words. We can't add any offset or rounding bits, lest we
2998 // take a size -1 of bytes and make it positive. Use an unsigned
2999 // compare, so negative sizes look hugely positive.
3000 int fast_size_limit = FastAllocateSizeLimit;
3001 if (layout_is_con) {
3002 assert(!StressReflectiveCode, "stress mode does not use these paths");
3003 // Increase the size limit if we have exact knowledge of array type.
3004 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3005 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3006 }
3007
3008 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3009 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3010 if (initial_slow_test->is_Bool()) {
3011 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3012 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3013 }
3014
3015 // --- Size Computation ---
3016 // array_size = round_to_heap(array_header + (length << elem_shift));
3017 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3018 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3019 // The rounding mask is strength-reduced, if possible.
3020 int round_mask = MinObjAlignmentInBytes - 1;
3021 Node* header_size = NULL;
3022 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3023 // (T_BYTE has the weakest alignment and size restrictions...)
3024 if (layout_is_con) {
3025 int hsize = Klass::layout_helper_header_size(layout_con);
3026 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3027 BasicType etype = Klass::layout_helper_element_type(layout_con);
3028 if ((round_mask & ~right_n_bits(eshift)) == 0)
3029 round_mask = 0; // strength-reduce it if it goes away completely
3030 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3031 assert(header_size_min <= hsize, "generic minimum is smallest");
3032 header_size_min = hsize;
3033 header_size = intcon(hsize + round_mask);
3034 } else {
3035 Node* hss = intcon(Klass::_lh_header_size_shift);
3036 Node* hsm = intcon(Klass::_lh_header_size_mask);
3037 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3038 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3039 Node* mask = intcon(round_mask);
3040 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3041 }
3042
3043 Node* elem_shift = NULL;
3044 if (layout_is_con) {
3045 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3046 if (eshift != 0)
3047 elem_shift = intcon(eshift);
3048 } else {
3049 // There is no need to mask or shift this value.
3050 // The semantics of LShiftINode include an implicit mask to 0x1F.
3051 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3052 elem_shift = layout_val;
3053 }
3054
3055 // Transition to native address size for all offset calculations:
3056 Node* lengthx = ConvI2X(length);
3057 Node* headerx = ConvI2X(header_size);
3058 #ifdef _LP64
3059 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3060 if (tllen != NULL && tllen->_lo < 0) {
3061 // Add a manual constraint to a positive range. Cf. array_element_address.
3062 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3063 if (size_max > tllen->_hi) size_max = tllen->_hi;
3064 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3065 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3066 }
3067 }
3068 #endif
3069
3070 // Combine header size (plus rounding) and body size. Then round down.
3071 // This computation cannot overflow, because it is used only in two
3072 // places, one where the length is sharply limited, and the other
3073 // after a successful allocation.
3074 Node* abody = lengthx;
3075 if (elem_shift != NULL)
3076 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3077 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3078 if (round_mask != 0) {
3079 Node* mask = MakeConX(~round_mask);
3080 size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3081 }
3082 // else if round_mask == 0, the size computation is self-rounding
3083
3084 if (return_size_val != NULL) {
3085 // This is the size
3086 (*return_size_val) = size;
3087 }
3088
3089 // Now generate allocation code
3090
3091 // With escape analysis, the entire memory state is needed to be able to
3092 // eliminate the allocation. If the allocations cannot be eliminated, this
3093 // will be optimized to the raw slice when the allocation is expanded.
3094 Node *mem;
3095 if (C->do_escape_analysis()) {
3096 mem = reset_memory();
3097 set_all_memory(mem);
3098 } else {
3099 mem = memory(Compile::AliasIdxRaw);
3100 }
3101
3102 // Create the AllocateArrayNode and its result projections
3103 AllocateArrayNode* alloc
3104 = new (C, AllocateArrayNode::ParmLimit)
3105 AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3106 control(), mem, i_o(),
3107 size, klass_node,
3108 initial_slow_test,
3109 length);
3110
3111 // Cast to correct type. Note that the klass_node may be constant or not,
3112 // and in the latter case the actual array type will be inexact also.
3113 // (This happens via a non-constant argument to inline_native_newArray.)
3114 // In any case, the value of klass_node provides the desired array type.
3115 const TypeInt* length_type = _gvn.find_int_type(length);
3116 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3117 if (ary_type->isa_aryptr() && length_type != NULL) {
3118 // Try to get a better type than POS for the size
3119 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3120 }
3121
3122 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
3123
3124 // Cast length on remaining path to be as narrow as possible
3125 if (map()->find_edge(length) >= 0) {
3126 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3127 if (ccast != length) {
3128 _gvn.set_type_bottom(ccast);
3129 record_for_igvn(ccast);
3130 replace_in_map(length, ccast);
3131 }
3132 }
3133
3134 return javaoop;
3135 }
3136
3137 // The following "Ideal_foo" functions are placed here because they recognize
3138 // the graph shapes created by the functions immediately above.
3139
3140 //---------------------------Ideal_allocation----------------------------------
3141 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3142 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3143 if (ptr == NULL) { // reduce dumb test in callers
3144 return NULL;
3145 }
3146 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3147 ptr = ptr->in(1);
3148 if (ptr == NULL) return NULL;
3149 }
3150 if (ptr->is_Proj()) {
3151 Node* allo = ptr->in(0);
3152 if (allo != NULL && allo->is_Allocate()) {
3153 return allo->as_Allocate();
3154 }
3155 }
3156 // Report failure to match.
3157 return NULL;
3158 }
3159
3160 // Fancy version which also strips off an offset (and reports it to caller).
3161 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3162 intptr_t& offset) {
3163 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3164 if (base == NULL) return NULL;
3165 return Ideal_allocation(base, phase);
3166 }
3167
3168 // Trace Initialize <- Proj[Parm] <- Allocate
3169 AllocateNode* InitializeNode::allocation() {
3170 Node* rawoop = in(InitializeNode::RawAddress);
3171 if (rawoop->is_Proj()) {
3172 Node* alloc = rawoop->in(0);
3173 if (alloc->is_Allocate()) {
3174 return alloc->as_Allocate();
3175 }
3176 }
3177 return NULL;
3178 }
3179
3180 // Trace Allocate -> Proj[Parm] -> Initialize
3181 InitializeNode* AllocateNode::initialization() {
3182 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3183 if (rawoop == NULL) return NULL;
3184 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3185 Node* init = rawoop->fast_out(i);
3186 if (init->is_Initialize()) {
3187 assert(init->as_Initialize()->allocation() == this, "2-way link");
3188 return init->as_Initialize();
3189 }
3190 }
3191 return NULL;
3192 }
3193
3194 void GraphKit::g1_write_barrier_pre(Node* obj,
3195 Node* adr,
3196 uint alias_idx,
3197 Node* val,
3198 const Type* val_type,
3199 BasicType bt) {
3200 IdealKit ideal(gvn(), control(), merged_memory(), true);
3201 #define __ ideal.
3202 __ declares_done();
3203
3204 Node* thread = __ thread();
3205
3206 Node* no_ctrl = NULL;
3207 Node* no_base = __ top();
3208 Node* zero = __ ConI(0);
3209
3210 float likely = PROB_LIKELY(0.999);
3211 float unlikely = PROB_UNLIKELY(0.999);
3212
3213 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3214 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3215
3216 // Offsets into the thread
3217 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3218 PtrQueue::byte_offset_of_active());
3219 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3220 PtrQueue::byte_offset_of_index());
3221 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3222 PtrQueue::byte_offset_of_buf());
3223 // Now the actual pointers into the thread
3224
3225 // set_control( ctl);
3226
3227 Node* marking_adr = __ AddP(no_base, thread, __ ConX(marking_offset));
3228 Node* buffer_adr = __ AddP(no_base, thread, __ ConX(buffer_offset));
3229 Node* index_adr = __ AddP(no_base, thread, __ ConX(index_offset));
3230
3231 // Now some of the values
3232
3233 Node* marking = __ load(no_ctrl, marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3234 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3235 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3236
3237 // if (!marking)
3238 __ if_then(marking, BoolTest::ne, zero); {
3239
3240 const Type* t1 = adr->bottom_type();
3241 const Type* t2 = val->bottom_type();
3242
3243 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3244 // if (orig != NULL)
3245 __ if_then(orig, BoolTest::ne, null()); {
3246
3247 // load original value
3248 // alias_idx correct??
3249
3250 // is the queue for this thread full?
3251 __ if_then(index, BoolTest::ne, zero, likely); {
3252
3253 // decrement the index
3254 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3255 Node* next_indexX = next_index;
3256 #ifdef _LP64
3257 // We could refine the type for what it's worth
3258 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3259 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3260 #endif // _LP64
3261
3262 // Now get the buffer location we will log the original value into and store it
3263
3264 Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3265 // __ store(__ ctrl(), log_addr, orig, T_OBJECT, C->get_alias_index(TypeOopPtr::BOTTOM));
3266 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw);
3267
3268
3269 // update the index
3270 // __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3271 // This is a hack to force this store to occur before the oop store that is coming up
3272 __ store(__ ctrl(), index_adr, next_index, T_INT, C->get_alias_index(TypeOopPtr::BOTTOM));
3273
3274 } __ else_(); {
3275
3276 // logging buffer is full, call the runtime
3277 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3278 // __ make_leaf_call(tf, OptoRuntime::g1_wb_pre_Java(), "g1_wb_pre", orig, thread);
3279 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, thread);
3280 } __ end_if();
3281 } __ end_if();
3282 } __ end_if();
3283
3284 __ drain_delay_transform();
3285 set_control( __ ctrl());
3286 set_all_memory( __ merged_memory());
3287
3288 #undef __
3289 }
3290
3291 //
3292 // Update the card table and add card address to the queue
3293 //
3294 void GraphKit::g1_mark_card(IdealKit* ideal, Node* card_adr, Node* store, Node* index, Node* index_adr, Node* buffer, const TypeFunc* tf) {
3295 #define __ ideal->
3296 Node* zero = __ ConI(0);
3297 Node* no_base = __ top();
3298 BasicType card_bt = T_BYTE;
3299 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3300 __ storeCM(__ ctrl(), card_adr, zero, store, card_bt, Compile::AliasIdxRaw);
3301
3302 // Now do the queue work
3303 __ if_then(index, BoolTest::ne, zero); {
3304
3305 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3306 Node* next_indexX = next_index;
3307 #ifdef _LP64
3308 // We could refine the type for what it's worth
3309 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3310 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3311 #endif // _LP64
3312 Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3313
3314 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3315 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3316
3317 } __ else_(); {
3318 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3319 } __ end_if();
3320 #undef __
3321 }
3322
3323 void GraphKit::g1_write_barrier_post(Node* store,
3324 Node* obj,
3325 Node* adr,
3326 uint alias_idx,
3327 Node* val,
3328 BasicType bt,
3329 bool use_precise) {
3330 // If we are writing a NULL then we need no post barrier
3331
3332 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3333 // Must be NULL
3334 const Type* t = val->bottom_type();
3335 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3336 // No post barrier if writing NULLx
3337 return;
3338 }
3339
3340 if (!use_precise) {
3341 // All card marks for a (non-array) instance are in one place:
3342 adr = obj;
3343 }
3344 // (Else it's an array (or unknown), and we want more precise card marks.)
3345 assert(adr != NULL, "");
3346
3347 IdealKit ideal(gvn(), control(), merged_memory(), true);
3348 #define __ ideal.
3349 __ declares_done();
3350
3351 Node* thread = __ thread();
3352
3353 Node* no_ctrl = NULL;
3354 Node* no_base = __ top();
3355 float likely = PROB_LIKELY(0.999);
3356 float unlikely = PROB_UNLIKELY(0.999);
3357 Node* zero = __ ConI(0);
3358 Node* zeroX = __ ConX(0);
3359
3360 // Get the alias_index for raw card-mark memory
3361 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3362
3363 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3364
3365 // Offsets into the thread
3366 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3367 PtrQueue::byte_offset_of_index());
3368 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3369 PtrQueue::byte_offset_of_buf());
3370
3371 // Pointers into the thread
3372
3373 Node* buffer_adr = __ AddP(no_base, thread, __ ConX(buffer_offset));
3374 Node* index_adr = __ AddP(no_base, thread, __ ConX(index_offset));
3375
3376 // Now some values
3377
3378 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3379 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3380
3381
3382 // Convert the store obj pointer to an int prior to doing math on it
3383 // Use addr not obj gets accurate card marks
3384
3385 // Node* cast = __ CastPX(no_ctrl, adr /* obj */);
3386
3387 // Must use ctrl to prevent "integerized oop" existing across safepoint
3388 Node* cast = __ CastPX(__ ctrl(), ( use_precise ? adr : obj ));
3389
3390 // Divide pointer by card size
3391 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3392
3393 // Combine card table base and card offset
3394 Node *card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3395
3396 // If we know the value being stored does it cross regions?
3397
3398 if (val != NULL) {
3399 // Does the store cause us to cross regions?
3400
3401 // Should be able to do an unsigned compare of region_size instead of
3402 // and extra shift. Do we have an unsigned compare??
3403 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3404 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3405
3406 // if (xor_res == 0) same region so skip
3407 __ if_then(xor_res, BoolTest::ne, zeroX); {
3408
3409 // No barrier if we are storing a NULL
3410 __ if_then(val, BoolTest::ne, null(), unlikely); {
3411
3412 // Ok must mark the card if not already dirty
3413
3414 // load the original value of the card
3415 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3416
3417 __ if_then(card_val, BoolTest::ne, zero); {
3418 g1_mark_card(&ideal, card_adr, store, index, index_adr, buffer, tf);
3419 } __ end_if();
3420 } __ end_if();
3421 } __ end_if();
3422 } else {
3423 g1_mark_card(&ideal, card_adr, store, index, index_adr, buffer, tf);
3424 }
3425
3426
3427 __ drain_delay_transform();
3428 set_control( __ ctrl());
3429 set_all_memory( __ merged_memory());
3430 #undef __
3431
3432 }