Wed, 01 Jul 2009 20:22:18 -0700
6840775: Multiple JVM crashes seen with 1.6.0_10 through 1.6.0_14
Summary: Put missed reference to allocated array in copyOf() intrinsic into OopMap for the call slow_arraycopy().
Reviewed-by: never
1 /*
2 * Copyright 2001-2009 Sun Microsystems, Inc. All Rights Reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_graphKit.cpp.incl"
28 //----------------------------GraphKit-----------------------------------------
29 // Main utility constructor.
30 GraphKit::GraphKit(JVMState* jvms)
31 : Phase(Phase::Parser),
32 _env(C->env()),
33 _gvn(*C->initial_gvn())
34 {
35 _exceptions = jvms->map()->next_exception();
36 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
37 set_jvms(jvms);
38 }
40 // Private constructor for parser.
41 GraphKit::GraphKit()
42 : Phase(Phase::Parser),
43 _env(C->env()),
44 _gvn(*C->initial_gvn())
45 {
46 _exceptions = NULL;
47 set_map(NULL);
48 debug_only(_sp = -99);
49 debug_only(set_bci(-99));
50 }
54 //---------------------------clean_stack---------------------------------------
55 // Clear away rubbish from the stack area of the JVM state.
56 // This destroys any arguments that may be waiting on the stack.
57 void GraphKit::clean_stack(int from_sp) {
58 SafePointNode* map = this->map();
59 JVMState* jvms = this->jvms();
60 int stk_size = jvms->stk_size();
61 int stkoff = jvms->stkoff();
62 Node* top = this->top();
63 for (int i = from_sp; i < stk_size; i++) {
64 if (map->in(stkoff + i) != top) {
65 map->set_req(stkoff + i, top);
66 }
67 }
68 }
71 //--------------------------------sync_jvms-----------------------------------
72 // Make sure our current jvms agrees with our parse state.
73 JVMState* GraphKit::sync_jvms() const {
74 JVMState* jvms = this->jvms();
75 jvms->set_bci(bci()); // Record the new bci in the JVMState
76 jvms->set_sp(sp()); // Record the new sp in the JVMState
77 assert(jvms_in_sync(), "jvms is now in sync");
78 return jvms;
79 }
81 #ifdef ASSERT
82 bool GraphKit::jvms_in_sync() const {
83 Parse* parse = is_Parse();
84 if (parse == NULL) {
85 if (bci() != jvms()->bci()) return false;
86 if (sp() != (int)jvms()->sp()) return false;
87 return true;
88 }
89 if (jvms()->method() != parse->method()) return false;
90 if (jvms()->bci() != parse->bci()) return false;
91 int jvms_sp = jvms()->sp();
92 if (jvms_sp != parse->sp()) return false;
93 int jvms_depth = jvms()->depth();
94 if (jvms_depth != parse->depth()) return false;
95 return true;
96 }
98 // Local helper checks for special internal merge points
99 // used to accumulate and merge exception states.
100 // They are marked by the region's in(0) edge being the map itself.
101 // Such merge points must never "escape" into the parser at large,
102 // until they have been handed to gvn.transform.
103 static bool is_hidden_merge(Node* reg) {
104 if (reg == NULL) return false;
105 if (reg->is_Phi()) {
106 reg = reg->in(0);
107 if (reg == NULL) return false;
108 }
109 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
110 }
112 void GraphKit::verify_map() const {
113 if (map() == NULL) return; // null map is OK
114 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
115 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
116 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
117 }
119 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
120 assert(ex_map->next_exception() == NULL, "not already part of a chain");
121 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
122 }
123 #endif
125 //---------------------------stop_and_kill_map---------------------------------
126 // Set _map to NULL, signalling a stop to further bytecode execution.
127 // First smash the current map's control to a constant, to mark it dead.
128 void GraphKit::stop_and_kill_map() {
129 SafePointNode* dead_map = stop();
130 if (dead_map != NULL) {
131 dead_map->disconnect_inputs(NULL); // Mark the map as killed.
132 assert(dead_map->is_killed(), "must be so marked");
133 }
134 }
137 //--------------------------------stopped--------------------------------------
138 // Tell if _map is NULL, or control is top.
139 bool GraphKit::stopped() {
140 if (map() == NULL) return true;
141 else if (control() == top()) return true;
142 else return false;
143 }
146 //-----------------------------has_ex_handler----------------------------------
147 // Tell if this method or any caller method has exception handlers.
148 bool GraphKit::has_ex_handler() {
149 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
150 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
151 return true;
152 }
153 }
154 return false;
155 }
157 //------------------------------save_ex_oop------------------------------------
158 // Save an exception without blowing stack contents or other JVM state.
159 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
160 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
161 ex_map->add_req(ex_oop);
162 debug_only(verify_exception_state(ex_map));
163 }
165 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
166 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
167 Node* ex_oop = ex_map->in(ex_map->req()-1);
168 if (clear_it) ex_map->del_req(ex_map->req()-1);
169 return ex_oop;
170 }
172 //-----------------------------saved_ex_oop------------------------------------
173 // Recover a saved exception from its map.
174 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
175 return common_saved_ex_oop(ex_map, false);
176 }
178 //--------------------------clear_saved_ex_oop---------------------------------
179 // Erase a previously saved exception from its map.
180 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
181 return common_saved_ex_oop(ex_map, true);
182 }
184 #ifdef ASSERT
185 //---------------------------has_saved_ex_oop----------------------------------
186 // Erase a previously saved exception from its map.
187 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
188 return ex_map->req() == ex_map->jvms()->endoff()+1;
189 }
190 #endif
192 //-------------------------make_exception_state--------------------------------
193 // Turn the current JVM state into an exception state, appending the ex_oop.
194 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
195 sync_jvms();
196 SafePointNode* ex_map = stop(); // do not manipulate this map any more
197 set_saved_ex_oop(ex_map, ex_oop);
198 return ex_map;
199 }
202 //--------------------------add_exception_state--------------------------------
203 // Add an exception to my list of exceptions.
204 void GraphKit::add_exception_state(SafePointNode* ex_map) {
205 if (ex_map == NULL || ex_map->control() == top()) {
206 return;
207 }
208 #ifdef ASSERT
209 verify_exception_state(ex_map);
210 if (has_exceptions()) {
211 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
212 }
213 #endif
215 // If there is already an exception of exactly this type, merge with it.
216 // In particular, null-checks and other low-level exceptions common up here.
217 Node* ex_oop = saved_ex_oop(ex_map);
218 const Type* ex_type = _gvn.type(ex_oop);
219 if (ex_oop == top()) {
220 // No action needed.
221 return;
222 }
223 assert(ex_type->isa_instptr(), "exception must be an instance");
224 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
225 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
226 // We check sp also because call bytecodes can generate exceptions
227 // both before and after arguments are popped!
228 if (ex_type2 == ex_type
229 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
230 combine_exception_states(ex_map, e2);
231 return;
232 }
233 }
235 // No pre-existing exception of the same type. Chain it on the list.
236 push_exception_state(ex_map);
237 }
239 //-----------------------add_exception_states_from-----------------------------
240 void GraphKit::add_exception_states_from(JVMState* jvms) {
241 SafePointNode* ex_map = jvms->map()->next_exception();
242 if (ex_map != NULL) {
243 jvms->map()->set_next_exception(NULL);
244 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
245 next_map = ex_map->next_exception();
246 ex_map->set_next_exception(NULL);
247 add_exception_state(ex_map);
248 }
249 }
250 }
252 //-----------------------transfer_exceptions_into_jvms-------------------------
253 JVMState* GraphKit::transfer_exceptions_into_jvms() {
254 if (map() == NULL) {
255 // We need a JVMS to carry the exceptions, but the map has gone away.
256 // Create a scratch JVMS, cloned from any of the exception states...
257 if (has_exceptions()) {
258 _map = _exceptions;
259 _map = clone_map();
260 _map->set_next_exception(NULL);
261 clear_saved_ex_oop(_map);
262 debug_only(verify_map());
263 } else {
264 // ...or created from scratch
265 JVMState* jvms = new (C) JVMState(_method, NULL);
266 jvms->set_bci(_bci);
267 jvms->set_sp(_sp);
268 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms));
269 set_jvms(jvms);
270 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
271 set_all_memory(top());
272 while (map()->req() < jvms->endoff()) map()->add_req(top());
273 }
274 // (This is a kludge, in case you didn't notice.)
275 set_control(top());
276 }
277 JVMState* jvms = sync_jvms();
278 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
279 jvms->map()->set_next_exception(_exceptions);
280 _exceptions = NULL; // done with this set of exceptions
281 return jvms;
282 }
284 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
285 assert(is_hidden_merge(dstphi), "must be a special merge node");
286 assert(is_hidden_merge(srcphi), "must be a special merge node");
287 uint limit = srcphi->req();
288 for (uint i = PhiNode::Input; i < limit; i++) {
289 dstphi->add_req(srcphi->in(i));
290 }
291 }
292 static inline void add_one_req(Node* dstphi, Node* src) {
293 assert(is_hidden_merge(dstphi), "must be a special merge node");
294 assert(!is_hidden_merge(src), "must not be a special merge node");
295 dstphi->add_req(src);
296 }
298 //-----------------------combine_exception_states------------------------------
299 // This helper function combines exception states by building phis on a
300 // specially marked state-merging region. These regions and phis are
301 // untransformed, and can build up gradually. The region is marked by
302 // having a control input of its exception map, rather than NULL. Such
303 // regions do not appear except in this function, and in use_exception_state.
304 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
305 if (failing()) return; // dying anyway...
306 JVMState* ex_jvms = ex_map->_jvms;
307 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
308 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
309 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
310 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
311 assert(ex_map->req() == phi_map->req(), "matching maps");
312 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
313 Node* hidden_merge_mark = root();
314 Node* region = phi_map->control();
315 MergeMemNode* phi_mem = phi_map->merged_memory();
316 MergeMemNode* ex_mem = ex_map->merged_memory();
317 if (region->in(0) != hidden_merge_mark) {
318 // The control input is not (yet) a specially-marked region in phi_map.
319 // Make it so, and build some phis.
320 region = new (C, 2) RegionNode(2);
321 _gvn.set_type(region, Type::CONTROL);
322 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
323 region->init_req(1, phi_map->control());
324 phi_map->set_control(region);
325 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
326 record_for_igvn(io_phi);
327 _gvn.set_type(io_phi, Type::ABIO);
328 phi_map->set_i_o(io_phi);
329 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
330 Node* m = mms.memory();
331 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
332 record_for_igvn(m_phi);
333 _gvn.set_type(m_phi, Type::MEMORY);
334 mms.set_memory(m_phi);
335 }
336 }
338 // Either or both of phi_map and ex_map might already be converted into phis.
339 Node* ex_control = ex_map->control();
340 // if there is special marking on ex_map also, we add multiple edges from src
341 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
342 // how wide was the destination phi_map, originally?
343 uint orig_width = region->req();
345 if (add_multiple) {
346 add_n_reqs(region, ex_control);
347 add_n_reqs(phi_map->i_o(), ex_map->i_o());
348 } else {
349 // ex_map has no merges, so we just add single edges everywhere
350 add_one_req(region, ex_control);
351 add_one_req(phi_map->i_o(), ex_map->i_o());
352 }
353 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
354 if (mms.is_empty()) {
355 // get a copy of the base memory, and patch some inputs into it
356 const TypePtr* adr_type = mms.adr_type(C);
357 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
358 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
359 mms.set_memory(phi);
360 // Prepare to append interesting stuff onto the newly sliced phi:
361 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
362 }
363 // Append stuff from ex_map:
364 if (add_multiple) {
365 add_n_reqs(mms.memory(), mms.memory2());
366 } else {
367 add_one_req(mms.memory(), mms.memory2());
368 }
369 }
370 uint limit = ex_map->req();
371 for (uint i = TypeFunc::Parms; i < limit; i++) {
372 // Skip everything in the JVMS after tos. (The ex_oop follows.)
373 if (i == tos) i = ex_jvms->monoff();
374 Node* src = ex_map->in(i);
375 Node* dst = phi_map->in(i);
376 if (src != dst) {
377 PhiNode* phi;
378 if (dst->in(0) != region) {
379 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
380 record_for_igvn(phi);
381 _gvn.set_type(phi, phi->type());
382 phi_map->set_req(i, dst);
383 // Prepare to append interesting stuff onto the new phi:
384 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
385 } else {
386 assert(dst->is_Phi(), "nobody else uses a hidden region");
387 phi = (PhiNode*)dst;
388 }
389 if (add_multiple && src->in(0) == ex_control) {
390 // Both are phis.
391 add_n_reqs(dst, src);
392 } else {
393 while (dst->req() < region->req()) add_one_req(dst, src);
394 }
395 const Type* srctype = _gvn.type(src);
396 if (phi->type() != srctype) {
397 const Type* dsttype = phi->type()->meet(srctype);
398 if (phi->type() != dsttype) {
399 phi->set_type(dsttype);
400 _gvn.set_type(phi, dsttype);
401 }
402 }
403 }
404 }
405 }
407 //--------------------------use_exception_state--------------------------------
408 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
409 if (failing()) { stop(); return top(); }
410 Node* region = phi_map->control();
411 Node* hidden_merge_mark = root();
412 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
413 Node* ex_oop = clear_saved_ex_oop(phi_map);
414 if (region->in(0) == hidden_merge_mark) {
415 // Special marking for internal ex-states. Process the phis now.
416 region->set_req(0, region); // now it's an ordinary region
417 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
418 // Note: Setting the jvms also sets the bci and sp.
419 set_control(_gvn.transform(region));
420 uint tos = jvms()->stkoff() + sp();
421 for (uint i = 1; i < tos; i++) {
422 Node* x = phi_map->in(i);
423 if (x->in(0) == region) {
424 assert(x->is_Phi(), "expected a special phi");
425 phi_map->set_req(i, _gvn.transform(x));
426 }
427 }
428 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
429 Node* x = mms.memory();
430 if (x->in(0) == region) {
431 assert(x->is_Phi(), "nobody else uses a hidden region");
432 mms.set_memory(_gvn.transform(x));
433 }
434 }
435 if (ex_oop->in(0) == region) {
436 assert(ex_oop->is_Phi(), "expected a special phi");
437 ex_oop = _gvn.transform(ex_oop);
438 }
439 } else {
440 set_jvms(phi_map->jvms());
441 }
443 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
444 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
445 return ex_oop;
446 }
448 //---------------------------------java_bc-------------------------------------
449 Bytecodes::Code GraphKit::java_bc() const {
450 ciMethod* method = this->method();
451 int bci = this->bci();
452 if (method != NULL && bci != InvocationEntryBci)
453 return method->java_code_at_bci(bci);
454 else
455 return Bytecodes::_illegal;
456 }
458 //------------------------------builtin_throw----------------------------------
459 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
460 bool must_throw = true;
462 if (env()->jvmti_can_post_exceptions()) {
463 // Do not try anything fancy if we're notifying the VM on every throw.
464 // Cf. case Bytecodes::_athrow in parse2.cpp.
465 uncommon_trap(reason, Deoptimization::Action_none,
466 (ciKlass*)NULL, (char*)NULL, must_throw);
467 return;
468 }
470 // If this particular condition has not yet happened at this
471 // bytecode, then use the uncommon trap mechanism, and allow for
472 // a future recompilation if several traps occur here.
473 // If the throw is hot, try to use a more complicated inline mechanism
474 // which keeps execution inside the compiled code.
475 bool treat_throw_as_hot = false;
476 ciMethodData* md = method()->method_data();
478 if (ProfileTraps) {
479 if (too_many_traps(reason)) {
480 treat_throw_as_hot = true;
481 }
482 // (If there is no MDO at all, assume it is early in
483 // execution, and that any deopts are part of the
484 // startup transient, and don't need to be remembered.)
486 // Also, if there is a local exception handler, treat all throws
487 // as hot if there has been at least one in this method.
488 if (C->trap_count(reason) != 0
489 && method()->method_data()->trap_count(reason) != 0
490 && has_ex_handler()) {
491 treat_throw_as_hot = true;
492 }
493 }
495 // If this throw happens frequently, an uncommon trap might cause
496 // a performance pothole. If there is a local exception handler,
497 // and if this particular bytecode appears to be deoptimizing often,
498 // let us handle the throw inline, with a preconstructed instance.
499 // Note: If the deopt count has blown up, the uncommon trap
500 // runtime is going to flush this nmethod, not matter what.
501 if (treat_throw_as_hot
502 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
503 // If the throw is local, we use a pre-existing instance and
504 // punt on the backtrace. This would lead to a missing backtrace
505 // (a repeat of 4292742) if the backtrace object is ever asked
506 // for its backtrace.
507 // Fixing this remaining case of 4292742 requires some flavor of
508 // escape analysis. Leave that for the future.
509 ciInstance* ex_obj = NULL;
510 switch (reason) {
511 case Deoptimization::Reason_null_check:
512 ex_obj = env()->NullPointerException_instance();
513 break;
514 case Deoptimization::Reason_div0_check:
515 ex_obj = env()->ArithmeticException_instance();
516 break;
517 case Deoptimization::Reason_range_check:
518 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
519 break;
520 case Deoptimization::Reason_class_check:
521 if (java_bc() == Bytecodes::_aastore) {
522 ex_obj = env()->ArrayStoreException_instance();
523 } else {
524 ex_obj = env()->ClassCastException_instance();
525 }
526 break;
527 }
528 if (failing()) { stop(); return; } // exception allocation might fail
529 if (ex_obj != NULL) {
530 // Cheat with a preallocated exception object.
531 if (C->log() != NULL)
532 C->log()->elem("hot_throw preallocated='1' reason='%s'",
533 Deoptimization::trap_reason_name(reason));
534 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
535 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
537 // Clear the detail message of the preallocated exception object.
538 // Weblogic sometimes mutates the detail message of exceptions
539 // using reflection.
540 int offset = java_lang_Throwable::get_detailMessage_offset();
541 const TypePtr* adr_typ = ex_con->add_offset(offset);
543 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
544 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT);
546 add_exception_state(make_exception_state(ex_node));
547 return;
548 }
549 }
551 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
552 // It won't be much cheaper than bailing to the interp., since we'll
553 // have to pass up all the debug-info, and the runtime will have to
554 // create the stack trace.
556 // Usual case: Bail to interpreter.
557 // Reserve the right to recompile if we haven't seen anything yet.
559 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
560 if (treat_throw_as_hot
561 && (method()->method_data()->trap_recompiled_at(bci())
562 || C->too_many_traps(reason))) {
563 // We cannot afford to take more traps here. Suffer in the interpreter.
564 if (C->log() != NULL)
565 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
566 Deoptimization::trap_reason_name(reason),
567 C->trap_count(reason));
568 action = Deoptimization::Action_none;
569 }
571 // "must_throw" prunes the JVM state to include only the stack, if there
572 // are no local exception handlers. This should cut down on register
573 // allocation time and code size, by drastically reducing the number
574 // of in-edges on the call to the uncommon trap.
576 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
577 }
580 //----------------------------PreserveJVMState---------------------------------
581 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
582 debug_only(kit->verify_map());
583 _kit = kit;
584 _map = kit->map(); // preserve the map
585 _sp = kit->sp();
586 kit->set_map(clone_map ? kit->clone_map() : NULL);
587 #ifdef ASSERT
588 _bci = kit->bci();
589 Parse* parser = kit->is_Parse();
590 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->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 }
607 //-----------------------------BuildCutout-------------------------------------
608 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
609 : PreserveJVMState(kit)
610 {
611 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
612 SafePointNode* outer_map = _map; // preserved map is caller's
613 SafePointNode* inner_map = kit->map();
614 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
615 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
616 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
617 }
618 BuildCutout::~BuildCutout() {
619 GraphKit* kit = _kit;
620 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
621 }
624 //------------------------------clone_map--------------------------------------
625 // Implementation of PreserveJVMState
626 //
627 // Only clone_map(...) here. If this function is only used in the
628 // PreserveJVMState class we may want to get rid of this extra
629 // function eventually and do it all there.
631 SafePointNode* GraphKit::clone_map() {
632 if (map() == NULL) return NULL;
634 // Clone the memory edge first
635 Node* mem = MergeMemNode::make(C, map()->memory());
636 gvn().set_type_bottom(mem);
638 SafePointNode *clonemap = (SafePointNode*)map()->clone();
639 JVMState* jvms = this->jvms();
640 JVMState* clonejvms = jvms->clone_shallow(C);
641 clonemap->set_memory(mem);
642 clonemap->set_jvms(clonejvms);
643 clonejvms->set_map(clonemap);
644 record_for_igvn(clonemap);
645 gvn().set_type_bottom(clonemap);
646 return clonemap;
647 }
650 //-----------------------------set_map_clone-----------------------------------
651 void GraphKit::set_map_clone(SafePointNode* m) {
652 _map = m;
653 _map = clone_map();
654 _map->set_next_exception(NULL);
655 debug_only(verify_map());
656 }
659 //----------------------------kill_dead_locals---------------------------------
660 // Detect any locals which are known to be dead, and force them to top.
661 void GraphKit::kill_dead_locals() {
662 // Consult the liveness information for the locals. If any
663 // of them are unused, then they can be replaced by top(). This
664 // should help register allocation time and cut down on the size
665 // of the deoptimization information.
667 // This call is made from many of the bytecode handling
668 // subroutines called from the Big Switch in do_one_bytecode.
669 // Every bytecode which might include a slow path is responsible
670 // for killing its dead locals. The more consistent we
671 // are about killing deads, the fewer useless phis will be
672 // constructed for them at various merge points.
674 // bci can be -1 (InvocationEntryBci). We return the entry
675 // liveness for the method.
677 if (method() == NULL || method()->code_size() == 0) {
678 // We are building a graph for a call to a native method.
679 // All locals are live.
680 return;
681 }
683 ResourceMark rm;
685 // Consult the liveness information for the locals. If any
686 // of them are unused, then they can be replaced by top(). This
687 // should help register allocation time and cut down on the size
688 // of the deoptimization information.
689 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
691 int len = (int)live_locals.size();
692 assert(len <= jvms()->loc_size(), "too many live locals");
693 for (int local = 0; local < len; local++) {
694 if (!live_locals.at(local)) {
695 set_local(local, top());
696 }
697 }
698 }
700 #ifdef ASSERT
701 //-------------------------dead_locals_are_killed------------------------------
702 // Return true if all dead locals are set to top in the map.
703 // Used to assert "clean" debug info at various points.
704 bool GraphKit::dead_locals_are_killed() {
705 if (method() == NULL || method()->code_size() == 0) {
706 // No locals need to be dead, so all is as it should be.
707 return true;
708 }
710 // Make sure somebody called kill_dead_locals upstream.
711 ResourceMark rm;
712 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
713 if (jvms->loc_size() == 0) continue; // no locals to consult
714 SafePointNode* map = jvms->map();
715 ciMethod* method = jvms->method();
716 int bci = jvms->bci();
717 if (jvms == this->jvms()) {
718 bci = this->bci(); // it might not yet be synched
719 }
720 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
721 int len = (int)live_locals.size();
722 if (!live_locals.is_valid() || len == 0)
723 // This method is trivial, or is poisoned by a breakpoint.
724 return true;
725 assert(len == jvms->loc_size(), "live map consistent with locals map");
726 for (int local = 0; local < len; local++) {
727 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
728 if (PrintMiscellaneous && (Verbose || WizardMode)) {
729 tty->print_cr("Zombie local %d: ", local);
730 jvms->dump();
731 }
732 return false;
733 }
734 }
735 }
736 return true;
737 }
739 #endif //ASSERT
741 // Helper function for adding JVMState and debug information to node
742 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
743 // Add the safepoint edges to the call (or other safepoint).
745 // Make sure dead locals are set to top. This
746 // should help register allocation time and cut down on the size
747 // of the deoptimization information.
748 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
750 // Walk the inline list to fill in the correct set of JVMState's
751 // Also fill in the associated edges for each JVMState.
753 JVMState* youngest_jvms = sync_jvms();
755 // Do we need debug info here? If it is a SafePoint and this method
756 // cannot de-opt, then we do NOT need any debug info.
757 bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint);
759 // If we are guaranteed to throw, we can prune everything but the
760 // input to the current bytecode.
761 bool can_prune_locals = false;
762 uint stack_slots_not_pruned = 0;
763 int inputs = 0, depth = 0;
764 if (must_throw) {
765 assert(method() == youngest_jvms->method(), "sanity");
766 if (compute_stack_effects(inputs, depth)) {
767 can_prune_locals = true;
768 stack_slots_not_pruned = inputs;
769 }
770 }
772 if (env()->jvmti_can_examine_or_deopt_anywhere()) {
773 // At any safepoint, this method can get breakpointed, which would
774 // then require an immediate deoptimization.
775 full_info = true;
776 can_prune_locals = false; // do not prune locals
777 stack_slots_not_pruned = 0;
778 }
780 // do not scribble on the input jvms
781 JVMState* out_jvms = youngest_jvms->clone_deep(C);
782 call->set_jvms(out_jvms); // Start jvms list for call node
784 // Presize the call:
785 debug_only(uint non_debug_edges = call->req());
786 call->add_req_batch(top(), youngest_jvms->debug_depth());
787 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
789 // Set up edges so that the call looks like this:
790 // Call [state:] ctl io mem fptr retadr
791 // [parms:] parm0 ... parmN
792 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
793 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
794 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
795 // Note that caller debug info precedes callee debug info.
797 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
798 uint debug_ptr = call->req();
800 // Loop over the map input edges associated with jvms, add them
801 // to the call node, & reset all offsets to match call node array.
802 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
803 uint debug_end = debug_ptr;
804 uint debug_start = debug_ptr - in_jvms->debug_size();
805 debug_ptr = debug_start; // back up the ptr
807 uint p = debug_start; // walks forward in [debug_start, debug_end)
808 uint j, k, l;
809 SafePointNode* in_map = in_jvms->map();
810 out_jvms->set_map(call);
812 if (can_prune_locals) {
813 assert(in_jvms->method() == out_jvms->method(), "sanity");
814 // If the current throw can reach an exception handler in this JVMS,
815 // then we must keep everything live that can reach that handler.
816 // As a quick and dirty approximation, we look for any handlers at all.
817 if (in_jvms->method()->has_exception_handlers()) {
818 can_prune_locals = false;
819 }
820 }
822 // Add the Locals
823 k = in_jvms->locoff();
824 l = in_jvms->loc_size();
825 out_jvms->set_locoff(p);
826 if (full_info && !can_prune_locals) {
827 for (j = 0; j < l; j++)
828 call->set_req(p++, in_map->in(k+j));
829 } else {
830 p += l; // already set to top above by add_req_batch
831 }
833 // Add the Expression Stack
834 k = in_jvms->stkoff();
835 l = in_jvms->sp();
836 out_jvms->set_stkoff(p);
837 if (full_info && !can_prune_locals) {
838 for (j = 0; j < l; j++)
839 call->set_req(p++, in_map->in(k+j));
840 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
841 // Divide stack into {S0,...,S1}, where S0 is set to top.
842 uint s1 = stack_slots_not_pruned;
843 stack_slots_not_pruned = 0; // for next iteration
844 if (s1 > l) s1 = l;
845 uint s0 = l - s1;
846 p += s0; // skip the tops preinstalled by add_req_batch
847 for (j = s0; j < l; j++)
848 call->set_req(p++, in_map->in(k+j));
849 } else {
850 p += l; // already set to top above by add_req_batch
851 }
853 // Add the Monitors
854 k = in_jvms->monoff();
855 l = in_jvms->mon_size();
856 out_jvms->set_monoff(p);
857 for (j = 0; j < l; j++)
858 call->set_req(p++, in_map->in(k+j));
860 // Copy any scalar object fields.
861 k = in_jvms->scloff();
862 l = in_jvms->scl_size();
863 out_jvms->set_scloff(p);
864 for (j = 0; j < l; j++)
865 call->set_req(p++, in_map->in(k+j));
867 // Finish the new jvms.
868 out_jvms->set_endoff(p);
870 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
871 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
872 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
873 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
874 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
875 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
877 // Update the two tail pointers in parallel.
878 out_jvms = out_jvms->caller();
879 in_jvms = in_jvms->caller();
880 }
882 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
884 // Test the correctness of JVMState::debug_xxx accessors:
885 assert(call->jvms()->debug_start() == non_debug_edges, "");
886 assert(call->jvms()->debug_end() == call->req(), "");
887 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
888 }
890 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
891 Bytecodes::Code code = java_bc();
892 if (code == Bytecodes::_wide) {
893 code = method()->java_code_at_bci(bci() + 1);
894 }
896 BasicType rtype = T_ILLEGAL;
897 int rsize = 0;
899 if (code != Bytecodes::_illegal) {
900 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
901 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
902 if (rtype < T_CONFLICT)
903 rsize = type2size[rtype];
904 }
906 switch (code) {
907 case Bytecodes::_illegal:
908 return false;
910 case Bytecodes::_ldc:
911 case Bytecodes::_ldc_w:
912 case Bytecodes::_ldc2_w:
913 inputs = 0;
914 break;
916 case Bytecodes::_dup: inputs = 1; break;
917 case Bytecodes::_dup_x1: inputs = 2; break;
918 case Bytecodes::_dup_x2: inputs = 3; break;
919 case Bytecodes::_dup2: inputs = 2; break;
920 case Bytecodes::_dup2_x1: inputs = 3; break;
921 case Bytecodes::_dup2_x2: inputs = 4; break;
922 case Bytecodes::_swap: inputs = 2; break;
923 case Bytecodes::_arraylength: inputs = 1; break;
925 case Bytecodes::_getstatic:
926 case Bytecodes::_putstatic:
927 case Bytecodes::_getfield:
928 case Bytecodes::_putfield:
929 {
930 bool is_get = (depth >= 0), is_static = (depth & 1);
931 bool ignore;
932 ciBytecodeStream iter(method());
933 iter.reset_to_bci(bci());
934 iter.next();
935 ciField* field = iter.get_field(ignore);
936 int size = field->type()->size();
937 inputs = (is_static ? 0 : 1);
938 if (is_get) {
939 depth = size - inputs;
940 } else {
941 inputs += size; // putxxx pops the value from the stack
942 depth = - inputs;
943 }
944 }
945 break;
947 case Bytecodes::_invokevirtual:
948 case Bytecodes::_invokespecial:
949 case Bytecodes::_invokestatic:
950 case Bytecodes::_invokedynamic:
951 case Bytecodes::_invokeinterface:
952 {
953 bool is_static = (depth == 0);
954 bool ignore;
955 ciBytecodeStream iter(method());
956 iter.reset_to_bci(bci());
957 iter.next();
958 ciMethod* method = iter.get_method(ignore);
959 inputs = method->arg_size_no_receiver();
960 if (!is_static) inputs += 1;
961 int size = method->return_type()->size();
962 depth = size - inputs;
963 }
964 break;
966 case Bytecodes::_multianewarray:
967 {
968 ciBytecodeStream iter(method());
969 iter.reset_to_bci(bci());
970 iter.next();
971 inputs = iter.get_dimensions();
972 assert(rsize == 1, "");
973 depth = rsize - inputs;
974 }
975 break;
977 case Bytecodes::_ireturn:
978 case Bytecodes::_lreturn:
979 case Bytecodes::_freturn:
980 case Bytecodes::_dreturn:
981 case Bytecodes::_areturn:
982 assert(rsize = -depth, "");
983 inputs = rsize;
984 break;
986 case Bytecodes::_jsr:
987 case Bytecodes::_jsr_w:
988 inputs = 0;
989 depth = 1; // S.B. depth=1, not zero
990 break;
992 default:
993 // bytecode produces a typed result
994 inputs = rsize - depth;
995 assert(inputs >= 0, "");
996 break;
997 }
999 #ifdef ASSERT
1000 // spot check
1001 int outputs = depth + inputs;
1002 assert(outputs >= 0, "sanity");
1003 switch (code) {
1004 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1005 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1006 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1007 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1008 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1009 }
1010 #endif //ASSERT
1012 return true;
1013 }
1017 //------------------------------basic_plus_adr---------------------------------
1018 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1019 // short-circuit a common case
1020 if (offset == intcon(0)) return ptr;
1021 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1022 }
1024 Node* GraphKit::ConvI2L(Node* offset) {
1025 // short-circuit a common case
1026 jint offset_con = find_int_con(offset, Type::OffsetBot);
1027 if (offset_con != Type::OffsetBot) {
1028 return longcon((long) offset_con);
1029 }
1030 return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1031 }
1032 Node* GraphKit::ConvL2I(Node* offset) {
1033 // short-circuit a common case
1034 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1035 if (offset_con != (jlong)Type::OffsetBot) {
1036 return intcon((int) offset_con);
1037 }
1038 return _gvn.transform( new (C, 2) ConvL2INode(offset));
1039 }
1041 //-------------------------load_object_klass-----------------------------------
1042 Node* GraphKit::load_object_klass(Node* obj) {
1043 // Special-case a fresh allocation to avoid building nodes:
1044 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1045 if (akls != NULL) return akls;
1046 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1047 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1048 }
1050 //-------------------------load_array_length-----------------------------------
1051 Node* GraphKit::load_array_length(Node* array) {
1052 // Special-case a fresh allocation to avoid building nodes:
1053 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1054 Node *alen;
1055 if (alloc == NULL) {
1056 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1057 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1058 } else {
1059 alen = alloc->Ideal_length();
1060 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_aryptr(), &_gvn);
1061 if (ccast != alen) {
1062 alen = _gvn.transform(ccast);
1063 }
1064 }
1065 return alen;
1066 }
1068 //------------------------------do_null_check----------------------------------
1069 // Helper function to do a NULL pointer check. Returned value is
1070 // the incoming address with NULL casted away. You are allowed to use the
1071 // not-null value only if you are control dependent on the test.
1072 extern int explicit_null_checks_inserted,
1073 explicit_null_checks_elided;
1074 Node* GraphKit::null_check_common(Node* value, BasicType type,
1075 // optional arguments for variations:
1076 bool assert_null,
1077 Node* *null_control) {
1078 assert(!assert_null || null_control == NULL, "not both at once");
1079 if (stopped()) return top();
1080 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1081 // For some performance testing, we may wish to suppress null checking.
1082 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1083 return value;
1084 }
1085 explicit_null_checks_inserted++;
1087 // Construct NULL check
1088 Node *chk = NULL;
1089 switch(type) {
1090 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1091 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1092 case T_ARRAY : // fall through
1093 type = T_OBJECT; // simplify further tests
1094 case T_OBJECT : {
1095 const Type *t = _gvn.type( value );
1097 const TypeInstPtr* tp = t->isa_instptr();
1098 if (tp != NULL && !tp->klass()->is_loaded()
1099 // Only for do_null_check, not any of its siblings:
1100 && !assert_null && null_control == NULL) {
1101 // Usually, any field access or invocation on an unloaded oop type
1102 // will simply fail to link, since the statically linked class is
1103 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1104 // the static class is loaded but the sharper oop type is not.
1105 // Rather than checking for this obscure case in lots of places,
1106 // we simply observe that a null check on an unloaded class
1107 // will always be followed by a nonsense operation, so we
1108 // can just issue the uncommon trap here.
1109 // Our access to the unloaded class will only be correct
1110 // after it has been loaded and initialized, which requires
1111 // a trip through the interpreter.
1112 #ifndef PRODUCT
1113 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1114 #endif
1115 uncommon_trap(Deoptimization::Reason_unloaded,
1116 Deoptimization::Action_reinterpret,
1117 tp->klass(), "!loaded");
1118 return top();
1119 }
1121 if (assert_null) {
1122 // See if the type is contained in NULL_PTR.
1123 // If so, then the value is already null.
1124 if (t->higher_equal(TypePtr::NULL_PTR)) {
1125 explicit_null_checks_elided++;
1126 return value; // Elided null assert quickly!
1127 }
1128 } else {
1129 // See if mixing in the NULL pointer changes type.
1130 // If so, then the NULL pointer was not allowed in the original
1131 // type. In other words, "value" was not-null.
1132 if (t->meet(TypePtr::NULL_PTR) != t) {
1133 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1134 explicit_null_checks_elided++;
1135 return value; // Elided null check quickly!
1136 }
1137 }
1138 chk = new (C, 3) CmpPNode( value, null() );
1139 break;
1140 }
1142 default : ShouldNotReachHere();
1143 }
1144 assert(chk != NULL, "sanity check");
1145 chk = _gvn.transform(chk);
1147 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1148 BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1149 Node *tst = _gvn.transform( btst );
1151 //-----------
1152 // if peephole optimizations occurred, a prior test existed.
1153 // If a prior test existed, maybe it dominates as we can avoid this test.
1154 if (tst != btst && type == T_OBJECT) {
1155 // At this point we want to scan up the CFG to see if we can
1156 // find an identical test (and so avoid this test altogether).
1157 Node *cfg = control();
1158 int depth = 0;
1159 while( depth < 16 ) { // Limit search depth for speed
1160 if( cfg->Opcode() == Op_IfTrue &&
1161 cfg->in(0)->in(1) == tst ) {
1162 // Found prior test. Use "cast_not_null" to construct an identical
1163 // CastPP (and hence hash to) as already exists for the prior test.
1164 // Return that casted value.
1165 if (assert_null) {
1166 replace_in_map(value, null());
1167 return null(); // do not issue the redundant test
1168 }
1169 Node *oldcontrol = control();
1170 set_control(cfg);
1171 Node *res = cast_not_null(value);
1172 set_control(oldcontrol);
1173 explicit_null_checks_elided++;
1174 return res;
1175 }
1176 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1177 if (cfg == NULL) break; // Quit at region nodes
1178 depth++;
1179 }
1180 }
1182 //-----------
1183 // Branch to failure if null
1184 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1185 Deoptimization::DeoptReason reason;
1186 if (assert_null)
1187 reason = Deoptimization::Reason_null_assert;
1188 else if (type == T_OBJECT)
1189 reason = Deoptimization::Reason_null_check;
1190 else
1191 reason = Deoptimization::Reason_div0_check;
1193 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1194 // ciMethodData::has_trap_at will return a conservative -1 if any
1195 // must-be-null assertion has failed. This could cause performance
1196 // problems for a method after its first do_null_assert failure.
1197 // Consider using 'Reason_class_check' instead?
1199 // To cause an implicit null check, we set the not-null probability
1200 // to the maximum (PROB_MAX). For an explicit check the probability
1201 // is set to a smaller value.
1202 if (null_control != NULL || too_many_traps(reason)) {
1203 // probability is less likely
1204 ok_prob = PROB_LIKELY_MAG(3);
1205 } else if (!assert_null &&
1206 (ImplicitNullCheckThreshold > 0) &&
1207 method() != NULL &&
1208 (method()->method_data()->trap_count(reason)
1209 >= (uint)ImplicitNullCheckThreshold)) {
1210 ok_prob = PROB_LIKELY_MAG(3);
1211 }
1213 if (null_control != NULL) {
1214 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1215 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1216 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
1217 if (null_true == top())
1218 explicit_null_checks_elided++;
1219 (*null_control) = null_true;
1220 } else {
1221 BuildCutout unless(this, tst, ok_prob);
1222 // Check for optimizer eliding test at parse time
1223 if (stopped()) {
1224 // Failure not possible; do not bother making uncommon trap.
1225 explicit_null_checks_elided++;
1226 } else if (assert_null) {
1227 uncommon_trap(reason,
1228 Deoptimization::Action_make_not_entrant,
1229 NULL, "assert_null");
1230 } else {
1231 replace_in_map(value, zerocon(type));
1232 builtin_throw(reason);
1233 }
1234 }
1236 // Must throw exception, fall-thru not possible?
1237 if (stopped()) {
1238 return top(); // No result
1239 }
1241 if (assert_null) {
1242 // Cast obj to null on this path.
1243 replace_in_map(value, zerocon(type));
1244 return zerocon(type);
1245 }
1247 // Cast obj to not-null on this path, if there is no null_control.
1248 // (If there is a null_control, a non-null value may come back to haunt us.)
1249 if (type == T_OBJECT) {
1250 Node* cast = cast_not_null(value, false);
1251 if (null_control == NULL || (*null_control) == top())
1252 replace_in_map(value, cast);
1253 value = cast;
1254 }
1256 return value;
1257 }
1260 //------------------------------cast_not_null----------------------------------
1261 // Cast obj to not-null on this path
1262 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1263 const Type *t = _gvn.type(obj);
1264 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1265 // Object is already not-null?
1266 if( t == t_not_null ) return obj;
1268 Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1269 cast->init_req(0, control());
1270 cast = _gvn.transform( cast );
1272 // Scan for instances of 'obj' in the current JVM mapping.
1273 // These instances are known to be not-null after the test.
1274 if (do_replace_in_map)
1275 replace_in_map(obj, cast);
1277 return cast; // Return casted value
1278 }
1281 //--------------------------replace_in_map-------------------------------------
1282 void GraphKit::replace_in_map(Node* old, Node* neww) {
1283 this->map()->replace_edge(old, neww);
1285 // Note: This operation potentially replaces any edge
1286 // on the map. This includes locals, stack, and monitors
1287 // of the current (innermost) JVM state.
1289 // We can consider replacing in caller maps.
1290 // The idea would be that an inlined function's null checks
1291 // can be shared with the entire inlining tree.
1292 // The expense of doing this is that the PreserveJVMState class
1293 // would have to preserve caller states too, with a deep copy.
1294 }
1298 //=============================================================================
1299 //--------------------------------memory---------------------------------------
1300 Node* GraphKit::memory(uint alias_idx) {
1301 MergeMemNode* mem = merged_memory();
1302 Node* p = mem->memory_at(alias_idx);
1303 _gvn.set_type(p, Type::MEMORY); // must be mapped
1304 return p;
1305 }
1307 //-----------------------------reset_memory------------------------------------
1308 Node* GraphKit::reset_memory() {
1309 Node* mem = map()->memory();
1310 // do not use this node for any more parsing!
1311 debug_only( map()->set_memory((Node*)NULL) );
1312 return _gvn.transform( mem );
1313 }
1315 //------------------------------set_all_memory---------------------------------
1316 void GraphKit::set_all_memory(Node* newmem) {
1317 Node* mergemem = MergeMemNode::make(C, newmem);
1318 gvn().set_type_bottom(mergemem);
1319 map()->set_memory(mergemem);
1320 }
1322 //------------------------------set_all_memory_call----------------------------
1323 void GraphKit::set_all_memory_call(Node* call) {
1324 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1325 set_all_memory(newmem);
1326 }
1328 //=============================================================================
1329 //
1330 // parser factory methods for MemNodes
1331 //
1332 // These are layered on top of the factory methods in LoadNode and StoreNode,
1333 // and integrate with the parser's memory state and _gvn engine.
1334 //
1336 // factory methods in "int adr_idx"
1337 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1338 int adr_idx,
1339 bool require_atomic_access) {
1340 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1341 const TypePtr* adr_type = NULL; // debug-mode-only argument
1342 debug_only(adr_type = C->get_adr_type(adr_idx));
1343 Node* mem = memory(adr_idx);
1344 Node* ld;
1345 if (require_atomic_access && bt == T_LONG) {
1346 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1347 } else {
1348 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1349 }
1350 return _gvn.transform(ld);
1351 }
1353 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1354 int adr_idx,
1355 bool require_atomic_access) {
1356 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1357 const TypePtr* adr_type = NULL;
1358 debug_only(adr_type = C->get_adr_type(adr_idx));
1359 Node *mem = memory(adr_idx);
1360 Node* st;
1361 if (require_atomic_access && bt == T_LONG) {
1362 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1363 } else {
1364 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1365 }
1366 st = _gvn.transform(st);
1367 set_memory(st, adr_idx);
1368 // Back-to-back stores can only remove intermediate store with DU info
1369 // so push on worklist for optimizer.
1370 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1371 record_for_igvn(st);
1373 return st;
1374 }
1376 void GraphKit::pre_barrier(Node* ctl,
1377 Node* obj,
1378 Node* adr,
1379 uint adr_idx,
1380 Node *val,
1381 const TypeOopPtr* val_type,
1382 BasicType bt) {
1383 BarrierSet* bs = Universe::heap()->barrier_set();
1384 set_control(ctl);
1385 switch (bs->kind()) {
1386 case BarrierSet::G1SATBCT:
1387 case BarrierSet::G1SATBCTLogging:
1388 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt);
1389 break;
1391 case BarrierSet::CardTableModRef:
1392 case BarrierSet::CardTableExtension:
1393 case BarrierSet::ModRef:
1394 break;
1396 case BarrierSet::Other:
1397 default :
1398 ShouldNotReachHere();
1400 }
1401 }
1403 void GraphKit::post_barrier(Node* ctl,
1404 Node* store,
1405 Node* obj,
1406 Node* adr,
1407 uint adr_idx,
1408 Node *val,
1409 BasicType bt,
1410 bool use_precise) {
1411 BarrierSet* bs = Universe::heap()->barrier_set();
1412 set_control(ctl);
1413 switch (bs->kind()) {
1414 case BarrierSet::G1SATBCT:
1415 case BarrierSet::G1SATBCTLogging:
1416 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1417 break;
1419 case BarrierSet::CardTableModRef:
1420 case BarrierSet::CardTableExtension:
1421 write_barrier_post(store, obj, adr, val, use_precise);
1422 break;
1424 case BarrierSet::ModRef:
1425 break;
1427 case BarrierSet::Other:
1428 default :
1429 ShouldNotReachHere();
1431 }
1432 }
1434 Node* GraphKit::store_oop_to_object(Node* ctl,
1435 Node* obj,
1436 Node* adr,
1437 const TypePtr* adr_type,
1438 Node *val,
1439 const TypeOopPtr* val_type,
1440 BasicType bt) {
1441 uint adr_idx = C->get_alias_index(adr_type);
1442 Node* store;
1443 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
1444 store = store_to_memory(control(), adr, val, bt, adr_idx);
1445 post_barrier(control(), store, obj, adr, adr_idx, val, bt, false);
1446 return store;
1447 }
1449 Node* GraphKit::store_oop_to_array(Node* ctl,
1450 Node* obj,
1451 Node* adr,
1452 const TypePtr* adr_type,
1453 Node *val,
1454 const TypeOopPtr* val_type,
1455 BasicType bt) {
1456 uint adr_idx = C->get_alias_index(adr_type);
1457 Node* store;
1458 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
1459 store = store_to_memory(control(), adr, val, bt, adr_idx);
1460 post_barrier(control(), store, obj, adr, adr_idx, val, bt, true);
1461 return store;
1462 }
1464 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1465 Node* obj,
1466 Node* adr,
1467 const TypePtr* adr_type,
1468 Node *val,
1469 BasicType bt) {
1470 Compile::AliasType* at = C->alias_type(adr_type);
1471 const TypeOopPtr* val_type = NULL;
1472 if (adr_type->isa_instptr()) {
1473 if (at->field() != NULL) {
1474 // known field. This code is a copy of the do_put_xxx logic.
1475 ciField* field = at->field();
1476 if (!field->type()->is_loaded()) {
1477 val_type = TypeInstPtr::BOTTOM;
1478 } else {
1479 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1480 }
1481 }
1482 } else if (adr_type->isa_aryptr()) {
1483 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1484 }
1485 if (val_type == NULL) {
1486 val_type = TypeInstPtr::BOTTOM;
1487 }
1489 uint adr_idx = at->index();
1490 pre_barrier(ctl, obj, adr, adr_idx, val, val_type, bt);
1491 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1492 post_barrier(control(), store, obj, adr, adr_idx, val, bt, true);
1493 return store;
1494 }
1497 //-------------------------array_element_address-------------------------
1498 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1499 const TypeInt* sizetype) {
1500 uint shift = exact_log2(type2aelembytes(elembt));
1501 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1503 // short-circuit a common case (saves lots of confusing waste motion)
1504 jint idx_con = find_int_con(idx, -1);
1505 if (idx_con >= 0) {
1506 intptr_t offset = header + ((intptr_t)idx_con << shift);
1507 return basic_plus_adr(ary, offset);
1508 }
1510 // must be correct type for alignment purposes
1511 Node* base = basic_plus_adr(ary, header);
1512 #ifdef _LP64
1513 // The scaled index operand to AddP must be a clean 64-bit value.
1514 // Java allows a 32-bit int to be incremented to a negative
1515 // value, which appears in a 64-bit register as a large
1516 // positive number. Using that large positive number as an
1517 // operand in pointer arithmetic has bad consequences.
1518 // On the other hand, 32-bit overflow is rare, and the possibility
1519 // can often be excluded, if we annotate the ConvI2L node with
1520 // a type assertion that its value is known to be a small positive
1521 // number. (The prior range check has ensured this.)
1522 // This assertion is used by ConvI2LNode::Ideal.
1523 int index_max = max_jint - 1; // array size is max_jint, index is one less
1524 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1525 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1526 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1527 #endif
1528 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1529 return basic_plus_adr(ary, base, scale);
1530 }
1532 //-------------------------load_array_element-------------------------
1533 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1534 const Type* elemtype = arytype->elem();
1535 BasicType elembt = elemtype->array_element_basic_type();
1536 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1537 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1538 return ld;
1539 }
1541 //-------------------------set_arguments_for_java_call-------------------------
1542 // Arguments (pre-popped from the stack) are taken from the JVMS.
1543 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1544 // Add the call arguments:
1545 uint nargs = call->method()->arg_size();
1546 for (uint i = 0; i < nargs; i++) {
1547 Node* arg = argument(i);
1548 call->init_req(i + TypeFunc::Parms, arg);
1549 }
1550 }
1552 //---------------------------set_edges_for_java_call---------------------------
1553 // Connect a newly created call into the current JVMS.
1554 // A return value node (if any) is returned from set_edges_for_java_call.
1555 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw) {
1557 // Add the predefined inputs:
1558 call->init_req( TypeFunc::Control, control() );
1559 call->init_req( TypeFunc::I_O , i_o() );
1560 call->init_req( TypeFunc::Memory , reset_memory() );
1561 call->init_req( TypeFunc::FramePtr, frameptr() );
1562 call->init_req( TypeFunc::ReturnAdr, top() );
1564 add_safepoint_edges(call, must_throw);
1566 Node* xcall = _gvn.transform(call);
1568 if (xcall == top()) {
1569 set_control(top());
1570 return;
1571 }
1572 assert(xcall == call, "call identity is stable");
1574 // Re-use the current map to produce the result.
1576 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1577 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O )));
1578 set_all_memory_call(xcall);
1580 //return xcall; // no need, caller already has it
1581 }
1583 Node* GraphKit::set_results_for_java_call(CallJavaNode* call) {
1584 if (stopped()) return top(); // maybe the call folded up?
1586 // Capture the return value, if any.
1587 Node* ret;
1588 if (call->method() == NULL ||
1589 call->method()->return_type()->basic_type() == T_VOID)
1590 ret = top();
1591 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1593 // Note: Since any out-of-line call can produce an exception,
1594 // we always insert an I_O projection from the call into the result.
1596 make_slow_call_ex(call, env()->Throwable_klass(), false);
1598 return ret;
1599 }
1601 //--------------------set_predefined_input_for_runtime_call--------------------
1602 // Reading and setting the memory state is way conservative here.
1603 // The real problem is that I am not doing real Type analysis on memory,
1604 // so I cannot distinguish card mark stores from other stores. Across a GC
1605 // point the Store Barrier and the card mark memory has to agree. I cannot
1606 // have a card mark store and its barrier split across the GC point from
1607 // either above or below. Here I get that to happen by reading ALL of memory.
1608 // A better answer would be to separate out card marks from other memory.
1609 // For now, return the input memory state, so that it can be reused
1610 // after the call, if this call has restricted memory effects.
1611 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1612 // Set fixed predefined input arguments
1613 Node* memory = reset_memory();
1614 call->init_req( TypeFunc::Control, control() );
1615 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1616 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1617 call->init_req( TypeFunc::FramePtr, frameptr() );
1618 call->init_req( TypeFunc::ReturnAdr, top() );
1619 return memory;
1620 }
1622 //-------------------set_predefined_output_for_runtime_call--------------------
1623 // Set control and memory (not i_o) from the call.
1624 // If keep_mem is not NULL, use it for the output state,
1625 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1626 // If hook_mem is NULL, this call produces no memory effects at all.
1627 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1628 // then only that memory slice is taken from the call.
1629 // In the last case, we must put an appropriate memory barrier before
1630 // the call, so as to create the correct anti-dependencies on loads
1631 // preceding the call.
1632 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1633 Node* keep_mem,
1634 const TypePtr* hook_mem) {
1635 // no i/o
1636 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1637 if (keep_mem) {
1638 // First clone the existing memory state
1639 set_all_memory(keep_mem);
1640 if (hook_mem != NULL) {
1641 // Make memory for the call
1642 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1643 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1644 // We also use hook_mem to extract specific effects from arraycopy stubs.
1645 set_memory(mem, hook_mem);
1646 }
1647 // ...else the call has NO memory effects.
1649 // Make sure the call advertises its memory effects precisely.
1650 // This lets us build accurate anti-dependences in gcm.cpp.
1651 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1652 "call node must be constructed correctly");
1653 } else {
1654 assert(hook_mem == NULL, "");
1655 // This is not a "slow path" call; all memory comes from the call.
1656 set_all_memory_call(call);
1657 }
1658 }
1660 //------------------------------increment_counter------------------------------
1661 // for statistics: increment a VM counter by 1
1663 void GraphKit::increment_counter(address counter_addr) {
1664 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1665 increment_counter(adr1);
1666 }
1668 void GraphKit::increment_counter(Node* counter_addr) {
1669 int adr_type = Compile::AliasIdxRaw;
1670 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
1671 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1672 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type );
1673 }
1676 //------------------------------uncommon_trap----------------------------------
1677 // Bail out to the interpreter in mid-method. Implemented by calling the
1678 // uncommon_trap blob. This helper function inserts a runtime call with the
1679 // right debug info.
1680 void GraphKit::uncommon_trap(int trap_request,
1681 ciKlass* klass, const char* comment,
1682 bool must_throw,
1683 bool keep_exact_action) {
1684 if (failing()) stop();
1685 if (stopped()) return; // trap reachable?
1687 // Note: If ProfileTraps is true, and if a deopt. actually
1688 // occurs here, the runtime will make sure an MDO exists. There is
1689 // no need to call method()->build_method_data() at this point.
1691 #ifdef ASSERT
1692 if (!must_throw) {
1693 // Make sure the stack has at least enough depth to execute
1694 // the current bytecode.
1695 int inputs, ignore;
1696 if (compute_stack_effects(inputs, ignore)) {
1697 assert(sp() >= inputs, "must have enough JVMS stack to execute");
1698 // It is a frequent error in library_call.cpp to issue an
1699 // uncommon trap with the _sp value already popped.
1700 }
1701 }
1702 #endif
1704 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1705 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1707 switch (action) {
1708 case Deoptimization::Action_maybe_recompile:
1709 case Deoptimization::Action_reinterpret:
1710 // Temporary fix for 6529811 to allow virtual calls to be sure they
1711 // get the chance to go from mono->bi->mega
1712 if (!keep_exact_action &&
1713 Deoptimization::trap_request_index(trap_request) < 0 &&
1714 too_many_recompiles(reason)) {
1715 // This BCI is causing too many recompilations.
1716 action = Deoptimization::Action_none;
1717 trap_request = Deoptimization::make_trap_request(reason, action);
1718 } else {
1719 C->set_trap_can_recompile(true);
1720 }
1721 break;
1722 case Deoptimization::Action_make_not_entrant:
1723 C->set_trap_can_recompile(true);
1724 break;
1725 #ifdef ASSERT
1726 case Deoptimization::Action_none:
1727 case Deoptimization::Action_make_not_compilable:
1728 break;
1729 default:
1730 assert(false, "bad action");
1731 #endif
1732 }
1734 if (TraceOptoParse) {
1735 char buf[100];
1736 tty->print_cr("Uncommon trap %s at bci:%d",
1737 Deoptimization::format_trap_request(buf, sizeof(buf),
1738 trap_request), bci());
1739 }
1741 CompileLog* log = C->log();
1742 if (log != NULL) {
1743 int kid = (klass == NULL)? -1: log->identify(klass);
1744 log->begin_elem("uncommon_trap bci='%d'", bci());
1745 char buf[100];
1746 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1747 trap_request));
1748 if (kid >= 0) log->print(" klass='%d'", kid);
1749 if (comment != NULL) log->print(" comment='%s'", comment);
1750 log->end_elem();
1751 }
1753 // Make sure any guarding test views this path as very unlikely
1754 Node *i0 = control()->in(0);
1755 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1756 IfNode *iff = i0->as_If();
1757 float f = iff->_prob; // Get prob
1758 if (control()->Opcode() == Op_IfTrue) {
1759 if (f > PROB_UNLIKELY_MAG(4))
1760 iff->_prob = PROB_MIN;
1761 } else {
1762 if (f < PROB_LIKELY_MAG(4))
1763 iff->_prob = PROB_MAX;
1764 }
1765 }
1767 // Clear out dead values from the debug info.
1768 kill_dead_locals();
1770 // Now insert the uncommon trap subroutine call
1771 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin();
1772 const TypePtr* no_memory_effects = NULL;
1773 // Pass the index of the class to be loaded
1774 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1775 (must_throw ? RC_MUST_THROW : 0),
1776 OptoRuntime::uncommon_trap_Type(),
1777 call_addr, "uncommon_trap", no_memory_effects,
1778 intcon(trap_request));
1779 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1780 "must extract request correctly from the graph");
1781 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1783 call->set_req(TypeFunc::ReturnAdr, returnadr());
1784 // The debug info is the only real input to this call.
1786 // Halt-and-catch fire here. The above call should never return!
1787 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1788 _gvn.set_type_bottom(halt);
1789 root()->add_req(halt);
1791 stop_and_kill_map();
1792 }
1795 //--------------------------just_allocated_object------------------------------
1796 // Report the object that was just allocated.
1797 // It must be the case that there are no intervening safepoints.
1798 // We use this to determine if an object is so "fresh" that
1799 // it does not require card marks.
1800 Node* GraphKit::just_allocated_object(Node* current_control) {
1801 if (C->recent_alloc_ctl() == current_control)
1802 return C->recent_alloc_obj();
1803 return NULL;
1804 }
1807 //------------------------------store_barrier----------------------------------
1808 // Insert a write-barrier store. This is to let generational GC work; we have
1809 // to flag all oop-stores before the next GC point.
1810 void GraphKit::write_barrier_post(Node* oop_store, Node* obj, Node* adr,
1811 Node* val, bool use_precise) {
1812 // No store check needed if we're storing a NULL or an old object
1813 // (latter case is probably a string constant). The concurrent
1814 // mark sweep garbage collector, however, needs to have all nonNull
1815 // oop updates flagged via card-marks.
1816 if (val != NULL && val->is_Con()) {
1817 // must be either an oop or NULL
1818 const Type* t = val->bottom_type();
1819 if (t == TypePtr::NULL_PTR || t == Type::TOP)
1820 // stores of null never (?) need barriers
1821 return;
1822 ciObject* con = t->is_oopptr()->const_oop();
1823 if (con != NULL
1824 && con->is_perm()
1825 && Universe::heap()->can_elide_permanent_oop_store_barriers())
1826 // no store barrier needed, because no old-to-new ref created
1827 return;
1828 }
1830 if (use_ReduceInitialCardMarks()
1831 && obj == just_allocated_object(control())) {
1832 // We can skip marks on a freshly-allocated object.
1833 // Keep this code in sync with do_eager_card_mark in runtime.cpp.
1834 // That routine eagerly marks the occasional object which is produced
1835 // by the slow path, so that we don't have to do it here.
1836 return;
1837 }
1839 if (!use_precise) {
1840 // All card marks for a (non-array) instance are in one place:
1841 adr = obj;
1842 }
1843 // (Else it's an array (or unknown), and we want more precise card marks.)
1844 assert(adr != NULL, "");
1846 // Get the alias_index for raw card-mark memory
1847 int adr_type = Compile::AliasIdxRaw;
1848 // Convert the pointer to an int prior to doing math on it
1849 Node* cast = _gvn.transform(new (C, 2) CastP2XNode(control(), adr));
1850 // Divide by card size
1851 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
1852 "Only one we handle so far.");
1853 CardTableModRefBS* ct =
1854 (CardTableModRefBS*)(Universe::heap()->barrier_set());
1855 Node *b = _gvn.transform(new (C, 3) URShiftXNode( cast, _gvn.intcon(CardTableModRefBS::card_shift) ));
1856 // We store into a byte array, so do not bother to left-shift by zero
1857 Node *c = byte_map_base_node();
1858 // Combine
1859 Node *sb_ctl = control();
1860 Node *sb_adr = _gvn.transform(new (C, 4) AddPNode( top()/*no base ptr*/, c, b ));
1861 Node *sb_val = _gvn.intcon(0);
1862 // Smash zero into card
1863 if( !UseConcMarkSweepGC ) {
1864 BasicType bt = T_BYTE;
1865 store_to_memory(sb_ctl, sb_adr, sb_val, bt, adr_type);
1866 } else {
1867 // Specialized path for CM store barrier
1868 cms_card_mark( sb_ctl, sb_adr, sb_val, oop_store);
1869 }
1870 }
1872 // Specialized path for CMS store barrier
1873 void GraphKit::cms_card_mark(Node* ctl, Node* adr, Node* val, Node *oop_store) {
1874 BasicType bt = T_BYTE;
1875 int adr_idx = Compile::AliasIdxRaw;
1876 Node* mem = memory(adr_idx);
1878 // The type input is NULL in PRODUCT builds
1879 const TypePtr* type = NULL;
1880 debug_only(type = C->get_adr_type(adr_idx));
1882 // Add required edge to oop_store, optimizer does not support precedence edges.
1883 // Convert required edge to precedence edge before allocation.
1884 Node *store = _gvn.transform( new (C, 5) StoreCMNode(ctl, mem, adr, type, val, oop_store) );
1885 set_memory(store, adr_idx);
1887 // For CMS, back-to-back card-marks can only remove the first one
1888 // and this requires DU info. Push on worklist for optimizer.
1889 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1890 record_for_igvn(store);
1891 }
1894 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1895 // (Note: TypeFunc::make has a cache that makes this fast.)
1896 const TypeFunc* tf = TypeFunc::make(dest_method);
1897 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
1898 for (int j = 0; j < nargs; j++) {
1899 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1900 if( targ->basic_type() == T_DOUBLE ) {
1901 // If any parameters are doubles, they must be rounded before
1902 // the call, dstore_rounding does gvn.transform
1903 Node *arg = argument(j);
1904 arg = dstore_rounding(arg);
1905 set_argument(j, arg);
1906 }
1907 }
1908 }
1910 void GraphKit::round_double_result(ciMethod* dest_method) {
1911 // A non-strict method may return a double value which has an extended
1912 // exponent, but this must not be visible in a caller which is 'strict'
1913 // If a strict caller invokes a non-strict callee, round a double result
1915 BasicType result_type = dest_method->return_type()->basic_type();
1916 assert( method() != NULL, "must have caller context");
1917 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
1918 // Destination method's return value is on top of stack
1919 // dstore_rounding() does gvn.transform
1920 Node *result = pop_pair();
1921 result = dstore_rounding(result);
1922 push_pair(result);
1923 }
1924 }
1926 // rounding for strict float precision conformance
1927 Node* GraphKit::precision_rounding(Node* n) {
1928 return UseStrictFP && _method->flags().is_strict()
1929 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
1930 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
1931 : n;
1932 }
1934 // rounding for strict double precision conformance
1935 Node* GraphKit::dprecision_rounding(Node *n) {
1936 return UseStrictFP && _method->flags().is_strict()
1937 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
1938 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1939 : n;
1940 }
1942 // rounding for non-strict double stores
1943 Node* GraphKit::dstore_rounding(Node* n) {
1944 return Matcher::strict_fp_requires_explicit_rounding
1945 && UseSSE <= 1
1946 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1947 : n;
1948 }
1950 //=============================================================================
1951 // Generate a fast path/slow path idiom. Graph looks like:
1952 // [foo] indicates that 'foo' is a parameter
1953 //
1954 // [in] NULL
1955 // \ /
1956 // CmpP
1957 // Bool ne
1958 // If
1959 // / \
1960 // True False-<2>
1961 // / |
1962 // / cast_not_null
1963 // Load | | ^
1964 // [fast_test] | |
1965 // gvn to opt_test | |
1966 // / \ | <1>
1967 // True False |
1968 // | \\ |
1969 // [slow_call] \[fast_result]
1970 // Ctl Val \ \
1971 // | \ \
1972 // Catch <1> \ \
1973 // / \ ^ \ \
1974 // Ex No_Ex | \ \
1975 // | \ \ | \ <2> \
1976 // ... \ [slow_res] | | \ [null_result]
1977 // \ \--+--+--- | |
1978 // \ | / \ | /
1979 // --------Region Phi
1980 //
1981 //=============================================================================
1982 // Code is structured as a series of driver functions all called 'do_XXX' that
1983 // call a set of helper functions. Helper functions first, then drivers.
1985 //------------------------------null_check_oop---------------------------------
1986 // Null check oop. Set null-path control into Region in slot 3.
1987 // Make a cast-not-nullness use the other not-null control. Return cast.
1988 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
1989 bool never_see_null) {
1990 // Initial NULL check taken path
1991 (*null_control) = top();
1992 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
1994 // Generate uncommon_trap:
1995 if (never_see_null && (*null_control) != top()) {
1996 // If we see an unexpected null at a check-cast we record it and force a
1997 // recompile; the offending check-cast will be compiled to handle NULLs.
1998 // If we see more than one offending BCI, then all checkcasts in the
1999 // method will be compiled to handle NULLs.
2000 PreserveJVMState pjvms(this);
2001 set_control(*null_control);
2002 replace_in_map(value, null());
2003 uncommon_trap(Deoptimization::Reason_null_check,
2004 Deoptimization::Action_make_not_entrant);
2005 (*null_control) = top(); // NULL path is dead
2006 }
2008 // Cast away null-ness on the result
2009 return cast;
2010 }
2012 //------------------------------opt_iff----------------------------------------
2013 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2014 // Return slow-path control.
2015 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2016 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2018 // Fast path taken; set region slot 2
2019 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2020 region->init_req(2,fast_taken); // Capture fast-control
2022 // Fast path not-taken, i.e. slow path
2023 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2024 return slow_taken;
2025 }
2027 //-----------------------------make_runtime_call-------------------------------
2028 Node* GraphKit::make_runtime_call(int flags,
2029 const TypeFunc* call_type, address call_addr,
2030 const char* call_name,
2031 const TypePtr* adr_type,
2032 // The following parms are all optional.
2033 // The first NULL ends the list.
2034 Node* parm0, Node* parm1,
2035 Node* parm2, Node* parm3,
2036 Node* parm4, Node* parm5,
2037 Node* parm6, Node* parm7) {
2038 // Slow-path call
2039 int size = call_type->domain()->cnt();
2040 bool is_leaf = !(flags & RC_NO_LEAF);
2041 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2042 if (call_name == NULL) {
2043 assert(!is_leaf, "must supply name for leaf");
2044 call_name = OptoRuntime::stub_name(call_addr);
2045 }
2046 CallNode* call;
2047 if (!is_leaf) {
2048 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2049 bci(), adr_type);
2050 } else if (flags & RC_NO_FP) {
2051 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2052 } else {
2053 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2054 }
2056 // The following is similar to set_edges_for_java_call,
2057 // except that the memory effects of the call are restricted to AliasIdxRaw.
2059 // Slow path call has no side-effects, uses few values
2060 bool wide_in = !(flags & RC_NARROW_MEM);
2061 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2063 Node* prev_mem = NULL;
2064 if (wide_in) {
2065 prev_mem = set_predefined_input_for_runtime_call(call);
2066 } else {
2067 assert(!wide_out, "narrow in => narrow out");
2068 Node* narrow_mem = memory(adr_type);
2069 prev_mem = reset_memory();
2070 map()->set_memory(narrow_mem);
2071 set_predefined_input_for_runtime_call(call);
2072 }
2074 // Hook each parm in order. Stop looking at the first NULL.
2075 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2076 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2077 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2078 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2079 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2080 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2081 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2082 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2083 /* close each nested if ===> */ } } } } } } } }
2084 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2086 if (!is_leaf) {
2087 // Non-leaves can block and take safepoints:
2088 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2089 }
2090 // Non-leaves can throw exceptions:
2091 if (has_io) {
2092 call->set_req(TypeFunc::I_O, i_o());
2093 }
2095 if (flags & RC_UNCOMMON) {
2096 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2097 // (An "if" probability corresponds roughly to an unconditional count.
2098 // Sort of.)
2099 call->set_cnt(PROB_UNLIKELY_MAG(4));
2100 }
2102 Node* c = _gvn.transform(call);
2103 assert(c == call, "cannot disappear");
2105 if (wide_out) {
2106 // Slow path call has full side-effects.
2107 set_predefined_output_for_runtime_call(call);
2108 } else {
2109 // Slow path call has few side-effects, and/or sets few values.
2110 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2111 }
2113 if (has_io) {
2114 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2115 }
2116 return call;
2118 }
2120 //------------------------------merge_memory-----------------------------------
2121 // Merge memory from one path into the current memory state.
2122 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2123 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2124 Node* old_slice = mms.force_memory();
2125 Node* new_slice = mms.memory2();
2126 if (old_slice != new_slice) {
2127 PhiNode* phi;
2128 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2129 phi = new_slice->as_Phi();
2130 #ifdef ASSERT
2131 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2132 old_slice = old_slice->in(new_path);
2133 // Caller is responsible for ensuring that any pre-existing
2134 // phis are already aware of old memory.
2135 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2136 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2137 #endif
2138 mms.set_memory(phi);
2139 } else {
2140 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2141 _gvn.set_type(phi, Type::MEMORY);
2142 phi->set_req(new_path, new_slice);
2143 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2144 }
2145 }
2146 }
2147 }
2149 //------------------------------make_slow_call_ex------------------------------
2150 // Make the exception handler hookups for the slow call
2151 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2152 if (stopped()) return;
2154 // Make a catch node with just two handlers: fall-through and catch-all
2155 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2156 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2157 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2158 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2160 { PreserveJVMState pjvms(this);
2161 set_control(excp);
2162 set_i_o(i_o);
2164 if (excp != top()) {
2165 // Create an exception state also.
2166 // Use an exact type if the caller has specified a specific exception.
2167 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2168 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
2169 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2170 }
2171 }
2173 // Get the no-exception control from the CatchNode.
2174 set_control(norm);
2175 }
2178 //-------------------------------gen_subtype_check-----------------------------
2179 // Generate a subtyping check. Takes as input the subtype and supertype.
2180 // Returns 2 values: sets the default control() to the true path and returns
2181 // the false path. Only reads invariant memory; sets no (visible) memory.
2182 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2183 // but that's not exposed to the optimizer. This call also doesn't take in an
2184 // Object; if you wish to check an Object you need to load the Object's class
2185 // prior to coming here.
2186 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2187 // Fast check for identical types, perhaps identical constants.
2188 // The types can even be identical non-constants, in cases
2189 // involving Array.newInstance, Object.clone, etc.
2190 if (subklass == superklass)
2191 return top(); // false path is dead; no test needed.
2193 if (_gvn.type(superklass)->singleton()) {
2194 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2195 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2197 // In the common case of an exact superklass, try to fold up the
2198 // test before generating code. You may ask, why not just generate
2199 // the code and then let it fold up? The answer is that the generated
2200 // code will necessarily include null checks, which do not always
2201 // completely fold away. If they are also needless, then they turn
2202 // into a performance loss. Example:
2203 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2204 // Here, the type of 'fa' is often exact, so the store check
2205 // of fa[1]=x will fold up, without testing the nullness of x.
2206 switch (static_subtype_check(superk, subk)) {
2207 case SSC_always_false:
2208 {
2209 Node* always_fail = control();
2210 set_control(top());
2211 return always_fail;
2212 }
2213 case SSC_always_true:
2214 return top();
2215 case SSC_easy_test:
2216 {
2217 // Just do a direct pointer compare and be done.
2218 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2219 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2220 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2221 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2222 return _gvn.transform( new(C, 1) IfFalseNode(iff) );
2223 }
2224 case SSC_full_test:
2225 break;
2226 default:
2227 ShouldNotReachHere();
2228 }
2229 }
2231 // %%% Possible further optimization: Even if the superklass is not exact,
2232 // if the subklass is the unique subtype of the superklass, the check
2233 // will always succeed. We could leave a dependency behind to ensure this.
2235 // First load the super-klass's check-offset
2236 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
2237 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2238 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
2239 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2241 // Load from the sub-klass's super-class display list, or a 1-word cache of
2242 // the secondary superclass list, or a failing value with a sentinel offset
2243 // if the super-klass is an interface or exceptionally deep in the Java
2244 // hierarchy and we have to scan the secondary superclass list the hard way.
2245 // Worst-case type is a little odd: NULL is allowed as a result (usually
2246 // klass loads can never produce a NULL).
2247 Node *chk_off_X = ConvI2X(chk_off);
2248 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2249 // For some types like interfaces the following loadKlass is from a 1-word
2250 // cache which is mutable so can't use immutable memory. Other
2251 // types load from the super-class display table which is immutable.
2252 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2253 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2255 // Compile speed common case: ARE a subtype and we canNOT fail
2256 if( superklass == nkls )
2257 return top(); // false path is dead; no test needed.
2259 // See if we get an immediate positive hit. Happens roughly 83% of the
2260 // time. Test to see if the value loaded just previously from the subklass
2261 // is exactly the superklass.
2262 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2263 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2264 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2265 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2266 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2268 // Compile speed common case: Check for being deterministic right now. If
2269 // chk_off is a constant and not equal to cacheoff then we are NOT a
2270 // subklass. In this case we need exactly the 1 test above and we can
2271 // return those results immediately.
2272 if (!might_be_cache) {
2273 Node* not_subtype_ctrl = control();
2274 set_control(iftrue1); // We need exactly the 1 test above
2275 return not_subtype_ctrl;
2276 }
2278 // Gather the various success & failures here
2279 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2280 record_for_igvn(r_ok_subtype);
2281 RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2282 record_for_igvn(r_not_subtype);
2284 r_ok_subtype->init_req(1, iftrue1);
2286 // Check for immediate negative hit. Happens roughly 11% of the time (which
2287 // is roughly 63% of the remaining cases). Test to see if the loaded
2288 // check-offset points into the subklass display list or the 1-element
2289 // cache. If it points to the display (and NOT the cache) and the display
2290 // missed then it's not a subtype.
2291 Node *cacheoff = _gvn.intcon(cacheoff_con);
2292 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2293 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2294 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2295 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2296 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2298 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2299 // No performance impact (too rare) but allows sharing of secondary arrays
2300 // which has some footprint reduction.
2301 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2302 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2303 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2304 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2305 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2307 // -- Roads not taken here: --
2308 // We could also have chosen to perform the self-check at the beginning
2309 // of this code sequence, as the assembler does. This would not pay off
2310 // the same way, since the optimizer, unlike the assembler, can perform
2311 // static type analysis to fold away many successful self-checks.
2312 // Non-foldable self checks work better here in second position, because
2313 // the initial primary superclass check subsumes a self-check for most
2314 // types. An exception would be a secondary type like array-of-interface,
2315 // which does not appear in its own primary supertype display.
2316 // Finally, we could have chosen to move the self-check into the
2317 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2318 // dependent manner. But it is worthwhile to have the check here,
2319 // where it can be perhaps be optimized. The cost in code space is
2320 // small (register compare, branch).
2322 // Now do a linear scan of the secondary super-klass array. Again, no real
2323 // performance impact (too rare) but it's gotta be done.
2324 // Since the code is rarely used, there is no penalty for moving it
2325 // out of line, and it can only improve I-cache density.
2326 // The decision to inline or out-of-line this final check is platform
2327 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2328 Node* psc = _gvn.transform(
2329 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2331 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2332 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2333 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2334 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2335 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2337 // Return false path; set default control to true path.
2338 set_control( _gvn.transform(r_ok_subtype) );
2339 return _gvn.transform(r_not_subtype);
2340 }
2342 //----------------------------static_subtype_check-----------------------------
2343 // Shortcut important common cases when superklass is exact:
2344 // (0) superklass is java.lang.Object (can occur in reflective code)
2345 // (1) subklass is already limited to a subtype of superklass => always ok
2346 // (2) subklass does not overlap with superklass => always fail
2347 // (3) superklass has NO subtypes and we can check with a simple compare.
2348 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2349 if (StressReflectiveCode) {
2350 return SSC_full_test; // Let caller generate the general case.
2351 }
2353 if (superk == env()->Object_klass()) {
2354 return SSC_always_true; // (0) this test cannot fail
2355 }
2357 ciType* superelem = superk;
2358 if (superelem->is_array_klass())
2359 superelem = superelem->as_array_klass()->base_element_type();
2361 if (!subk->is_interface()) { // cannot trust static interface types yet
2362 if (subk->is_subtype_of(superk)) {
2363 return SSC_always_true; // (1) false path dead; no dynamic test needed
2364 }
2365 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2366 !superk->is_subtype_of(subk)) {
2367 return SSC_always_false;
2368 }
2369 }
2371 // If casting to an instance klass, it must have no subtypes
2372 if (superk->is_interface()) {
2373 // Cannot trust interfaces yet.
2374 // %%% S.B. superk->nof_implementors() == 1
2375 } else if (superelem->is_instance_klass()) {
2376 ciInstanceKlass* ik = superelem->as_instance_klass();
2377 if (!ik->has_subklass() && !ik->is_interface()) {
2378 if (!ik->is_final()) {
2379 // Add a dependency if there is a chance of a later subclass.
2380 C->dependencies()->assert_leaf_type(ik);
2381 }
2382 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2383 }
2384 } else {
2385 // A primitive array type has no subtypes.
2386 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2387 }
2389 return SSC_full_test;
2390 }
2392 // Profile-driven exact type check:
2393 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2394 float prob,
2395 Node* *casted_receiver) {
2396 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2397 Node* recv_klass = load_object_klass(receiver);
2398 Node* want_klass = makecon(tklass);
2399 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2400 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2401 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2402 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2403 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2405 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2406 assert(recv_xtype->klass_is_exact(), "");
2408 // Subsume downstream occurrences of receiver with a cast to
2409 // recv_xtype, since now we know what the type will be.
2410 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2411 (*casted_receiver) = _gvn.transform(cast);
2412 // (User must make the replace_in_map call.)
2414 return fail;
2415 }
2418 //-------------------------------gen_instanceof--------------------------------
2419 // Generate an instance-of idiom. Used by both the instance-of bytecode
2420 // and the reflective instance-of call.
2421 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) {
2422 C->set_has_split_ifs(true); // Has chance for split-if optimization
2423 assert( !stopped(), "dead parse path should be checked in callers" );
2424 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2425 "must check for not-null not-dead klass in callers");
2427 // Make the merge point
2428 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2429 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2430 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2431 C->set_has_split_ifs(true); // Has chance for split-if optimization
2433 // Null check; get casted pointer; set region slot 3
2434 Node* null_ctl = top();
2435 Node* not_null_obj = null_check_oop(subobj, &null_ctl);
2437 // If not_null_obj is dead, only null-path is taken
2438 if (stopped()) { // Doing instance-of on a NULL?
2439 set_control(null_ctl);
2440 return intcon(0);
2441 }
2442 region->init_req(_null_path, null_ctl);
2443 phi ->init_req(_null_path, intcon(0)); // Set null path value
2445 // Load the object's klass
2446 Node* obj_klass = load_object_klass(not_null_obj);
2448 // Generate the subtype check
2449 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2451 // Plug in the success path to the general merge in slot 1.
2452 region->init_req(_obj_path, control());
2453 phi ->init_req(_obj_path, intcon(1));
2455 // Plug in the failing path to the general merge in slot 2.
2456 region->init_req(_fail_path, not_subtype_ctrl);
2457 phi ->init_req(_fail_path, intcon(0));
2459 // Return final merged results
2460 set_control( _gvn.transform(region) );
2461 record_for_igvn(region);
2462 return _gvn.transform(phi);
2463 }
2465 //-------------------------------gen_checkcast---------------------------------
2466 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2467 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2468 // uncommon-trap paths work. Adjust stack after this call.
2469 // If failure_control is supplied and not null, it is filled in with
2470 // the control edge for the cast failure. Otherwise, an appropriate
2471 // uncommon trap or exception is thrown.
2472 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2473 Node* *failure_control) {
2474 kill_dead_locals(); // Benefit all the uncommon traps
2475 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2476 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2478 // Fast cutout: Check the case that the cast is vacuously true.
2479 // This detects the common cases where the test will short-circuit
2480 // away completely. We do this before we perform the null check,
2481 // because if the test is going to turn into zero code, we don't
2482 // want a residual null check left around. (Causes a slowdown,
2483 // for example, in some objArray manipulations, such as a[i]=a[j].)
2484 if (tk->singleton()) {
2485 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2486 if (objtp != NULL && objtp->klass() != NULL) {
2487 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2488 case SSC_always_true:
2489 return obj;
2490 case SSC_always_false:
2491 // It needs a null check because a null will *pass* the cast check.
2492 // A non-null value will always produce an exception.
2493 return do_null_assert(obj, T_OBJECT);
2494 }
2495 }
2496 }
2498 ciProfileData* data = NULL;
2499 if (failure_control == NULL) { // use MDO in regular case only
2500 assert(java_bc() == Bytecodes::_aastore ||
2501 java_bc() == Bytecodes::_checkcast,
2502 "interpreter profiles type checks only for these BCs");
2503 data = method()->method_data()->bci_to_data(bci());
2504 }
2506 // Make the merge point
2507 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2508 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2509 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
2510 C->set_has_split_ifs(true); // Has chance for split-if optimization
2512 // Use null-cast information if it is available
2513 bool never_see_null = false;
2514 // If we see an unexpected null at a check-cast we record it and force a
2515 // recompile; the offending check-cast will be compiled to handle NULLs.
2516 // If we see several offending BCIs, then all checkcasts in the
2517 // method will be compiled to handle NULLs.
2518 if (UncommonNullCast // Cutout for this technique
2519 && failure_control == NULL // regular case
2520 && obj != null() // And not the -Xcomp stupid case?
2521 && !too_many_traps(Deoptimization::Reason_null_check)) {
2522 // Finally, check the "null_seen" bit from the interpreter.
2523 if (data == NULL || !data->as_BitData()->null_seen()) {
2524 never_see_null = true;
2525 }
2526 }
2528 // Null check; get casted pointer; set region slot 3
2529 Node* null_ctl = top();
2530 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2532 // If not_null_obj is dead, only null-path is taken
2533 if (stopped()) { // Doing instance-of on a NULL?
2534 set_control(null_ctl);
2535 return null();
2536 }
2537 region->init_req(_null_path, null_ctl);
2538 phi ->init_req(_null_path, null()); // Set null path value
2540 Node* cast_obj = NULL; // the casted version of the object
2542 // If the profile has seen exactly one type, narrow to that type.
2543 // (The subsequent subtype check will always fold up.)
2544 if (UseTypeProfile && TypeProfileCasts && data != NULL &&
2545 // Counter has never been decremented (due to cast failure).
2546 // ...This is a reasonable thing to expect. It is true of
2547 // all casts inserted by javac to implement generic types.
2548 data->as_CounterData()->count() >= 0 &&
2549 !too_many_traps(Deoptimization::Reason_class_check)) {
2550 // (No, this isn't a call, but it's enough like a virtual call
2551 // to use the same ciMethod accessor to get the profile info...)
2552 ciCallProfile profile = method()->call_profile_at_bci(bci());
2553 if (profile.count() >= 0 && // no cast failures here
2554 profile.has_receiver(0) &&
2555 profile.morphism() == 1) {
2556 ciKlass* exact_kls = profile.receiver(0);
2557 int ssc = static_subtype_check(tk->klass(), exact_kls);
2558 if (ssc == SSC_always_true) {
2559 // If we narrow the type to match what the type profile sees,
2560 // we can then remove the rest of the cast.
2561 // This is a win, even if the exact_kls is very specific,
2562 // because downstream operations, such as method calls,
2563 // will often benefit from the sharper type.
2564 Node* exact_obj = not_null_obj; // will get updated in place...
2565 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2566 &exact_obj);
2567 { PreserveJVMState pjvms(this);
2568 set_control(slow_ctl);
2569 uncommon_trap(Deoptimization::Reason_class_check,
2570 Deoptimization::Action_maybe_recompile);
2571 }
2572 if (failure_control != NULL) // failure is now impossible
2573 (*failure_control) = top();
2574 replace_in_map(not_null_obj, exact_obj);
2575 // adjust the type of the phi to the exact klass:
2576 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR));
2577 cast_obj = exact_obj;
2578 }
2579 // assert(cast_obj != NULL)... except maybe the profile lied to us.
2580 }
2581 }
2583 if (cast_obj == NULL) {
2584 // Load the object's klass
2585 Node* obj_klass = load_object_klass(not_null_obj);
2587 // Generate the subtype check
2588 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2590 // Plug in success path into the merge
2591 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2592 not_null_obj, toop));
2593 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2594 if (failure_control == NULL) {
2595 if (not_subtype_ctrl != top()) { // If failure is possible
2596 PreserveJVMState pjvms(this);
2597 set_control(not_subtype_ctrl);
2598 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2599 }
2600 } else {
2601 (*failure_control) = not_subtype_ctrl;
2602 }
2603 }
2605 region->init_req(_obj_path, control());
2606 phi ->init_req(_obj_path, cast_obj);
2608 // A merge of NULL or Casted-NotNull obj
2609 Node* res = _gvn.transform(phi);
2611 // Note I do NOT always 'replace_in_map(obj,result)' here.
2612 // if( tk->klass()->can_be_primary_super() )
2613 // This means that if I successfully store an Object into an array-of-String
2614 // I 'forget' that the Object is really now known to be a String. I have to
2615 // do this because we don't have true union types for interfaces - if I store
2616 // a Baz into an array-of-Interface and then tell the optimizer it's an
2617 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2618 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2619 // replace_in_map( obj, res );
2621 // Return final merged results
2622 set_control( _gvn.transform(region) );
2623 record_for_igvn(region);
2624 return res;
2625 }
2627 //------------------------------next_monitor-----------------------------------
2628 // What number should be given to the next monitor?
2629 int GraphKit::next_monitor() {
2630 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2631 int next = current + C->sync_stack_slots();
2632 // Keep the toplevel high water mark current:
2633 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2634 return current;
2635 }
2637 //------------------------------insert_mem_bar---------------------------------
2638 // Memory barrier to avoid floating things around
2639 // The membar serves as a pinch point between both control and all memory slices.
2640 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2641 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2642 mb->init_req(TypeFunc::Control, control());
2643 mb->init_req(TypeFunc::Memory, reset_memory());
2644 Node* membar = _gvn.transform(mb);
2645 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2646 set_all_memory_call(membar);
2647 return membar;
2648 }
2650 //-------------------------insert_mem_bar_volatile----------------------------
2651 // Memory barrier to avoid floating things around
2652 // The membar serves as a pinch point between both control and memory(alias_idx).
2653 // If you want to make a pinch point on all memory slices, do not use this
2654 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2655 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2656 // When Parse::do_put_xxx updates a volatile field, it appends a series
2657 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2658 // The first membar is on the same memory slice as the field store opcode.
2659 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2660 // All the other membars (for other volatile slices, including AliasIdxBot,
2661 // which stands for all unknown volatile slices) are control-dependent
2662 // on the first membar. This prevents later volatile loads or stores
2663 // from sliding up past the just-emitted store.
2665 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2666 mb->set_req(TypeFunc::Control,control());
2667 if (alias_idx == Compile::AliasIdxBot) {
2668 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2669 } else {
2670 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2671 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2672 }
2673 Node* membar = _gvn.transform(mb);
2674 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2675 if (alias_idx == Compile::AliasIdxBot) {
2676 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2677 } else {
2678 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2679 }
2680 return membar;
2681 }
2683 //------------------------------shared_lock------------------------------------
2684 // Emit locking code.
2685 FastLockNode* GraphKit::shared_lock(Node* obj) {
2686 // bci is either a monitorenter bc or InvocationEntryBci
2687 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2688 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2690 if( !GenerateSynchronizationCode )
2691 return NULL; // Not locking things?
2692 if (stopped()) // Dead monitor?
2693 return NULL;
2695 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2697 // Box the stack location
2698 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2699 Node* mem = reset_memory();
2701 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2702 if (PrintPreciseBiasedLockingStatistics) {
2703 // Create the counters for this fast lock.
2704 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2705 }
2706 // Add monitor to debug info for the slow path. If we block inside the
2707 // slow path and de-opt, we need the monitor hanging around
2708 map()->push_monitor( flock );
2710 const TypeFunc *tf = LockNode::lock_type();
2711 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2713 lock->init_req( TypeFunc::Control, control() );
2714 lock->init_req( TypeFunc::Memory , mem );
2715 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2716 lock->init_req( TypeFunc::FramePtr, frameptr() );
2717 lock->init_req( TypeFunc::ReturnAdr, top() );
2719 lock->init_req(TypeFunc::Parms + 0, obj);
2720 lock->init_req(TypeFunc::Parms + 1, box);
2721 lock->init_req(TypeFunc::Parms + 2, flock);
2722 add_safepoint_edges(lock);
2724 lock = _gvn.transform( lock )->as_Lock();
2726 // lock has no side-effects, sets few values
2727 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2729 insert_mem_bar(Op_MemBarAcquire);
2731 // Add this to the worklist so that the lock can be eliminated
2732 record_for_igvn(lock);
2734 #ifndef PRODUCT
2735 if (PrintLockStatistics) {
2736 // Update the counter for this lock. Don't bother using an atomic
2737 // operation since we don't require absolute accuracy.
2738 lock->create_lock_counter(map()->jvms());
2739 int adr_type = Compile::AliasIdxRaw;
2740 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr()));
2741 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
2742 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
2743 store_to_memory(control(), counter_addr, incr, T_INT, adr_type);
2744 }
2745 #endif
2747 return flock;
2748 }
2751 //------------------------------shared_unlock----------------------------------
2752 // Emit unlocking code.
2753 void GraphKit::shared_unlock(Node* box, Node* obj) {
2754 // bci is either a monitorenter bc or InvocationEntryBci
2755 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2756 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2758 if( !GenerateSynchronizationCode )
2759 return;
2760 if (stopped()) { // Dead monitor?
2761 map()->pop_monitor(); // Kill monitor from debug info
2762 return;
2763 }
2765 // Memory barrier to avoid floating things down past the locked region
2766 insert_mem_bar(Op_MemBarRelease);
2768 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2769 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2770 uint raw_idx = Compile::AliasIdxRaw;
2771 unlock->init_req( TypeFunc::Control, control() );
2772 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2773 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2774 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2775 unlock->init_req( TypeFunc::ReturnAdr, top() );
2777 unlock->init_req(TypeFunc::Parms + 0, obj);
2778 unlock->init_req(TypeFunc::Parms + 1, box);
2779 unlock = _gvn.transform(unlock)->as_Unlock();
2781 Node* mem = reset_memory();
2783 // unlock has no side-effects, sets few values
2784 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2786 // Kill monitor from debug info
2787 map()->pop_monitor( );
2788 }
2790 //-------------------------------get_layout_helper-----------------------------
2791 // If the given klass is a constant or known to be an array,
2792 // fetch the constant layout helper value into constant_value
2793 // and return (Node*)NULL. Otherwise, load the non-constant
2794 // layout helper value, and return the node which represents it.
2795 // This two-faced routine is useful because allocation sites
2796 // almost always feature constant types.
2797 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2798 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2799 if (!StressReflectiveCode && inst_klass != NULL) {
2800 ciKlass* klass = inst_klass->klass();
2801 bool xklass = inst_klass->klass_is_exact();
2802 if (xklass || klass->is_array_klass()) {
2803 jint lhelper = klass->layout_helper();
2804 if (lhelper != Klass::_lh_neutral_value) {
2805 constant_value = lhelper;
2806 return (Node*) NULL;
2807 }
2808 }
2809 }
2810 constant_value = Klass::_lh_neutral_value; // put in a known value
2811 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
2812 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2813 }
2815 // We just put in an allocate/initialize with a big raw-memory effect.
2816 // Hook selected additional alias categories on the initialization.
2817 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2818 MergeMemNode* init_in_merge,
2819 Node* init_out_raw) {
2820 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2821 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2823 Node* prevmem = kit.memory(alias_idx);
2824 init_in_merge->set_memory_at(alias_idx, prevmem);
2825 kit.set_memory(init_out_raw, alias_idx);
2826 }
2828 //---------------------------set_output_for_allocation-------------------------
2829 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2830 const TypeOopPtr* oop_type,
2831 bool raw_mem_only) {
2832 int rawidx = Compile::AliasIdxRaw;
2833 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2834 add_safepoint_edges(alloc);
2835 Node* allocx = _gvn.transform(alloc);
2836 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2837 // create memory projection for i_o
2838 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2839 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2841 // create a memory projection as for the normal control path
2842 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2843 set_memory(malloc, rawidx);
2845 // a normal slow-call doesn't change i_o, but an allocation does
2846 // we create a separate i_o projection for the normal control path
2847 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2848 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2850 // put in an initialization barrier
2851 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2852 rawoop)->as_Initialize();
2853 assert(alloc->initialization() == init, "2-way macro link must work");
2854 assert(init ->allocation() == alloc, "2-way macro link must work");
2855 if (ReduceFieldZeroing && !raw_mem_only) {
2856 // Extract memory strands which may participate in the new object's
2857 // initialization, and source them from the new InitializeNode.
2858 // This will allow us to observe initializations when they occur,
2859 // and link them properly (as a group) to the InitializeNode.
2860 assert(init->in(InitializeNode::Memory) == malloc, "");
2861 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
2862 init->set_req(InitializeNode::Memory, minit_in);
2863 record_for_igvn(minit_in); // fold it up later, if possible
2864 Node* minit_out = memory(rawidx);
2865 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
2866 if (oop_type->isa_aryptr()) {
2867 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
2868 int elemidx = C->get_alias_index(telemref);
2869 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
2870 } else if (oop_type->isa_instptr()) {
2871 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
2872 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
2873 ciField* field = ik->nonstatic_field_at(i);
2874 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
2875 continue; // do not bother to track really large numbers of fields
2876 // Find (or create) the alias category for this field:
2877 int fieldidx = C->alias_type(field)->index();
2878 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
2879 }
2880 }
2881 }
2883 // Cast raw oop to the real thing...
2884 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
2885 javaoop = _gvn.transform(javaoop);
2886 C->set_recent_alloc(control(), javaoop);
2887 assert(just_allocated_object(control()) == javaoop, "just allocated");
2889 #ifdef ASSERT
2890 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
2891 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
2892 "Ideal_allocation works");
2893 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
2894 "Ideal_allocation works");
2895 if (alloc->is_AllocateArray()) {
2896 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
2897 "Ideal_allocation works");
2898 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
2899 "Ideal_allocation works");
2900 } else {
2901 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
2902 }
2903 }
2904 #endif //ASSERT
2906 return javaoop;
2907 }
2909 //---------------------------new_instance--------------------------------------
2910 // This routine takes a klass_node which may be constant (for a static type)
2911 // or may be non-constant (for reflective code). It will work equally well
2912 // for either, and the graph will fold nicely if the optimizer later reduces
2913 // the type to a constant.
2914 // The optional arguments are for specialized use by intrinsics:
2915 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
2916 // - If 'raw_mem_only', do not cast the result to an oop.
2917 // - If 'return_size_val', report the the total object size to the caller.
2918 Node* GraphKit::new_instance(Node* klass_node,
2919 Node* extra_slow_test,
2920 bool raw_mem_only, // affect only raw memory
2921 Node* *return_size_val) {
2922 // Compute size in doublewords
2923 // The size is always an integral number of doublewords, represented
2924 // as a positive bytewise size stored in the klass's layout_helper.
2925 // The layout_helper also encodes (in a low bit) the need for a slow path.
2926 jint layout_con = Klass::_lh_neutral_value;
2927 Node* layout_val = get_layout_helper(klass_node, layout_con);
2928 int layout_is_con = (layout_val == NULL);
2930 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
2931 // Generate the initial go-slow test. It's either ALWAYS (return a
2932 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
2933 // case) a computed value derived from the layout_helper.
2934 Node* initial_slow_test = NULL;
2935 if (layout_is_con) {
2936 assert(!StressReflectiveCode, "stress mode does not use these paths");
2937 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
2938 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
2940 } else { // reflective case
2941 // This reflective path is used by Unsafe.allocateInstance.
2942 // (It may be stress-tested by specifying StressReflectiveCode.)
2943 // Basically, we want to get into the VM is there's an illegal argument.
2944 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
2945 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
2946 if (extra_slow_test != intcon(0)) {
2947 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
2948 }
2949 // (Macro-expander will further convert this to a Bool, if necessary.)
2950 }
2952 // Find the size in bytes. This is easy; it's the layout_helper.
2953 // The size value must be valid even if the slow path is taken.
2954 Node* size = NULL;
2955 if (layout_is_con) {
2956 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
2957 } else { // reflective case
2958 // This reflective path is used by clone and Unsafe.allocateInstance.
2959 size = ConvI2X(layout_val);
2961 // Clear the low bits to extract layout_helper_size_in_bytes:
2962 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
2963 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
2964 size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
2965 }
2966 if (return_size_val != NULL) {
2967 (*return_size_val) = size;
2968 }
2970 // This is a precise notnull oop of the klass.
2971 // (Actually, it need not be precise if this is a reflective allocation.)
2972 // It's what we cast the result to.
2973 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
2974 if (!tklass) tklass = TypeKlassPtr::OBJECT;
2975 const TypeOopPtr* oop_type = tklass->as_instance_type();
2977 // Now generate allocation code
2979 // The entire memory state is needed for slow path of the allocation
2980 // since GC and deoptimization can happened.
2981 Node *mem = reset_memory();
2982 set_all_memory(mem); // Create new memory state
2984 AllocateNode* alloc
2985 = new (C, AllocateNode::ParmLimit)
2986 AllocateNode(C, AllocateNode::alloc_type(),
2987 control(), mem, i_o(),
2988 size, klass_node,
2989 initial_slow_test);
2991 return set_output_for_allocation(alloc, oop_type, raw_mem_only);
2992 }
2994 //-------------------------------new_array-------------------------------------
2995 // helper for both newarray and anewarray
2996 // The 'length' parameter is (obviously) the length of the array.
2997 // See comments on new_instance for the meaning of the other arguments.
2998 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
2999 Node* length, // number of array elements
3000 int nargs, // number of arguments to push back for uncommon trap
3001 bool raw_mem_only, // affect only raw memory
3002 Node* *return_size_val) {
3003 jint layout_con = Klass::_lh_neutral_value;
3004 Node* layout_val = get_layout_helper(klass_node, layout_con);
3005 int layout_is_con = (layout_val == NULL);
3007 if (!layout_is_con && !StressReflectiveCode &&
3008 !too_many_traps(Deoptimization::Reason_class_check)) {
3009 // This is a reflective array creation site.
3010 // Optimistically assume that it is a subtype of Object[],
3011 // so that we can fold up all the address arithmetic.
3012 layout_con = Klass::array_layout_helper(T_OBJECT);
3013 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
3014 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
3015 { BuildCutout unless(this, bol_lh, PROB_MAX);
3016 _sp += nargs;
3017 uncommon_trap(Deoptimization::Reason_class_check,
3018 Deoptimization::Action_maybe_recompile);
3019 }
3020 layout_val = NULL;
3021 layout_is_con = true;
3022 }
3024 // Generate the initial go-slow test. Make sure we do not overflow
3025 // if length is huge (near 2Gig) or negative! We do not need
3026 // exact double-words here, just a close approximation of needed
3027 // double-words. We can't add any offset or rounding bits, lest we
3028 // take a size -1 of bytes and make it positive. Use an unsigned
3029 // compare, so negative sizes look hugely positive.
3030 int fast_size_limit = FastAllocateSizeLimit;
3031 if (layout_is_con) {
3032 assert(!StressReflectiveCode, "stress mode does not use these paths");
3033 // Increase the size limit if we have exact knowledge of array type.
3034 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3035 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3036 }
3038 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3039 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3040 if (initial_slow_test->is_Bool()) {
3041 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3042 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3043 }
3045 // --- Size Computation ---
3046 // array_size = round_to_heap(array_header + (length << elem_shift));
3047 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3048 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3049 // The rounding mask is strength-reduced, if possible.
3050 int round_mask = MinObjAlignmentInBytes - 1;
3051 Node* header_size = NULL;
3052 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3053 // (T_BYTE has the weakest alignment and size restrictions...)
3054 if (layout_is_con) {
3055 int hsize = Klass::layout_helper_header_size(layout_con);
3056 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3057 BasicType etype = Klass::layout_helper_element_type(layout_con);
3058 if ((round_mask & ~right_n_bits(eshift)) == 0)
3059 round_mask = 0; // strength-reduce it if it goes away completely
3060 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3061 assert(header_size_min <= hsize, "generic minimum is smallest");
3062 header_size_min = hsize;
3063 header_size = intcon(hsize + round_mask);
3064 } else {
3065 Node* hss = intcon(Klass::_lh_header_size_shift);
3066 Node* hsm = intcon(Klass::_lh_header_size_mask);
3067 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3068 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3069 Node* mask = intcon(round_mask);
3070 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3071 }
3073 Node* elem_shift = NULL;
3074 if (layout_is_con) {
3075 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3076 if (eshift != 0)
3077 elem_shift = intcon(eshift);
3078 } else {
3079 // There is no need to mask or shift this value.
3080 // The semantics of LShiftINode include an implicit mask to 0x1F.
3081 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3082 elem_shift = layout_val;
3083 }
3085 // Transition to native address size for all offset calculations:
3086 Node* lengthx = ConvI2X(length);
3087 Node* headerx = ConvI2X(header_size);
3088 #ifdef _LP64
3089 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3090 if (tllen != NULL && tllen->_lo < 0) {
3091 // Add a manual constraint to a positive range. Cf. array_element_address.
3092 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3093 if (size_max > tllen->_hi) size_max = tllen->_hi;
3094 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3095 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3096 }
3097 }
3098 #endif
3100 // Combine header size (plus rounding) and body size. Then round down.
3101 // This computation cannot overflow, because it is used only in two
3102 // places, one where the length is sharply limited, and the other
3103 // after a successful allocation.
3104 Node* abody = lengthx;
3105 if (elem_shift != NULL)
3106 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3107 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3108 if (round_mask != 0) {
3109 Node* mask = MakeConX(~round_mask);
3110 size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3111 }
3112 // else if round_mask == 0, the size computation is self-rounding
3114 if (return_size_val != NULL) {
3115 // This is the size
3116 (*return_size_val) = size;
3117 }
3119 // Now generate allocation code
3121 // The entire memory state is needed for slow path of the allocation
3122 // since GC and deoptimization can happened.
3123 Node *mem = reset_memory();
3124 set_all_memory(mem); // Create new memory state
3126 // Create the AllocateArrayNode and its result projections
3127 AllocateArrayNode* alloc
3128 = new (C, AllocateArrayNode::ParmLimit)
3129 AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3130 control(), mem, i_o(),
3131 size, klass_node,
3132 initial_slow_test,
3133 length);
3135 // Cast to correct type. Note that the klass_node may be constant or not,
3136 // and in the latter case the actual array type will be inexact also.
3137 // (This happens via a non-constant argument to inline_native_newArray.)
3138 // In any case, the value of klass_node provides the desired array type.
3139 const TypeInt* length_type = _gvn.find_int_type(length);
3140 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3141 if (ary_type->isa_aryptr() && length_type != NULL) {
3142 // Try to get a better type than POS for the size
3143 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3144 }
3146 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
3148 // Cast length on remaining path to be as narrow as possible
3149 if (map()->find_edge(length) >= 0) {
3150 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3151 if (ccast != length) {
3152 _gvn.set_type_bottom(ccast);
3153 record_for_igvn(ccast);
3154 replace_in_map(length, ccast);
3155 }
3156 }
3158 return javaoop;
3159 }
3161 // The following "Ideal_foo" functions are placed here because they recognize
3162 // the graph shapes created by the functions immediately above.
3164 //---------------------------Ideal_allocation----------------------------------
3165 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3166 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3167 if (ptr == NULL) { // reduce dumb test in callers
3168 return NULL;
3169 }
3170 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3171 ptr = ptr->in(1);
3172 if (ptr == NULL) return NULL;
3173 }
3174 if (ptr->is_Proj()) {
3175 Node* allo = ptr->in(0);
3176 if (allo != NULL && allo->is_Allocate()) {
3177 return allo->as_Allocate();
3178 }
3179 }
3180 // Report failure to match.
3181 return NULL;
3182 }
3184 // Fancy version which also strips off an offset (and reports it to caller).
3185 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3186 intptr_t& offset) {
3187 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3188 if (base == NULL) return NULL;
3189 return Ideal_allocation(base, phase);
3190 }
3192 // Trace Initialize <- Proj[Parm] <- Allocate
3193 AllocateNode* InitializeNode::allocation() {
3194 Node* rawoop = in(InitializeNode::RawAddress);
3195 if (rawoop->is_Proj()) {
3196 Node* alloc = rawoop->in(0);
3197 if (alloc->is_Allocate()) {
3198 return alloc->as_Allocate();
3199 }
3200 }
3201 return NULL;
3202 }
3204 // Trace Allocate -> Proj[Parm] -> Initialize
3205 InitializeNode* AllocateNode::initialization() {
3206 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3207 if (rawoop == NULL) return NULL;
3208 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3209 Node* init = rawoop->fast_out(i);
3210 if (init->is_Initialize()) {
3211 assert(init->as_Initialize()->allocation() == this, "2-way link");
3212 return init->as_Initialize();
3213 }
3214 }
3215 return NULL;
3216 }
3218 void GraphKit::g1_write_barrier_pre(Node* obj,
3219 Node* adr,
3220 uint alias_idx,
3221 Node* val,
3222 const TypeOopPtr* val_type,
3223 BasicType bt) {
3224 IdealKit ideal(gvn(), control(), merged_memory(), true);
3225 #define __ ideal.
3226 __ declares_done();
3228 Node* thread = __ thread();
3230 Node* no_ctrl = NULL;
3231 Node* no_base = __ top();
3232 Node* zero = __ ConI(0);
3234 float likely = PROB_LIKELY(0.999);
3235 float unlikely = PROB_UNLIKELY(0.999);
3237 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3238 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3240 // Offsets into the thread
3241 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3242 PtrQueue::byte_offset_of_active());
3243 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3244 PtrQueue::byte_offset_of_index());
3245 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3246 PtrQueue::byte_offset_of_buf());
3247 // Now the actual pointers into the thread
3249 // set_control( ctl);
3251 Node* marking_adr = __ AddP(no_base, thread, __ ConX(marking_offset));
3252 Node* buffer_adr = __ AddP(no_base, thread, __ ConX(buffer_offset));
3253 Node* index_adr = __ AddP(no_base, thread, __ ConX(index_offset));
3255 // Now some of the values
3257 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3259 // if (!marking)
3260 __ if_then(marking, BoolTest::ne, zero); {
3261 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3263 const Type* t1 = adr->bottom_type();
3264 const Type* t2 = val->bottom_type();
3266 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3267 // if (orig != NULL)
3268 __ if_then(orig, BoolTest::ne, null()); {
3269 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3271 // load original value
3272 // alias_idx correct??
3274 // is the queue for this thread full?
3275 __ if_then(index, BoolTest::ne, zero, likely); {
3277 // decrement the index
3278 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3279 Node* next_indexX = next_index;
3280 #ifdef _LP64
3281 // We could refine the type for what it's worth
3282 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3283 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3284 #endif // _LP64
3286 // Now get the buffer location we will log the original value into and store it
3288 Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3289 // __ store(__ ctrl(), log_addr, orig, T_OBJECT, C->get_alias_index(TypeOopPtr::BOTTOM));
3290 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw);
3293 // update the index
3294 // __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3295 // This is a hack to force this store to occur before the oop store that is coming up
3296 __ store(__ ctrl(), index_adr, next_index, T_INT, C->get_alias_index(TypeOopPtr::BOTTOM));
3298 } __ else_(); {
3300 // logging buffer is full, call the runtime
3301 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3302 // __ make_leaf_call(tf, OptoRuntime::g1_wb_pre_Java(), "g1_wb_pre", orig, thread);
3303 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, thread);
3304 } __ end_if();
3305 } __ end_if();
3306 } __ end_if();
3308 __ drain_delay_transform();
3309 set_control( __ ctrl());
3310 set_all_memory( __ merged_memory());
3312 #undef __
3313 }
3315 //
3316 // Update the card table and add card address to the queue
3317 //
3318 void GraphKit::g1_mark_card(IdealKit* ideal, Node* card_adr, Node* store, Node* index, Node* index_adr, Node* buffer, const TypeFunc* tf) {
3319 #define __ ideal->
3320 Node* zero = __ ConI(0);
3321 Node* no_base = __ top();
3322 BasicType card_bt = T_BYTE;
3323 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3324 __ storeCM(__ ctrl(), card_adr, zero, store, card_bt, Compile::AliasIdxRaw);
3326 // Now do the queue work
3327 __ if_then(index, BoolTest::ne, zero); {
3329 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3330 Node* next_indexX = next_index;
3331 #ifdef _LP64
3332 // We could refine the type for what it's worth
3333 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3334 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3335 #endif // _LP64
3336 Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3338 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3339 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3341 } __ else_(); {
3342 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3343 } __ end_if();
3344 #undef __
3345 }
3347 void GraphKit::g1_write_barrier_post(Node* store,
3348 Node* obj,
3349 Node* adr,
3350 uint alias_idx,
3351 Node* val,
3352 BasicType bt,
3353 bool use_precise) {
3354 // If we are writing a NULL then we need no post barrier
3356 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3357 // Must be NULL
3358 const Type* t = val->bottom_type();
3359 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3360 // No post barrier if writing NULLx
3361 return;
3362 }
3364 if (!use_precise) {
3365 // All card marks for a (non-array) instance are in one place:
3366 adr = obj;
3367 }
3368 // (Else it's an array (or unknown), and we want more precise card marks.)
3369 assert(adr != NULL, "");
3371 IdealKit ideal(gvn(), control(), merged_memory(), true);
3372 #define __ ideal.
3373 __ declares_done();
3375 Node* thread = __ thread();
3377 Node* no_ctrl = NULL;
3378 Node* no_base = __ top();
3379 float likely = PROB_LIKELY(0.999);
3380 float unlikely = PROB_UNLIKELY(0.999);
3381 Node* zero = __ ConI(0);
3382 Node* zeroX = __ ConX(0);
3384 // Get the alias_index for raw card-mark memory
3385 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3387 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3389 // Offsets into the thread
3390 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3391 PtrQueue::byte_offset_of_index());
3392 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3393 PtrQueue::byte_offset_of_buf());
3395 // Pointers into the thread
3397 Node* buffer_adr = __ AddP(no_base, thread, __ ConX(buffer_offset));
3398 Node* index_adr = __ AddP(no_base, thread, __ ConX(index_offset));
3400 // Now some values
3402 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3403 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3406 // Convert the store obj pointer to an int prior to doing math on it
3407 // Use addr not obj gets accurate card marks
3409 // Node* cast = __ CastPX(no_ctrl, adr /* obj */);
3411 // Must use ctrl to prevent "integerized oop" existing across safepoint
3412 Node* cast = __ CastPX(__ ctrl(), ( use_precise ? adr : obj ));
3414 // Divide pointer by card size
3415 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3417 // Combine card table base and card offset
3418 Node *card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3420 // If we know the value being stored does it cross regions?
3422 if (val != NULL) {
3423 // Does the store cause us to cross regions?
3425 // Should be able to do an unsigned compare of region_size instead of
3426 // and extra shift. Do we have an unsigned compare??
3427 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3428 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3430 // if (xor_res == 0) same region so skip
3431 __ if_then(xor_res, BoolTest::ne, zeroX); {
3433 // No barrier if we are storing a NULL
3434 __ if_then(val, BoolTest::ne, null(), unlikely); {
3436 // Ok must mark the card if not already dirty
3438 // load the original value of the card
3439 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3441 __ if_then(card_val, BoolTest::ne, zero); {
3442 g1_mark_card(&ideal, card_adr, store, index, index_adr, buffer, tf);
3443 } __ end_if();
3444 } __ end_if();
3445 } __ end_if();
3446 } else {
3447 g1_mark_card(&ideal, card_adr, store, index, index_adr, buffer, tf);
3448 }
3451 __ drain_delay_transform();
3452 set_control( __ ctrl());
3453 set_all_memory( __ merged_memory());
3454 #undef __
3456 }