Tue, 05 Jan 2010 13:05:58 +0100
6829187: compiler optimizations required for JSR 292
Summary: C2 implementation for invokedynamic support.
Reviewed-by: kvn, 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 }
623 //---------------------------PreserveReexecuteState----------------------------
624 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
625 assert(!kit->stopped(), "must call stopped() before");
626 _kit = kit;
627 _sp = kit->sp();
628 _reexecute = kit->jvms()->_reexecute;
629 }
630 PreserveReexecuteState::~PreserveReexecuteState() {
631 if (_kit->stopped()) return;
632 _kit->jvms()->_reexecute = _reexecute;
633 _kit->set_sp(_sp);
634 }
636 //------------------------------clone_map--------------------------------------
637 // Implementation of PreserveJVMState
638 //
639 // Only clone_map(...) here. If this function is only used in the
640 // PreserveJVMState class we may want to get rid of this extra
641 // function eventually and do it all there.
643 SafePointNode* GraphKit::clone_map() {
644 if (map() == NULL) return NULL;
646 // Clone the memory edge first
647 Node* mem = MergeMemNode::make(C, map()->memory());
648 gvn().set_type_bottom(mem);
650 SafePointNode *clonemap = (SafePointNode*)map()->clone();
651 JVMState* jvms = this->jvms();
652 JVMState* clonejvms = jvms->clone_shallow(C);
653 clonemap->set_memory(mem);
654 clonemap->set_jvms(clonejvms);
655 clonejvms->set_map(clonemap);
656 record_for_igvn(clonemap);
657 gvn().set_type_bottom(clonemap);
658 return clonemap;
659 }
662 //-----------------------------set_map_clone-----------------------------------
663 void GraphKit::set_map_clone(SafePointNode* m) {
664 _map = m;
665 _map = clone_map();
666 _map->set_next_exception(NULL);
667 debug_only(verify_map());
668 }
671 //----------------------------kill_dead_locals---------------------------------
672 // Detect any locals which are known to be dead, and force them to top.
673 void GraphKit::kill_dead_locals() {
674 // Consult the liveness information for the locals. If any
675 // of them are unused, then they can be replaced by top(). This
676 // should help register allocation time and cut down on the size
677 // of the deoptimization information.
679 // This call is made from many of the bytecode handling
680 // subroutines called from the Big Switch in do_one_bytecode.
681 // Every bytecode which might include a slow path is responsible
682 // for killing its dead locals. The more consistent we
683 // are about killing deads, the fewer useless phis will be
684 // constructed for them at various merge points.
686 // bci can be -1 (InvocationEntryBci). We return the entry
687 // liveness for the method.
689 if (method() == NULL || method()->code_size() == 0) {
690 // We are building a graph for a call to a native method.
691 // All locals are live.
692 return;
693 }
695 ResourceMark rm;
697 // Consult the liveness information for the locals. If any
698 // of them are unused, then they can be replaced by top(). This
699 // should help register allocation time and cut down on the size
700 // of the deoptimization information.
701 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
703 int len = (int)live_locals.size();
704 assert(len <= jvms()->loc_size(), "too many live locals");
705 for (int local = 0; local < len; local++) {
706 if (!live_locals.at(local)) {
707 set_local(local, top());
708 }
709 }
710 }
712 #ifdef ASSERT
713 //-------------------------dead_locals_are_killed------------------------------
714 // Return true if all dead locals are set to top in the map.
715 // Used to assert "clean" debug info at various points.
716 bool GraphKit::dead_locals_are_killed() {
717 if (method() == NULL || method()->code_size() == 0) {
718 // No locals need to be dead, so all is as it should be.
719 return true;
720 }
722 // Make sure somebody called kill_dead_locals upstream.
723 ResourceMark rm;
724 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
725 if (jvms->loc_size() == 0) continue; // no locals to consult
726 SafePointNode* map = jvms->map();
727 ciMethod* method = jvms->method();
728 int bci = jvms->bci();
729 if (jvms == this->jvms()) {
730 bci = this->bci(); // it might not yet be synched
731 }
732 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
733 int len = (int)live_locals.size();
734 if (!live_locals.is_valid() || len == 0)
735 // This method is trivial, or is poisoned by a breakpoint.
736 return true;
737 assert(len == jvms->loc_size(), "live map consistent with locals map");
738 for (int local = 0; local < len; local++) {
739 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
740 if (PrintMiscellaneous && (Verbose || WizardMode)) {
741 tty->print_cr("Zombie local %d: ", local);
742 jvms->dump();
743 }
744 return false;
745 }
746 }
747 }
748 return true;
749 }
751 #endif //ASSERT
753 // Helper function for enforcing certain bytecodes to reexecute if
754 // deoptimization happens
755 static bool should_reexecute_implied_by_bytecode(JVMState *jvms) {
756 ciMethod* cur_method = jvms->method();
757 int cur_bci = jvms->bci();
758 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
759 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
760 return Interpreter::bytecode_should_reexecute(code);
761 } else
762 return false;
763 }
765 // Helper function for adding JVMState and debug information to node
766 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
767 // Add the safepoint edges to the call (or other safepoint).
769 // Make sure dead locals are set to top. This
770 // should help register allocation time and cut down on the size
771 // of the deoptimization information.
772 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
774 // Walk the inline list to fill in the correct set of JVMState's
775 // Also fill in the associated edges for each JVMState.
777 JVMState* youngest_jvms = sync_jvms();
779 // Do we need debug info here? If it is a SafePoint and this method
780 // cannot de-opt, then we do NOT need any debug info.
781 bool full_info = (C->deopt_happens() || call->Opcode() != Op_SafePoint);
783 // If we are guaranteed to throw, we can prune everything but the
784 // input to the current bytecode.
785 bool can_prune_locals = false;
786 uint stack_slots_not_pruned = 0;
787 int inputs = 0, depth = 0;
788 if (must_throw) {
789 assert(method() == youngest_jvms->method(), "sanity");
790 if (compute_stack_effects(inputs, depth)) {
791 can_prune_locals = true;
792 stack_slots_not_pruned = inputs;
793 }
794 }
796 if (env()->jvmti_can_examine_or_deopt_anywhere()) {
797 // At any safepoint, this method can get breakpointed, which would
798 // then require an immediate deoptimization.
799 full_info = true;
800 can_prune_locals = false; // do not prune locals
801 stack_slots_not_pruned = 0;
802 }
804 // do not scribble on the input jvms
805 JVMState* out_jvms = youngest_jvms->clone_deep(C);
806 call->set_jvms(out_jvms); // Start jvms list for call node
808 // For a known set of bytecodes, the interpreter should reexecute them if
809 // deoptimization happens. We set the reexecute state for them here
810 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
811 should_reexecute_implied_by_bytecode(out_jvms)) {
812 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
813 }
815 // Presize the call:
816 debug_only(uint non_debug_edges = call->req());
817 call->add_req_batch(top(), youngest_jvms->debug_depth());
818 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
820 // Set up edges so that the call looks like this:
821 // Call [state:] ctl io mem fptr retadr
822 // [parms:] parm0 ... parmN
823 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
824 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
825 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
826 // Note that caller debug info precedes callee debug info.
828 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
829 uint debug_ptr = call->req();
831 // Loop over the map input edges associated with jvms, add them
832 // to the call node, & reset all offsets to match call node array.
833 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
834 uint debug_end = debug_ptr;
835 uint debug_start = debug_ptr - in_jvms->debug_size();
836 debug_ptr = debug_start; // back up the ptr
838 uint p = debug_start; // walks forward in [debug_start, debug_end)
839 uint j, k, l;
840 SafePointNode* in_map = in_jvms->map();
841 out_jvms->set_map(call);
843 if (can_prune_locals) {
844 assert(in_jvms->method() == out_jvms->method(), "sanity");
845 // If the current throw can reach an exception handler in this JVMS,
846 // then we must keep everything live that can reach that handler.
847 // As a quick and dirty approximation, we look for any handlers at all.
848 if (in_jvms->method()->has_exception_handlers()) {
849 can_prune_locals = false;
850 }
851 }
853 // Add the Locals
854 k = in_jvms->locoff();
855 l = in_jvms->loc_size();
856 out_jvms->set_locoff(p);
857 if (full_info && !can_prune_locals) {
858 for (j = 0; j < l; j++)
859 call->set_req(p++, in_map->in(k+j));
860 } else {
861 p += l; // already set to top above by add_req_batch
862 }
864 // Add the Expression Stack
865 k = in_jvms->stkoff();
866 l = in_jvms->sp();
867 out_jvms->set_stkoff(p);
868 if (full_info && !can_prune_locals) {
869 for (j = 0; j < l; j++)
870 call->set_req(p++, in_map->in(k+j));
871 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
872 // Divide stack into {S0,...,S1}, where S0 is set to top.
873 uint s1 = stack_slots_not_pruned;
874 stack_slots_not_pruned = 0; // for next iteration
875 if (s1 > l) s1 = l;
876 uint s0 = l - s1;
877 p += s0; // skip the tops preinstalled by add_req_batch
878 for (j = s0; j < l; j++)
879 call->set_req(p++, in_map->in(k+j));
880 } else {
881 p += l; // already set to top above by add_req_batch
882 }
884 // Add the Monitors
885 k = in_jvms->monoff();
886 l = in_jvms->mon_size();
887 out_jvms->set_monoff(p);
888 for (j = 0; j < l; j++)
889 call->set_req(p++, in_map->in(k+j));
891 // Copy any scalar object fields.
892 k = in_jvms->scloff();
893 l = in_jvms->scl_size();
894 out_jvms->set_scloff(p);
895 for (j = 0; j < l; j++)
896 call->set_req(p++, in_map->in(k+j));
898 // Finish the new jvms.
899 out_jvms->set_endoff(p);
901 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
902 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
903 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
904 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
905 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
906 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
908 // Update the two tail pointers in parallel.
909 out_jvms = out_jvms->caller();
910 in_jvms = in_jvms->caller();
911 }
913 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
915 // Test the correctness of JVMState::debug_xxx accessors:
916 assert(call->jvms()->debug_start() == non_debug_edges, "");
917 assert(call->jvms()->debug_end() == call->req(), "");
918 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
919 }
921 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
922 Bytecodes::Code code = java_bc();
923 if (code == Bytecodes::_wide) {
924 code = method()->java_code_at_bci(bci() + 1);
925 }
927 BasicType rtype = T_ILLEGAL;
928 int rsize = 0;
930 if (code != Bytecodes::_illegal) {
931 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
932 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
933 if (rtype < T_CONFLICT)
934 rsize = type2size[rtype];
935 }
937 switch (code) {
938 case Bytecodes::_illegal:
939 return false;
941 case Bytecodes::_ldc:
942 case Bytecodes::_ldc_w:
943 case Bytecodes::_ldc2_w:
944 inputs = 0;
945 break;
947 case Bytecodes::_dup: inputs = 1; break;
948 case Bytecodes::_dup_x1: inputs = 2; break;
949 case Bytecodes::_dup_x2: inputs = 3; break;
950 case Bytecodes::_dup2: inputs = 2; break;
951 case Bytecodes::_dup2_x1: inputs = 3; break;
952 case Bytecodes::_dup2_x2: inputs = 4; break;
953 case Bytecodes::_swap: inputs = 2; break;
954 case Bytecodes::_arraylength: inputs = 1; break;
956 case Bytecodes::_getstatic:
957 case Bytecodes::_putstatic:
958 case Bytecodes::_getfield:
959 case Bytecodes::_putfield:
960 {
961 bool is_get = (depth >= 0), is_static = (depth & 1);
962 bool ignore;
963 ciBytecodeStream iter(method());
964 iter.reset_to_bci(bci());
965 iter.next();
966 ciField* field = iter.get_field(ignore);
967 int size = field->type()->size();
968 inputs = (is_static ? 0 : 1);
969 if (is_get) {
970 depth = size - inputs;
971 } else {
972 inputs += size; // putxxx pops the value from the stack
973 depth = - inputs;
974 }
975 }
976 break;
978 case Bytecodes::_invokevirtual:
979 case Bytecodes::_invokespecial:
980 case Bytecodes::_invokestatic:
981 case Bytecodes::_invokedynamic:
982 case Bytecodes::_invokeinterface:
983 {
984 bool ignore;
985 ciBytecodeStream iter(method());
986 iter.reset_to_bci(bci());
987 iter.next();
988 ciMethod* method = iter.get_method(ignore);
989 inputs = method->arg_size_no_receiver();
990 // Add a receiver argument, maybe:
991 if (code != Bytecodes::_invokestatic &&
992 code != Bytecodes::_invokedynamic)
993 inputs += 1;
994 // (Do not use ciMethod::arg_size(), because
995 // it might be an unloaded method, which doesn't
996 // know whether it is static or not.)
997 int size = method->return_type()->size();
998 depth = size - inputs;
999 }
1000 break;
1002 case Bytecodes::_multianewarray:
1003 {
1004 ciBytecodeStream iter(method());
1005 iter.reset_to_bci(bci());
1006 iter.next();
1007 inputs = iter.get_dimensions();
1008 assert(rsize == 1, "");
1009 depth = rsize - inputs;
1010 }
1011 break;
1013 case Bytecodes::_ireturn:
1014 case Bytecodes::_lreturn:
1015 case Bytecodes::_freturn:
1016 case Bytecodes::_dreturn:
1017 case Bytecodes::_areturn:
1018 assert(rsize = -depth, "");
1019 inputs = rsize;
1020 break;
1022 case Bytecodes::_jsr:
1023 case Bytecodes::_jsr_w:
1024 inputs = 0;
1025 depth = 1; // S.B. depth=1, not zero
1026 break;
1028 default:
1029 // bytecode produces a typed result
1030 inputs = rsize - depth;
1031 assert(inputs >= 0, "");
1032 break;
1033 }
1035 #ifdef ASSERT
1036 // spot check
1037 int outputs = depth + inputs;
1038 assert(outputs >= 0, "sanity");
1039 switch (code) {
1040 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1041 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1042 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1043 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1044 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1045 }
1046 #endif //ASSERT
1048 return true;
1049 }
1053 //------------------------------basic_plus_adr---------------------------------
1054 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1055 // short-circuit a common case
1056 if (offset == intcon(0)) return ptr;
1057 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1058 }
1060 Node* GraphKit::ConvI2L(Node* offset) {
1061 // short-circuit a common case
1062 jint offset_con = find_int_con(offset, Type::OffsetBot);
1063 if (offset_con != Type::OffsetBot) {
1064 return longcon((long) offset_con);
1065 }
1066 return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1067 }
1068 Node* GraphKit::ConvL2I(Node* offset) {
1069 // short-circuit a common case
1070 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1071 if (offset_con != (jlong)Type::OffsetBot) {
1072 return intcon((int) offset_con);
1073 }
1074 return _gvn.transform( new (C, 2) ConvL2INode(offset));
1075 }
1077 //-------------------------load_object_klass-----------------------------------
1078 Node* GraphKit::load_object_klass(Node* obj) {
1079 // Special-case a fresh allocation to avoid building nodes:
1080 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1081 if (akls != NULL) return akls;
1082 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1083 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1084 }
1086 //-------------------------load_array_length-----------------------------------
1087 Node* GraphKit::load_array_length(Node* array) {
1088 // Special-case a fresh allocation to avoid building nodes:
1089 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1090 Node *alen;
1091 if (alloc == NULL) {
1092 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1093 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1094 } else {
1095 alen = alloc->Ideal_length();
1096 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1097 if (ccast != alen) {
1098 alen = _gvn.transform(ccast);
1099 }
1100 }
1101 return alen;
1102 }
1104 //------------------------------do_null_check----------------------------------
1105 // Helper function to do a NULL pointer check. Returned value is
1106 // the incoming address with NULL casted away. You are allowed to use the
1107 // not-null value only if you are control dependent on the test.
1108 extern int explicit_null_checks_inserted,
1109 explicit_null_checks_elided;
1110 Node* GraphKit::null_check_common(Node* value, BasicType type,
1111 // optional arguments for variations:
1112 bool assert_null,
1113 Node* *null_control) {
1114 assert(!assert_null || null_control == NULL, "not both at once");
1115 if (stopped()) return top();
1116 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1117 // For some performance testing, we may wish to suppress null checking.
1118 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1119 return value;
1120 }
1121 explicit_null_checks_inserted++;
1123 // Construct NULL check
1124 Node *chk = NULL;
1125 switch(type) {
1126 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1127 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1128 case T_ARRAY : // fall through
1129 type = T_OBJECT; // simplify further tests
1130 case T_OBJECT : {
1131 const Type *t = _gvn.type( value );
1133 const TypeOopPtr* tp = t->isa_oopptr();
1134 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1135 // Only for do_null_check, not any of its siblings:
1136 && !assert_null && null_control == NULL) {
1137 // Usually, any field access or invocation on an unloaded oop type
1138 // will simply fail to link, since the statically linked class is
1139 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1140 // the static class is loaded but the sharper oop type is not.
1141 // Rather than checking for this obscure case in lots of places,
1142 // we simply observe that a null check on an unloaded class
1143 // will always be followed by a nonsense operation, so we
1144 // can just issue the uncommon trap here.
1145 // Our access to the unloaded class will only be correct
1146 // after it has been loaded and initialized, which requires
1147 // a trip through the interpreter.
1148 #ifndef PRODUCT
1149 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1150 #endif
1151 uncommon_trap(Deoptimization::Reason_unloaded,
1152 Deoptimization::Action_reinterpret,
1153 tp->klass(), "!loaded");
1154 return top();
1155 }
1157 if (assert_null) {
1158 // See if the type is contained in NULL_PTR.
1159 // If so, then the value is already null.
1160 if (t->higher_equal(TypePtr::NULL_PTR)) {
1161 explicit_null_checks_elided++;
1162 return value; // Elided null assert quickly!
1163 }
1164 } else {
1165 // See if mixing in the NULL pointer changes type.
1166 // If so, then the NULL pointer was not allowed in the original
1167 // type. In other words, "value" was not-null.
1168 if (t->meet(TypePtr::NULL_PTR) != t) {
1169 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1170 explicit_null_checks_elided++;
1171 return value; // Elided null check quickly!
1172 }
1173 }
1174 chk = new (C, 3) CmpPNode( value, null() );
1175 break;
1176 }
1178 default : ShouldNotReachHere();
1179 }
1180 assert(chk != NULL, "sanity check");
1181 chk = _gvn.transform(chk);
1183 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1184 BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1185 Node *tst = _gvn.transform( btst );
1187 //-----------
1188 // if peephole optimizations occurred, a prior test existed.
1189 // If a prior test existed, maybe it dominates as we can avoid this test.
1190 if (tst != btst && type == T_OBJECT) {
1191 // At this point we want to scan up the CFG to see if we can
1192 // find an identical test (and so avoid this test altogether).
1193 Node *cfg = control();
1194 int depth = 0;
1195 while( depth < 16 ) { // Limit search depth for speed
1196 if( cfg->Opcode() == Op_IfTrue &&
1197 cfg->in(0)->in(1) == tst ) {
1198 // Found prior test. Use "cast_not_null" to construct an identical
1199 // CastPP (and hence hash to) as already exists for the prior test.
1200 // Return that casted value.
1201 if (assert_null) {
1202 replace_in_map(value, null());
1203 return null(); // do not issue the redundant test
1204 }
1205 Node *oldcontrol = control();
1206 set_control(cfg);
1207 Node *res = cast_not_null(value);
1208 set_control(oldcontrol);
1209 explicit_null_checks_elided++;
1210 return res;
1211 }
1212 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1213 if (cfg == NULL) break; // Quit at region nodes
1214 depth++;
1215 }
1216 }
1218 //-----------
1219 // Branch to failure if null
1220 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1221 Deoptimization::DeoptReason reason;
1222 if (assert_null)
1223 reason = Deoptimization::Reason_null_assert;
1224 else if (type == T_OBJECT)
1225 reason = Deoptimization::Reason_null_check;
1226 else
1227 reason = Deoptimization::Reason_div0_check;
1229 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1230 // ciMethodData::has_trap_at will return a conservative -1 if any
1231 // must-be-null assertion has failed. This could cause performance
1232 // problems for a method after its first do_null_assert failure.
1233 // Consider using 'Reason_class_check' instead?
1235 // To cause an implicit null check, we set the not-null probability
1236 // to the maximum (PROB_MAX). For an explicit check the probability
1237 // is set to a smaller value.
1238 if (null_control != NULL || too_many_traps(reason)) {
1239 // probability is less likely
1240 ok_prob = PROB_LIKELY_MAG(3);
1241 } else if (!assert_null &&
1242 (ImplicitNullCheckThreshold > 0) &&
1243 method() != NULL &&
1244 (method()->method_data()->trap_count(reason)
1245 >= (uint)ImplicitNullCheckThreshold)) {
1246 ok_prob = PROB_LIKELY_MAG(3);
1247 }
1249 if (null_control != NULL) {
1250 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1251 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1252 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
1253 if (null_true == top())
1254 explicit_null_checks_elided++;
1255 (*null_control) = null_true;
1256 } else {
1257 BuildCutout unless(this, tst, ok_prob);
1258 // Check for optimizer eliding test at parse time
1259 if (stopped()) {
1260 // Failure not possible; do not bother making uncommon trap.
1261 explicit_null_checks_elided++;
1262 } else if (assert_null) {
1263 uncommon_trap(reason,
1264 Deoptimization::Action_make_not_entrant,
1265 NULL, "assert_null");
1266 } else {
1267 replace_in_map(value, zerocon(type));
1268 builtin_throw(reason);
1269 }
1270 }
1272 // Must throw exception, fall-thru not possible?
1273 if (stopped()) {
1274 return top(); // No result
1275 }
1277 if (assert_null) {
1278 // Cast obj to null on this path.
1279 replace_in_map(value, zerocon(type));
1280 return zerocon(type);
1281 }
1283 // Cast obj to not-null on this path, if there is no null_control.
1284 // (If there is a null_control, a non-null value may come back to haunt us.)
1285 if (type == T_OBJECT) {
1286 Node* cast = cast_not_null(value, false);
1287 if (null_control == NULL || (*null_control) == top())
1288 replace_in_map(value, cast);
1289 value = cast;
1290 }
1292 return value;
1293 }
1296 //------------------------------cast_not_null----------------------------------
1297 // Cast obj to not-null on this path
1298 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1299 const Type *t = _gvn.type(obj);
1300 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1301 // Object is already not-null?
1302 if( t == t_not_null ) return obj;
1304 Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1305 cast->init_req(0, control());
1306 cast = _gvn.transform( cast );
1308 // Scan for instances of 'obj' in the current JVM mapping.
1309 // These instances are known to be not-null after the test.
1310 if (do_replace_in_map)
1311 replace_in_map(obj, cast);
1313 return cast; // Return casted value
1314 }
1317 //--------------------------replace_in_map-------------------------------------
1318 void GraphKit::replace_in_map(Node* old, Node* neww) {
1319 this->map()->replace_edge(old, neww);
1321 // Note: This operation potentially replaces any edge
1322 // on the map. This includes locals, stack, and monitors
1323 // of the current (innermost) JVM state.
1325 // We can consider replacing in caller maps.
1326 // The idea would be that an inlined function's null checks
1327 // can be shared with the entire inlining tree.
1328 // The expense of doing this is that the PreserveJVMState class
1329 // would have to preserve caller states too, with a deep copy.
1330 }
1334 //=============================================================================
1335 //--------------------------------memory---------------------------------------
1336 Node* GraphKit::memory(uint alias_idx) {
1337 MergeMemNode* mem = merged_memory();
1338 Node* p = mem->memory_at(alias_idx);
1339 _gvn.set_type(p, Type::MEMORY); // must be mapped
1340 return p;
1341 }
1343 //-----------------------------reset_memory------------------------------------
1344 Node* GraphKit::reset_memory() {
1345 Node* mem = map()->memory();
1346 // do not use this node for any more parsing!
1347 debug_only( map()->set_memory((Node*)NULL) );
1348 return _gvn.transform( mem );
1349 }
1351 //------------------------------set_all_memory---------------------------------
1352 void GraphKit::set_all_memory(Node* newmem) {
1353 Node* mergemem = MergeMemNode::make(C, newmem);
1354 gvn().set_type_bottom(mergemem);
1355 map()->set_memory(mergemem);
1356 }
1358 //------------------------------set_all_memory_call----------------------------
1359 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1360 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1361 set_all_memory(newmem);
1362 }
1364 //=============================================================================
1365 //
1366 // parser factory methods for MemNodes
1367 //
1368 // These are layered on top of the factory methods in LoadNode and StoreNode,
1369 // and integrate with the parser's memory state and _gvn engine.
1370 //
1372 // factory methods in "int adr_idx"
1373 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1374 int adr_idx,
1375 bool require_atomic_access) {
1376 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1377 const TypePtr* adr_type = NULL; // debug-mode-only argument
1378 debug_only(adr_type = C->get_adr_type(adr_idx));
1379 Node* mem = memory(adr_idx);
1380 Node* ld;
1381 if (require_atomic_access && bt == T_LONG) {
1382 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1383 } else {
1384 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1385 }
1386 return _gvn.transform(ld);
1387 }
1389 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1390 int adr_idx,
1391 bool require_atomic_access) {
1392 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1393 const TypePtr* adr_type = NULL;
1394 debug_only(adr_type = C->get_adr_type(adr_idx));
1395 Node *mem = memory(adr_idx);
1396 Node* st;
1397 if (require_atomic_access && bt == T_LONG) {
1398 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1399 } else {
1400 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1401 }
1402 st = _gvn.transform(st);
1403 set_memory(st, adr_idx);
1404 // Back-to-back stores can only remove intermediate store with DU info
1405 // so push on worklist for optimizer.
1406 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1407 record_for_igvn(st);
1409 return st;
1410 }
1413 void GraphKit::pre_barrier(Node* ctl,
1414 Node* obj,
1415 Node* adr,
1416 uint adr_idx,
1417 Node* val,
1418 const TypeOopPtr* val_type,
1419 BasicType bt) {
1420 BarrierSet* bs = Universe::heap()->barrier_set();
1421 set_control(ctl);
1422 switch (bs->kind()) {
1423 case BarrierSet::G1SATBCT:
1424 case BarrierSet::G1SATBCTLogging:
1425 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt);
1426 break;
1428 case BarrierSet::CardTableModRef:
1429 case BarrierSet::CardTableExtension:
1430 case BarrierSet::ModRef:
1431 break;
1433 case BarrierSet::Other:
1434 default :
1435 ShouldNotReachHere();
1437 }
1438 }
1440 void GraphKit::post_barrier(Node* ctl,
1441 Node* store,
1442 Node* obj,
1443 Node* adr,
1444 uint adr_idx,
1445 Node* val,
1446 BasicType bt,
1447 bool use_precise) {
1448 BarrierSet* bs = Universe::heap()->barrier_set();
1449 set_control(ctl);
1450 switch (bs->kind()) {
1451 case BarrierSet::G1SATBCT:
1452 case BarrierSet::G1SATBCTLogging:
1453 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1454 break;
1456 case BarrierSet::CardTableModRef:
1457 case BarrierSet::CardTableExtension:
1458 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1459 break;
1461 case BarrierSet::ModRef:
1462 break;
1464 case BarrierSet::Other:
1465 default :
1466 ShouldNotReachHere();
1468 }
1469 }
1471 Node* GraphKit::store_oop(Node* ctl,
1472 Node* obj,
1473 Node* adr,
1474 const TypePtr* adr_type,
1475 Node* val,
1476 const TypeOopPtr* val_type,
1477 BasicType bt,
1478 bool use_precise) {
1480 set_control(ctl);
1481 if (stopped()) return top(); // Dead path ?
1483 assert(bt == T_OBJECT, "sanity");
1484 assert(val != NULL, "not dead path");
1485 uint adr_idx = C->get_alias_index(adr_type);
1486 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1488 pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt);
1489 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1490 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1491 return store;
1492 }
1494 // Could be an array or object we don't know at compile time (unsafe ref.)
1495 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1496 Node* obj, // containing obj
1497 Node* adr, // actual adress to store val at
1498 const TypePtr* adr_type,
1499 Node* val,
1500 BasicType bt) {
1501 Compile::AliasType* at = C->alias_type(adr_type);
1502 const TypeOopPtr* val_type = NULL;
1503 if (adr_type->isa_instptr()) {
1504 if (at->field() != NULL) {
1505 // known field. This code is a copy of the do_put_xxx logic.
1506 ciField* field = at->field();
1507 if (!field->type()->is_loaded()) {
1508 val_type = TypeInstPtr::BOTTOM;
1509 } else {
1510 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1511 }
1512 }
1513 } else if (adr_type->isa_aryptr()) {
1514 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1515 }
1516 if (val_type == NULL) {
1517 val_type = TypeInstPtr::BOTTOM;
1518 }
1519 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1520 }
1523 //-------------------------array_element_address-------------------------
1524 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1525 const TypeInt* sizetype) {
1526 uint shift = exact_log2(type2aelembytes(elembt));
1527 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1529 // short-circuit a common case (saves lots of confusing waste motion)
1530 jint idx_con = find_int_con(idx, -1);
1531 if (idx_con >= 0) {
1532 intptr_t offset = header + ((intptr_t)idx_con << shift);
1533 return basic_plus_adr(ary, offset);
1534 }
1536 // must be correct type for alignment purposes
1537 Node* base = basic_plus_adr(ary, header);
1538 #ifdef _LP64
1539 // The scaled index operand to AddP must be a clean 64-bit value.
1540 // Java allows a 32-bit int to be incremented to a negative
1541 // value, which appears in a 64-bit register as a large
1542 // positive number. Using that large positive number as an
1543 // operand in pointer arithmetic has bad consequences.
1544 // On the other hand, 32-bit overflow is rare, and the possibility
1545 // can often be excluded, if we annotate the ConvI2L node with
1546 // a type assertion that its value is known to be a small positive
1547 // number. (The prior range check has ensured this.)
1548 // This assertion is used by ConvI2LNode::Ideal.
1549 int index_max = max_jint - 1; // array size is max_jint, index is one less
1550 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1551 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1552 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1553 #endif
1554 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1555 return basic_plus_adr(ary, base, scale);
1556 }
1558 //-------------------------load_array_element-------------------------
1559 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1560 const Type* elemtype = arytype->elem();
1561 BasicType elembt = elemtype->array_element_basic_type();
1562 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1563 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1564 return ld;
1565 }
1567 //-------------------------set_arguments_for_java_call-------------------------
1568 // Arguments (pre-popped from the stack) are taken from the JVMS.
1569 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1570 // Add the call arguments:
1571 uint nargs = call->method()->arg_size();
1572 for (uint i = 0; i < nargs; i++) {
1573 Node* arg = argument(i);
1574 call->init_req(i + TypeFunc::Parms, arg);
1575 }
1576 }
1578 //---------------------------set_edges_for_java_call---------------------------
1579 // Connect a newly created call into the current JVMS.
1580 // A return value node (if any) is returned from set_edges_for_java_call.
1581 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1583 // Add the predefined inputs:
1584 call->init_req( TypeFunc::Control, control() );
1585 call->init_req( TypeFunc::I_O , i_o() );
1586 call->init_req( TypeFunc::Memory , reset_memory() );
1587 call->init_req( TypeFunc::FramePtr, frameptr() );
1588 call->init_req( TypeFunc::ReturnAdr, top() );
1590 add_safepoint_edges(call, must_throw);
1592 Node* xcall = _gvn.transform(call);
1594 if (xcall == top()) {
1595 set_control(top());
1596 return;
1597 }
1598 assert(xcall == call, "call identity is stable");
1600 // Re-use the current map to produce the result.
1602 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1603 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1604 set_all_memory_call(xcall, separate_io_proj);
1606 //return xcall; // no need, caller already has it
1607 }
1609 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1610 if (stopped()) return top(); // maybe the call folded up?
1612 // Capture the return value, if any.
1613 Node* ret;
1614 if (call->method() == NULL ||
1615 call->method()->return_type()->basic_type() == T_VOID)
1616 ret = top();
1617 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1619 // Note: Since any out-of-line call can produce an exception,
1620 // we always insert an I_O projection from the call into the result.
1622 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1624 if (separate_io_proj) {
1625 // The caller requested separate projections be used by the fall
1626 // through and exceptional paths, so replace the projections for
1627 // the fall through path.
1628 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) ));
1629 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ));
1630 }
1631 return ret;
1632 }
1634 //--------------------set_predefined_input_for_runtime_call--------------------
1635 // Reading and setting the memory state is way conservative here.
1636 // The real problem is that I am not doing real Type analysis on memory,
1637 // so I cannot distinguish card mark stores from other stores. Across a GC
1638 // point the Store Barrier and the card mark memory has to agree. I cannot
1639 // have a card mark store and its barrier split across the GC point from
1640 // either above or below. Here I get that to happen by reading ALL of memory.
1641 // A better answer would be to separate out card marks from other memory.
1642 // For now, return the input memory state, so that it can be reused
1643 // after the call, if this call has restricted memory effects.
1644 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1645 // Set fixed predefined input arguments
1646 Node* memory = reset_memory();
1647 call->init_req( TypeFunc::Control, control() );
1648 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1649 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1650 call->init_req( TypeFunc::FramePtr, frameptr() );
1651 call->init_req( TypeFunc::ReturnAdr, top() );
1652 return memory;
1653 }
1655 //-------------------set_predefined_output_for_runtime_call--------------------
1656 // Set control and memory (not i_o) from the call.
1657 // If keep_mem is not NULL, use it for the output state,
1658 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1659 // If hook_mem is NULL, this call produces no memory effects at all.
1660 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1661 // then only that memory slice is taken from the call.
1662 // In the last case, we must put an appropriate memory barrier before
1663 // the call, so as to create the correct anti-dependencies on loads
1664 // preceding the call.
1665 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1666 Node* keep_mem,
1667 const TypePtr* hook_mem) {
1668 // no i/o
1669 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1670 if (keep_mem) {
1671 // First clone the existing memory state
1672 set_all_memory(keep_mem);
1673 if (hook_mem != NULL) {
1674 // Make memory for the call
1675 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1676 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1677 // We also use hook_mem to extract specific effects from arraycopy stubs.
1678 set_memory(mem, hook_mem);
1679 }
1680 // ...else the call has NO memory effects.
1682 // Make sure the call advertises its memory effects precisely.
1683 // This lets us build accurate anti-dependences in gcm.cpp.
1684 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1685 "call node must be constructed correctly");
1686 } else {
1687 assert(hook_mem == NULL, "");
1688 // This is not a "slow path" call; all memory comes from the call.
1689 set_all_memory_call(call);
1690 }
1691 }
1694 // Replace the call with the current state of the kit.
1695 void GraphKit::replace_call(CallNode* call, Node* result) {
1696 JVMState* ejvms = NULL;
1697 if (has_exceptions()) {
1698 ejvms = transfer_exceptions_into_jvms();
1699 }
1701 SafePointNode* final_state = stop();
1703 // Find all the needed outputs of this call
1704 CallProjections callprojs;
1705 call->extract_projections(&callprojs, true);
1707 // Replace all the old call edges with the edges from the inlining result
1708 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control));
1709 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory));
1710 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O));
1712 // Replace the result with the new result if it exists and is used
1713 if (callprojs.resproj != NULL && result != NULL) {
1714 C->gvn_replace_by(callprojs.resproj, result);
1715 }
1717 if (ejvms == NULL) {
1718 // No exception edges to simply kill off those paths
1719 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1720 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1721 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1723 // Replace the old exception object with top
1724 if (callprojs.exobj != NULL) {
1725 C->gvn_replace_by(callprojs.exobj, C->top());
1726 }
1727 } else {
1728 GraphKit ekit(ejvms);
1730 // Load my combined exception state into the kit, with all phis transformed:
1731 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1733 Node* ex_oop = ekit.use_exception_state(ex_map);
1735 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1736 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1737 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1739 // Replace the old exception object with the newly created one
1740 if (callprojs.exobj != NULL) {
1741 C->gvn_replace_by(callprojs.exobj, ex_oop);
1742 }
1743 }
1745 // Disconnect the call from the graph
1746 call->disconnect_inputs(NULL);
1747 C->gvn_replace_by(call, C->top());
1748 }
1751 //------------------------------increment_counter------------------------------
1752 // for statistics: increment a VM counter by 1
1754 void GraphKit::increment_counter(address counter_addr) {
1755 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1756 increment_counter(adr1);
1757 }
1759 void GraphKit::increment_counter(Node* counter_addr) {
1760 int adr_type = Compile::AliasIdxRaw;
1761 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
1762 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1763 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type );
1764 }
1767 //------------------------------uncommon_trap----------------------------------
1768 // Bail out to the interpreter in mid-method. Implemented by calling the
1769 // uncommon_trap blob. This helper function inserts a runtime call with the
1770 // right debug info.
1771 void GraphKit::uncommon_trap(int trap_request,
1772 ciKlass* klass, const char* comment,
1773 bool must_throw,
1774 bool keep_exact_action) {
1775 if (failing()) stop();
1776 if (stopped()) return; // trap reachable?
1778 // Note: If ProfileTraps is true, and if a deopt. actually
1779 // occurs here, the runtime will make sure an MDO exists. There is
1780 // no need to call method()->build_method_data() at this point.
1782 #ifdef ASSERT
1783 if (!must_throw) {
1784 // Make sure the stack has at least enough depth to execute
1785 // the current bytecode.
1786 int inputs, ignore;
1787 if (compute_stack_effects(inputs, ignore)) {
1788 assert(sp() >= inputs, "must have enough JVMS stack to execute");
1789 // It is a frequent error in library_call.cpp to issue an
1790 // uncommon trap with the _sp value already popped.
1791 }
1792 }
1793 #endif
1795 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1796 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1798 switch (action) {
1799 case Deoptimization::Action_maybe_recompile:
1800 case Deoptimization::Action_reinterpret:
1801 // Temporary fix for 6529811 to allow virtual calls to be sure they
1802 // get the chance to go from mono->bi->mega
1803 if (!keep_exact_action &&
1804 Deoptimization::trap_request_index(trap_request) < 0 &&
1805 too_many_recompiles(reason)) {
1806 // This BCI is causing too many recompilations.
1807 action = Deoptimization::Action_none;
1808 trap_request = Deoptimization::make_trap_request(reason, action);
1809 } else {
1810 C->set_trap_can_recompile(true);
1811 }
1812 break;
1813 case Deoptimization::Action_make_not_entrant:
1814 C->set_trap_can_recompile(true);
1815 break;
1816 #ifdef ASSERT
1817 case Deoptimization::Action_none:
1818 case Deoptimization::Action_make_not_compilable:
1819 break;
1820 default:
1821 assert(false, "bad action");
1822 #endif
1823 }
1825 if (TraceOptoParse) {
1826 char buf[100];
1827 tty->print_cr("Uncommon trap %s at bci:%d",
1828 Deoptimization::format_trap_request(buf, sizeof(buf),
1829 trap_request), bci());
1830 }
1832 CompileLog* log = C->log();
1833 if (log != NULL) {
1834 int kid = (klass == NULL)? -1: log->identify(klass);
1835 log->begin_elem("uncommon_trap bci='%d'", bci());
1836 char buf[100];
1837 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1838 trap_request));
1839 if (kid >= 0) log->print(" klass='%d'", kid);
1840 if (comment != NULL) log->print(" comment='%s'", comment);
1841 log->end_elem();
1842 }
1844 // Make sure any guarding test views this path as very unlikely
1845 Node *i0 = control()->in(0);
1846 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1847 IfNode *iff = i0->as_If();
1848 float f = iff->_prob; // Get prob
1849 if (control()->Opcode() == Op_IfTrue) {
1850 if (f > PROB_UNLIKELY_MAG(4))
1851 iff->_prob = PROB_MIN;
1852 } else {
1853 if (f < PROB_LIKELY_MAG(4))
1854 iff->_prob = PROB_MAX;
1855 }
1856 }
1858 // Clear out dead values from the debug info.
1859 kill_dead_locals();
1861 // Now insert the uncommon trap subroutine call
1862 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin();
1863 const TypePtr* no_memory_effects = NULL;
1864 // Pass the index of the class to be loaded
1865 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1866 (must_throw ? RC_MUST_THROW : 0),
1867 OptoRuntime::uncommon_trap_Type(),
1868 call_addr, "uncommon_trap", no_memory_effects,
1869 intcon(trap_request));
1870 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1871 "must extract request correctly from the graph");
1872 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1874 call->set_req(TypeFunc::ReturnAdr, returnadr());
1875 // The debug info is the only real input to this call.
1877 // Halt-and-catch fire here. The above call should never return!
1878 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1879 _gvn.set_type_bottom(halt);
1880 root()->add_req(halt);
1882 stop_and_kill_map();
1883 }
1886 //--------------------------just_allocated_object------------------------------
1887 // Report the object that was just allocated.
1888 // It must be the case that there are no intervening safepoints.
1889 // We use this to determine if an object is so "fresh" that
1890 // it does not require card marks.
1891 Node* GraphKit::just_allocated_object(Node* current_control) {
1892 if (C->recent_alloc_ctl() == current_control)
1893 return C->recent_alloc_obj();
1894 return NULL;
1895 }
1898 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1899 // (Note: TypeFunc::make has a cache that makes this fast.)
1900 const TypeFunc* tf = TypeFunc::make(dest_method);
1901 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
1902 for (int j = 0; j < nargs; j++) {
1903 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1904 if( targ->basic_type() == T_DOUBLE ) {
1905 // If any parameters are doubles, they must be rounded before
1906 // the call, dstore_rounding does gvn.transform
1907 Node *arg = argument(j);
1908 arg = dstore_rounding(arg);
1909 set_argument(j, arg);
1910 }
1911 }
1912 }
1914 void GraphKit::round_double_result(ciMethod* dest_method) {
1915 // A non-strict method may return a double value which has an extended
1916 // exponent, but this must not be visible in a caller which is 'strict'
1917 // If a strict caller invokes a non-strict callee, round a double result
1919 BasicType result_type = dest_method->return_type()->basic_type();
1920 assert( method() != NULL, "must have caller context");
1921 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
1922 // Destination method's return value is on top of stack
1923 // dstore_rounding() does gvn.transform
1924 Node *result = pop_pair();
1925 result = dstore_rounding(result);
1926 push_pair(result);
1927 }
1928 }
1930 // rounding for strict float precision conformance
1931 Node* GraphKit::precision_rounding(Node* n) {
1932 return UseStrictFP && _method->flags().is_strict()
1933 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
1934 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
1935 : n;
1936 }
1938 // rounding for strict double precision conformance
1939 Node* GraphKit::dprecision_rounding(Node *n) {
1940 return UseStrictFP && _method->flags().is_strict()
1941 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
1942 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1943 : n;
1944 }
1946 // rounding for non-strict double stores
1947 Node* GraphKit::dstore_rounding(Node* n) {
1948 return Matcher::strict_fp_requires_explicit_rounding
1949 && UseSSE <= 1
1950 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1951 : n;
1952 }
1954 //=============================================================================
1955 // Generate a fast path/slow path idiom. Graph looks like:
1956 // [foo] indicates that 'foo' is a parameter
1957 //
1958 // [in] NULL
1959 // \ /
1960 // CmpP
1961 // Bool ne
1962 // If
1963 // / \
1964 // True False-<2>
1965 // / |
1966 // / cast_not_null
1967 // Load | | ^
1968 // [fast_test] | |
1969 // gvn to opt_test | |
1970 // / \ | <1>
1971 // True False |
1972 // | \\ |
1973 // [slow_call] \[fast_result]
1974 // Ctl Val \ \
1975 // | \ \
1976 // Catch <1> \ \
1977 // / \ ^ \ \
1978 // Ex No_Ex | \ \
1979 // | \ \ | \ <2> \
1980 // ... \ [slow_res] | | \ [null_result]
1981 // \ \--+--+--- | |
1982 // \ | / \ | /
1983 // --------Region Phi
1984 //
1985 //=============================================================================
1986 // Code is structured as a series of driver functions all called 'do_XXX' that
1987 // call a set of helper functions. Helper functions first, then drivers.
1989 //------------------------------null_check_oop---------------------------------
1990 // Null check oop. Set null-path control into Region in slot 3.
1991 // Make a cast-not-nullness use the other not-null control. Return cast.
1992 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
1993 bool never_see_null) {
1994 // Initial NULL check taken path
1995 (*null_control) = top();
1996 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
1998 // Generate uncommon_trap:
1999 if (never_see_null && (*null_control) != top()) {
2000 // If we see an unexpected null at a check-cast we record it and force a
2001 // recompile; the offending check-cast will be compiled to handle NULLs.
2002 // If we see more than one offending BCI, then all checkcasts in the
2003 // method will be compiled to handle NULLs.
2004 PreserveJVMState pjvms(this);
2005 set_control(*null_control);
2006 replace_in_map(value, null());
2007 uncommon_trap(Deoptimization::Reason_null_check,
2008 Deoptimization::Action_make_not_entrant);
2009 (*null_control) = top(); // NULL path is dead
2010 }
2012 // Cast away null-ness on the result
2013 return cast;
2014 }
2016 //------------------------------opt_iff----------------------------------------
2017 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2018 // Return slow-path control.
2019 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2020 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2022 // Fast path taken; set region slot 2
2023 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2024 region->init_req(2,fast_taken); // Capture fast-control
2026 // Fast path not-taken, i.e. slow path
2027 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2028 return slow_taken;
2029 }
2031 //-----------------------------make_runtime_call-------------------------------
2032 Node* GraphKit::make_runtime_call(int flags,
2033 const TypeFunc* call_type, address call_addr,
2034 const char* call_name,
2035 const TypePtr* adr_type,
2036 // The following parms are all optional.
2037 // The first NULL ends the list.
2038 Node* parm0, Node* parm1,
2039 Node* parm2, Node* parm3,
2040 Node* parm4, Node* parm5,
2041 Node* parm6, Node* parm7) {
2042 // Slow-path call
2043 int size = call_type->domain()->cnt();
2044 bool is_leaf = !(flags & RC_NO_LEAF);
2045 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2046 if (call_name == NULL) {
2047 assert(!is_leaf, "must supply name for leaf");
2048 call_name = OptoRuntime::stub_name(call_addr);
2049 }
2050 CallNode* call;
2051 if (!is_leaf) {
2052 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2053 bci(), adr_type);
2054 } else if (flags & RC_NO_FP) {
2055 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2056 } else {
2057 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2058 }
2060 // The following is similar to set_edges_for_java_call,
2061 // except that the memory effects of the call are restricted to AliasIdxRaw.
2063 // Slow path call has no side-effects, uses few values
2064 bool wide_in = !(flags & RC_NARROW_MEM);
2065 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2067 Node* prev_mem = NULL;
2068 if (wide_in) {
2069 prev_mem = set_predefined_input_for_runtime_call(call);
2070 } else {
2071 assert(!wide_out, "narrow in => narrow out");
2072 Node* narrow_mem = memory(adr_type);
2073 prev_mem = reset_memory();
2074 map()->set_memory(narrow_mem);
2075 set_predefined_input_for_runtime_call(call);
2076 }
2078 // Hook each parm in order. Stop looking at the first NULL.
2079 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2080 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2081 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2082 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2083 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2084 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2085 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2086 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2087 /* close each nested if ===> */ } } } } } } } }
2088 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2090 if (!is_leaf) {
2091 // Non-leaves can block and take safepoints:
2092 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2093 }
2094 // Non-leaves can throw exceptions:
2095 if (has_io) {
2096 call->set_req(TypeFunc::I_O, i_o());
2097 }
2099 if (flags & RC_UNCOMMON) {
2100 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2101 // (An "if" probability corresponds roughly to an unconditional count.
2102 // Sort of.)
2103 call->set_cnt(PROB_UNLIKELY_MAG(4));
2104 }
2106 Node* c = _gvn.transform(call);
2107 assert(c == call, "cannot disappear");
2109 if (wide_out) {
2110 // Slow path call has full side-effects.
2111 set_predefined_output_for_runtime_call(call);
2112 } else {
2113 // Slow path call has few side-effects, and/or sets few values.
2114 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2115 }
2117 if (has_io) {
2118 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2119 }
2120 return call;
2122 }
2124 //------------------------------merge_memory-----------------------------------
2125 // Merge memory from one path into the current memory state.
2126 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2127 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2128 Node* old_slice = mms.force_memory();
2129 Node* new_slice = mms.memory2();
2130 if (old_slice != new_slice) {
2131 PhiNode* phi;
2132 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2133 phi = new_slice->as_Phi();
2134 #ifdef ASSERT
2135 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2136 old_slice = old_slice->in(new_path);
2137 // Caller is responsible for ensuring that any pre-existing
2138 // phis are already aware of old memory.
2139 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2140 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2141 #endif
2142 mms.set_memory(phi);
2143 } else {
2144 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2145 _gvn.set_type(phi, Type::MEMORY);
2146 phi->set_req(new_path, new_slice);
2147 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2148 }
2149 }
2150 }
2151 }
2153 //------------------------------make_slow_call_ex------------------------------
2154 // Make the exception handler hookups for the slow call
2155 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2156 if (stopped()) return;
2158 // Make a catch node with just two handlers: fall-through and catch-all
2159 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2160 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2161 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2162 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2164 { PreserveJVMState pjvms(this);
2165 set_control(excp);
2166 set_i_o(i_o);
2168 if (excp != top()) {
2169 // Create an exception state also.
2170 // Use an exact type if the caller has specified a specific exception.
2171 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2172 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
2173 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2174 }
2175 }
2177 // Get the no-exception control from the CatchNode.
2178 set_control(norm);
2179 }
2182 //-------------------------------gen_subtype_check-----------------------------
2183 // Generate a subtyping check. Takes as input the subtype and supertype.
2184 // Returns 2 values: sets the default control() to the true path and returns
2185 // the false path. Only reads invariant memory; sets no (visible) memory.
2186 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2187 // but that's not exposed to the optimizer. This call also doesn't take in an
2188 // Object; if you wish to check an Object you need to load the Object's class
2189 // prior to coming here.
2190 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2191 // Fast check for identical types, perhaps identical constants.
2192 // The types can even be identical non-constants, in cases
2193 // involving Array.newInstance, Object.clone, etc.
2194 if (subklass == superklass)
2195 return top(); // false path is dead; no test needed.
2197 if (_gvn.type(superklass)->singleton()) {
2198 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2199 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2201 // In the common case of an exact superklass, try to fold up the
2202 // test before generating code. You may ask, why not just generate
2203 // the code and then let it fold up? The answer is that the generated
2204 // code will necessarily include null checks, which do not always
2205 // completely fold away. If they are also needless, then they turn
2206 // into a performance loss. Example:
2207 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2208 // Here, the type of 'fa' is often exact, so the store check
2209 // of fa[1]=x will fold up, without testing the nullness of x.
2210 switch (static_subtype_check(superk, subk)) {
2211 case SSC_always_false:
2212 {
2213 Node* always_fail = control();
2214 set_control(top());
2215 return always_fail;
2216 }
2217 case SSC_always_true:
2218 return top();
2219 case SSC_easy_test:
2220 {
2221 // Just do a direct pointer compare and be done.
2222 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2223 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2224 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2225 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2226 return _gvn.transform( new(C, 1) IfFalseNode(iff) );
2227 }
2228 case SSC_full_test:
2229 break;
2230 default:
2231 ShouldNotReachHere();
2232 }
2233 }
2235 // %%% Possible further optimization: Even if the superklass is not exact,
2236 // if the subklass is the unique subtype of the superklass, the check
2237 // will always succeed. We could leave a dependency behind to ensure this.
2239 // First load the super-klass's check-offset
2240 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
2241 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2242 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
2243 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2245 // Load from the sub-klass's super-class display list, or a 1-word cache of
2246 // the secondary superclass list, or a failing value with a sentinel offset
2247 // if the super-klass is an interface or exceptionally deep in the Java
2248 // hierarchy and we have to scan the secondary superclass list the hard way.
2249 // Worst-case type is a little odd: NULL is allowed as a result (usually
2250 // klass loads can never produce a NULL).
2251 Node *chk_off_X = ConvI2X(chk_off);
2252 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2253 // For some types like interfaces the following loadKlass is from a 1-word
2254 // cache which is mutable so can't use immutable memory. Other
2255 // types load from the super-class display table which is immutable.
2256 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2257 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2259 // Compile speed common case: ARE a subtype and we canNOT fail
2260 if( superklass == nkls )
2261 return top(); // false path is dead; no test needed.
2263 // See if we get an immediate positive hit. Happens roughly 83% of the
2264 // time. Test to see if the value loaded just previously from the subklass
2265 // is exactly the superklass.
2266 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2267 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2268 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2269 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2270 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2272 // Compile speed common case: Check for being deterministic right now. If
2273 // chk_off is a constant and not equal to cacheoff then we are NOT a
2274 // subklass. In this case we need exactly the 1 test above and we can
2275 // return those results immediately.
2276 if (!might_be_cache) {
2277 Node* not_subtype_ctrl = control();
2278 set_control(iftrue1); // We need exactly the 1 test above
2279 return not_subtype_ctrl;
2280 }
2282 // Gather the various success & failures here
2283 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2284 record_for_igvn(r_ok_subtype);
2285 RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2286 record_for_igvn(r_not_subtype);
2288 r_ok_subtype->init_req(1, iftrue1);
2290 // Check for immediate negative hit. Happens roughly 11% of the time (which
2291 // is roughly 63% of the remaining cases). Test to see if the loaded
2292 // check-offset points into the subklass display list or the 1-element
2293 // cache. If it points to the display (and NOT the cache) and the display
2294 // missed then it's not a subtype.
2295 Node *cacheoff = _gvn.intcon(cacheoff_con);
2296 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2297 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2298 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2299 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2300 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2302 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2303 // No performance impact (too rare) but allows sharing of secondary arrays
2304 // which has some footprint reduction.
2305 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2306 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2307 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2308 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2309 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2311 // -- Roads not taken here: --
2312 // We could also have chosen to perform the self-check at the beginning
2313 // of this code sequence, as the assembler does. This would not pay off
2314 // the same way, since the optimizer, unlike the assembler, can perform
2315 // static type analysis to fold away many successful self-checks.
2316 // Non-foldable self checks work better here in second position, because
2317 // the initial primary superclass check subsumes a self-check for most
2318 // types. An exception would be a secondary type like array-of-interface,
2319 // which does not appear in its own primary supertype display.
2320 // Finally, we could have chosen to move the self-check into the
2321 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2322 // dependent manner. But it is worthwhile to have the check here,
2323 // where it can be perhaps be optimized. The cost in code space is
2324 // small (register compare, branch).
2326 // Now do a linear scan of the secondary super-klass array. Again, no real
2327 // performance impact (too rare) but it's gotta be done.
2328 // Since the code is rarely used, there is no penalty for moving it
2329 // out of line, and it can only improve I-cache density.
2330 // The decision to inline or out-of-line this final check is platform
2331 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2332 Node* psc = _gvn.transform(
2333 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2335 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2336 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2337 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2338 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2339 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2341 // Return false path; set default control to true path.
2342 set_control( _gvn.transform(r_ok_subtype) );
2343 return _gvn.transform(r_not_subtype);
2344 }
2346 //----------------------------static_subtype_check-----------------------------
2347 // Shortcut important common cases when superklass is exact:
2348 // (0) superklass is java.lang.Object (can occur in reflective code)
2349 // (1) subklass is already limited to a subtype of superklass => always ok
2350 // (2) subklass does not overlap with superklass => always fail
2351 // (3) superklass has NO subtypes and we can check with a simple compare.
2352 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2353 if (StressReflectiveCode) {
2354 return SSC_full_test; // Let caller generate the general case.
2355 }
2357 if (superk == env()->Object_klass()) {
2358 return SSC_always_true; // (0) this test cannot fail
2359 }
2361 ciType* superelem = superk;
2362 if (superelem->is_array_klass())
2363 superelem = superelem->as_array_klass()->base_element_type();
2365 if (!subk->is_interface()) { // cannot trust static interface types yet
2366 if (subk->is_subtype_of(superk)) {
2367 return SSC_always_true; // (1) false path dead; no dynamic test needed
2368 }
2369 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2370 !superk->is_subtype_of(subk)) {
2371 return SSC_always_false;
2372 }
2373 }
2375 // If casting to an instance klass, it must have no subtypes
2376 if (superk->is_interface()) {
2377 // Cannot trust interfaces yet.
2378 // %%% S.B. superk->nof_implementors() == 1
2379 } else if (superelem->is_instance_klass()) {
2380 ciInstanceKlass* ik = superelem->as_instance_klass();
2381 if (!ik->has_subklass() && !ik->is_interface()) {
2382 if (!ik->is_final()) {
2383 // Add a dependency if there is a chance of a later subclass.
2384 C->dependencies()->assert_leaf_type(ik);
2385 }
2386 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2387 }
2388 } else {
2389 // A primitive array type has no subtypes.
2390 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2391 }
2393 return SSC_full_test;
2394 }
2396 // Profile-driven exact type check:
2397 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2398 float prob,
2399 Node* *casted_receiver) {
2400 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2401 Node* recv_klass = load_object_klass(receiver);
2402 Node* want_klass = makecon(tklass);
2403 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2404 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2405 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2406 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2407 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2409 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2410 assert(recv_xtype->klass_is_exact(), "");
2412 // Subsume downstream occurrences of receiver with a cast to
2413 // recv_xtype, since now we know what the type will be.
2414 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2415 (*casted_receiver) = _gvn.transform(cast);
2416 // (User must make the replace_in_map call.)
2418 return fail;
2419 }
2422 //-------------------------------gen_instanceof--------------------------------
2423 // Generate an instance-of idiom. Used by both the instance-of bytecode
2424 // and the reflective instance-of call.
2425 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) {
2426 C->set_has_split_ifs(true); // Has chance for split-if optimization
2427 assert( !stopped(), "dead parse path should be checked in callers" );
2428 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2429 "must check for not-null not-dead klass in callers");
2431 // Make the merge point
2432 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2433 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2434 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2435 C->set_has_split_ifs(true); // Has chance for split-if optimization
2437 // Null check; get casted pointer; set region slot 3
2438 Node* null_ctl = top();
2439 Node* not_null_obj = null_check_oop(subobj, &null_ctl);
2441 // If not_null_obj is dead, only null-path is taken
2442 if (stopped()) { // Doing instance-of on a NULL?
2443 set_control(null_ctl);
2444 return intcon(0);
2445 }
2446 region->init_req(_null_path, null_ctl);
2447 phi ->init_req(_null_path, intcon(0)); // Set null path value
2449 // Load the object's klass
2450 Node* obj_klass = load_object_klass(not_null_obj);
2452 // Generate the subtype check
2453 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2455 // Plug in the success path to the general merge in slot 1.
2456 region->init_req(_obj_path, control());
2457 phi ->init_req(_obj_path, intcon(1));
2459 // Plug in the failing path to the general merge in slot 2.
2460 region->init_req(_fail_path, not_subtype_ctrl);
2461 phi ->init_req(_fail_path, intcon(0));
2463 // Return final merged results
2464 set_control( _gvn.transform(region) );
2465 record_for_igvn(region);
2466 return _gvn.transform(phi);
2467 }
2469 //-------------------------------gen_checkcast---------------------------------
2470 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2471 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2472 // uncommon-trap paths work. Adjust stack after this call.
2473 // If failure_control is supplied and not null, it is filled in with
2474 // the control edge for the cast failure. Otherwise, an appropriate
2475 // uncommon trap or exception is thrown.
2476 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2477 Node* *failure_control) {
2478 kill_dead_locals(); // Benefit all the uncommon traps
2479 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2480 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2482 // Fast cutout: Check the case that the cast is vacuously true.
2483 // This detects the common cases where the test will short-circuit
2484 // away completely. We do this before we perform the null check,
2485 // because if the test is going to turn into zero code, we don't
2486 // want a residual null check left around. (Causes a slowdown,
2487 // for example, in some objArray manipulations, such as a[i]=a[j].)
2488 if (tk->singleton()) {
2489 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2490 if (objtp != NULL && objtp->klass() != NULL) {
2491 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2492 case SSC_always_true:
2493 return obj;
2494 case SSC_always_false:
2495 // It needs a null check because a null will *pass* the cast check.
2496 // A non-null value will always produce an exception.
2497 return do_null_assert(obj, T_OBJECT);
2498 }
2499 }
2500 }
2502 ciProfileData* data = NULL;
2503 if (failure_control == NULL) { // use MDO in regular case only
2504 assert(java_bc() == Bytecodes::_aastore ||
2505 java_bc() == Bytecodes::_checkcast,
2506 "interpreter profiles type checks only for these BCs");
2507 data = method()->method_data()->bci_to_data(bci());
2508 }
2510 // Make the merge point
2511 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2512 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2513 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
2514 C->set_has_split_ifs(true); // Has chance for split-if optimization
2516 // Use null-cast information if it is available
2517 bool never_see_null = false;
2518 // If we see an unexpected null at a check-cast we record it and force a
2519 // recompile; the offending check-cast will be compiled to handle NULLs.
2520 // If we see several offending BCIs, then all checkcasts in the
2521 // method will be compiled to handle NULLs.
2522 if (UncommonNullCast // Cutout for this technique
2523 && failure_control == NULL // regular case
2524 && obj != null() // And not the -Xcomp stupid case?
2525 && !too_many_traps(Deoptimization::Reason_null_check)) {
2526 // Finally, check the "null_seen" bit from the interpreter.
2527 if (data == NULL || !data->as_BitData()->null_seen()) {
2528 never_see_null = true;
2529 }
2530 }
2532 // Null check; get casted pointer; set region slot 3
2533 Node* null_ctl = top();
2534 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2536 // If not_null_obj is dead, only null-path is taken
2537 if (stopped()) { // Doing instance-of on a NULL?
2538 set_control(null_ctl);
2539 return null();
2540 }
2541 region->init_req(_null_path, null_ctl);
2542 phi ->init_req(_null_path, null()); // Set null path value
2544 Node* cast_obj = NULL; // the casted version of the object
2546 // If the profile has seen exactly one type, narrow to that type.
2547 // (The subsequent subtype check will always fold up.)
2548 if (UseTypeProfile && TypeProfileCasts && data != NULL &&
2549 // Counter has never been decremented (due to cast failure).
2550 // ...This is a reasonable thing to expect. It is true of
2551 // all casts inserted by javac to implement generic types.
2552 data->as_CounterData()->count() >= 0 &&
2553 !too_many_traps(Deoptimization::Reason_class_check)) {
2554 // (No, this isn't a call, but it's enough like a virtual call
2555 // to use the same ciMethod accessor to get the profile info...)
2556 ciCallProfile profile = method()->call_profile_at_bci(bci());
2557 if (profile.count() >= 0 && // no cast failures here
2558 profile.has_receiver(0) &&
2559 profile.morphism() == 1) {
2560 ciKlass* exact_kls = profile.receiver(0);
2561 int ssc = static_subtype_check(tk->klass(), exact_kls);
2562 if (ssc == SSC_always_true) {
2563 // If we narrow the type to match what the type profile sees,
2564 // we can then remove the rest of the cast.
2565 // This is a win, even if the exact_kls is very specific,
2566 // because downstream operations, such as method calls,
2567 // will often benefit from the sharper type.
2568 Node* exact_obj = not_null_obj; // will get updated in place...
2569 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2570 &exact_obj);
2571 { PreserveJVMState pjvms(this);
2572 set_control(slow_ctl);
2573 uncommon_trap(Deoptimization::Reason_class_check,
2574 Deoptimization::Action_maybe_recompile);
2575 }
2576 if (failure_control != NULL) // failure is now impossible
2577 (*failure_control) = top();
2578 replace_in_map(not_null_obj, exact_obj);
2579 // adjust the type of the phi to the exact klass:
2580 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR));
2581 cast_obj = exact_obj;
2582 }
2583 // assert(cast_obj != NULL)... except maybe the profile lied to us.
2584 }
2585 }
2587 if (cast_obj == NULL) {
2588 // Load the object's klass
2589 Node* obj_klass = load_object_klass(not_null_obj);
2591 // Generate the subtype check
2592 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2594 // Plug in success path into the merge
2595 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2596 not_null_obj, toop));
2597 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2598 if (failure_control == NULL) {
2599 if (not_subtype_ctrl != top()) { // If failure is possible
2600 PreserveJVMState pjvms(this);
2601 set_control(not_subtype_ctrl);
2602 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2603 }
2604 } else {
2605 (*failure_control) = not_subtype_ctrl;
2606 }
2607 }
2609 region->init_req(_obj_path, control());
2610 phi ->init_req(_obj_path, cast_obj);
2612 // A merge of NULL or Casted-NotNull obj
2613 Node* res = _gvn.transform(phi);
2615 // Note I do NOT always 'replace_in_map(obj,result)' here.
2616 // if( tk->klass()->can_be_primary_super() )
2617 // This means that if I successfully store an Object into an array-of-String
2618 // I 'forget' that the Object is really now known to be a String. I have to
2619 // do this because we don't have true union types for interfaces - if I store
2620 // a Baz into an array-of-Interface and then tell the optimizer it's an
2621 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2622 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2623 // replace_in_map( obj, res );
2625 // Return final merged results
2626 set_control( _gvn.transform(region) );
2627 record_for_igvn(region);
2628 return res;
2629 }
2631 //------------------------------next_monitor-----------------------------------
2632 // What number should be given to the next monitor?
2633 int GraphKit::next_monitor() {
2634 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2635 int next = current + C->sync_stack_slots();
2636 // Keep the toplevel high water mark current:
2637 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2638 return current;
2639 }
2641 //------------------------------insert_mem_bar---------------------------------
2642 // Memory barrier to avoid floating things around
2643 // The membar serves as a pinch point between both control and all memory slices.
2644 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2645 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2646 mb->init_req(TypeFunc::Control, control());
2647 mb->init_req(TypeFunc::Memory, reset_memory());
2648 Node* membar = _gvn.transform(mb);
2649 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2650 set_all_memory_call(membar);
2651 return membar;
2652 }
2654 //-------------------------insert_mem_bar_volatile----------------------------
2655 // Memory barrier to avoid floating things around
2656 // The membar serves as a pinch point between both control and memory(alias_idx).
2657 // If you want to make a pinch point on all memory slices, do not use this
2658 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2659 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2660 // When Parse::do_put_xxx updates a volatile field, it appends a series
2661 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2662 // The first membar is on the same memory slice as the field store opcode.
2663 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2664 // All the other membars (for other volatile slices, including AliasIdxBot,
2665 // which stands for all unknown volatile slices) are control-dependent
2666 // on the first membar. This prevents later volatile loads or stores
2667 // from sliding up past the just-emitted store.
2669 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2670 mb->set_req(TypeFunc::Control,control());
2671 if (alias_idx == Compile::AliasIdxBot) {
2672 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2673 } else {
2674 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2675 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2676 }
2677 Node* membar = _gvn.transform(mb);
2678 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2679 if (alias_idx == Compile::AliasIdxBot) {
2680 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2681 } else {
2682 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2683 }
2684 return membar;
2685 }
2687 //------------------------------shared_lock------------------------------------
2688 // Emit locking code.
2689 FastLockNode* GraphKit::shared_lock(Node* obj) {
2690 // bci is either a monitorenter bc or InvocationEntryBci
2691 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2692 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2694 if( !GenerateSynchronizationCode )
2695 return NULL; // Not locking things?
2696 if (stopped()) // Dead monitor?
2697 return NULL;
2699 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2701 // Box the stack location
2702 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2703 Node* mem = reset_memory();
2705 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2706 if (PrintPreciseBiasedLockingStatistics) {
2707 // Create the counters for this fast lock.
2708 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2709 }
2710 // Add monitor to debug info for the slow path. If we block inside the
2711 // slow path and de-opt, we need the monitor hanging around
2712 map()->push_monitor( flock );
2714 const TypeFunc *tf = LockNode::lock_type();
2715 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2717 lock->init_req( TypeFunc::Control, control() );
2718 lock->init_req( TypeFunc::Memory , mem );
2719 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2720 lock->init_req( TypeFunc::FramePtr, frameptr() );
2721 lock->init_req( TypeFunc::ReturnAdr, top() );
2723 lock->init_req(TypeFunc::Parms + 0, obj);
2724 lock->init_req(TypeFunc::Parms + 1, box);
2725 lock->init_req(TypeFunc::Parms + 2, flock);
2726 add_safepoint_edges(lock);
2728 lock = _gvn.transform( lock )->as_Lock();
2730 // lock has no side-effects, sets few values
2731 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2733 insert_mem_bar(Op_MemBarAcquire);
2735 // Add this to the worklist so that the lock can be eliminated
2736 record_for_igvn(lock);
2738 #ifndef PRODUCT
2739 if (PrintLockStatistics) {
2740 // Update the counter for this lock. Don't bother using an atomic
2741 // operation since we don't require absolute accuracy.
2742 lock->create_lock_counter(map()->jvms());
2743 int adr_type = Compile::AliasIdxRaw;
2744 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr()));
2745 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
2746 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
2747 store_to_memory(control(), counter_addr, incr, T_INT, adr_type);
2748 }
2749 #endif
2751 return flock;
2752 }
2755 //------------------------------shared_unlock----------------------------------
2756 // Emit unlocking code.
2757 void GraphKit::shared_unlock(Node* box, Node* obj) {
2758 // bci is either a monitorenter bc or InvocationEntryBci
2759 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2760 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2762 if( !GenerateSynchronizationCode )
2763 return;
2764 if (stopped()) { // Dead monitor?
2765 map()->pop_monitor(); // Kill monitor from debug info
2766 return;
2767 }
2769 // Memory barrier to avoid floating things down past the locked region
2770 insert_mem_bar(Op_MemBarRelease);
2772 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2773 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2774 uint raw_idx = Compile::AliasIdxRaw;
2775 unlock->init_req( TypeFunc::Control, control() );
2776 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2777 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2778 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2779 unlock->init_req( TypeFunc::ReturnAdr, top() );
2781 unlock->init_req(TypeFunc::Parms + 0, obj);
2782 unlock->init_req(TypeFunc::Parms + 1, box);
2783 unlock = _gvn.transform(unlock)->as_Unlock();
2785 Node* mem = reset_memory();
2787 // unlock has no side-effects, sets few values
2788 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2790 // Kill monitor from debug info
2791 map()->pop_monitor( );
2792 }
2794 //-------------------------------get_layout_helper-----------------------------
2795 // If the given klass is a constant or known to be an array,
2796 // fetch the constant layout helper value into constant_value
2797 // and return (Node*)NULL. Otherwise, load the non-constant
2798 // layout helper value, and return the node which represents it.
2799 // This two-faced routine is useful because allocation sites
2800 // almost always feature constant types.
2801 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2802 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2803 if (!StressReflectiveCode && inst_klass != NULL) {
2804 ciKlass* klass = inst_klass->klass();
2805 bool xklass = inst_klass->klass_is_exact();
2806 if (xklass || klass->is_array_klass()) {
2807 jint lhelper = klass->layout_helper();
2808 if (lhelper != Klass::_lh_neutral_value) {
2809 constant_value = lhelper;
2810 return (Node*) NULL;
2811 }
2812 }
2813 }
2814 constant_value = Klass::_lh_neutral_value; // put in a known value
2815 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
2816 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2817 }
2819 // We just put in an allocate/initialize with a big raw-memory effect.
2820 // Hook selected additional alias categories on the initialization.
2821 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2822 MergeMemNode* init_in_merge,
2823 Node* init_out_raw) {
2824 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2825 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2827 Node* prevmem = kit.memory(alias_idx);
2828 init_in_merge->set_memory_at(alias_idx, prevmem);
2829 kit.set_memory(init_out_raw, alias_idx);
2830 }
2832 //---------------------------set_output_for_allocation-------------------------
2833 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2834 const TypeOopPtr* oop_type,
2835 bool raw_mem_only) {
2836 int rawidx = Compile::AliasIdxRaw;
2837 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2838 add_safepoint_edges(alloc);
2839 Node* allocx = _gvn.transform(alloc);
2840 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2841 // create memory projection for i_o
2842 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2843 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2845 // create a memory projection as for the normal control path
2846 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2847 set_memory(malloc, rawidx);
2849 // a normal slow-call doesn't change i_o, but an allocation does
2850 // we create a separate i_o projection for the normal control path
2851 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2852 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2854 // put in an initialization barrier
2855 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2856 rawoop)->as_Initialize();
2857 assert(alloc->initialization() == init, "2-way macro link must work");
2858 assert(init ->allocation() == alloc, "2-way macro link must work");
2859 if (ReduceFieldZeroing && !raw_mem_only) {
2860 // Extract memory strands which may participate in the new object's
2861 // initialization, and source them from the new InitializeNode.
2862 // This will allow us to observe initializations when they occur,
2863 // and link them properly (as a group) to the InitializeNode.
2864 assert(init->in(InitializeNode::Memory) == malloc, "");
2865 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
2866 init->set_req(InitializeNode::Memory, minit_in);
2867 record_for_igvn(minit_in); // fold it up later, if possible
2868 Node* minit_out = memory(rawidx);
2869 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
2870 if (oop_type->isa_aryptr()) {
2871 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
2872 int elemidx = C->get_alias_index(telemref);
2873 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
2874 } else if (oop_type->isa_instptr()) {
2875 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
2876 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
2877 ciField* field = ik->nonstatic_field_at(i);
2878 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
2879 continue; // do not bother to track really large numbers of fields
2880 // Find (or create) the alias category for this field:
2881 int fieldidx = C->alias_type(field)->index();
2882 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
2883 }
2884 }
2885 }
2887 // Cast raw oop to the real thing...
2888 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
2889 javaoop = _gvn.transform(javaoop);
2890 C->set_recent_alloc(control(), javaoop);
2891 assert(just_allocated_object(control()) == javaoop, "just allocated");
2893 #ifdef ASSERT
2894 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
2895 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
2896 "Ideal_allocation works");
2897 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
2898 "Ideal_allocation works");
2899 if (alloc->is_AllocateArray()) {
2900 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
2901 "Ideal_allocation works");
2902 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
2903 "Ideal_allocation works");
2904 } else {
2905 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
2906 }
2907 }
2908 #endif //ASSERT
2910 return javaoop;
2911 }
2913 //---------------------------new_instance--------------------------------------
2914 // This routine takes a klass_node which may be constant (for a static type)
2915 // or may be non-constant (for reflective code). It will work equally well
2916 // for either, and the graph will fold nicely if the optimizer later reduces
2917 // the type to a constant.
2918 // The optional arguments are for specialized use by intrinsics:
2919 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
2920 // - If 'raw_mem_only', do not cast the result to an oop.
2921 // - If 'return_size_val', report the the total object size to the caller.
2922 Node* GraphKit::new_instance(Node* klass_node,
2923 Node* extra_slow_test,
2924 bool raw_mem_only, // affect only raw memory
2925 Node* *return_size_val) {
2926 // Compute size in doublewords
2927 // The size is always an integral number of doublewords, represented
2928 // as a positive bytewise size stored in the klass's layout_helper.
2929 // The layout_helper also encodes (in a low bit) the need for a slow path.
2930 jint layout_con = Klass::_lh_neutral_value;
2931 Node* layout_val = get_layout_helper(klass_node, layout_con);
2932 int layout_is_con = (layout_val == NULL);
2934 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
2935 // Generate the initial go-slow test. It's either ALWAYS (return a
2936 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
2937 // case) a computed value derived from the layout_helper.
2938 Node* initial_slow_test = NULL;
2939 if (layout_is_con) {
2940 assert(!StressReflectiveCode, "stress mode does not use these paths");
2941 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
2942 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
2944 } else { // reflective case
2945 // This reflective path is used by Unsafe.allocateInstance.
2946 // (It may be stress-tested by specifying StressReflectiveCode.)
2947 // Basically, we want to get into the VM is there's an illegal argument.
2948 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
2949 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
2950 if (extra_slow_test != intcon(0)) {
2951 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
2952 }
2953 // (Macro-expander will further convert this to a Bool, if necessary.)
2954 }
2956 // Find the size in bytes. This is easy; it's the layout_helper.
2957 // The size value must be valid even if the slow path is taken.
2958 Node* size = NULL;
2959 if (layout_is_con) {
2960 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
2961 } else { // reflective case
2962 // This reflective path is used by clone and Unsafe.allocateInstance.
2963 size = ConvI2X(layout_val);
2965 // Clear the low bits to extract layout_helper_size_in_bytes:
2966 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
2967 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
2968 size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
2969 }
2970 if (return_size_val != NULL) {
2971 (*return_size_val) = size;
2972 }
2974 // This is a precise notnull oop of the klass.
2975 // (Actually, it need not be precise if this is a reflective allocation.)
2976 // It's what we cast the result to.
2977 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
2978 if (!tklass) tklass = TypeKlassPtr::OBJECT;
2979 const TypeOopPtr* oop_type = tklass->as_instance_type();
2981 // Now generate allocation code
2983 // The entire memory state is needed for slow path of the allocation
2984 // since GC and deoptimization can happened.
2985 Node *mem = reset_memory();
2986 set_all_memory(mem); // Create new memory state
2988 AllocateNode* alloc
2989 = new (C, AllocateNode::ParmLimit)
2990 AllocateNode(C, AllocateNode::alloc_type(),
2991 control(), mem, i_o(),
2992 size, klass_node,
2993 initial_slow_test);
2995 return set_output_for_allocation(alloc, oop_type, raw_mem_only);
2996 }
2998 //-------------------------------new_array-------------------------------------
2999 // helper for both newarray and anewarray
3000 // The 'length' parameter is (obviously) the length of the array.
3001 // See comments on new_instance for the meaning of the other arguments.
3002 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3003 Node* length, // number of array elements
3004 int nargs, // number of arguments to push back for uncommon trap
3005 bool raw_mem_only, // affect only raw memory
3006 Node* *return_size_val) {
3007 jint layout_con = Klass::_lh_neutral_value;
3008 Node* layout_val = get_layout_helper(klass_node, layout_con);
3009 int layout_is_con = (layout_val == NULL);
3011 if (!layout_is_con && !StressReflectiveCode &&
3012 !too_many_traps(Deoptimization::Reason_class_check)) {
3013 // This is a reflective array creation site.
3014 // Optimistically assume that it is a subtype of Object[],
3015 // so that we can fold up all the address arithmetic.
3016 layout_con = Klass::array_layout_helper(T_OBJECT);
3017 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
3018 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
3019 { BuildCutout unless(this, bol_lh, PROB_MAX);
3020 _sp += nargs;
3021 uncommon_trap(Deoptimization::Reason_class_check,
3022 Deoptimization::Action_maybe_recompile);
3023 }
3024 layout_val = NULL;
3025 layout_is_con = true;
3026 }
3028 // Generate the initial go-slow test. Make sure we do not overflow
3029 // if length is huge (near 2Gig) or negative! We do not need
3030 // exact double-words here, just a close approximation of needed
3031 // double-words. We can't add any offset or rounding bits, lest we
3032 // take a size -1 of bytes and make it positive. Use an unsigned
3033 // compare, so negative sizes look hugely positive.
3034 int fast_size_limit = FastAllocateSizeLimit;
3035 if (layout_is_con) {
3036 assert(!StressReflectiveCode, "stress mode does not use these paths");
3037 // Increase the size limit if we have exact knowledge of array type.
3038 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3039 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3040 }
3042 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3043 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3044 if (initial_slow_test->is_Bool()) {
3045 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3046 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3047 }
3049 // --- Size Computation ---
3050 // array_size = round_to_heap(array_header + (length << elem_shift));
3051 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3052 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3053 // The rounding mask is strength-reduced, if possible.
3054 int round_mask = MinObjAlignmentInBytes - 1;
3055 Node* header_size = NULL;
3056 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3057 // (T_BYTE has the weakest alignment and size restrictions...)
3058 if (layout_is_con) {
3059 int hsize = Klass::layout_helper_header_size(layout_con);
3060 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3061 BasicType etype = Klass::layout_helper_element_type(layout_con);
3062 if ((round_mask & ~right_n_bits(eshift)) == 0)
3063 round_mask = 0; // strength-reduce it if it goes away completely
3064 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3065 assert(header_size_min <= hsize, "generic minimum is smallest");
3066 header_size_min = hsize;
3067 header_size = intcon(hsize + round_mask);
3068 } else {
3069 Node* hss = intcon(Klass::_lh_header_size_shift);
3070 Node* hsm = intcon(Klass::_lh_header_size_mask);
3071 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3072 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3073 Node* mask = intcon(round_mask);
3074 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3075 }
3077 Node* elem_shift = NULL;
3078 if (layout_is_con) {
3079 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3080 if (eshift != 0)
3081 elem_shift = intcon(eshift);
3082 } else {
3083 // There is no need to mask or shift this value.
3084 // The semantics of LShiftINode include an implicit mask to 0x1F.
3085 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3086 elem_shift = layout_val;
3087 }
3089 // Transition to native address size for all offset calculations:
3090 Node* lengthx = ConvI2X(length);
3091 Node* headerx = ConvI2X(header_size);
3092 #ifdef _LP64
3093 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3094 if (tllen != NULL && tllen->_lo < 0) {
3095 // Add a manual constraint to a positive range. Cf. array_element_address.
3096 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3097 if (size_max > tllen->_hi) size_max = tllen->_hi;
3098 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3099 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3100 }
3101 }
3102 #endif
3104 // Combine header size (plus rounding) and body size. Then round down.
3105 // This computation cannot overflow, because it is used only in two
3106 // places, one where the length is sharply limited, and the other
3107 // after a successful allocation.
3108 Node* abody = lengthx;
3109 if (elem_shift != NULL)
3110 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3111 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3112 if (round_mask != 0) {
3113 Node* mask = MakeConX(~round_mask);
3114 size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3115 }
3116 // else if round_mask == 0, the size computation is self-rounding
3118 if (return_size_val != NULL) {
3119 // This is the size
3120 (*return_size_val) = size;
3121 }
3123 // Now generate allocation code
3125 // The entire memory state is needed for slow path of the allocation
3126 // since GC and deoptimization can happened.
3127 Node *mem = reset_memory();
3128 set_all_memory(mem); // Create new memory state
3130 // Create the AllocateArrayNode and its result projections
3131 AllocateArrayNode* alloc
3132 = new (C, AllocateArrayNode::ParmLimit)
3133 AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3134 control(), mem, i_o(),
3135 size, klass_node,
3136 initial_slow_test,
3137 length);
3139 // Cast to correct type. Note that the klass_node may be constant or not,
3140 // and in the latter case the actual array type will be inexact also.
3141 // (This happens via a non-constant argument to inline_native_newArray.)
3142 // In any case, the value of klass_node provides the desired array type.
3143 const TypeInt* length_type = _gvn.find_int_type(length);
3144 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3145 if (ary_type->isa_aryptr() && length_type != NULL) {
3146 // Try to get a better type than POS for the size
3147 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3148 }
3150 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
3152 // Cast length on remaining path to be as narrow as possible
3153 if (map()->find_edge(length) >= 0) {
3154 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3155 if (ccast != length) {
3156 _gvn.set_type_bottom(ccast);
3157 record_for_igvn(ccast);
3158 replace_in_map(length, ccast);
3159 }
3160 }
3162 return javaoop;
3163 }
3165 // The following "Ideal_foo" functions are placed here because they recognize
3166 // the graph shapes created by the functions immediately above.
3168 //---------------------------Ideal_allocation----------------------------------
3169 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3170 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3171 if (ptr == NULL) { // reduce dumb test in callers
3172 return NULL;
3173 }
3174 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3175 ptr = ptr->in(1);
3176 if (ptr == NULL) return NULL;
3177 }
3178 if (ptr->is_Proj()) {
3179 Node* allo = ptr->in(0);
3180 if (allo != NULL && allo->is_Allocate()) {
3181 return allo->as_Allocate();
3182 }
3183 }
3184 // Report failure to match.
3185 return NULL;
3186 }
3188 // Fancy version which also strips off an offset (and reports it to caller).
3189 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3190 intptr_t& offset) {
3191 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3192 if (base == NULL) return NULL;
3193 return Ideal_allocation(base, phase);
3194 }
3196 // Trace Initialize <- Proj[Parm] <- Allocate
3197 AllocateNode* InitializeNode::allocation() {
3198 Node* rawoop = in(InitializeNode::RawAddress);
3199 if (rawoop->is_Proj()) {
3200 Node* alloc = rawoop->in(0);
3201 if (alloc->is_Allocate()) {
3202 return alloc->as_Allocate();
3203 }
3204 }
3205 return NULL;
3206 }
3208 // Trace Allocate -> Proj[Parm] -> Initialize
3209 InitializeNode* AllocateNode::initialization() {
3210 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3211 if (rawoop == NULL) return NULL;
3212 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3213 Node* init = rawoop->fast_out(i);
3214 if (init->is_Initialize()) {
3215 assert(init->as_Initialize()->allocation() == this, "2-way link");
3216 return init->as_Initialize();
3217 }
3218 }
3219 return NULL;
3220 }
3222 //----------------------------- store barriers ----------------------------
3223 #define __ ideal.
3225 void GraphKit::sync_kit(IdealKit& ideal) {
3226 // Final sync IdealKit and graphKit.
3227 __ drain_delay_transform();
3228 set_all_memory(__ merged_memory());
3229 set_control(__ ctrl());
3230 }
3232 // vanilla/CMS post barrier
3233 // Insert a write-barrier store. This is to let generational GC work; we have
3234 // to flag all oop-stores before the next GC point.
3235 void GraphKit::write_barrier_post(Node* oop_store,
3236 Node* obj,
3237 Node* adr,
3238 uint adr_idx,
3239 Node* val,
3240 bool use_precise) {
3241 // No store check needed if we're storing a NULL or an old object
3242 // (latter case is probably a string constant). The concurrent
3243 // mark sweep garbage collector, however, needs to have all nonNull
3244 // oop updates flagged via card-marks.
3245 if (val != NULL && val->is_Con()) {
3246 // must be either an oop or NULL
3247 const Type* t = val->bottom_type();
3248 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3249 // stores of null never (?) need barriers
3250 return;
3251 ciObject* con = t->is_oopptr()->const_oop();
3252 if (con != NULL
3253 && con->is_perm()
3254 && Universe::heap()->can_elide_permanent_oop_store_barriers())
3255 // no store barrier needed, because no old-to-new ref created
3256 return;
3257 }
3259 if (use_ReduceInitialCardMarks()
3260 && obj == just_allocated_object(control())) {
3261 // We can skip marks on a freshly-allocated object in Eden.
3262 // Keep this code in sync with maybe_defer_card_mark() in runtime.cpp.
3263 // That routine informs GC to take appropriate compensating steps
3264 // so as to make this card-mark elision safe.
3265 return;
3266 }
3268 if (!use_precise) {
3269 // All card marks for a (non-array) instance are in one place:
3270 adr = obj;
3271 }
3272 // (Else it's an array (or unknown), and we want more precise card marks.)
3273 assert(adr != NULL, "");
3275 IdealKit ideal(gvn(), control(), merged_memory(), true);
3277 // Convert the pointer to an int prior to doing math on it
3278 Node* cast = __ CastPX(__ ctrl(), adr);
3280 // Divide by card size
3281 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3282 "Only one we handle so far.");
3283 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3285 // Combine card table base and card offset
3286 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3288 // Get the alias_index for raw card-mark memory
3289 int adr_type = Compile::AliasIdxRaw;
3290 // Smash zero into card
3291 Node* zero = __ ConI(0);
3292 BasicType bt = T_BYTE;
3293 if( !UseConcMarkSweepGC ) {
3294 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3295 } else {
3296 // Specialized path for CM store barrier
3297 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3298 }
3300 // Final sync IdealKit and GraphKit.
3301 sync_kit(ideal);
3302 }
3304 // G1 pre/post barriers
3305 void GraphKit::g1_write_barrier_pre(Node* obj,
3306 Node* adr,
3307 uint alias_idx,
3308 Node* val,
3309 const TypeOopPtr* val_type,
3310 BasicType bt) {
3311 IdealKit ideal(gvn(), control(), merged_memory(), true);
3313 Node* tls = __ thread(); // ThreadLocalStorage
3315 Node* no_ctrl = NULL;
3316 Node* no_base = __ top();
3317 Node* zero = __ ConI(0);
3319 float likely = PROB_LIKELY(0.999);
3320 float unlikely = PROB_UNLIKELY(0.999);
3322 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3323 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3325 // Offsets into the thread
3326 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3327 PtrQueue::byte_offset_of_active());
3328 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3329 PtrQueue::byte_offset_of_index());
3330 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3331 PtrQueue::byte_offset_of_buf());
3332 // Now the actual pointers into the thread
3334 // set_control( ctl);
3336 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3337 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3338 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3340 // Now some of the values
3342 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3344 // if (!marking)
3345 __ if_then(marking, BoolTest::ne, zero); {
3346 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3348 const Type* t1 = adr->bottom_type();
3349 const Type* t2 = val->bottom_type();
3351 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3352 // if (orig != NULL)
3353 __ if_then(orig, BoolTest::ne, null()); {
3354 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3356 // load original value
3357 // alias_idx correct??
3359 // is the queue for this thread full?
3360 __ if_then(index, BoolTest::ne, zero, likely); {
3362 // decrement the index
3363 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3364 Node* next_indexX = next_index;
3365 #ifdef _LP64
3366 // We could refine the type for what it's worth
3367 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3368 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3369 #endif
3371 // Now get the buffer location we will log the original value into and store it
3372 Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3373 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw);
3375 // update the index
3376 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3378 } __ else_(); {
3380 // logging buffer is full, call the runtime
3381 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3382 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls);
3383 } __ end_if(); // (!index)
3384 } __ end_if(); // (orig != NULL)
3385 } __ end_if(); // (!marking)
3387 // Final sync IdealKit and GraphKit.
3388 sync_kit(ideal);
3389 }
3391 //
3392 // Update the card table and add card address to the queue
3393 //
3394 void GraphKit::g1_mark_card(IdealKit& ideal,
3395 Node* card_adr,
3396 Node* oop_store,
3397 uint oop_alias_idx,
3398 Node* index,
3399 Node* index_adr,
3400 Node* buffer,
3401 const TypeFunc* tf) {
3403 Node* zero = __ ConI(0);
3404 Node* no_base = __ top();
3405 BasicType card_bt = T_BYTE;
3406 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3407 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3409 // Now do the queue work
3410 __ if_then(index, BoolTest::ne, zero); {
3412 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3413 Node* next_indexX = next_index;
3414 #ifdef _LP64
3415 // We could refine the type for what it's worth
3416 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3417 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3418 #endif // _LP64
3419 Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3421 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3422 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3424 } __ else_(); {
3425 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3426 } __ end_if();
3428 }
3430 void GraphKit::g1_write_barrier_post(Node* oop_store,
3431 Node* obj,
3432 Node* adr,
3433 uint alias_idx,
3434 Node* val,
3435 BasicType bt,
3436 bool use_precise) {
3437 // If we are writing a NULL then we need no post barrier
3439 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3440 // Must be NULL
3441 const Type* t = val->bottom_type();
3442 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3443 // No post barrier if writing NULLx
3444 return;
3445 }
3447 if (!use_precise) {
3448 // All card marks for a (non-array) instance are in one place:
3449 adr = obj;
3450 }
3451 // (Else it's an array (or unknown), and we want more precise card marks.)
3452 assert(adr != NULL, "");
3454 IdealKit ideal(gvn(), control(), merged_memory(), true);
3456 Node* tls = __ thread(); // ThreadLocalStorage
3458 Node* no_ctrl = NULL;
3459 Node* no_base = __ top();
3460 float likely = PROB_LIKELY(0.999);
3461 float unlikely = PROB_UNLIKELY(0.999);
3462 Node* zero = __ ConI(0);
3463 Node* zeroX = __ ConX(0);
3465 // Get the alias_index for raw card-mark memory
3466 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3468 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3470 // Offsets into the thread
3471 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3472 PtrQueue::byte_offset_of_index());
3473 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3474 PtrQueue::byte_offset_of_buf());
3476 // Pointers into the thread
3478 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3479 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3481 // Now some values
3483 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3484 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3487 // Convert the store obj pointer to an int prior to doing math on it
3488 // Must use ctrl to prevent "integerized oop" existing across safepoint
3489 Node* cast = __ CastPX(__ ctrl(), adr);
3491 // Divide pointer by card size
3492 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3494 // Combine card table base and card offset
3495 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3497 // If we know the value being stored does it cross regions?
3499 if (val != NULL) {
3500 // Does the store cause us to cross regions?
3502 // Should be able to do an unsigned compare of region_size instead of
3503 // and extra shift. Do we have an unsigned compare??
3504 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3505 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3507 // if (xor_res == 0) same region so skip
3508 __ if_then(xor_res, BoolTest::ne, zeroX); {
3510 // No barrier if we are storing a NULL
3511 __ if_then(val, BoolTest::ne, null(), unlikely); {
3513 // Ok must mark the card if not already dirty
3515 // load the original value of the card
3516 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3518 __ if_then(card_val, BoolTest::ne, zero); {
3519 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3520 } __ end_if();
3521 } __ end_if();
3522 } __ end_if();
3523 } else {
3524 // Object.clone() instrinsic uses this path.
3525 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3526 }
3528 // Final sync IdealKit and GraphKit.
3529 sync_kit(ideal);
3530 }
3531 #undef __