Mon, 04 Jan 2010 18:38:08 +0100
6893081: method handle & invokedynamic code needs additional cleanup (post 6815692, 6858164)
Summary: During the work for 6829187 we have fixed a number of basic bugs which are logically grouped with 6815692 and 6858164 but which must be reviewed and pushed separately.
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 is_static = (depth == 0);
985 bool ignore;
986 ciBytecodeStream iter(method());
987 iter.reset_to_bci(bci());
988 iter.next();
989 ciMethod* method = iter.get_method(ignore);
990 inputs = method->arg_size_no_receiver();
991 if (!is_static) inputs += 1;
992 int size = method->return_type()->size();
993 depth = size - inputs;
994 }
995 break;
997 case Bytecodes::_multianewarray:
998 {
999 ciBytecodeStream iter(method());
1000 iter.reset_to_bci(bci());
1001 iter.next();
1002 inputs = iter.get_dimensions();
1003 assert(rsize == 1, "");
1004 depth = rsize - inputs;
1005 }
1006 break;
1008 case Bytecodes::_ireturn:
1009 case Bytecodes::_lreturn:
1010 case Bytecodes::_freturn:
1011 case Bytecodes::_dreturn:
1012 case Bytecodes::_areturn:
1013 assert(rsize = -depth, "");
1014 inputs = rsize;
1015 break;
1017 case Bytecodes::_jsr:
1018 case Bytecodes::_jsr_w:
1019 inputs = 0;
1020 depth = 1; // S.B. depth=1, not zero
1021 break;
1023 default:
1024 // bytecode produces a typed result
1025 inputs = rsize - depth;
1026 assert(inputs >= 0, "");
1027 break;
1028 }
1030 #ifdef ASSERT
1031 // spot check
1032 int outputs = depth + inputs;
1033 assert(outputs >= 0, "sanity");
1034 switch (code) {
1035 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1036 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1037 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1038 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1039 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1040 }
1041 #endif //ASSERT
1043 return true;
1044 }
1048 //------------------------------basic_plus_adr---------------------------------
1049 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1050 // short-circuit a common case
1051 if (offset == intcon(0)) return ptr;
1052 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1053 }
1055 Node* GraphKit::ConvI2L(Node* offset) {
1056 // short-circuit a common case
1057 jint offset_con = find_int_con(offset, Type::OffsetBot);
1058 if (offset_con != Type::OffsetBot) {
1059 return longcon((long) offset_con);
1060 }
1061 return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1062 }
1063 Node* GraphKit::ConvL2I(Node* offset) {
1064 // short-circuit a common case
1065 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1066 if (offset_con != (jlong)Type::OffsetBot) {
1067 return intcon((int) offset_con);
1068 }
1069 return _gvn.transform( new (C, 2) ConvL2INode(offset));
1070 }
1072 //-------------------------load_object_klass-----------------------------------
1073 Node* GraphKit::load_object_klass(Node* obj) {
1074 // Special-case a fresh allocation to avoid building nodes:
1075 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1076 if (akls != NULL) return akls;
1077 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1078 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1079 }
1081 //-------------------------load_array_length-----------------------------------
1082 Node* GraphKit::load_array_length(Node* array) {
1083 // Special-case a fresh allocation to avoid building nodes:
1084 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1085 Node *alen;
1086 if (alloc == NULL) {
1087 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1088 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1089 } else {
1090 alen = alloc->Ideal_length();
1091 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1092 if (ccast != alen) {
1093 alen = _gvn.transform(ccast);
1094 }
1095 }
1096 return alen;
1097 }
1099 //------------------------------do_null_check----------------------------------
1100 // Helper function to do a NULL pointer check. Returned value is
1101 // the incoming address with NULL casted away. You are allowed to use the
1102 // not-null value only if you are control dependent on the test.
1103 extern int explicit_null_checks_inserted,
1104 explicit_null_checks_elided;
1105 Node* GraphKit::null_check_common(Node* value, BasicType type,
1106 // optional arguments for variations:
1107 bool assert_null,
1108 Node* *null_control) {
1109 assert(!assert_null || null_control == NULL, "not both at once");
1110 if (stopped()) return top();
1111 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1112 // For some performance testing, we may wish to suppress null checking.
1113 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1114 return value;
1115 }
1116 explicit_null_checks_inserted++;
1118 // Construct NULL check
1119 Node *chk = NULL;
1120 switch(type) {
1121 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1122 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1123 case T_ARRAY : // fall through
1124 type = T_OBJECT; // simplify further tests
1125 case T_OBJECT : {
1126 const Type *t = _gvn.type( value );
1128 const TypeOopPtr* tp = t->isa_oopptr();
1129 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1130 // Only for do_null_check, not any of its siblings:
1131 && !assert_null && null_control == NULL) {
1132 // Usually, any field access or invocation on an unloaded oop type
1133 // will simply fail to link, since the statically linked class is
1134 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1135 // the static class is loaded but the sharper oop type is not.
1136 // Rather than checking for this obscure case in lots of places,
1137 // we simply observe that a null check on an unloaded class
1138 // will always be followed by a nonsense operation, so we
1139 // can just issue the uncommon trap here.
1140 // Our access to the unloaded class will only be correct
1141 // after it has been loaded and initialized, which requires
1142 // a trip through the interpreter.
1143 #ifndef PRODUCT
1144 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1145 #endif
1146 uncommon_trap(Deoptimization::Reason_unloaded,
1147 Deoptimization::Action_reinterpret,
1148 tp->klass(), "!loaded");
1149 return top();
1150 }
1152 if (assert_null) {
1153 // See if the type is contained in NULL_PTR.
1154 // If so, then the value is already null.
1155 if (t->higher_equal(TypePtr::NULL_PTR)) {
1156 explicit_null_checks_elided++;
1157 return value; // Elided null assert quickly!
1158 }
1159 } else {
1160 // See if mixing in the NULL pointer changes type.
1161 // If so, then the NULL pointer was not allowed in the original
1162 // type. In other words, "value" was not-null.
1163 if (t->meet(TypePtr::NULL_PTR) != t) {
1164 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1165 explicit_null_checks_elided++;
1166 return value; // Elided null check quickly!
1167 }
1168 }
1169 chk = new (C, 3) CmpPNode( value, null() );
1170 break;
1171 }
1173 default : ShouldNotReachHere();
1174 }
1175 assert(chk != NULL, "sanity check");
1176 chk = _gvn.transform(chk);
1178 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1179 BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1180 Node *tst = _gvn.transform( btst );
1182 //-----------
1183 // if peephole optimizations occurred, a prior test existed.
1184 // If a prior test existed, maybe it dominates as we can avoid this test.
1185 if (tst != btst && type == T_OBJECT) {
1186 // At this point we want to scan up the CFG to see if we can
1187 // find an identical test (and so avoid this test altogether).
1188 Node *cfg = control();
1189 int depth = 0;
1190 while( depth < 16 ) { // Limit search depth for speed
1191 if( cfg->Opcode() == Op_IfTrue &&
1192 cfg->in(0)->in(1) == tst ) {
1193 // Found prior test. Use "cast_not_null" to construct an identical
1194 // CastPP (and hence hash to) as already exists for the prior test.
1195 // Return that casted value.
1196 if (assert_null) {
1197 replace_in_map(value, null());
1198 return null(); // do not issue the redundant test
1199 }
1200 Node *oldcontrol = control();
1201 set_control(cfg);
1202 Node *res = cast_not_null(value);
1203 set_control(oldcontrol);
1204 explicit_null_checks_elided++;
1205 return res;
1206 }
1207 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1208 if (cfg == NULL) break; // Quit at region nodes
1209 depth++;
1210 }
1211 }
1213 //-----------
1214 // Branch to failure if null
1215 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1216 Deoptimization::DeoptReason reason;
1217 if (assert_null)
1218 reason = Deoptimization::Reason_null_assert;
1219 else if (type == T_OBJECT)
1220 reason = Deoptimization::Reason_null_check;
1221 else
1222 reason = Deoptimization::Reason_div0_check;
1224 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1225 // ciMethodData::has_trap_at will return a conservative -1 if any
1226 // must-be-null assertion has failed. This could cause performance
1227 // problems for a method after its first do_null_assert failure.
1228 // Consider using 'Reason_class_check' instead?
1230 // To cause an implicit null check, we set the not-null probability
1231 // to the maximum (PROB_MAX). For an explicit check the probability
1232 // is set to a smaller value.
1233 if (null_control != NULL || too_many_traps(reason)) {
1234 // probability is less likely
1235 ok_prob = PROB_LIKELY_MAG(3);
1236 } else if (!assert_null &&
1237 (ImplicitNullCheckThreshold > 0) &&
1238 method() != NULL &&
1239 (method()->method_data()->trap_count(reason)
1240 >= (uint)ImplicitNullCheckThreshold)) {
1241 ok_prob = PROB_LIKELY_MAG(3);
1242 }
1244 if (null_control != NULL) {
1245 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1246 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1247 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
1248 if (null_true == top())
1249 explicit_null_checks_elided++;
1250 (*null_control) = null_true;
1251 } else {
1252 BuildCutout unless(this, tst, ok_prob);
1253 // Check for optimizer eliding test at parse time
1254 if (stopped()) {
1255 // Failure not possible; do not bother making uncommon trap.
1256 explicit_null_checks_elided++;
1257 } else if (assert_null) {
1258 uncommon_trap(reason,
1259 Deoptimization::Action_make_not_entrant,
1260 NULL, "assert_null");
1261 } else {
1262 replace_in_map(value, zerocon(type));
1263 builtin_throw(reason);
1264 }
1265 }
1267 // Must throw exception, fall-thru not possible?
1268 if (stopped()) {
1269 return top(); // No result
1270 }
1272 if (assert_null) {
1273 // Cast obj to null on this path.
1274 replace_in_map(value, zerocon(type));
1275 return zerocon(type);
1276 }
1278 // Cast obj to not-null on this path, if there is no null_control.
1279 // (If there is a null_control, a non-null value may come back to haunt us.)
1280 if (type == T_OBJECT) {
1281 Node* cast = cast_not_null(value, false);
1282 if (null_control == NULL || (*null_control) == top())
1283 replace_in_map(value, cast);
1284 value = cast;
1285 }
1287 return value;
1288 }
1291 //------------------------------cast_not_null----------------------------------
1292 // Cast obj to not-null on this path
1293 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1294 const Type *t = _gvn.type(obj);
1295 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1296 // Object is already not-null?
1297 if( t == t_not_null ) return obj;
1299 Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1300 cast->init_req(0, control());
1301 cast = _gvn.transform( cast );
1303 // Scan for instances of 'obj' in the current JVM mapping.
1304 // These instances are known to be not-null after the test.
1305 if (do_replace_in_map)
1306 replace_in_map(obj, cast);
1308 return cast; // Return casted value
1309 }
1312 //--------------------------replace_in_map-------------------------------------
1313 void GraphKit::replace_in_map(Node* old, Node* neww) {
1314 this->map()->replace_edge(old, neww);
1316 // Note: This operation potentially replaces any edge
1317 // on the map. This includes locals, stack, and monitors
1318 // of the current (innermost) JVM state.
1320 // We can consider replacing in caller maps.
1321 // The idea would be that an inlined function's null checks
1322 // can be shared with the entire inlining tree.
1323 // The expense of doing this is that the PreserveJVMState class
1324 // would have to preserve caller states too, with a deep copy.
1325 }
1329 //=============================================================================
1330 //--------------------------------memory---------------------------------------
1331 Node* GraphKit::memory(uint alias_idx) {
1332 MergeMemNode* mem = merged_memory();
1333 Node* p = mem->memory_at(alias_idx);
1334 _gvn.set_type(p, Type::MEMORY); // must be mapped
1335 return p;
1336 }
1338 //-----------------------------reset_memory------------------------------------
1339 Node* GraphKit::reset_memory() {
1340 Node* mem = map()->memory();
1341 // do not use this node for any more parsing!
1342 debug_only( map()->set_memory((Node*)NULL) );
1343 return _gvn.transform( mem );
1344 }
1346 //------------------------------set_all_memory---------------------------------
1347 void GraphKit::set_all_memory(Node* newmem) {
1348 Node* mergemem = MergeMemNode::make(C, newmem);
1349 gvn().set_type_bottom(mergemem);
1350 map()->set_memory(mergemem);
1351 }
1353 //------------------------------set_all_memory_call----------------------------
1354 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1355 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1356 set_all_memory(newmem);
1357 }
1359 //=============================================================================
1360 //
1361 // parser factory methods for MemNodes
1362 //
1363 // These are layered on top of the factory methods in LoadNode and StoreNode,
1364 // and integrate with the parser's memory state and _gvn engine.
1365 //
1367 // factory methods in "int adr_idx"
1368 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1369 int adr_idx,
1370 bool require_atomic_access) {
1371 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1372 const TypePtr* adr_type = NULL; // debug-mode-only argument
1373 debug_only(adr_type = C->get_adr_type(adr_idx));
1374 Node* mem = memory(adr_idx);
1375 Node* ld;
1376 if (require_atomic_access && bt == T_LONG) {
1377 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1378 } else {
1379 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1380 }
1381 return _gvn.transform(ld);
1382 }
1384 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1385 int adr_idx,
1386 bool require_atomic_access) {
1387 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1388 const TypePtr* adr_type = NULL;
1389 debug_only(adr_type = C->get_adr_type(adr_idx));
1390 Node *mem = memory(adr_idx);
1391 Node* st;
1392 if (require_atomic_access && bt == T_LONG) {
1393 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1394 } else {
1395 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1396 }
1397 st = _gvn.transform(st);
1398 set_memory(st, adr_idx);
1399 // Back-to-back stores can only remove intermediate store with DU info
1400 // so push on worklist for optimizer.
1401 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1402 record_for_igvn(st);
1404 return st;
1405 }
1408 void GraphKit::pre_barrier(Node* ctl,
1409 Node* obj,
1410 Node* adr,
1411 uint adr_idx,
1412 Node* val,
1413 const TypeOopPtr* val_type,
1414 BasicType bt) {
1415 BarrierSet* bs = Universe::heap()->barrier_set();
1416 set_control(ctl);
1417 switch (bs->kind()) {
1418 case BarrierSet::G1SATBCT:
1419 case BarrierSet::G1SATBCTLogging:
1420 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt);
1421 break;
1423 case BarrierSet::CardTableModRef:
1424 case BarrierSet::CardTableExtension:
1425 case BarrierSet::ModRef:
1426 break;
1428 case BarrierSet::Other:
1429 default :
1430 ShouldNotReachHere();
1432 }
1433 }
1435 void GraphKit::post_barrier(Node* ctl,
1436 Node* store,
1437 Node* obj,
1438 Node* adr,
1439 uint adr_idx,
1440 Node* val,
1441 BasicType bt,
1442 bool use_precise) {
1443 BarrierSet* bs = Universe::heap()->barrier_set();
1444 set_control(ctl);
1445 switch (bs->kind()) {
1446 case BarrierSet::G1SATBCT:
1447 case BarrierSet::G1SATBCTLogging:
1448 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1449 break;
1451 case BarrierSet::CardTableModRef:
1452 case BarrierSet::CardTableExtension:
1453 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1454 break;
1456 case BarrierSet::ModRef:
1457 break;
1459 case BarrierSet::Other:
1460 default :
1461 ShouldNotReachHere();
1463 }
1464 }
1466 Node* GraphKit::store_oop(Node* ctl,
1467 Node* obj,
1468 Node* adr,
1469 const TypePtr* adr_type,
1470 Node* val,
1471 const TypeOopPtr* val_type,
1472 BasicType bt,
1473 bool use_precise) {
1475 set_control(ctl);
1476 if (stopped()) return top(); // Dead path ?
1478 assert(bt == T_OBJECT, "sanity");
1479 assert(val != NULL, "not dead path");
1480 uint adr_idx = C->get_alias_index(adr_type);
1481 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1483 pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt);
1484 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1485 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1486 return store;
1487 }
1489 // Could be an array or object we don't know at compile time (unsafe ref.)
1490 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1491 Node* obj, // containing obj
1492 Node* adr, // actual adress to store val at
1493 const TypePtr* adr_type,
1494 Node* val,
1495 BasicType bt) {
1496 Compile::AliasType* at = C->alias_type(adr_type);
1497 const TypeOopPtr* val_type = NULL;
1498 if (adr_type->isa_instptr()) {
1499 if (at->field() != NULL) {
1500 // known field. This code is a copy of the do_put_xxx logic.
1501 ciField* field = at->field();
1502 if (!field->type()->is_loaded()) {
1503 val_type = TypeInstPtr::BOTTOM;
1504 } else {
1505 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1506 }
1507 }
1508 } else if (adr_type->isa_aryptr()) {
1509 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1510 }
1511 if (val_type == NULL) {
1512 val_type = TypeInstPtr::BOTTOM;
1513 }
1514 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1515 }
1518 //-------------------------array_element_address-------------------------
1519 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1520 const TypeInt* sizetype) {
1521 uint shift = exact_log2(type2aelembytes(elembt));
1522 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1524 // short-circuit a common case (saves lots of confusing waste motion)
1525 jint idx_con = find_int_con(idx, -1);
1526 if (idx_con >= 0) {
1527 intptr_t offset = header + ((intptr_t)idx_con << shift);
1528 return basic_plus_adr(ary, offset);
1529 }
1531 // must be correct type for alignment purposes
1532 Node* base = basic_plus_adr(ary, header);
1533 #ifdef _LP64
1534 // The scaled index operand to AddP must be a clean 64-bit value.
1535 // Java allows a 32-bit int to be incremented to a negative
1536 // value, which appears in a 64-bit register as a large
1537 // positive number. Using that large positive number as an
1538 // operand in pointer arithmetic has bad consequences.
1539 // On the other hand, 32-bit overflow is rare, and the possibility
1540 // can often be excluded, if we annotate the ConvI2L node with
1541 // a type assertion that its value is known to be a small positive
1542 // number. (The prior range check has ensured this.)
1543 // This assertion is used by ConvI2LNode::Ideal.
1544 int index_max = max_jint - 1; // array size is max_jint, index is one less
1545 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1546 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1547 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1548 #endif
1549 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1550 return basic_plus_adr(ary, base, scale);
1551 }
1553 //-------------------------load_array_element-------------------------
1554 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1555 const Type* elemtype = arytype->elem();
1556 BasicType elembt = elemtype->array_element_basic_type();
1557 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1558 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1559 return ld;
1560 }
1562 //-------------------------set_arguments_for_java_call-------------------------
1563 // Arguments (pre-popped from the stack) are taken from the JVMS.
1564 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1565 // Add the call arguments:
1566 uint nargs = call->method()->arg_size();
1567 for (uint i = 0; i < nargs; i++) {
1568 Node* arg = argument(i);
1569 call->init_req(i + TypeFunc::Parms, arg);
1570 }
1571 }
1573 //---------------------------set_edges_for_java_call---------------------------
1574 // Connect a newly created call into the current JVMS.
1575 // A return value node (if any) is returned from set_edges_for_java_call.
1576 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1578 // Add the predefined inputs:
1579 call->init_req( TypeFunc::Control, control() );
1580 call->init_req( TypeFunc::I_O , i_o() );
1581 call->init_req( TypeFunc::Memory , reset_memory() );
1582 call->init_req( TypeFunc::FramePtr, frameptr() );
1583 call->init_req( TypeFunc::ReturnAdr, top() );
1585 add_safepoint_edges(call, must_throw);
1587 Node* xcall = _gvn.transform(call);
1589 if (xcall == top()) {
1590 set_control(top());
1591 return;
1592 }
1593 assert(xcall == call, "call identity is stable");
1595 // Re-use the current map to produce the result.
1597 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1598 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1599 set_all_memory_call(xcall, separate_io_proj);
1601 //return xcall; // no need, caller already has it
1602 }
1604 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1605 if (stopped()) return top(); // maybe the call folded up?
1607 // Capture the return value, if any.
1608 Node* ret;
1609 if (call->method() == NULL ||
1610 call->method()->return_type()->basic_type() == T_VOID)
1611 ret = top();
1612 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1614 // Note: Since any out-of-line call can produce an exception,
1615 // we always insert an I_O projection from the call into the result.
1617 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1619 if (separate_io_proj) {
1620 // The caller requested separate projections be used by the fall
1621 // through and exceptional paths, so replace the projections for
1622 // the fall through path.
1623 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) ));
1624 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ));
1625 }
1626 return ret;
1627 }
1629 //--------------------set_predefined_input_for_runtime_call--------------------
1630 // Reading and setting the memory state is way conservative here.
1631 // The real problem is that I am not doing real Type analysis on memory,
1632 // so I cannot distinguish card mark stores from other stores. Across a GC
1633 // point the Store Barrier and the card mark memory has to agree. I cannot
1634 // have a card mark store and its barrier split across the GC point from
1635 // either above or below. Here I get that to happen by reading ALL of memory.
1636 // A better answer would be to separate out card marks from other memory.
1637 // For now, return the input memory state, so that it can be reused
1638 // after the call, if this call has restricted memory effects.
1639 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1640 // Set fixed predefined input arguments
1641 Node* memory = reset_memory();
1642 call->init_req( TypeFunc::Control, control() );
1643 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1644 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1645 call->init_req( TypeFunc::FramePtr, frameptr() );
1646 call->init_req( TypeFunc::ReturnAdr, top() );
1647 return memory;
1648 }
1650 //-------------------set_predefined_output_for_runtime_call--------------------
1651 // Set control and memory (not i_o) from the call.
1652 // If keep_mem is not NULL, use it for the output state,
1653 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1654 // If hook_mem is NULL, this call produces no memory effects at all.
1655 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1656 // then only that memory slice is taken from the call.
1657 // In the last case, we must put an appropriate memory barrier before
1658 // the call, so as to create the correct anti-dependencies on loads
1659 // preceding the call.
1660 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1661 Node* keep_mem,
1662 const TypePtr* hook_mem) {
1663 // no i/o
1664 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1665 if (keep_mem) {
1666 // First clone the existing memory state
1667 set_all_memory(keep_mem);
1668 if (hook_mem != NULL) {
1669 // Make memory for the call
1670 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1671 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1672 // We also use hook_mem to extract specific effects from arraycopy stubs.
1673 set_memory(mem, hook_mem);
1674 }
1675 // ...else the call has NO memory effects.
1677 // Make sure the call advertises its memory effects precisely.
1678 // This lets us build accurate anti-dependences in gcm.cpp.
1679 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1680 "call node must be constructed correctly");
1681 } else {
1682 assert(hook_mem == NULL, "");
1683 // This is not a "slow path" call; all memory comes from the call.
1684 set_all_memory_call(call);
1685 }
1686 }
1689 // Replace the call with the current state of the kit.
1690 void GraphKit::replace_call(CallNode* call, Node* result) {
1691 JVMState* ejvms = NULL;
1692 if (has_exceptions()) {
1693 ejvms = transfer_exceptions_into_jvms();
1694 }
1696 SafePointNode* final_state = stop();
1698 // Find all the needed outputs of this call
1699 CallProjections callprojs;
1700 call->extract_projections(&callprojs, true);
1702 // Replace all the old call edges with the edges from the inlining result
1703 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control));
1704 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory));
1705 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O));
1707 // Replace the result with the new result if it exists and is used
1708 if (callprojs.resproj != NULL && result != NULL) {
1709 C->gvn_replace_by(callprojs.resproj, result);
1710 }
1712 if (ejvms == NULL) {
1713 // No exception edges to simply kill off those paths
1714 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1715 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1716 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1718 // Replace the old exception object with top
1719 if (callprojs.exobj != NULL) {
1720 C->gvn_replace_by(callprojs.exobj, C->top());
1721 }
1722 } else {
1723 GraphKit ekit(ejvms);
1725 // Load my combined exception state into the kit, with all phis transformed:
1726 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1728 Node* ex_oop = ekit.use_exception_state(ex_map);
1730 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1731 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1732 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1734 // Replace the old exception object with the newly created one
1735 if (callprojs.exobj != NULL) {
1736 C->gvn_replace_by(callprojs.exobj, ex_oop);
1737 }
1738 }
1740 // Disconnect the call from the graph
1741 call->disconnect_inputs(NULL);
1742 C->gvn_replace_by(call, C->top());
1743 }
1746 //------------------------------increment_counter------------------------------
1747 // for statistics: increment a VM counter by 1
1749 void GraphKit::increment_counter(address counter_addr) {
1750 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1751 increment_counter(adr1);
1752 }
1754 void GraphKit::increment_counter(Node* counter_addr) {
1755 int adr_type = Compile::AliasIdxRaw;
1756 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
1757 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1758 store_to_memory( NULL, counter_addr, incr, T_INT, adr_type );
1759 }
1762 //------------------------------uncommon_trap----------------------------------
1763 // Bail out to the interpreter in mid-method. Implemented by calling the
1764 // uncommon_trap blob. This helper function inserts a runtime call with the
1765 // right debug info.
1766 void GraphKit::uncommon_trap(int trap_request,
1767 ciKlass* klass, const char* comment,
1768 bool must_throw,
1769 bool keep_exact_action) {
1770 if (failing()) stop();
1771 if (stopped()) return; // trap reachable?
1773 // Note: If ProfileTraps is true, and if a deopt. actually
1774 // occurs here, the runtime will make sure an MDO exists. There is
1775 // no need to call method()->build_method_data() at this point.
1777 #ifdef ASSERT
1778 if (!must_throw) {
1779 // Make sure the stack has at least enough depth to execute
1780 // the current bytecode.
1781 int inputs, ignore;
1782 if (compute_stack_effects(inputs, ignore)) {
1783 assert(sp() >= inputs, "must have enough JVMS stack to execute");
1784 // It is a frequent error in library_call.cpp to issue an
1785 // uncommon trap with the _sp value already popped.
1786 }
1787 }
1788 #endif
1790 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1791 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1793 switch (action) {
1794 case Deoptimization::Action_maybe_recompile:
1795 case Deoptimization::Action_reinterpret:
1796 // Temporary fix for 6529811 to allow virtual calls to be sure they
1797 // get the chance to go from mono->bi->mega
1798 if (!keep_exact_action &&
1799 Deoptimization::trap_request_index(trap_request) < 0 &&
1800 too_many_recompiles(reason)) {
1801 // This BCI is causing too many recompilations.
1802 action = Deoptimization::Action_none;
1803 trap_request = Deoptimization::make_trap_request(reason, action);
1804 } else {
1805 C->set_trap_can_recompile(true);
1806 }
1807 break;
1808 case Deoptimization::Action_make_not_entrant:
1809 C->set_trap_can_recompile(true);
1810 break;
1811 #ifdef ASSERT
1812 case Deoptimization::Action_none:
1813 case Deoptimization::Action_make_not_compilable:
1814 break;
1815 default:
1816 assert(false, "bad action");
1817 #endif
1818 }
1820 if (TraceOptoParse) {
1821 char buf[100];
1822 tty->print_cr("Uncommon trap %s at bci:%d",
1823 Deoptimization::format_trap_request(buf, sizeof(buf),
1824 trap_request), bci());
1825 }
1827 CompileLog* log = C->log();
1828 if (log != NULL) {
1829 int kid = (klass == NULL)? -1: log->identify(klass);
1830 log->begin_elem("uncommon_trap bci='%d'", bci());
1831 char buf[100];
1832 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1833 trap_request));
1834 if (kid >= 0) log->print(" klass='%d'", kid);
1835 if (comment != NULL) log->print(" comment='%s'", comment);
1836 log->end_elem();
1837 }
1839 // Make sure any guarding test views this path as very unlikely
1840 Node *i0 = control()->in(0);
1841 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1842 IfNode *iff = i0->as_If();
1843 float f = iff->_prob; // Get prob
1844 if (control()->Opcode() == Op_IfTrue) {
1845 if (f > PROB_UNLIKELY_MAG(4))
1846 iff->_prob = PROB_MIN;
1847 } else {
1848 if (f < PROB_LIKELY_MAG(4))
1849 iff->_prob = PROB_MAX;
1850 }
1851 }
1853 // Clear out dead values from the debug info.
1854 kill_dead_locals();
1856 // Now insert the uncommon trap subroutine call
1857 address call_addr = SharedRuntime::uncommon_trap_blob()->instructions_begin();
1858 const TypePtr* no_memory_effects = NULL;
1859 // Pass the index of the class to be loaded
1860 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1861 (must_throw ? RC_MUST_THROW : 0),
1862 OptoRuntime::uncommon_trap_Type(),
1863 call_addr, "uncommon_trap", no_memory_effects,
1864 intcon(trap_request));
1865 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1866 "must extract request correctly from the graph");
1867 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1869 call->set_req(TypeFunc::ReturnAdr, returnadr());
1870 // The debug info is the only real input to this call.
1872 // Halt-and-catch fire here. The above call should never return!
1873 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1874 _gvn.set_type_bottom(halt);
1875 root()->add_req(halt);
1877 stop_and_kill_map();
1878 }
1881 //--------------------------just_allocated_object------------------------------
1882 // Report the object that was just allocated.
1883 // It must be the case that there are no intervening safepoints.
1884 // We use this to determine if an object is so "fresh" that
1885 // it does not require card marks.
1886 Node* GraphKit::just_allocated_object(Node* current_control) {
1887 if (C->recent_alloc_ctl() == current_control)
1888 return C->recent_alloc_obj();
1889 return NULL;
1890 }
1893 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1894 // (Note: TypeFunc::make has a cache that makes this fast.)
1895 const TypeFunc* tf = TypeFunc::make(dest_method);
1896 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
1897 for (int j = 0; j < nargs; j++) {
1898 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1899 if( targ->basic_type() == T_DOUBLE ) {
1900 // If any parameters are doubles, they must be rounded before
1901 // the call, dstore_rounding does gvn.transform
1902 Node *arg = argument(j);
1903 arg = dstore_rounding(arg);
1904 set_argument(j, arg);
1905 }
1906 }
1907 }
1909 void GraphKit::round_double_result(ciMethod* dest_method) {
1910 // A non-strict method may return a double value which has an extended
1911 // exponent, but this must not be visible in a caller which is 'strict'
1912 // If a strict caller invokes a non-strict callee, round a double result
1914 BasicType result_type = dest_method->return_type()->basic_type();
1915 assert( method() != NULL, "must have caller context");
1916 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
1917 // Destination method's return value is on top of stack
1918 // dstore_rounding() does gvn.transform
1919 Node *result = pop_pair();
1920 result = dstore_rounding(result);
1921 push_pair(result);
1922 }
1923 }
1925 // rounding for strict float precision conformance
1926 Node* GraphKit::precision_rounding(Node* n) {
1927 return UseStrictFP && _method->flags().is_strict()
1928 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
1929 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
1930 : n;
1931 }
1933 // rounding for strict double precision conformance
1934 Node* GraphKit::dprecision_rounding(Node *n) {
1935 return UseStrictFP && _method->flags().is_strict()
1936 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
1937 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1938 : n;
1939 }
1941 // rounding for non-strict double stores
1942 Node* GraphKit::dstore_rounding(Node* n) {
1943 return Matcher::strict_fp_requires_explicit_rounding
1944 && UseSSE <= 1
1945 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1946 : n;
1947 }
1949 //=============================================================================
1950 // Generate a fast path/slow path idiom. Graph looks like:
1951 // [foo] indicates that 'foo' is a parameter
1952 //
1953 // [in] NULL
1954 // \ /
1955 // CmpP
1956 // Bool ne
1957 // If
1958 // / \
1959 // True False-<2>
1960 // / |
1961 // / cast_not_null
1962 // Load | | ^
1963 // [fast_test] | |
1964 // gvn to opt_test | |
1965 // / \ | <1>
1966 // True False |
1967 // | \\ |
1968 // [slow_call] \[fast_result]
1969 // Ctl Val \ \
1970 // | \ \
1971 // Catch <1> \ \
1972 // / \ ^ \ \
1973 // Ex No_Ex | \ \
1974 // | \ \ | \ <2> \
1975 // ... \ [slow_res] | | \ [null_result]
1976 // \ \--+--+--- | |
1977 // \ | / \ | /
1978 // --------Region Phi
1979 //
1980 //=============================================================================
1981 // Code is structured as a series of driver functions all called 'do_XXX' that
1982 // call a set of helper functions. Helper functions first, then drivers.
1984 //------------------------------null_check_oop---------------------------------
1985 // Null check oop. Set null-path control into Region in slot 3.
1986 // Make a cast-not-nullness use the other not-null control. Return cast.
1987 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
1988 bool never_see_null) {
1989 // Initial NULL check taken path
1990 (*null_control) = top();
1991 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
1993 // Generate uncommon_trap:
1994 if (never_see_null && (*null_control) != top()) {
1995 // If we see an unexpected null at a check-cast we record it and force a
1996 // recompile; the offending check-cast will be compiled to handle NULLs.
1997 // If we see more than one offending BCI, then all checkcasts in the
1998 // method will be compiled to handle NULLs.
1999 PreserveJVMState pjvms(this);
2000 set_control(*null_control);
2001 replace_in_map(value, null());
2002 uncommon_trap(Deoptimization::Reason_null_check,
2003 Deoptimization::Action_make_not_entrant);
2004 (*null_control) = top(); // NULL path is dead
2005 }
2007 // Cast away null-ness on the result
2008 return cast;
2009 }
2011 //------------------------------opt_iff----------------------------------------
2012 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2013 // Return slow-path control.
2014 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2015 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2017 // Fast path taken; set region slot 2
2018 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2019 region->init_req(2,fast_taken); // Capture fast-control
2021 // Fast path not-taken, i.e. slow path
2022 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2023 return slow_taken;
2024 }
2026 //-----------------------------make_runtime_call-------------------------------
2027 Node* GraphKit::make_runtime_call(int flags,
2028 const TypeFunc* call_type, address call_addr,
2029 const char* call_name,
2030 const TypePtr* adr_type,
2031 // The following parms are all optional.
2032 // The first NULL ends the list.
2033 Node* parm0, Node* parm1,
2034 Node* parm2, Node* parm3,
2035 Node* parm4, Node* parm5,
2036 Node* parm6, Node* parm7) {
2037 // Slow-path call
2038 int size = call_type->domain()->cnt();
2039 bool is_leaf = !(flags & RC_NO_LEAF);
2040 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2041 if (call_name == NULL) {
2042 assert(!is_leaf, "must supply name for leaf");
2043 call_name = OptoRuntime::stub_name(call_addr);
2044 }
2045 CallNode* call;
2046 if (!is_leaf) {
2047 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2048 bci(), adr_type);
2049 } else if (flags & RC_NO_FP) {
2050 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2051 } else {
2052 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2053 }
2055 // The following is similar to set_edges_for_java_call,
2056 // except that the memory effects of the call are restricted to AliasIdxRaw.
2058 // Slow path call has no side-effects, uses few values
2059 bool wide_in = !(flags & RC_NARROW_MEM);
2060 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2062 Node* prev_mem = NULL;
2063 if (wide_in) {
2064 prev_mem = set_predefined_input_for_runtime_call(call);
2065 } else {
2066 assert(!wide_out, "narrow in => narrow out");
2067 Node* narrow_mem = memory(adr_type);
2068 prev_mem = reset_memory();
2069 map()->set_memory(narrow_mem);
2070 set_predefined_input_for_runtime_call(call);
2071 }
2073 // Hook each parm in order. Stop looking at the first NULL.
2074 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2075 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2076 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2077 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2078 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2079 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2080 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2081 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2082 /* close each nested if ===> */ } } } } } } } }
2083 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2085 if (!is_leaf) {
2086 // Non-leaves can block and take safepoints:
2087 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2088 }
2089 // Non-leaves can throw exceptions:
2090 if (has_io) {
2091 call->set_req(TypeFunc::I_O, i_o());
2092 }
2094 if (flags & RC_UNCOMMON) {
2095 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2096 // (An "if" probability corresponds roughly to an unconditional count.
2097 // Sort of.)
2098 call->set_cnt(PROB_UNLIKELY_MAG(4));
2099 }
2101 Node* c = _gvn.transform(call);
2102 assert(c == call, "cannot disappear");
2104 if (wide_out) {
2105 // Slow path call has full side-effects.
2106 set_predefined_output_for_runtime_call(call);
2107 } else {
2108 // Slow path call has few side-effects, and/or sets few values.
2109 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2110 }
2112 if (has_io) {
2113 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2114 }
2115 return call;
2117 }
2119 //------------------------------merge_memory-----------------------------------
2120 // Merge memory from one path into the current memory state.
2121 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2122 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2123 Node* old_slice = mms.force_memory();
2124 Node* new_slice = mms.memory2();
2125 if (old_slice != new_slice) {
2126 PhiNode* phi;
2127 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2128 phi = new_slice->as_Phi();
2129 #ifdef ASSERT
2130 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2131 old_slice = old_slice->in(new_path);
2132 // Caller is responsible for ensuring that any pre-existing
2133 // phis are already aware of old memory.
2134 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2135 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2136 #endif
2137 mms.set_memory(phi);
2138 } else {
2139 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2140 _gvn.set_type(phi, Type::MEMORY);
2141 phi->set_req(new_path, new_slice);
2142 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2143 }
2144 }
2145 }
2146 }
2148 //------------------------------make_slow_call_ex------------------------------
2149 // Make the exception handler hookups for the slow call
2150 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2151 if (stopped()) return;
2153 // Make a catch node with just two handlers: fall-through and catch-all
2154 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2155 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2156 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2157 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2159 { PreserveJVMState pjvms(this);
2160 set_control(excp);
2161 set_i_o(i_o);
2163 if (excp != top()) {
2164 // Create an exception state also.
2165 // Use an exact type if the caller has specified a specific exception.
2166 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2167 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
2168 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2169 }
2170 }
2172 // Get the no-exception control from the CatchNode.
2173 set_control(norm);
2174 }
2177 //-------------------------------gen_subtype_check-----------------------------
2178 // Generate a subtyping check. Takes as input the subtype and supertype.
2179 // Returns 2 values: sets the default control() to the true path and returns
2180 // the false path. Only reads invariant memory; sets no (visible) memory.
2181 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2182 // but that's not exposed to the optimizer. This call also doesn't take in an
2183 // Object; if you wish to check an Object you need to load the Object's class
2184 // prior to coming here.
2185 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2186 // Fast check for identical types, perhaps identical constants.
2187 // The types can even be identical non-constants, in cases
2188 // involving Array.newInstance, Object.clone, etc.
2189 if (subklass == superklass)
2190 return top(); // false path is dead; no test needed.
2192 if (_gvn.type(superklass)->singleton()) {
2193 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2194 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2196 // In the common case of an exact superklass, try to fold up the
2197 // test before generating code. You may ask, why not just generate
2198 // the code and then let it fold up? The answer is that the generated
2199 // code will necessarily include null checks, which do not always
2200 // completely fold away. If they are also needless, then they turn
2201 // into a performance loss. Example:
2202 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2203 // Here, the type of 'fa' is often exact, so the store check
2204 // of fa[1]=x will fold up, without testing the nullness of x.
2205 switch (static_subtype_check(superk, subk)) {
2206 case SSC_always_false:
2207 {
2208 Node* always_fail = control();
2209 set_control(top());
2210 return always_fail;
2211 }
2212 case SSC_always_true:
2213 return top();
2214 case SSC_easy_test:
2215 {
2216 // Just do a direct pointer compare and be done.
2217 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2218 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2219 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2220 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2221 return _gvn.transform( new(C, 1) IfFalseNode(iff) );
2222 }
2223 case SSC_full_test:
2224 break;
2225 default:
2226 ShouldNotReachHere();
2227 }
2228 }
2230 // %%% Possible further optimization: Even if the superklass is not exact,
2231 // if the subklass is the unique subtype of the superklass, the check
2232 // will always succeed. We could leave a dependency behind to ensure this.
2234 // First load the super-klass's check-offset
2235 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
2236 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2237 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
2238 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2240 // Load from the sub-klass's super-class display list, or a 1-word cache of
2241 // the secondary superclass list, or a failing value with a sentinel offset
2242 // if the super-klass is an interface or exceptionally deep in the Java
2243 // hierarchy and we have to scan the secondary superclass list the hard way.
2244 // Worst-case type is a little odd: NULL is allowed as a result (usually
2245 // klass loads can never produce a NULL).
2246 Node *chk_off_X = ConvI2X(chk_off);
2247 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2248 // For some types like interfaces the following loadKlass is from a 1-word
2249 // cache which is mutable so can't use immutable memory. Other
2250 // types load from the super-class display table which is immutable.
2251 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2252 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2254 // Compile speed common case: ARE a subtype and we canNOT fail
2255 if( superklass == nkls )
2256 return top(); // false path is dead; no test needed.
2258 // See if we get an immediate positive hit. Happens roughly 83% of the
2259 // time. Test to see if the value loaded just previously from the subklass
2260 // is exactly the superklass.
2261 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2262 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2263 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2264 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2265 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2267 // Compile speed common case: Check for being deterministic right now. If
2268 // chk_off is a constant and not equal to cacheoff then we are NOT a
2269 // subklass. In this case we need exactly the 1 test above and we can
2270 // return those results immediately.
2271 if (!might_be_cache) {
2272 Node* not_subtype_ctrl = control();
2273 set_control(iftrue1); // We need exactly the 1 test above
2274 return not_subtype_ctrl;
2275 }
2277 // Gather the various success & failures here
2278 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2279 record_for_igvn(r_ok_subtype);
2280 RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2281 record_for_igvn(r_not_subtype);
2283 r_ok_subtype->init_req(1, iftrue1);
2285 // Check for immediate negative hit. Happens roughly 11% of the time (which
2286 // is roughly 63% of the remaining cases). Test to see if the loaded
2287 // check-offset points into the subklass display list or the 1-element
2288 // cache. If it points to the display (and NOT the cache) and the display
2289 // missed then it's not a subtype.
2290 Node *cacheoff = _gvn.intcon(cacheoff_con);
2291 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2292 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2293 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2294 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2295 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2297 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2298 // No performance impact (too rare) but allows sharing of secondary arrays
2299 // which has some footprint reduction.
2300 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2301 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2302 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2303 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2304 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2306 // -- Roads not taken here: --
2307 // We could also have chosen to perform the self-check at the beginning
2308 // of this code sequence, as the assembler does. This would not pay off
2309 // the same way, since the optimizer, unlike the assembler, can perform
2310 // static type analysis to fold away many successful self-checks.
2311 // Non-foldable self checks work better here in second position, because
2312 // the initial primary superclass check subsumes a self-check for most
2313 // types. An exception would be a secondary type like array-of-interface,
2314 // which does not appear in its own primary supertype display.
2315 // Finally, we could have chosen to move the self-check into the
2316 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2317 // dependent manner. But it is worthwhile to have the check here,
2318 // where it can be perhaps be optimized. The cost in code space is
2319 // small (register compare, branch).
2321 // Now do a linear scan of the secondary super-klass array. Again, no real
2322 // performance impact (too rare) but it's gotta be done.
2323 // Since the code is rarely used, there is no penalty for moving it
2324 // out of line, and it can only improve I-cache density.
2325 // The decision to inline or out-of-line this final check is platform
2326 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2327 Node* psc = _gvn.transform(
2328 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2330 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2331 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2332 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2333 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2334 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2336 // Return false path; set default control to true path.
2337 set_control( _gvn.transform(r_ok_subtype) );
2338 return _gvn.transform(r_not_subtype);
2339 }
2341 //----------------------------static_subtype_check-----------------------------
2342 // Shortcut important common cases when superklass is exact:
2343 // (0) superklass is java.lang.Object (can occur in reflective code)
2344 // (1) subklass is already limited to a subtype of superklass => always ok
2345 // (2) subklass does not overlap with superklass => always fail
2346 // (3) superklass has NO subtypes and we can check with a simple compare.
2347 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2348 if (StressReflectiveCode) {
2349 return SSC_full_test; // Let caller generate the general case.
2350 }
2352 if (superk == env()->Object_klass()) {
2353 return SSC_always_true; // (0) this test cannot fail
2354 }
2356 ciType* superelem = superk;
2357 if (superelem->is_array_klass())
2358 superelem = superelem->as_array_klass()->base_element_type();
2360 if (!subk->is_interface()) { // cannot trust static interface types yet
2361 if (subk->is_subtype_of(superk)) {
2362 return SSC_always_true; // (1) false path dead; no dynamic test needed
2363 }
2364 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2365 !superk->is_subtype_of(subk)) {
2366 return SSC_always_false;
2367 }
2368 }
2370 // If casting to an instance klass, it must have no subtypes
2371 if (superk->is_interface()) {
2372 // Cannot trust interfaces yet.
2373 // %%% S.B. superk->nof_implementors() == 1
2374 } else if (superelem->is_instance_klass()) {
2375 ciInstanceKlass* ik = superelem->as_instance_klass();
2376 if (!ik->has_subklass() && !ik->is_interface()) {
2377 if (!ik->is_final()) {
2378 // Add a dependency if there is a chance of a later subclass.
2379 C->dependencies()->assert_leaf_type(ik);
2380 }
2381 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2382 }
2383 } else {
2384 // A primitive array type has no subtypes.
2385 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2386 }
2388 return SSC_full_test;
2389 }
2391 // Profile-driven exact type check:
2392 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2393 float prob,
2394 Node* *casted_receiver) {
2395 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2396 Node* recv_klass = load_object_klass(receiver);
2397 Node* want_klass = makecon(tklass);
2398 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2399 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2400 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2401 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2402 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2404 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2405 assert(recv_xtype->klass_is_exact(), "");
2407 // Subsume downstream occurrences of receiver with a cast to
2408 // recv_xtype, since now we know what the type will be.
2409 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2410 (*casted_receiver) = _gvn.transform(cast);
2411 // (User must make the replace_in_map call.)
2413 return fail;
2414 }
2417 //-------------------------------gen_instanceof--------------------------------
2418 // Generate an instance-of idiom. Used by both the instance-of bytecode
2419 // and the reflective instance-of call.
2420 Node* GraphKit::gen_instanceof( Node *subobj, Node* superklass ) {
2421 C->set_has_split_ifs(true); // Has chance for split-if optimization
2422 assert( !stopped(), "dead parse path should be checked in callers" );
2423 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2424 "must check for not-null not-dead klass in callers");
2426 // Make the merge point
2427 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2428 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2429 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2430 C->set_has_split_ifs(true); // Has chance for split-if optimization
2432 // Null check; get casted pointer; set region slot 3
2433 Node* null_ctl = top();
2434 Node* not_null_obj = null_check_oop(subobj, &null_ctl);
2436 // If not_null_obj is dead, only null-path is taken
2437 if (stopped()) { // Doing instance-of on a NULL?
2438 set_control(null_ctl);
2439 return intcon(0);
2440 }
2441 region->init_req(_null_path, null_ctl);
2442 phi ->init_req(_null_path, intcon(0)); // Set null path value
2444 // Load the object's klass
2445 Node* obj_klass = load_object_klass(not_null_obj);
2447 // Generate the subtype check
2448 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2450 // Plug in the success path to the general merge in slot 1.
2451 region->init_req(_obj_path, control());
2452 phi ->init_req(_obj_path, intcon(1));
2454 // Plug in the failing path to the general merge in slot 2.
2455 region->init_req(_fail_path, not_subtype_ctrl);
2456 phi ->init_req(_fail_path, intcon(0));
2458 // Return final merged results
2459 set_control( _gvn.transform(region) );
2460 record_for_igvn(region);
2461 return _gvn.transform(phi);
2462 }
2464 //-------------------------------gen_checkcast---------------------------------
2465 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2466 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2467 // uncommon-trap paths work. Adjust stack after this call.
2468 // If failure_control is supplied and not null, it is filled in with
2469 // the control edge for the cast failure. Otherwise, an appropriate
2470 // uncommon trap or exception is thrown.
2471 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2472 Node* *failure_control) {
2473 kill_dead_locals(); // Benefit all the uncommon traps
2474 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2475 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2477 // Fast cutout: Check the case that the cast is vacuously true.
2478 // This detects the common cases where the test will short-circuit
2479 // away completely. We do this before we perform the null check,
2480 // because if the test is going to turn into zero code, we don't
2481 // want a residual null check left around. (Causes a slowdown,
2482 // for example, in some objArray manipulations, such as a[i]=a[j].)
2483 if (tk->singleton()) {
2484 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2485 if (objtp != NULL && objtp->klass() != NULL) {
2486 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2487 case SSC_always_true:
2488 return obj;
2489 case SSC_always_false:
2490 // It needs a null check because a null will *pass* the cast check.
2491 // A non-null value will always produce an exception.
2492 return do_null_assert(obj, T_OBJECT);
2493 }
2494 }
2495 }
2497 ciProfileData* data = NULL;
2498 if (failure_control == NULL) { // use MDO in regular case only
2499 assert(java_bc() == Bytecodes::_aastore ||
2500 java_bc() == Bytecodes::_checkcast,
2501 "interpreter profiles type checks only for these BCs");
2502 data = method()->method_data()->bci_to_data(bci());
2503 }
2505 // Make the merge point
2506 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2507 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2508 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
2509 C->set_has_split_ifs(true); // Has chance for split-if optimization
2511 // Use null-cast information if it is available
2512 bool never_see_null = false;
2513 // If we see an unexpected null at a check-cast we record it and force a
2514 // recompile; the offending check-cast will be compiled to handle NULLs.
2515 // If we see several offending BCIs, then all checkcasts in the
2516 // method will be compiled to handle NULLs.
2517 if (UncommonNullCast // Cutout for this technique
2518 && failure_control == NULL // regular case
2519 && obj != null() // And not the -Xcomp stupid case?
2520 && !too_many_traps(Deoptimization::Reason_null_check)) {
2521 // Finally, check the "null_seen" bit from the interpreter.
2522 if (data == NULL || !data->as_BitData()->null_seen()) {
2523 never_see_null = true;
2524 }
2525 }
2527 // Null check; get casted pointer; set region slot 3
2528 Node* null_ctl = top();
2529 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2531 // If not_null_obj is dead, only null-path is taken
2532 if (stopped()) { // Doing instance-of on a NULL?
2533 set_control(null_ctl);
2534 return null();
2535 }
2536 region->init_req(_null_path, null_ctl);
2537 phi ->init_req(_null_path, null()); // Set null path value
2539 Node* cast_obj = NULL; // the casted version of the object
2541 // If the profile has seen exactly one type, narrow to that type.
2542 // (The subsequent subtype check will always fold up.)
2543 if (UseTypeProfile && TypeProfileCasts && data != NULL &&
2544 // Counter has never been decremented (due to cast failure).
2545 // ...This is a reasonable thing to expect. It is true of
2546 // all casts inserted by javac to implement generic types.
2547 data->as_CounterData()->count() >= 0 &&
2548 !too_many_traps(Deoptimization::Reason_class_check)) {
2549 // (No, this isn't a call, but it's enough like a virtual call
2550 // to use the same ciMethod accessor to get the profile info...)
2551 ciCallProfile profile = method()->call_profile_at_bci(bci());
2552 if (profile.count() >= 0 && // no cast failures here
2553 profile.has_receiver(0) &&
2554 profile.morphism() == 1) {
2555 ciKlass* exact_kls = profile.receiver(0);
2556 int ssc = static_subtype_check(tk->klass(), exact_kls);
2557 if (ssc == SSC_always_true) {
2558 // If we narrow the type to match what the type profile sees,
2559 // we can then remove the rest of the cast.
2560 // This is a win, even if the exact_kls is very specific,
2561 // because downstream operations, such as method calls,
2562 // will often benefit from the sharper type.
2563 Node* exact_obj = not_null_obj; // will get updated in place...
2564 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2565 &exact_obj);
2566 { PreserveJVMState pjvms(this);
2567 set_control(slow_ctl);
2568 uncommon_trap(Deoptimization::Reason_class_check,
2569 Deoptimization::Action_maybe_recompile);
2570 }
2571 if (failure_control != NULL) // failure is now impossible
2572 (*failure_control) = top();
2573 replace_in_map(not_null_obj, exact_obj);
2574 // adjust the type of the phi to the exact klass:
2575 phi->raise_bottom_type(_gvn.type(exact_obj)->meet(TypePtr::NULL_PTR));
2576 cast_obj = exact_obj;
2577 }
2578 // assert(cast_obj != NULL)... except maybe the profile lied to us.
2579 }
2580 }
2582 if (cast_obj == NULL) {
2583 // Load the object's klass
2584 Node* obj_klass = load_object_klass(not_null_obj);
2586 // Generate the subtype check
2587 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2589 // Plug in success path into the merge
2590 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2591 not_null_obj, toop));
2592 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2593 if (failure_control == NULL) {
2594 if (not_subtype_ctrl != top()) { // If failure is possible
2595 PreserveJVMState pjvms(this);
2596 set_control(not_subtype_ctrl);
2597 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2598 }
2599 } else {
2600 (*failure_control) = not_subtype_ctrl;
2601 }
2602 }
2604 region->init_req(_obj_path, control());
2605 phi ->init_req(_obj_path, cast_obj);
2607 // A merge of NULL or Casted-NotNull obj
2608 Node* res = _gvn.transform(phi);
2610 // Note I do NOT always 'replace_in_map(obj,result)' here.
2611 // if( tk->klass()->can_be_primary_super() )
2612 // This means that if I successfully store an Object into an array-of-String
2613 // I 'forget' that the Object is really now known to be a String. I have to
2614 // do this because we don't have true union types for interfaces - if I store
2615 // a Baz into an array-of-Interface and then tell the optimizer it's an
2616 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2617 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2618 // replace_in_map( obj, res );
2620 // Return final merged results
2621 set_control( _gvn.transform(region) );
2622 record_for_igvn(region);
2623 return res;
2624 }
2626 //------------------------------next_monitor-----------------------------------
2627 // What number should be given to the next monitor?
2628 int GraphKit::next_monitor() {
2629 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2630 int next = current + C->sync_stack_slots();
2631 // Keep the toplevel high water mark current:
2632 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2633 return current;
2634 }
2636 //------------------------------insert_mem_bar---------------------------------
2637 // Memory barrier to avoid floating things around
2638 // The membar serves as a pinch point between both control and all memory slices.
2639 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2640 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2641 mb->init_req(TypeFunc::Control, control());
2642 mb->init_req(TypeFunc::Memory, reset_memory());
2643 Node* membar = _gvn.transform(mb);
2644 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2645 set_all_memory_call(membar);
2646 return membar;
2647 }
2649 //-------------------------insert_mem_bar_volatile----------------------------
2650 // Memory barrier to avoid floating things around
2651 // The membar serves as a pinch point between both control and memory(alias_idx).
2652 // If you want to make a pinch point on all memory slices, do not use this
2653 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2654 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2655 // When Parse::do_put_xxx updates a volatile field, it appends a series
2656 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2657 // The first membar is on the same memory slice as the field store opcode.
2658 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2659 // All the other membars (for other volatile slices, including AliasIdxBot,
2660 // which stands for all unknown volatile slices) are control-dependent
2661 // on the first membar. This prevents later volatile loads or stores
2662 // from sliding up past the just-emitted store.
2664 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2665 mb->set_req(TypeFunc::Control,control());
2666 if (alias_idx == Compile::AliasIdxBot) {
2667 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2668 } else {
2669 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2670 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2671 }
2672 Node* membar = _gvn.transform(mb);
2673 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2674 if (alias_idx == Compile::AliasIdxBot) {
2675 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2676 } else {
2677 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2678 }
2679 return membar;
2680 }
2682 //------------------------------shared_lock------------------------------------
2683 // Emit locking code.
2684 FastLockNode* GraphKit::shared_lock(Node* obj) {
2685 // bci is either a monitorenter bc or InvocationEntryBci
2686 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2687 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2689 if( !GenerateSynchronizationCode )
2690 return NULL; // Not locking things?
2691 if (stopped()) // Dead monitor?
2692 return NULL;
2694 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2696 // Box the stack location
2697 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2698 Node* mem = reset_memory();
2700 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2701 if (PrintPreciseBiasedLockingStatistics) {
2702 // Create the counters for this fast lock.
2703 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2704 }
2705 // Add monitor to debug info for the slow path. If we block inside the
2706 // slow path and de-opt, we need the monitor hanging around
2707 map()->push_monitor( flock );
2709 const TypeFunc *tf = LockNode::lock_type();
2710 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2712 lock->init_req( TypeFunc::Control, control() );
2713 lock->init_req( TypeFunc::Memory , mem );
2714 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2715 lock->init_req( TypeFunc::FramePtr, frameptr() );
2716 lock->init_req( TypeFunc::ReturnAdr, top() );
2718 lock->init_req(TypeFunc::Parms + 0, obj);
2719 lock->init_req(TypeFunc::Parms + 1, box);
2720 lock->init_req(TypeFunc::Parms + 2, flock);
2721 add_safepoint_edges(lock);
2723 lock = _gvn.transform( lock )->as_Lock();
2725 // lock has no side-effects, sets few values
2726 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2728 insert_mem_bar(Op_MemBarAcquire);
2730 // Add this to the worklist so that the lock can be eliminated
2731 record_for_igvn(lock);
2733 #ifndef PRODUCT
2734 if (PrintLockStatistics) {
2735 // Update the counter for this lock. Don't bother using an atomic
2736 // operation since we don't require absolute accuracy.
2737 lock->create_lock_counter(map()->jvms());
2738 int adr_type = Compile::AliasIdxRaw;
2739 Node* counter_addr = makecon(TypeRawPtr::make(lock->counter()->addr()));
2740 Node* cnt = make_load(NULL, counter_addr, TypeInt::INT, T_INT, adr_type);
2741 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
2742 store_to_memory(control(), counter_addr, incr, T_INT, adr_type);
2743 }
2744 #endif
2746 return flock;
2747 }
2750 //------------------------------shared_unlock----------------------------------
2751 // Emit unlocking code.
2752 void GraphKit::shared_unlock(Node* box, Node* obj) {
2753 // bci is either a monitorenter bc or InvocationEntryBci
2754 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2755 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2757 if( !GenerateSynchronizationCode )
2758 return;
2759 if (stopped()) { // Dead monitor?
2760 map()->pop_monitor(); // Kill monitor from debug info
2761 return;
2762 }
2764 // Memory barrier to avoid floating things down past the locked region
2765 insert_mem_bar(Op_MemBarRelease);
2767 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2768 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2769 uint raw_idx = Compile::AliasIdxRaw;
2770 unlock->init_req( TypeFunc::Control, control() );
2771 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2772 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2773 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2774 unlock->init_req( TypeFunc::ReturnAdr, top() );
2776 unlock->init_req(TypeFunc::Parms + 0, obj);
2777 unlock->init_req(TypeFunc::Parms + 1, box);
2778 unlock = _gvn.transform(unlock)->as_Unlock();
2780 Node* mem = reset_memory();
2782 // unlock has no side-effects, sets few values
2783 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2785 // Kill monitor from debug info
2786 map()->pop_monitor( );
2787 }
2789 //-------------------------------get_layout_helper-----------------------------
2790 // If the given klass is a constant or known to be an array,
2791 // fetch the constant layout helper value into constant_value
2792 // and return (Node*)NULL. Otherwise, load the non-constant
2793 // layout helper value, and return the node which represents it.
2794 // This two-faced routine is useful because allocation sites
2795 // almost always feature constant types.
2796 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2797 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2798 if (!StressReflectiveCode && inst_klass != NULL) {
2799 ciKlass* klass = inst_klass->klass();
2800 bool xklass = inst_klass->klass_is_exact();
2801 if (xklass || klass->is_array_klass()) {
2802 jint lhelper = klass->layout_helper();
2803 if (lhelper != Klass::_lh_neutral_value) {
2804 constant_value = lhelper;
2805 return (Node*) NULL;
2806 }
2807 }
2808 }
2809 constant_value = Klass::_lh_neutral_value; // put in a known value
2810 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
2811 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2812 }
2814 // We just put in an allocate/initialize with a big raw-memory effect.
2815 // Hook selected additional alias categories on the initialization.
2816 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2817 MergeMemNode* init_in_merge,
2818 Node* init_out_raw) {
2819 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2820 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2822 Node* prevmem = kit.memory(alias_idx);
2823 init_in_merge->set_memory_at(alias_idx, prevmem);
2824 kit.set_memory(init_out_raw, alias_idx);
2825 }
2827 //---------------------------set_output_for_allocation-------------------------
2828 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2829 const TypeOopPtr* oop_type,
2830 bool raw_mem_only) {
2831 int rawidx = Compile::AliasIdxRaw;
2832 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2833 add_safepoint_edges(alloc);
2834 Node* allocx = _gvn.transform(alloc);
2835 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2836 // create memory projection for i_o
2837 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2838 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2840 // create a memory projection as for the normal control path
2841 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2842 set_memory(malloc, rawidx);
2844 // a normal slow-call doesn't change i_o, but an allocation does
2845 // we create a separate i_o projection for the normal control path
2846 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2847 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2849 // put in an initialization barrier
2850 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2851 rawoop)->as_Initialize();
2852 assert(alloc->initialization() == init, "2-way macro link must work");
2853 assert(init ->allocation() == alloc, "2-way macro link must work");
2854 if (ReduceFieldZeroing && !raw_mem_only) {
2855 // Extract memory strands which may participate in the new object's
2856 // initialization, and source them from the new InitializeNode.
2857 // This will allow us to observe initializations when they occur,
2858 // and link them properly (as a group) to the InitializeNode.
2859 assert(init->in(InitializeNode::Memory) == malloc, "");
2860 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
2861 init->set_req(InitializeNode::Memory, minit_in);
2862 record_for_igvn(minit_in); // fold it up later, if possible
2863 Node* minit_out = memory(rawidx);
2864 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
2865 if (oop_type->isa_aryptr()) {
2866 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
2867 int elemidx = C->get_alias_index(telemref);
2868 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
2869 } else if (oop_type->isa_instptr()) {
2870 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
2871 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
2872 ciField* field = ik->nonstatic_field_at(i);
2873 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
2874 continue; // do not bother to track really large numbers of fields
2875 // Find (or create) the alias category for this field:
2876 int fieldidx = C->alias_type(field)->index();
2877 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
2878 }
2879 }
2880 }
2882 // Cast raw oop to the real thing...
2883 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
2884 javaoop = _gvn.transform(javaoop);
2885 C->set_recent_alloc(control(), javaoop);
2886 assert(just_allocated_object(control()) == javaoop, "just allocated");
2888 #ifdef ASSERT
2889 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
2890 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
2891 "Ideal_allocation works");
2892 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
2893 "Ideal_allocation works");
2894 if (alloc->is_AllocateArray()) {
2895 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
2896 "Ideal_allocation works");
2897 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
2898 "Ideal_allocation works");
2899 } else {
2900 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
2901 }
2902 }
2903 #endif //ASSERT
2905 return javaoop;
2906 }
2908 //---------------------------new_instance--------------------------------------
2909 // This routine takes a klass_node which may be constant (for a static type)
2910 // or may be non-constant (for reflective code). It will work equally well
2911 // for either, and the graph will fold nicely if the optimizer later reduces
2912 // the type to a constant.
2913 // The optional arguments are for specialized use by intrinsics:
2914 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
2915 // - If 'raw_mem_only', do not cast the result to an oop.
2916 // - If 'return_size_val', report the the total object size to the caller.
2917 Node* GraphKit::new_instance(Node* klass_node,
2918 Node* extra_slow_test,
2919 bool raw_mem_only, // affect only raw memory
2920 Node* *return_size_val) {
2921 // Compute size in doublewords
2922 // The size is always an integral number of doublewords, represented
2923 // as a positive bytewise size stored in the klass's layout_helper.
2924 // The layout_helper also encodes (in a low bit) the need for a slow path.
2925 jint layout_con = Klass::_lh_neutral_value;
2926 Node* layout_val = get_layout_helper(klass_node, layout_con);
2927 int layout_is_con = (layout_val == NULL);
2929 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
2930 // Generate the initial go-slow test. It's either ALWAYS (return a
2931 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
2932 // case) a computed value derived from the layout_helper.
2933 Node* initial_slow_test = NULL;
2934 if (layout_is_con) {
2935 assert(!StressReflectiveCode, "stress mode does not use these paths");
2936 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
2937 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
2939 } else { // reflective case
2940 // This reflective path is used by Unsafe.allocateInstance.
2941 // (It may be stress-tested by specifying StressReflectiveCode.)
2942 // Basically, we want to get into the VM is there's an illegal argument.
2943 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
2944 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
2945 if (extra_slow_test != intcon(0)) {
2946 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
2947 }
2948 // (Macro-expander will further convert this to a Bool, if necessary.)
2949 }
2951 // Find the size in bytes. This is easy; it's the layout_helper.
2952 // The size value must be valid even if the slow path is taken.
2953 Node* size = NULL;
2954 if (layout_is_con) {
2955 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
2956 } else { // reflective case
2957 // This reflective path is used by clone and Unsafe.allocateInstance.
2958 size = ConvI2X(layout_val);
2960 // Clear the low bits to extract layout_helper_size_in_bytes:
2961 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
2962 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
2963 size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
2964 }
2965 if (return_size_val != NULL) {
2966 (*return_size_val) = size;
2967 }
2969 // This is a precise notnull oop of the klass.
2970 // (Actually, it need not be precise if this is a reflective allocation.)
2971 // It's what we cast the result to.
2972 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
2973 if (!tklass) tklass = TypeKlassPtr::OBJECT;
2974 const TypeOopPtr* oop_type = tklass->as_instance_type();
2976 // Now generate allocation code
2978 // The entire memory state is needed for slow path of the allocation
2979 // since GC and deoptimization can happened.
2980 Node *mem = reset_memory();
2981 set_all_memory(mem); // Create new memory state
2983 AllocateNode* alloc
2984 = new (C, AllocateNode::ParmLimit)
2985 AllocateNode(C, AllocateNode::alloc_type(),
2986 control(), mem, i_o(),
2987 size, klass_node,
2988 initial_slow_test);
2990 return set_output_for_allocation(alloc, oop_type, raw_mem_only);
2991 }
2993 //-------------------------------new_array-------------------------------------
2994 // helper for both newarray and anewarray
2995 // The 'length' parameter is (obviously) the length of the array.
2996 // See comments on new_instance for the meaning of the other arguments.
2997 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
2998 Node* length, // number of array elements
2999 int nargs, // number of arguments to push back for uncommon trap
3000 bool raw_mem_only, // affect only raw memory
3001 Node* *return_size_val) {
3002 jint layout_con = Klass::_lh_neutral_value;
3003 Node* layout_val = get_layout_helper(klass_node, layout_con);
3004 int layout_is_con = (layout_val == NULL);
3006 if (!layout_is_con && !StressReflectiveCode &&
3007 !too_many_traps(Deoptimization::Reason_class_check)) {
3008 // This is a reflective array creation site.
3009 // Optimistically assume that it is a subtype of Object[],
3010 // so that we can fold up all the address arithmetic.
3011 layout_con = Klass::array_layout_helper(T_OBJECT);
3012 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
3013 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
3014 { BuildCutout unless(this, bol_lh, PROB_MAX);
3015 _sp += nargs;
3016 uncommon_trap(Deoptimization::Reason_class_check,
3017 Deoptimization::Action_maybe_recompile);
3018 }
3019 layout_val = NULL;
3020 layout_is_con = true;
3021 }
3023 // Generate the initial go-slow test. Make sure we do not overflow
3024 // if length is huge (near 2Gig) or negative! We do not need
3025 // exact double-words here, just a close approximation of needed
3026 // double-words. We can't add any offset or rounding bits, lest we
3027 // take a size -1 of bytes and make it positive. Use an unsigned
3028 // compare, so negative sizes look hugely positive.
3029 int fast_size_limit = FastAllocateSizeLimit;
3030 if (layout_is_con) {
3031 assert(!StressReflectiveCode, "stress mode does not use these paths");
3032 // Increase the size limit if we have exact knowledge of array type.
3033 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3034 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3035 }
3037 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3038 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3039 if (initial_slow_test->is_Bool()) {
3040 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3041 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3042 }
3044 // --- Size Computation ---
3045 // array_size = round_to_heap(array_header + (length << elem_shift));
3046 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3047 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3048 // The rounding mask is strength-reduced, if possible.
3049 int round_mask = MinObjAlignmentInBytes - 1;
3050 Node* header_size = NULL;
3051 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3052 // (T_BYTE has the weakest alignment and size restrictions...)
3053 if (layout_is_con) {
3054 int hsize = Klass::layout_helper_header_size(layout_con);
3055 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3056 BasicType etype = Klass::layout_helper_element_type(layout_con);
3057 if ((round_mask & ~right_n_bits(eshift)) == 0)
3058 round_mask = 0; // strength-reduce it if it goes away completely
3059 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3060 assert(header_size_min <= hsize, "generic minimum is smallest");
3061 header_size_min = hsize;
3062 header_size = intcon(hsize + round_mask);
3063 } else {
3064 Node* hss = intcon(Klass::_lh_header_size_shift);
3065 Node* hsm = intcon(Klass::_lh_header_size_mask);
3066 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3067 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3068 Node* mask = intcon(round_mask);
3069 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3070 }
3072 Node* elem_shift = NULL;
3073 if (layout_is_con) {
3074 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3075 if (eshift != 0)
3076 elem_shift = intcon(eshift);
3077 } else {
3078 // There is no need to mask or shift this value.
3079 // The semantics of LShiftINode include an implicit mask to 0x1F.
3080 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3081 elem_shift = layout_val;
3082 }
3084 // Transition to native address size for all offset calculations:
3085 Node* lengthx = ConvI2X(length);
3086 Node* headerx = ConvI2X(header_size);
3087 #ifdef _LP64
3088 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3089 if (tllen != NULL && tllen->_lo < 0) {
3090 // Add a manual constraint to a positive range. Cf. array_element_address.
3091 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3092 if (size_max > tllen->_hi) size_max = tllen->_hi;
3093 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3094 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3095 }
3096 }
3097 #endif
3099 // Combine header size (plus rounding) and body size. Then round down.
3100 // This computation cannot overflow, because it is used only in two
3101 // places, one where the length is sharply limited, and the other
3102 // after a successful allocation.
3103 Node* abody = lengthx;
3104 if (elem_shift != NULL)
3105 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3106 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3107 if (round_mask != 0) {
3108 Node* mask = MakeConX(~round_mask);
3109 size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3110 }
3111 // else if round_mask == 0, the size computation is self-rounding
3113 if (return_size_val != NULL) {
3114 // This is the size
3115 (*return_size_val) = size;
3116 }
3118 // Now generate allocation code
3120 // The entire memory state is needed for slow path of the allocation
3121 // since GC and deoptimization can happened.
3122 Node *mem = reset_memory();
3123 set_all_memory(mem); // Create new memory state
3125 // Create the AllocateArrayNode and its result projections
3126 AllocateArrayNode* alloc
3127 = new (C, AllocateArrayNode::ParmLimit)
3128 AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3129 control(), mem, i_o(),
3130 size, klass_node,
3131 initial_slow_test,
3132 length);
3134 // Cast to correct type. Note that the klass_node may be constant or not,
3135 // and in the latter case the actual array type will be inexact also.
3136 // (This happens via a non-constant argument to inline_native_newArray.)
3137 // In any case, the value of klass_node provides the desired array type.
3138 const TypeInt* length_type = _gvn.find_int_type(length);
3139 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3140 if (ary_type->isa_aryptr() && length_type != NULL) {
3141 // Try to get a better type than POS for the size
3142 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3143 }
3145 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
3147 // Cast length on remaining path to be as narrow as possible
3148 if (map()->find_edge(length) >= 0) {
3149 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3150 if (ccast != length) {
3151 _gvn.set_type_bottom(ccast);
3152 record_for_igvn(ccast);
3153 replace_in_map(length, ccast);
3154 }
3155 }
3157 return javaoop;
3158 }
3160 // The following "Ideal_foo" functions are placed here because they recognize
3161 // the graph shapes created by the functions immediately above.
3163 //---------------------------Ideal_allocation----------------------------------
3164 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3165 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3166 if (ptr == NULL) { // reduce dumb test in callers
3167 return NULL;
3168 }
3169 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3170 ptr = ptr->in(1);
3171 if (ptr == NULL) return NULL;
3172 }
3173 if (ptr->is_Proj()) {
3174 Node* allo = ptr->in(0);
3175 if (allo != NULL && allo->is_Allocate()) {
3176 return allo->as_Allocate();
3177 }
3178 }
3179 // Report failure to match.
3180 return NULL;
3181 }
3183 // Fancy version which also strips off an offset (and reports it to caller).
3184 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3185 intptr_t& offset) {
3186 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3187 if (base == NULL) return NULL;
3188 return Ideal_allocation(base, phase);
3189 }
3191 // Trace Initialize <- Proj[Parm] <- Allocate
3192 AllocateNode* InitializeNode::allocation() {
3193 Node* rawoop = in(InitializeNode::RawAddress);
3194 if (rawoop->is_Proj()) {
3195 Node* alloc = rawoop->in(0);
3196 if (alloc->is_Allocate()) {
3197 return alloc->as_Allocate();
3198 }
3199 }
3200 return NULL;
3201 }
3203 // Trace Allocate -> Proj[Parm] -> Initialize
3204 InitializeNode* AllocateNode::initialization() {
3205 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3206 if (rawoop == NULL) return NULL;
3207 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3208 Node* init = rawoop->fast_out(i);
3209 if (init->is_Initialize()) {
3210 assert(init->as_Initialize()->allocation() == this, "2-way link");
3211 return init->as_Initialize();
3212 }
3213 }
3214 return NULL;
3215 }
3217 //----------------------------- store barriers ----------------------------
3218 #define __ ideal.
3220 void GraphKit::sync_kit(IdealKit& ideal) {
3221 // Final sync IdealKit and graphKit.
3222 __ drain_delay_transform();
3223 set_all_memory(__ merged_memory());
3224 set_control(__ ctrl());
3225 }
3227 // vanilla/CMS post barrier
3228 // Insert a write-barrier store. This is to let generational GC work; we have
3229 // to flag all oop-stores before the next GC point.
3230 void GraphKit::write_barrier_post(Node* oop_store,
3231 Node* obj,
3232 Node* adr,
3233 uint adr_idx,
3234 Node* val,
3235 bool use_precise) {
3236 // No store check needed if we're storing a NULL or an old object
3237 // (latter case is probably a string constant). The concurrent
3238 // mark sweep garbage collector, however, needs to have all nonNull
3239 // oop updates flagged via card-marks.
3240 if (val != NULL && val->is_Con()) {
3241 // must be either an oop or NULL
3242 const Type* t = val->bottom_type();
3243 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3244 // stores of null never (?) need barriers
3245 return;
3246 ciObject* con = t->is_oopptr()->const_oop();
3247 if (con != NULL
3248 && con->is_perm()
3249 && Universe::heap()->can_elide_permanent_oop_store_barriers())
3250 // no store barrier needed, because no old-to-new ref created
3251 return;
3252 }
3254 if (use_ReduceInitialCardMarks()
3255 && obj == just_allocated_object(control())) {
3256 // We can skip marks on a freshly-allocated object in Eden.
3257 // Keep this code in sync with maybe_defer_card_mark() in runtime.cpp.
3258 // That routine informs GC to take appropriate compensating steps
3259 // so as to make this card-mark elision safe.
3260 return;
3261 }
3263 if (!use_precise) {
3264 // All card marks for a (non-array) instance are in one place:
3265 adr = obj;
3266 }
3267 // (Else it's an array (or unknown), and we want more precise card marks.)
3268 assert(adr != NULL, "");
3270 IdealKit ideal(gvn(), control(), merged_memory(), true);
3272 // Convert the pointer to an int prior to doing math on it
3273 Node* cast = __ CastPX(__ ctrl(), adr);
3275 // Divide by card size
3276 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3277 "Only one we handle so far.");
3278 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3280 // Combine card table base and card offset
3281 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3283 // Get the alias_index for raw card-mark memory
3284 int adr_type = Compile::AliasIdxRaw;
3285 // Smash zero into card
3286 Node* zero = __ ConI(0);
3287 BasicType bt = T_BYTE;
3288 if( !UseConcMarkSweepGC ) {
3289 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3290 } else {
3291 // Specialized path for CM store barrier
3292 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3293 }
3295 // Final sync IdealKit and GraphKit.
3296 sync_kit(ideal);
3297 }
3299 // G1 pre/post barriers
3300 void GraphKit::g1_write_barrier_pre(Node* obj,
3301 Node* adr,
3302 uint alias_idx,
3303 Node* val,
3304 const TypeOopPtr* val_type,
3305 BasicType bt) {
3306 IdealKit ideal(gvn(), control(), merged_memory(), true);
3308 Node* tls = __ thread(); // ThreadLocalStorage
3310 Node* no_ctrl = NULL;
3311 Node* no_base = __ top();
3312 Node* zero = __ ConI(0);
3314 float likely = PROB_LIKELY(0.999);
3315 float unlikely = PROB_UNLIKELY(0.999);
3317 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3318 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3320 // Offsets into the thread
3321 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3322 PtrQueue::byte_offset_of_active());
3323 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3324 PtrQueue::byte_offset_of_index());
3325 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3326 PtrQueue::byte_offset_of_buf());
3327 // Now the actual pointers into the thread
3329 // set_control( ctl);
3331 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3332 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3333 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3335 // Now some of the values
3337 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3339 // if (!marking)
3340 __ if_then(marking, BoolTest::ne, zero); {
3341 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3343 const Type* t1 = adr->bottom_type();
3344 const Type* t2 = val->bottom_type();
3346 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3347 // if (orig != NULL)
3348 __ if_then(orig, BoolTest::ne, null()); {
3349 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3351 // load original value
3352 // alias_idx correct??
3354 // is the queue for this thread full?
3355 __ if_then(index, BoolTest::ne, zero, likely); {
3357 // decrement the index
3358 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3359 Node* next_indexX = next_index;
3360 #ifdef _LP64
3361 // We could refine the type for what it's worth
3362 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3363 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3364 #endif
3366 // Now get the buffer location we will log the original value into and store it
3367 Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3368 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw);
3370 // update the index
3371 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3373 } __ else_(); {
3375 // logging buffer is full, call the runtime
3376 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3377 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls);
3378 } __ end_if(); // (!index)
3379 } __ end_if(); // (orig != NULL)
3380 } __ end_if(); // (!marking)
3382 // Final sync IdealKit and GraphKit.
3383 sync_kit(ideal);
3384 }
3386 //
3387 // Update the card table and add card address to the queue
3388 //
3389 void GraphKit::g1_mark_card(IdealKit& ideal,
3390 Node* card_adr,
3391 Node* oop_store,
3392 uint oop_alias_idx,
3393 Node* index,
3394 Node* index_adr,
3395 Node* buffer,
3396 const TypeFunc* tf) {
3398 Node* zero = __ ConI(0);
3399 Node* no_base = __ top();
3400 BasicType card_bt = T_BYTE;
3401 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3402 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3404 // Now do the queue work
3405 __ if_then(index, BoolTest::ne, zero); {
3407 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3408 Node* next_indexX = next_index;
3409 #ifdef _LP64
3410 // We could refine the type for what it's worth
3411 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3412 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3413 #endif // _LP64
3414 Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3416 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3417 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3419 } __ else_(); {
3420 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3421 } __ end_if();
3423 }
3425 void GraphKit::g1_write_barrier_post(Node* oop_store,
3426 Node* obj,
3427 Node* adr,
3428 uint alias_idx,
3429 Node* val,
3430 BasicType bt,
3431 bool use_precise) {
3432 // If we are writing a NULL then we need no post barrier
3434 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3435 // Must be NULL
3436 const Type* t = val->bottom_type();
3437 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3438 // No post barrier if writing NULLx
3439 return;
3440 }
3442 if (!use_precise) {
3443 // All card marks for a (non-array) instance are in one place:
3444 adr = obj;
3445 }
3446 // (Else it's an array (or unknown), and we want more precise card marks.)
3447 assert(adr != NULL, "");
3449 IdealKit ideal(gvn(), control(), merged_memory(), true);
3451 Node* tls = __ thread(); // ThreadLocalStorage
3453 Node* no_ctrl = NULL;
3454 Node* no_base = __ top();
3455 float likely = PROB_LIKELY(0.999);
3456 float unlikely = PROB_UNLIKELY(0.999);
3457 Node* zero = __ ConI(0);
3458 Node* zeroX = __ ConX(0);
3460 // Get the alias_index for raw card-mark memory
3461 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3463 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3465 // Offsets into the thread
3466 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3467 PtrQueue::byte_offset_of_index());
3468 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3469 PtrQueue::byte_offset_of_buf());
3471 // Pointers into the thread
3473 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3474 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3476 // Now some values
3478 Node* index = __ load(no_ctrl, index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3479 Node* buffer = __ load(no_ctrl, buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3482 // Convert the store obj pointer to an int prior to doing math on it
3483 // Must use ctrl to prevent "integerized oop" existing across safepoint
3484 Node* cast = __ CastPX(__ ctrl(), adr);
3486 // Divide pointer by card size
3487 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3489 // Combine card table base and card offset
3490 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3492 // If we know the value being stored does it cross regions?
3494 if (val != NULL) {
3495 // Does the store cause us to cross regions?
3497 // Should be able to do an unsigned compare of region_size instead of
3498 // and extra shift. Do we have an unsigned compare??
3499 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3500 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3502 // if (xor_res == 0) same region so skip
3503 __ if_then(xor_res, BoolTest::ne, zeroX); {
3505 // No barrier if we are storing a NULL
3506 __ if_then(val, BoolTest::ne, null(), unlikely); {
3508 // Ok must mark the card if not already dirty
3510 // load the original value of the card
3511 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3513 __ if_then(card_val, BoolTest::ne, zero); {
3514 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3515 } __ end_if();
3516 } __ end_if();
3517 } __ end_if();
3518 } else {
3519 // Object.clone() instrinsic uses this path.
3520 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3521 }
3523 // Final sync IdealKit and GraphKit.
3524 sync_kit(ideal);
3525 }
3526 #undef __