Wed, 08 Sep 2010 20:28:57 -0700
6965815: OptimizeStringConcat: assert(!q->is_MergeMem()) failed with specjbb2000
Reviewed-by: kvn
1 /*
2 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * 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 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
459 bool must_throw) {
460 // if the exception capability is set, then we will generate code
461 // to check the JavaThread.should_post_on_exceptions flag to see
462 // if we actually need to report exception events (for this
463 // thread). If we don't need to report exception events, we will
464 // take the normal fast path provided by add_exception_events. If
465 // exception event reporting is enabled for this thread, we will
466 // take the uncommon_trap in the BuildCutout below.
468 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
469 Node* jthread = _gvn.transform(new (C, 1) ThreadLocalNode());
470 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
471 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false);
473 // Test the should_post_on_exceptions_flag vs. 0
474 Node* chk = _gvn.transform( new (C, 3) CmpINode(should_post_flag, intcon(0)) );
475 Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, BoolTest::eq) );
477 // Branch to slow_path if should_post_on_exceptions_flag was true
478 { BuildCutout unless(this, tst, PROB_MAX);
479 // Do not try anything fancy if we're notifying the VM on every throw.
480 // Cf. case Bytecodes::_athrow in parse2.cpp.
481 uncommon_trap(reason, Deoptimization::Action_none,
482 (ciKlass*)NULL, (char*)NULL, must_throw);
483 }
485 }
487 //------------------------------builtin_throw----------------------------------
488 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
489 bool must_throw = true;
491 if (env()->jvmti_can_post_on_exceptions()) {
492 // check if we must post exception events, take uncommon trap if so
493 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
494 // here if should_post_on_exceptions is false
495 // continue on with the normal codegen
496 }
498 // If this particular condition has not yet happened at this
499 // bytecode, then use the uncommon trap mechanism, and allow for
500 // a future recompilation if several traps occur here.
501 // If the throw is hot, try to use a more complicated inline mechanism
502 // which keeps execution inside the compiled code.
503 bool treat_throw_as_hot = false;
504 ciMethodData* md = method()->method_data();
506 if (ProfileTraps) {
507 if (too_many_traps(reason)) {
508 treat_throw_as_hot = true;
509 }
510 // (If there is no MDO at all, assume it is early in
511 // execution, and that any deopts are part of the
512 // startup transient, and don't need to be remembered.)
514 // Also, if there is a local exception handler, treat all throws
515 // as hot if there has been at least one in this method.
516 if (C->trap_count(reason) != 0
517 && method()->method_data()->trap_count(reason) != 0
518 && has_ex_handler()) {
519 treat_throw_as_hot = true;
520 }
521 }
523 // If this throw happens frequently, an uncommon trap might cause
524 // a performance pothole. If there is a local exception handler,
525 // and if this particular bytecode appears to be deoptimizing often,
526 // let us handle the throw inline, with a preconstructed instance.
527 // Note: If the deopt count has blown up, the uncommon trap
528 // runtime is going to flush this nmethod, not matter what.
529 if (treat_throw_as_hot
530 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
531 // If the throw is local, we use a pre-existing instance and
532 // punt on the backtrace. This would lead to a missing backtrace
533 // (a repeat of 4292742) if the backtrace object is ever asked
534 // for its backtrace.
535 // Fixing this remaining case of 4292742 requires some flavor of
536 // escape analysis. Leave that for the future.
537 ciInstance* ex_obj = NULL;
538 switch (reason) {
539 case Deoptimization::Reason_null_check:
540 ex_obj = env()->NullPointerException_instance();
541 break;
542 case Deoptimization::Reason_div0_check:
543 ex_obj = env()->ArithmeticException_instance();
544 break;
545 case Deoptimization::Reason_range_check:
546 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
547 break;
548 case Deoptimization::Reason_class_check:
549 if (java_bc() == Bytecodes::_aastore) {
550 ex_obj = env()->ArrayStoreException_instance();
551 } else {
552 ex_obj = env()->ClassCastException_instance();
553 }
554 break;
555 }
556 if (failing()) { stop(); return; } // exception allocation might fail
557 if (ex_obj != NULL) {
558 // Cheat with a preallocated exception object.
559 if (C->log() != NULL)
560 C->log()->elem("hot_throw preallocated='1' reason='%s'",
561 Deoptimization::trap_reason_name(reason));
562 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
563 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
565 // Clear the detail message of the preallocated exception object.
566 // Weblogic sometimes mutates the detail message of exceptions
567 // using reflection.
568 int offset = java_lang_Throwable::get_detailMessage_offset();
569 const TypePtr* adr_typ = ex_con->add_offset(offset);
571 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
572 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), ex_con, T_OBJECT);
574 add_exception_state(make_exception_state(ex_node));
575 return;
576 }
577 }
579 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
580 // It won't be much cheaper than bailing to the interp., since we'll
581 // have to pass up all the debug-info, and the runtime will have to
582 // create the stack trace.
584 // Usual case: Bail to interpreter.
585 // Reserve the right to recompile if we haven't seen anything yet.
587 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
588 if (treat_throw_as_hot
589 && (method()->method_data()->trap_recompiled_at(bci())
590 || C->too_many_traps(reason))) {
591 // We cannot afford to take more traps here. Suffer in the interpreter.
592 if (C->log() != NULL)
593 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
594 Deoptimization::trap_reason_name(reason),
595 C->trap_count(reason));
596 action = Deoptimization::Action_none;
597 }
599 // "must_throw" prunes the JVM state to include only the stack, if there
600 // are no local exception handlers. This should cut down on register
601 // allocation time and code size, by drastically reducing the number
602 // of in-edges on the call to the uncommon trap.
604 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
605 }
608 //----------------------------PreserveJVMState---------------------------------
609 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
610 debug_only(kit->verify_map());
611 _kit = kit;
612 _map = kit->map(); // preserve the map
613 _sp = kit->sp();
614 kit->set_map(clone_map ? kit->clone_map() : NULL);
615 #ifdef ASSERT
616 _bci = kit->bci();
617 Parse* parser = kit->is_Parse();
618 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
619 _block = block;
620 #endif
621 }
622 PreserveJVMState::~PreserveJVMState() {
623 GraphKit* kit = _kit;
624 #ifdef ASSERT
625 assert(kit->bci() == _bci, "bci must not shift");
626 Parse* parser = kit->is_Parse();
627 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
628 assert(block == _block, "block must not shift");
629 #endif
630 kit->set_map(_map);
631 kit->set_sp(_sp);
632 }
635 //-----------------------------BuildCutout-------------------------------------
636 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
637 : PreserveJVMState(kit)
638 {
639 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
640 SafePointNode* outer_map = _map; // preserved map is caller's
641 SafePointNode* inner_map = kit->map();
642 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
643 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
644 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
645 }
646 BuildCutout::~BuildCutout() {
647 GraphKit* kit = _kit;
648 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
649 }
651 //---------------------------PreserveReexecuteState----------------------------
652 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
653 assert(!kit->stopped(), "must call stopped() before");
654 _kit = kit;
655 _sp = kit->sp();
656 _reexecute = kit->jvms()->_reexecute;
657 }
658 PreserveReexecuteState::~PreserveReexecuteState() {
659 if (_kit->stopped()) return;
660 _kit->jvms()->_reexecute = _reexecute;
661 _kit->set_sp(_sp);
662 }
664 //------------------------------clone_map--------------------------------------
665 // Implementation of PreserveJVMState
666 //
667 // Only clone_map(...) here. If this function is only used in the
668 // PreserveJVMState class we may want to get rid of this extra
669 // function eventually and do it all there.
671 SafePointNode* GraphKit::clone_map() {
672 if (map() == NULL) return NULL;
674 // Clone the memory edge first
675 Node* mem = MergeMemNode::make(C, map()->memory());
676 gvn().set_type_bottom(mem);
678 SafePointNode *clonemap = (SafePointNode*)map()->clone();
679 JVMState* jvms = this->jvms();
680 JVMState* clonejvms = jvms->clone_shallow(C);
681 clonemap->set_memory(mem);
682 clonemap->set_jvms(clonejvms);
683 clonejvms->set_map(clonemap);
684 record_for_igvn(clonemap);
685 gvn().set_type_bottom(clonemap);
686 return clonemap;
687 }
690 //-----------------------------set_map_clone-----------------------------------
691 void GraphKit::set_map_clone(SafePointNode* m) {
692 _map = m;
693 _map = clone_map();
694 _map->set_next_exception(NULL);
695 debug_only(verify_map());
696 }
699 //----------------------------kill_dead_locals---------------------------------
700 // Detect any locals which are known to be dead, and force them to top.
701 void GraphKit::kill_dead_locals() {
702 // Consult the liveness information for the locals. If any
703 // of them are unused, then they can be replaced by top(). This
704 // should help register allocation time and cut down on the size
705 // of the deoptimization information.
707 // This call is made from many of the bytecode handling
708 // subroutines called from the Big Switch in do_one_bytecode.
709 // Every bytecode which might include a slow path is responsible
710 // for killing its dead locals. The more consistent we
711 // are about killing deads, the fewer useless phis will be
712 // constructed for them at various merge points.
714 // bci can be -1 (InvocationEntryBci). We return the entry
715 // liveness for the method.
717 if (method() == NULL || method()->code_size() == 0) {
718 // We are building a graph for a call to a native method.
719 // All locals are live.
720 return;
721 }
723 ResourceMark rm;
725 // Consult the liveness information for the locals. If any
726 // of them are unused, then they can be replaced by top(). This
727 // should help register allocation time and cut down on the size
728 // of the deoptimization information.
729 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
731 int len = (int)live_locals.size();
732 assert(len <= jvms()->loc_size(), "too many live locals");
733 for (int local = 0; local < len; local++) {
734 if (!live_locals.at(local)) {
735 set_local(local, top());
736 }
737 }
738 }
740 #ifdef ASSERT
741 //-------------------------dead_locals_are_killed------------------------------
742 // Return true if all dead locals are set to top in the map.
743 // Used to assert "clean" debug info at various points.
744 bool GraphKit::dead_locals_are_killed() {
745 if (method() == NULL || method()->code_size() == 0) {
746 // No locals need to be dead, so all is as it should be.
747 return true;
748 }
750 // Make sure somebody called kill_dead_locals upstream.
751 ResourceMark rm;
752 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
753 if (jvms->loc_size() == 0) continue; // no locals to consult
754 SafePointNode* map = jvms->map();
755 ciMethod* method = jvms->method();
756 int bci = jvms->bci();
757 if (jvms == this->jvms()) {
758 bci = this->bci(); // it might not yet be synched
759 }
760 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
761 int len = (int)live_locals.size();
762 if (!live_locals.is_valid() || len == 0)
763 // This method is trivial, or is poisoned by a breakpoint.
764 return true;
765 assert(len == jvms->loc_size(), "live map consistent with locals map");
766 for (int local = 0; local < len; local++) {
767 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
768 if (PrintMiscellaneous && (Verbose || WizardMode)) {
769 tty->print_cr("Zombie local %d: ", local);
770 jvms->dump();
771 }
772 return false;
773 }
774 }
775 }
776 return true;
777 }
779 #endif //ASSERT
781 // Helper function for enforcing certain bytecodes to reexecute if
782 // deoptimization happens
783 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
784 ciMethod* cur_method = jvms->method();
785 int cur_bci = jvms->bci();
786 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
787 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
788 return Interpreter::bytecode_should_reexecute(code) ||
789 is_anewarray && code == Bytecodes::_multianewarray;
790 // Reexecute _multianewarray bytecode which was replaced with
791 // sequence of [a]newarray. See Parse::do_multianewarray().
792 //
793 // Note: interpreter should not have it set since this optimization
794 // is limited by dimensions and guarded by flag so in some cases
795 // multianewarray() runtime calls will be generated and
796 // the bytecode should not be reexecutes (stack will not be reset).
797 } else
798 return false;
799 }
801 // Helper function for adding JVMState and debug information to node
802 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
803 // Add the safepoint edges to the call (or other safepoint).
805 // Make sure dead locals are set to top. This
806 // should help register allocation time and cut down on the size
807 // of the deoptimization information.
808 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
810 // Walk the inline list to fill in the correct set of JVMState's
811 // Also fill in the associated edges for each JVMState.
813 JVMState* youngest_jvms = sync_jvms();
815 // If we are guaranteed to throw, we can prune everything but the
816 // input to the current bytecode.
817 bool can_prune_locals = false;
818 uint stack_slots_not_pruned = 0;
819 int inputs = 0, depth = 0;
820 if (must_throw) {
821 assert(method() == youngest_jvms->method(), "sanity");
822 if (compute_stack_effects(inputs, depth)) {
823 can_prune_locals = true;
824 stack_slots_not_pruned = inputs;
825 }
826 }
828 if (env()->jvmti_can_access_local_variables()) {
829 // At any safepoint, this method can get breakpointed, which would
830 // then require an immediate deoptimization.
831 can_prune_locals = false; // do not prune locals
832 stack_slots_not_pruned = 0;
833 }
835 // do not scribble on the input jvms
836 JVMState* out_jvms = youngest_jvms->clone_deep(C);
837 call->set_jvms(out_jvms); // Start jvms list for call node
839 // For a known set of bytecodes, the interpreter should reexecute them if
840 // deoptimization happens. We set the reexecute state for them here
841 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
842 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
843 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
844 }
846 // Presize the call:
847 debug_only(uint non_debug_edges = call->req());
848 call->add_req_batch(top(), youngest_jvms->debug_depth());
849 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
851 // Set up edges so that the call looks like this:
852 // Call [state:] ctl io mem fptr retadr
853 // [parms:] parm0 ... parmN
854 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
855 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
856 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
857 // Note that caller debug info precedes callee debug info.
859 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
860 uint debug_ptr = call->req();
862 // Loop over the map input edges associated with jvms, add them
863 // to the call node, & reset all offsets to match call node array.
864 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
865 uint debug_end = debug_ptr;
866 uint debug_start = debug_ptr - in_jvms->debug_size();
867 debug_ptr = debug_start; // back up the ptr
869 uint p = debug_start; // walks forward in [debug_start, debug_end)
870 uint j, k, l;
871 SafePointNode* in_map = in_jvms->map();
872 out_jvms->set_map(call);
874 if (can_prune_locals) {
875 assert(in_jvms->method() == out_jvms->method(), "sanity");
876 // If the current throw can reach an exception handler in this JVMS,
877 // then we must keep everything live that can reach that handler.
878 // As a quick and dirty approximation, we look for any handlers at all.
879 if (in_jvms->method()->has_exception_handlers()) {
880 can_prune_locals = false;
881 }
882 }
884 // Add the Locals
885 k = in_jvms->locoff();
886 l = in_jvms->loc_size();
887 out_jvms->set_locoff(p);
888 if (!can_prune_locals) {
889 for (j = 0; j < l; j++)
890 call->set_req(p++, in_map->in(k+j));
891 } else {
892 p += l; // already set to top above by add_req_batch
893 }
895 // Add the Expression Stack
896 k = in_jvms->stkoff();
897 l = in_jvms->sp();
898 out_jvms->set_stkoff(p);
899 if (!can_prune_locals) {
900 for (j = 0; j < l; j++)
901 call->set_req(p++, in_map->in(k+j));
902 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
903 // Divide stack into {S0,...,S1}, where S0 is set to top.
904 uint s1 = stack_slots_not_pruned;
905 stack_slots_not_pruned = 0; // for next iteration
906 if (s1 > l) s1 = l;
907 uint s0 = l - s1;
908 p += s0; // skip the tops preinstalled by add_req_batch
909 for (j = s0; j < l; j++)
910 call->set_req(p++, in_map->in(k+j));
911 } else {
912 p += l; // already set to top above by add_req_batch
913 }
915 // Add the Monitors
916 k = in_jvms->monoff();
917 l = in_jvms->mon_size();
918 out_jvms->set_monoff(p);
919 for (j = 0; j < l; j++)
920 call->set_req(p++, in_map->in(k+j));
922 // Copy any scalar object fields.
923 k = in_jvms->scloff();
924 l = in_jvms->scl_size();
925 out_jvms->set_scloff(p);
926 for (j = 0; j < l; j++)
927 call->set_req(p++, in_map->in(k+j));
929 // Finish the new jvms.
930 out_jvms->set_endoff(p);
932 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
933 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
934 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
935 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
936 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
937 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
939 // Update the two tail pointers in parallel.
940 out_jvms = out_jvms->caller();
941 in_jvms = in_jvms->caller();
942 }
944 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
946 // Test the correctness of JVMState::debug_xxx accessors:
947 assert(call->jvms()->debug_start() == non_debug_edges, "");
948 assert(call->jvms()->debug_end() == call->req(), "");
949 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
950 }
952 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
953 Bytecodes::Code code = java_bc();
954 if (code == Bytecodes::_wide) {
955 code = method()->java_code_at_bci(bci() + 1);
956 }
958 BasicType rtype = T_ILLEGAL;
959 int rsize = 0;
961 if (code != Bytecodes::_illegal) {
962 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
963 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
964 if (rtype < T_CONFLICT)
965 rsize = type2size[rtype];
966 }
968 switch (code) {
969 case Bytecodes::_illegal:
970 return false;
972 case Bytecodes::_ldc:
973 case Bytecodes::_ldc_w:
974 case Bytecodes::_ldc2_w:
975 inputs = 0;
976 break;
978 case Bytecodes::_dup: inputs = 1; break;
979 case Bytecodes::_dup_x1: inputs = 2; break;
980 case Bytecodes::_dup_x2: inputs = 3; break;
981 case Bytecodes::_dup2: inputs = 2; break;
982 case Bytecodes::_dup2_x1: inputs = 3; break;
983 case Bytecodes::_dup2_x2: inputs = 4; break;
984 case Bytecodes::_swap: inputs = 2; break;
985 case Bytecodes::_arraylength: inputs = 1; break;
987 case Bytecodes::_getstatic:
988 case Bytecodes::_putstatic:
989 case Bytecodes::_getfield:
990 case Bytecodes::_putfield:
991 {
992 bool is_get = (depth >= 0), is_static = (depth & 1);
993 bool ignore;
994 ciBytecodeStream iter(method());
995 iter.reset_to_bci(bci());
996 iter.next();
997 ciField* field = iter.get_field(ignore);
998 int size = field->type()->size();
999 inputs = (is_static ? 0 : 1);
1000 if (is_get) {
1001 depth = size - inputs;
1002 } else {
1003 inputs += size; // putxxx pops the value from the stack
1004 depth = - inputs;
1005 }
1006 }
1007 break;
1009 case Bytecodes::_invokevirtual:
1010 case Bytecodes::_invokespecial:
1011 case Bytecodes::_invokestatic:
1012 case Bytecodes::_invokedynamic:
1013 case Bytecodes::_invokeinterface:
1014 {
1015 bool ignore;
1016 ciBytecodeStream iter(method());
1017 iter.reset_to_bci(bci());
1018 iter.next();
1019 ciMethod* method = iter.get_method(ignore);
1020 inputs = method->arg_size_no_receiver();
1021 // Add a receiver argument, maybe:
1022 if (code != Bytecodes::_invokestatic &&
1023 code != Bytecodes::_invokedynamic)
1024 inputs += 1;
1025 // (Do not use ciMethod::arg_size(), because
1026 // it might be an unloaded method, which doesn't
1027 // know whether it is static or not.)
1028 int size = method->return_type()->size();
1029 depth = size - inputs;
1030 }
1031 break;
1033 case Bytecodes::_multianewarray:
1034 {
1035 ciBytecodeStream iter(method());
1036 iter.reset_to_bci(bci());
1037 iter.next();
1038 inputs = iter.get_dimensions();
1039 assert(rsize == 1, "");
1040 depth = rsize - inputs;
1041 }
1042 break;
1044 case Bytecodes::_ireturn:
1045 case Bytecodes::_lreturn:
1046 case Bytecodes::_freturn:
1047 case Bytecodes::_dreturn:
1048 case Bytecodes::_areturn:
1049 assert(rsize = -depth, "");
1050 inputs = rsize;
1051 break;
1053 case Bytecodes::_jsr:
1054 case Bytecodes::_jsr_w:
1055 inputs = 0;
1056 depth = 1; // S.B. depth=1, not zero
1057 break;
1059 default:
1060 // bytecode produces a typed result
1061 inputs = rsize - depth;
1062 assert(inputs >= 0, "");
1063 break;
1064 }
1066 #ifdef ASSERT
1067 // spot check
1068 int outputs = depth + inputs;
1069 assert(outputs >= 0, "sanity");
1070 switch (code) {
1071 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1072 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1073 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1074 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1075 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1076 }
1077 #endif //ASSERT
1079 return true;
1080 }
1084 //------------------------------basic_plus_adr---------------------------------
1085 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1086 // short-circuit a common case
1087 if (offset == intcon(0)) return ptr;
1088 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1089 }
1091 Node* GraphKit::ConvI2L(Node* offset) {
1092 // short-circuit a common case
1093 jint offset_con = find_int_con(offset, Type::OffsetBot);
1094 if (offset_con != Type::OffsetBot) {
1095 return longcon((long) offset_con);
1096 }
1097 return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1098 }
1099 Node* GraphKit::ConvL2I(Node* offset) {
1100 // short-circuit a common case
1101 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1102 if (offset_con != (jlong)Type::OffsetBot) {
1103 return intcon((int) offset_con);
1104 }
1105 return _gvn.transform( new (C, 2) ConvL2INode(offset));
1106 }
1108 //-------------------------load_object_klass-----------------------------------
1109 Node* GraphKit::load_object_klass(Node* obj) {
1110 // Special-case a fresh allocation to avoid building nodes:
1111 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1112 if (akls != NULL) return akls;
1113 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1114 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1115 }
1117 //-------------------------load_array_length-----------------------------------
1118 Node* GraphKit::load_array_length(Node* array) {
1119 // Special-case a fresh allocation to avoid building nodes:
1120 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1121 Node *alen;
1122 if (alloc == NULL) {
1123 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1124 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1125 } else {
1126 alen = alloc->Ideal_length();
1127 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1128 if (ccast != alen) {
1129 alen = _gvn.transform(ccast);
1130 }
1131 }
1132 return alen;
1133 }
1135 //------------------------------do_null_check----------------------------------
1136 // Helper function to do a NULL pointer check. Returned value is
1137 // the incoming address with NULL casted away. You are allowed to use the
1138 // not-null value only if you are control dependent on the test.
1139 extern int explicit_null_checks_inserted,
1140 explicit_null_checks_elided;
1141 Node* GraphKit::null_check_common(Node* value, BasicType type,
1142 // optional arguments for variations:
1143 bool assert_null,
1144 Node* *null_control) {
1145 assert(!assert_null || null_control == NULL, "not both at once");
1146 if (stopped()) return top();
1147 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1148 // For some performance testing, we may wish to suppress null checking.
1149 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1150 return value;
1151 }
1152 explicit_null_checks_inserted++;
1154 // Construct NULL check
1155 Node *chk = NULL;
1156 switch(type) {
1157 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1158 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1159 case T_ARRAY : // fall through
1160 type = T_OBJECT; // simplify further tests
1161 case T_OBJECT : {
1162 const Type *t = _gvn.type( value );
1164 const TypeOopPtr* tp = t->isa_oopptr();
1165 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1166 // Only for do_null_check, not any of its siblings:
1167 && !assert_null && null_control == NULL) {
1168 // Usually, any field access or invocation on an unloaded oop type
1169 // will simply fail to link, since the statically linked class is
1170 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1171 // the static class is loaded but the sharper oop type is not.
1172 // Rather than checking for this obscure case in lots of places,
1173 // we simply observe that a null check on an unloaded class
1174 // will always be followed by a nonsense operation, so we
1175 // can just issue the uncommon trap here.
1176 // Our access to the unloaded class will only be correct
1177 // after it has been loaded and initialized, which requires
1178 // a trip through the interpreter.
1179 #ifndef PRODUCT
1180 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1181 #endif
1182 uncommon_trap(Deoptimization::Reason_unloaded,
1183 Deoptimization::Action_reinterpret,
1184 tp->klass(), "!loaded");
1185 return top();
1186 }
1188 if (assert_null) {
1189 // See if the type is contained in NULL_PTR.
1190 // If so, then the value is already null.
1191 if (t->higher_equal(TypePtr::NULL_PTR)) {
1192 explicit_null_checks_elided++;
1193 return value; // Elided null assert quickly!
1194 }
1195 } else {
1196 // See if mixing in the NULL pointer changes type.
1197 // If so, then the NULL pointer was not allowed in the original
1198 // type. In other words, "value" was not-null.
1199 if (t->meet(TypePtr::NULL_PTR) != t) {
1200 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1201 explicit_null_checks_elided++;
1202 return value; // Elided null check quickly!
1203 }
1204 }
1205 chk = new (C, 3) CmpPNode( value, null() );
1206 break;
1207 }
1209 default : ShouldNotReachHere();
1210 }
1211 assert(chk != NULL, "sanity check");
1212 chk = _gvn.transform(chk);
1214 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1215 BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1216 Node *tst = _gvn.transform( btst );
1218 //-----------
1219 // if peephole optimizations occurred, a prior test existed.
1220 // If a prior test existed, maybe it dominates as we can avoid this test.
1221 if (tst != btst && type == T_OBJECT) {
1222 // At this point we want to scan up the CFG to see if we can
1223 // find an identical test (and so avoid this test altogether).
1224 Node *cfg = control();
1225 int depth = 0;
1226 while( depth < 16 ) { // Limit search depth for speed
1227 if( cfg->Opcode() == Op_IfTrue &&
1228 cfg->in(0)->in(1) == tst ) {
1229 // Found prior test. Use "cast_not_null" to construct an identical
1230 // CastPP (and hence hash to) as already exists for the prior test.
1231 // Return that casted value.
1232 if (assert_null) {
1233 replace_in_map(value, null());
1234 return null(); // do not issue the redundant test
1235 }
1236 Node *oldcontrol = control();
1237 set_control(cfg);
1238 Node *res = cast_not_null(value);
1239 set_control(oldcontrol);
1240 explicit_null_checks_elided++;
1241 return res;
1242 }
1243 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1244 if (cfg == NULL) break; // Quit at region nodes
1245 depth++;
1246 }
1247 }
1249 //-----------
1250 // Branch to failure if null
1251 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1252 Deoptimization::DeoptReason reason;
1253 if (assert_null)
1254 reason = Deoptimization::Reason_null_assert;
1255 else if (type == T_OBJECT)
1256 reason = Deoptimization::Reason_null_check;
1257 else
1258 reason = Deoptimization::Reason_div0_check;
1260 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1261 // ciMethodData::has_trap_at will return a conservative -1 if any
1262 // must-be-null assertion has failed. This could cause performance
1263 // problems for a method after its first do_null_assert failure.
1264 // Consider using 'Reason_class_check' instead?
1266 // To cause an implicit null check, we set the not-null probability
1267 // to the maximum (PROB_MAX). For an explicit check the probability
1268 // is set to a smaller value.
1269 if (null_control != NULL || too_many_traps(reason)) {
1270 // probability is less likely
1271 ok_prob = PROB_LIKELY_MAG(3);
1272 } else if (!assert_null &&
1273 (ImplicitNullCheckThreshold > 0) &&
1274 method() != NULL &&
1275 (method()->method_data()->trap_count(reason)
1276 >= (uint)ImplicitNullCheckThreshold)) {
1277 ok_prob = PROB_LIKELY_MAG(3);
1278 }
1280 if (null_control != NULL) {
1281 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1282 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1283 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
1284 if (null_true == top())
1285 explicit_null_checks_elided++;
1286 (*null_control) = null_true;
1287 } else {
1288 BuildCutout unless(this, tst, ok_prob);
1289 // Check for optimizer eliding test at parse time
1290 if (stopped()) {
1291 // Failure not possible; do not bother making uncommon trap.
1292 explicit_null_checks_elided++;
1293 } else if (assert_null) {
1294 uncommon_trap(reason,
1295 Deoptimization::Action_make_not_entrant,
1296 NULL, "assert_null");
1297 } else {
1298 replace_in_map(value, zerocon(type));
1299 builtin_throw(reason);
1300 }
1301 }
1303 // Must throw exception, fall-thru not possible?
1304 if (stopped()) {
1305 return top(); // No result
1306 }
1308 if (assert_null) {
1309 // Cast obj to null on this path.
1310 replace_in_map(value, zerocon(type));
1311 return zerocon(type);
1312 }
1314 // Cast obj to not-null on this path, if there is no null_control.
1315 // (If there is a null_control, a non-null value may come back to haunt us.)
1316 if (type == T_OBJECT) {
1317 Node* cast = cast_not_null(value, false);
1318 if (null_control == NULL || (*null_control) == top())
1319 replace_in_map(value, cast);
1320 value = cast;
1321 }
1323 return value;
1324 }
1327 //------------------------------cast_not_null----------------------------------
1328 // Cast obj to not-null on this path
1329 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1330 const Type *t = _gvn.type(obj);
1331 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1332 // Object is already not-null?
1333 if( t == t_not_null ) return obj;
1335 Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1336 cast->init_req(0, control());
1337 cast = _gvn.transform( cast );
1339 // Scan for instances of 'obj' in the current JVM mapping.
1340 // These instances are known to be not-null after the test.
1341 if (do_replace_in_map)
1342 replace_in_map(obj, cast);
1344 return cast; // Return casted value
1345 }
1348 //--------------------------replace_in_map-------------------------------------
1349 void GraphKit::replace_in_map(Node* old, Node* neww) {
1350 this->map()->replace_edge(old, neww);
1352 // Note: This operation potentially replaces any edge
1353 // on the map. This includes locals, stack, and monitors
1354 // of the current (innermost) JVM state.
1356 // We can consider replacing in caller maps.
1357 // The idea would be that an inlined function's null checks
1358 // can be shared with the entire inlining tree.
1359 // The expense of doing this is that the PreserveJVMState class
1360 // would have to preserve caller states too, with a deep copy.
1361 }
1365 //=============================================================================
1366 //--------------------------------memory---------------------------------------
1367 Node* GraphKit::memory(uint alias_idx) {
1368 MergeMemNode* mem = merged_memory();
1369 Node* p = mem->memory_at(alias_idx);
1370 _gvn.set_type(p, Type::MEMORY); // must be mapped
1371 return p;
1372 }
1374 //-----------------------------reset_memory------------------------------------
1375 Node* GraphKit::reset_memory() {
1376 Node* mem = map()->memory();
1377 // do not use this node for any more parsing!
1378 debug_only( map()->set_memory((Node*)NULL) );
1379 return _gvn.transform( mem );
1380 }
1382 //------------------------------set_all_memory---------------------------------
1383 void GraphKit::set_all_memory(Node* newmem) {
1384 Node* mergemem = MergeMemNode::make(C, newmem);
1385 gvn().set_type_bottom(mergemem);
1386 map()->set_memory(mergemem);
1387 }
1389 //------------------------------set_all_memory_call----------------------------
1390 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1391 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1392 set_all_memory(newmem);
1393 }
1395 //=============================================================================
1396 //
1397 // parser factory methods for MemNodes
1398 //
1399 // These are layered on top of the factory methods in LoadNode and StoreNode,
1400 // and integrate with the parser's memory state and _gvn engine.
1401 //
1403 // factory methods in "int adr_idx"
1404 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1405 int adr_idx,
1406 bool require_atomic_access) {
1407 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1408 const TypePtr* adr_type = NULL; // debug-mode-only argument
1409 debug_only(adr_type = C->get_adr_type(adr_idx));
1410 Node* mem = memory(adr_idx);
1411 Node* ld;
1412 if (require_atomic_access && bt == T_LONG) {
1413 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1414 } else {
1415 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1416 }
1417 return _gvn.transform(ld);
1418 }
1420 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1421 int adr_idx,
1422 bool require_atomic_access) {
1423 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1424 const TypePtr* adr_type = NULL;
1425 debug_only(adr_type = C->get_adr_type(adr_idx));
1426 Node *mem = memory(adr_idx);
1427 Node* st;
1428 if (require_atomic_access && bt == T_LONG) {
1429 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1430 } else {
1431 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1432 }
1433 st = _gvn.transform(st);
1434 set_memory(st, adr_idx);
1435 // Back-to-back stores can only remove intermediate store with DU info
1436 // so push on worklist for optimizer.
1437 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1438 record_for_igvn(st);
1440 return st;
1441 }
1444 void GraphKit::pre_barrier(Node* ctl,
1445 Node* obj,
1446 Node* adr,
1447 uint adr_idx,
1448 Node* val,
1449 const TypeOopPtr* val_type,
1450 BasicType bt) {
1451 BarrierSet* bs = Universe::heap()->barrier_set();
1452 set_control(ctl);
1453 switch (bs->kind()) {
1454 case BarrierSet::G1SATBCT:
1455 case BarrierSet::G1SATBCTLogging:
1456 g1_write_barrier_pre(obj, adr, adr_idx, val, val_type, bt);
1457 break;
1459 case BarrierSet::CardTableModRef:
1460 case BarrierSet::CardTableExtension:
1461 case BarrierSet::ModRef:
1462 break;
1464 case BarrierSet::Other:
1465 default :
1466 ShouldNotReachHere();
1468 }
1469 }
1471 void GraphKit::post_barrier(Node* ctl,
1472 Node* store,
1473 Node* obj,
1474 Node* adr,
1475 uint adr_idx,
1476 Node* val,
1477 BasicType bt,
1478 bool use_precise) {
1479 BarrierSet* bs = Universe::heap()->barrier_set();
1480 set_control(ctl);
1481 switch (bs->kind()) {
1482 case BarrierSet::G1SATBCT:
1483 case BarrierSet::G1SATBCTLogging:
1484 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1485 break;
1487 case BarrierSet::CardTableModRef:
1488 case BarrierSet::CardTableExtension:
1489 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1490 break;
1492 case BarrierSet::ModRef:
1493 break;
1495 case BarrierSet::Other:
1496 default :
1497 ShouldNotReachHere();
1499 }
1500 }
1502 Node* GraphKit::store_oop(Node* ctl,
1503 Node* obj,
1504 Node* adr,
1505 const TypePtr* adr_type,
1506 Node* val,
1507 const TypeOopPtr* val_type,
1508 BasicType bt,
1509 bool use_precise) {
1511 set_control(ctl);
1512 if (stopped()) return top(); // Dead path ?
1514 assert(bt == T_OBJECT, "sanity");
1515 assert(val != NULL, "not dead path");
1516 uint adr_idx = C->get_alias_index(adr_type);
1517 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1519 pre_barrier(control(), obj, adr, adr_idx, val, val_type, bt);
1520 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1521 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1522 return store;
1523 }
1525 // Could be an array or object we don't know at compile time (unsafe ref.)
1526 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1527 Node* obj, // containing obj
1528 Node* adr, // actual adress to store val at
1529 const TypePtr* adr_type,
1530 Node* val,
1531 BasicType bt) {
1532 Compile::AliasType* at = C->alias_type(adr_type);
1533 const TypeOopPtr* val_type = NULL;
1534 if (adr_type->isa_instptr()) {
1535 if (at->field() != NULL) {
1536 // known field. This code is a copy of the do_put_xxx logic.
1537 ciField* field = at->field();
1538 if (!field->type()->is_loaded()) {
1539 val_type = TypeInstPtr::BOTTOM;
1540 } else {
1541 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1542 }
1543 }
1544 } else if (adr_type->isa_aryptr()) {
1545 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1546 }
1547 if (val_type == NULL) {
1548 val_type = TypeInstPtr::BOTTOM;
1549 }
1550 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1551 }
1554 //-------------------------array_element_address-------------------------
1555 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1556 const TypeInt* sizetype) {
1557 uint shift = exact_log2(type2aelembytes(elembt));
1558 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1560 // short-circuit a common case (saves lots of confusing waste motion)
1561 jint idx_con = find_int_con(idx, -1);
1562 if (idx_con >= 0) {
1563 intptr_t offset = header + ((intptr_t)idx_con << shift);
1564 return basic_plus_adr(ary, offset);
1565 }
1567 // must be correct type for alignment purposes
1568 Node* base = basic_plus_adr(ary, header);
1569 #ifdef _LP64
1570 // The scaled index operand to AddP must be a clean 64-bit value.
1571 // Java allows a 32-bit int to be incremented to a negative
1572 // value, which appears in a 64-bit register as a large
1573 // positive number. Using that large positive number as an
1574 // operand in pointer arithmetic has bad consequences.
1575 // On the other hand, 32-bit overflow is rare, and the possibility
1576 // can often be excluded, if we annotate the ConvI2L node with
1577 // a type assertion that its value is known to be a small positive
1578 // number. (The prior range check has ensured this.)
1579 // This assertion is used by ConvI2LNode::Ideal.
1580 int index_max = max_jint - 1; // array size is max_jint, index is one less
1581 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1582 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1583 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1584 #endif
1585 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1586 return basic_plus_adr(ary, base, scale);
1587 }
1589 //-------------------------load_array_element-------------------------
1590 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1591 const Type* elemtype = arytype->elem();
1592 BasicType elembt = elemtype->array_element_basic_type();
1593 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1594 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1595 return ld;
1596 }
1598 //-------------------------set_arguments_for_java_call-------------------------
1599 // Arguments (pre-popped from the stack) are taken from the JVMS.
1600 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1601 // Add the call arguments:
1602 uint nargs = call->method()->arg_size();
1603 for (uint i = 0; i < nargs; i++) {
1604 Node* arg = argument(i);
1605 call->init_req(i + TypeFunc::Parms, arg);
1606 }
1607 }
1609 //---------------------------set_edges_for_java_call---------------------------
1610 // Connect a newly created call into the current JVMS.
1611 // A return value node (if any) is returned from set_edges_for_java_call.
1612 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1614 // Add the predefined inputs:
1615 call->init_req( TypeFunc::Control, control() );
1616 call->init_req( TypeFunc::I_O , i_o() );
1617 call->init_req( TypeFunc::Memory , reset_memory() );
1618 call->init_req( TypeFunc::FramePtr, frameptr() );
1619 call->init_req( TypeFunc::ReturnAdr, top() );
1621 add_safepoint_edges(call, must_throw);
1623 Node* xcall = _gvn.transform(call);
1625 if (xcall == top()) {
1626 set_control(top());
1627 return;
1628 }
1629 assert(xcall == call, "call identity is stable");
1631 // Re-use the current map to produce the result.
1633 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1634 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1635 set_all_memory_call(xcall, separate_io_proj);
1637 //return xcall; // no need, caller already has it
1638 }
1640 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1641 if (stopped()) return top(); // maybe the call folded up?
1643 // Capture the return value, if any.
1644 Node* ret;
1645 if (call->method() == NULL ||
1646 call->method()->return_type()->basic_type() == T_VOID)
1647 ret = top();
1648 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1650 // Note: Since any out-of-line call can produce an exception,
1651 // we always insert an I_O projection from the call into the result.
1653 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1655 if (separate_io_proj) {
1656 // The caller requested separate projections be used by the fall
1657 // through and exceptional paths, so replace the projections for
1658 // the fall through path.
1659 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) ));
1660 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ));
1661 }
1662 return ret;
1663 }
1665 //--------------------set_predefined_input_for_runtime_call--------------------
1666 // Reading and setting the memory state is way conservative here.
1667 // The real problem is that I am not doing real Type analysis on memory,
1668 // so I cannot distinguish card mark stores from other stores. Across a GC
1669 // point the Store Barrier and the card mark memory has to agree. I cannot
1670 // have a card mark store and its barrier split across the GC point from
1671 // either above or below. Here I get that to happen by reading ALL of memory.
1672 // A better answer would be to separate out card marks from other memory.
1673 // For now, return the input memory state, so that it can be reused
1674 // after the call, if this call has restricted memory effects.
1675 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1676 // Set fixed predefined input arguments
1677 Node* memory = reset_memory();
1678 call->init_req( TypeFunc::Control, control() );
1679 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1680 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1681 call->init_req( TypeFunc::FramePtr, frameptr() );
1682 call->init_req( TypeFunc::ReturnAdr, top() );
1683 return memory;
1684 }
1686 //-------------------set_predefined_output_for_runtime_call--------------------
1687 // Set control and memory (not i_o) from the call.
1688 // If keep_mem is not NULL, use it for the output state,
1689 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1690 // If hook_mem is NULL, this call produces no memory effects at all.
1691 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1692 // then only that memory slice is taken from the call.
1693 // In the last case, we must put an appropriate memory barrier before
1694 // the call, so as to create the correct anti-dependencies on loads
1695 // preceding the call.
1696 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1697 Node* keep_mem,
1698 const TypePtr* hook_mem) {
1699 // no i/o
1700 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1701 if (keep_mem) {
1702 // First clone the existing memory state
1703 set_all_memory(keep_mem);
1704 if (hook_mem != NULL) {
1705 // Make memory for the call
1706 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1707 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1708 // We also use hook_mem to extract specific effects from arraycopy stubs.
1709 set_memory(mem, hook_mem);
1710 }
1711 // ...else the call has NO memory effects.
1713 // Make sure the call advertises its memory effects precisely.
1714 // This lets us build accurate anti-dependences in gcm.cpp.
1715 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1716 "call node must be constructed correctly");
1717 } else {
1718 assert(hook_mem == NULL, "");
1719 // This is not a "slow path" call; all memory comes from the call.
1720 set_all_memory_call(call);
1721 }
1722 }
1725 // Replace the call with the current state of the kit.
1726 void GraphKit::replace_call(CallNode* call, Node* result) {
1727 JVMState* ejvms = NULL;
1728 if (has_exceptions()) {
1729 ejvms = transfer_exceptions_into_jvms();
1730 }
1732 SafePointNode* final_state = stop();
1734 // Find all the needed outputs of this call
1735 CallProjections callprojs;
1736 call->extract_projections(&callprojs, true);
1738 // Replace all the old call edges with the edges from the inlining result
1739 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control));
1740 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory));
1741 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O));
1742 Node* final_mem = final_state->in(TypeFunc::Memory);
1744 // Replace the result with the new result if it exists and is used
1745 if (callprojs.resproj != NULL && result != NULL) {
1746 C->gvn_replace_by(callprojs.resproj, result);
1747 }
1749 if (ejvms == NULL) {
1750 // No exception edges to simply kill off those paths
1751 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1752 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1753 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1755 // Replace the old exception object with top
1756 if (callprojs.exobj != NULL) {
1757 C->gvn_replace_by(callprojs.exobj, C->top());
1758 }
1759 } else {
1760 GraphKit ekit(ejvms);
1762 // Load my combined exception state into the kit, with all phis transformed:
1763 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1765 Node* ex_oop = ekit.use_exception_state(ex_map);
1767 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1768 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1769 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1771 // Replace the old exception object with the newly created one
1772 if (callprojs.exobj != NULL) {
1773 C->gvn_replace_by(callprojs.exobj, ex_oop);
1774 }
1775 }
1777 // Disconnect the call from the graph
1778 call->disconnect_inputs(NULL);
1779 C->gvn_replace_by(call, C->top());
1781 // Clean up any MergeMems that feed other MergeMems since the
1782 // optimizer doesn't like that.
1783 if (final_mem->is_MergeMem()) {
1784 Node_List wl;
1785 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1786 Node* m = i.get();
1787 if (m->is_MergeMem() && !wl.contains(m)) {
1788 wl.push(m);
1789 }
1790 }
1791 while (wl.size() > 0) {
1792 _gvn.transform(wl.pop());
1793 }
1794 }
1795 }
1798 //------------------------------increment_counter------------------------------
1799 // for statistics: increment a VM counter by 1
1801 void GraphKit::increment_counter(address counter_addr) {
1802 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1803 increment_counter(adr1);
1804 }
1806 void GraphKit::increment_counter(Node* counter_addr) {
1807 int adr_type = Compile::AliasIdxRaw;
1808 Node* ctrl = control();
1809 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type);
1810 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1811 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type );
1812 }
1815 //------------------------------uncommon_trap----------------------------------
1816 // Bail out to the interpreter in mid-method. Implemented by calling the
1817 // uncommon_trap blob. This helper function inserts a runtime call with the
1818 // right debug info.
1819 void GraphKit::uncommon_trap(int trap_request,
1820 ciKlass* klass, const char* comment,
1821 bool must_throw,
1822 bool keep_exact_action) {
1823 if (failing()) stop();
1824 if (stopped()) return; // trap reachable?
1826 // Note: If ProfileTraps is true, and if a deopt. actually
1827 // occurs here, the runtime will make sure an MDO exists. There is
1828 // no need to call method()->build_method_data() at this point.
1830 #ifdef ASSERT
1831 if (!must_throw) {
1832 // Make sure the stack has at least enough depth to execute
1833 // the current bytecode.
1834 int inputs, ignore;
1835 if (compute_stack_effects(inputs, ignore)) {
1836 assert(sp() >= inputs, "must have enough JVMS stack to execute");
1837 // It is a frequent error in library_call.cpp to issue an
1838 // uncommon trap with the _sp value already popped.
1839 }
1840 }
1841 #endif
1843 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1844 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1846 switch (action) {
1847 case Deoptimization::Action_maybe_recompile:
1848 case Deoptimization::Action_reinterpret:
1849 // Temporary fix for 6529811 to allow virtual calls to be sure they
1850 // get the chance to go from mono->bi->mega
1851 if (!keep_exact_action &&
1852 Deoptimization::trap_request_index(trap_request) < 0 &&
1853 too_many_recompiles(reason)) {
1854 // This BCI is causing too many recompilations.
1855 action = Deoptimization::Action_none;
1856 trap_request = Deoptimization::make_trap_request(reason, action);
1857 } else {
1858 C->set_trap_can_recompile(true);
1859 }
1860 break;
1861 case Deoptimization::Action_make_not_entrant:
1862 C->set_trap_can_recompile(true);
1863 break;
1864 #ifdef ASSERT
1865 case Deoptimization::Action_none:
1866 case Deoptimization::Action_make_not_compilable:
1867 break;
1868 default:
1869 assert(false, "bad action");
1870 #endif
1871 }
1873 if (TraceOptoParse) {
1874 char buf[100];
1875 tty->print_cr("Uncommon trap %s at bci:%d",
1876 Deoptimization::format_trap_request(buf, sizeof(buf),
1877 trap_request), bci());
1878 }
1880 CompileLog* log = C->log();
1881 if (log != NULL) {
1882 int kid = (klass == NULL)? -1: log->identify(klass);
1883 log->begin_elem("uncommon_trap bci='%d'", bci());
1884 char buf[100];
1885 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1886 trap_request));
1887 if (kid >= 0) log->print(" klass='%d'", kid);
1888 if (comment != NULL) log->print(" comment='%s'", comment);
1889 log->end_elem();
1890 }
1892 // Make sure any guarding test views this path as very unlikely
1893 Node *i0 = control()->in(0);
1894 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1895 IfNode *iff = i0->as_If();
1896 float f = iff->_prob; // Get prob
1897 if (control()->Opcode() == Op_IfTrue) {
1898 if (f > PROB_UNLIKELY_MAG(4))
1899 iff->_prob = PROB_MIN;
1900 } else {
1901 if (f < PROB_LIKELY_MAG(4))
1902 iff->_prob = PROB_MAX;
1903 }
1904 }
1906 // Clear out dead values from the debug info.
1907 kill_dead_locals();
1909 // Now insert the uncommon trap subroutine call
1910 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
1911 const TypePtr* no_memory_effects = NULL;
1912 // Pass the index of the class to be loaded
1913 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1914 (must_throw ? RC_MUST_THROW : 0),
1915 OptoRuntime::uncommon_trap_Type(),
1916 call_addr, "uncommon_trap", no_memory_effects,
1917 intcon(trap_request));
1918 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1919 "must extract request correctly from the graph");
1920 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1922 call->set_req(TypeFunc::ReturnAdr, returnadr());
1923 // The debug info is the only real input to this call.
1925 // Halt-and-catch fire here. The above call should never return!
1926 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1927 _gvn.set_type_bottom(halt);
1928 root()->add_req(halt);
1930 stop_and_kill_map();
1931 }
1934 //--------------------------just_allocated_object------------------------------
1935 // Report the object that was just allocated.
1936 // It must be the case that there are no intervening safepoints.
1937 // We use this to determine if an object is so "fresh" that
1938 // it does not require card marks.
1939 Node* GraphKit::just_allocated_object(Node* current_control) {
1940 if (C->recent_alloc_ctl() == current_control)
1941 return C->recent_alloc_obj();
1942 return NULL;
1943 }
1946 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1947 // (Note: TypeFunc::make has a cache that makes this fast.)
1948 const TypeFunc* tf = TypeFunc::make(dest_method);
1949 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
1950 for (int j = 0; j < nargs; j++) {
1951 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1952 if( targ->basic_type() == T_DOUBLE ) {
1953 // If any parameters are doubles, they must be rounded before
1954 // the call, dstore_rounding does gvn.transform
1955 Node *arg = argument(j);
1956 arg = dstore_rounding(arg);
1957 set_argument(j, arg);
1958 }
1959 }
1960 }
1962 void GraphKit::round_double_result(ciMethod* dest_method) {
1963 // A non-strict method may return a double value which has an extended
1964 // exponent, but this must not be visible in a caller which is 'strict'
1965 // If a strict caller invokes a non-strict callee, round a double result
1967 BasicType result_type = dest_method->return_type()->basic_type();
1968 assert( method() != NULL, "must have caller context");
1969 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
1970 // Destination method's return value is on top of stack
1971 // dstore_rounding() does gvn.transform
1972 Node *result = pop_pair();
1973 result = dstore_rounding(result);
1974 push_pair(result);
1975 }
1976 }
1978 // rounding for strict float precision conformance
1979 Node* GraphKit::precision_rounding(Node* n) {
1980 return UseStrictFP && _method->flags().is_strict()
1981 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
1982 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
1983 : n;
1984 }
1986 // rounding for strict double precision conformance
1987 Node* GraphKit::dprecision_rounding(Node *n) {
1988 return UseStrictFP && _method->flags().is_strict()
1989 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
1990 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1991 : n;
1992 }
1994 // rounding for non-strict double stores
1995 Node* GraphKit::dstore_rounding(Node* n) {
1996 return Matcher::strict_fp_requires_explicit_rounding
1997 && UseSSE <= 1
1998 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
1999 : n;
2000 }
2002 //=============================================================================
2003 // Generate a fast path/slow path idiom. Graph looks like:
2004 // [foo] indicates that 'foo' is a parameter
2005 //
2006 // [in] NULL
2007 // \ /
2008 // CmpP
2009 // Bool ne
2010 // If
2011 // / \
2012 // True False-<2>
2013 // / |
2014 // / cast_not_null
2015 // Load | | ^
2016 // [fast_test] | |
2017 // gvn to opt_test | |
2018 // / \ | <1>
2019 // True False |
2020 // | \\ |
2021 // [slow_call] \[fast_result]
2022 // Ctl Val \ \
2023 // | \ \
2024 // Catch <1> \ \
2025 // / \ ^ \ \
2026 // Ex No_Ex | \ \
2027 // | \ \ | \ <2> \
2028 // ... \ [slow_res] | | \ [null_result]
2029 // \ \--+--+--- | |
2030 // \ | / \ | /
2031 // --------Region Phi
2032 //
2033 //=============================================================================
2034 // Code is structured as a series of driver functions all called 'do_XXX' that
2035 // call a set of helper functions. Helper functions first, then drivers.
2037 //------------------------------null_check_oop---------------------------------
2038 // Null check oop. Set null-path control into Region in slot 3.
2039 // Make a cast-not-nullness use the other not-null control. Return cast.
2040 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2041 bool never_see_null) {
2042 // Initial NULL check taken path
2043 (*null_control) = top();
2044 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2046 // Generate uncommon_trap:
2047 if (never_see_null && (*null_control) != top()) {
2048 // If we see an unexpected null at a check-cast we record it and force a
2049 // recompile; the offending check-cast will be compiled to handle NULLs.
2050 // If we see more than one offending BCI, then all checkcasts in the
2051 // method will be compiled to handle NULLs.
2052 PreserveJVMState pjvms(this);
2053 set_control(*null_control);
2054 replace_in_map(value, null());
2055 uncommon_trap(Deoptimization::Reason_null_check,
2056 Deoptimization::Action_make_not_entrant);
2057 (*null_control) = top(); // NULL path is dead
2058 }
2060 // Cast away null-ness on the result
2061 return cast;
2062 }
2064 //------------------------------opt_iff----------------------------------------
2065 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2066 // Return slow-path control.
2067 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2068 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2070 // Fast path taken; set region slot 2
2071 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2072 region->init_req(2,fast_taken); // Capture fast-control
2074 // Fast path not-taken, i.e. slow path
2075 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2076 return slow_taken;
2077 }
2079 //-----------------------------make_runtime_call-------------------------------
2080 Node* GraphKit::make_runtime_call(int flags,
2081 const TypeFunc* call_type, address call_addr,
2082 const char* call_name,
2083 const TypePtr* adr_type,
2084 // The following parms are all optional.
2085 // The first NULL ends the list.
2086 Node* parm0, Node* parm1,
2087 Node* parm2, Node* parm3,
2088 Node* parm4, Node* parm5,
2089 Node* parm6, Node* parm7) {
2090 // Slow-path call
2091 int size = call_type->domain()->cnt();
2092 bool is_leaf = !(flags & RC_NO_LEAF);
2093 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2094 if (call_name == NULL) {
2095 assert(!is_leaf, "must supply name for leaf");
2096 call_name = OptoRuntime::stub_name(call_addr);
2097 }
2098 CallNode* call;
2099 if (!is_leaf) {
2100 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2101 bci(), adr_type);
2102 } else if (flags & RC_NO_FP) {
2103 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2104 } else {
2105 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2106 }
2108 // The following is similar to set_edges_for_java_call,
2109 // except that the memory effects of the call are restricted to AliasIdxRaw.
2111 // Slow path call has no side-effects, uses few values
2112 bool wide_in = !(flags & RC_NARROW_MEM);
2113 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2115 Node* prev_mem = NULL;
2116 if (wide_in) {
2117 prev_mem = set_predefined_input_for_runtime_call(call);
2118 } else {
2119 assert(!wide_out, "narrow in => narrow out");
2120 Node* narrow_mem = memory(adr_type);
2121 prev_mem = reset_memory();
2122 map()->set_memory(narrow_mem);
2123 set_predefined_input_for_runtime_call(call);
2124 }
2126 // Hook each parm in order. Stop looking at the first NULL.
2127 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2128 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2129 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2130 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2131 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2132 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2133 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2134 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2135 /* close each nested if ===> */ } } } } } } } }
2136 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2138 if (!is_leaf) {
2139 // Non-leaves can block and take safepoints:
2140 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2141 }
2142 // Non-leaves can throw exceptions:
2143 if (has_io) {
2144 call->set_req(TypeFunc::I_O, i_o());
2145 }
2147 if (flags & RC_UNCOMMON) {
2148 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2149 // (An "if" probability corresponds roughly to an unconditional count.
2150 // Sort of.)
2151 call->set_cnt(PROB_UNLIKELY_MAG(4));
2152 }
2154 Node* c = _gvn.transform(call);
2155 assert(c == call, "cannot disappear");
2157 if (wide_out) {
2158 // Slow path call has full side-effects.
2159 set_predefined_output_for_runtime_call(call);
2160 } else {
2161 // Slow path call has few side-effects, and/or sets few values.
2162 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2163 }
2165 if (has_io) {
2166 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2167 }
2168 return call;
2170 }
2172 //------------------------------merge_memory-----------------------------------
2173 // Merge memory from one path into the current memory state.
2174 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2175 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2176 Node* old_slice = mms.force_memory();
2177 Node* new_slice = mms.memory2();
2178 if (old_slice != new_slice) {
2179 PhiNode* phi;
2180 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2181 phi = new_slice->as_Phi();
2182 #ifdef ASSERT
2183 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2184 old_slice = old_slice->in(new_path);
2185 // Caller is responsible for ensuring that any pre-existing
2186 // phis are already aware of old memory.
2187 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2188 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2189 #endif
2190 mms.set_memory(phi);
2191 } else {
2192 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2193 _gvn.set_type(phi, Type::MEMORY);
2194 phi->set_req(new_path, new_slice);
2195 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2196 }
2197 }
2198 }
2199 }
2201 //------------------------------make_slow_call_ex------------------------------
2202 // Make the exception handler hookups for the slow call
2203 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2204 if (stopped()) return;
2206 // Make a catch node with just two handlers: fall-through and catch-all
2207 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2208 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2209 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2210 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2212 { PreserveJVMState pjvms(this);
2213 set_control(excp);
2214 set_i_o(i_o);
2216 if (excp != top()) {
2217 // Create an exception state also.
2218 // Use an exact type if the caller has specified a specific exception.
2219 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2220 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
2221 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2222 }
2223 }
2225 // Get the no-exception control from the CatchNode.
2226 set_control(norm);
2227 }
2230 //-------------------------------gen_subtype_check-----------------------------
2231 // Generate a subtyping check. Takes as input the subtype and supertype.
2232 // Returns 2 values: sets the default control() to the true path and returns
2233 // the false path. Only reads invariant memory; sets no (visible) memory.
2234 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2235 // but that's not exposed to the optimizer. This call also doesn't take in an
2236 // Object; if you wish to check an Object you need to load the Object's class
2237 // prior to coming here.
2238 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2239 // Fast check for identical types, perhaps identical constants.
2240 // The types can even be identical non-constants, in cases
2241 // involving Array.newInstance, Object.clone, etc.
2242 if (subklass == superklass)
2243 return top(); // false path is dead; no test needed.
2245 if (_gvn.type(superklass)->singleton()) {
2246 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2247 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2249 // In the common case of an exact superklass, try to fold up the
2250 // test before generating code. You may ask, why not just generate
2251 // the code and then let it fold up? The answer is that the generated
2252 // code will necessarily include null checks, which do not always
2253 // completely fold away. If they are also needless, then they turn
2254 // into a performance loss. Example:
2255 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2256 // Here, the type of 'fa' is often exact, so the store check
2257 // of fa[1]=x will fold up, without testing the nullness of x.
2258 switch (static_subtype_check(superk, subk)) {
2259 case SSC_always_false:
2260 {
2261 Node* always_fail = control();
2262 set_control(top());
2263 return always_fail;
2264 }
2265 case SSC_always_true:
2266 return top();
2267 case SSC_easy_test:
2268 {
2269 // Just do a direct pointer compare and be done.
2270 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2271 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2272 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2273 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2274 return _gvn.transform( new(C, 1) IfFalseNode(iff) );
2275 }
2276 case SSC_full_test:
2277 break;
2278 default:
2279 ShouldNotReachHere();
2280 }
2281 }
2283 // %%% Possible further optimization: Even if the superklass is not exact,
2284 // if the subklass is the unique subtype of the superklass, the check
2285 // will always succeed. We could leave a dependency behind to ensure this.
2287 // First load the super-klass's check-offset
2288 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
2289 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2290 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
2291 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2293 // Load from the sub-klass's super-class display list, or a 1-word cache of
2294 // the secondary superclass list, or a failing value with a sentinel offset
2295 // if the super-klass is an interface or exceptionally deep in the Java
2296 // hierarchy and we have to scan the secondary superclass list the hard way.
2297 // Worst-case type is a little odd: NULL is allowed as a result (usually
2298 // klass loads can never produce a NULL).
2299 Node *chk_off_X = ConvI2X(chk_off);
2300 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2301 // For some types like interfaces the following loadKlass is from a 1-word
2302 // cache which is mutable so can't use immutable memory. Other
2303 // types load from the super-class display table which is immutable.
2304 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2305 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2307 // Compile speed common case: ARE a subtype and we canNOT fail
2308 if( superklass == nkls )
2309 return top(); // false path is dead; no test needed.
2311 // See if we get an immediate positive hit. Happens roughly 83% of the
2312 // time. Test to see if the value loaded just previously from the subklass
2313 // is exactly the superklass.
2314 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2315 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2316 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2317 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2318 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2320 // Compile speed common case: Check for being deterministic right now. If
2321 // chk_off is a constant and not equal to cacheoff then we are NOT a
2322 // subklass. In this case we need exactly the 1 test above and we can
2323 // return those results immediately.
2324 if (!might_be_cache) {
2325 Node* not_subtype_ctrl = control();
2326 set_control(iftrue1); // We need exactly the 1 test above
2327 return not_subtype_ctrl;
2328 }
2330 // Gather the various success & failures here
2331 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2332 record_for_igvn(r_ok_subtype);
2333 RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2334 record_for_igvn(r_not_subtype);
2336 r_ok_subtype->init_req(1, iftrue1);
2338 // Check for immediate negative hit. Happens roughly 11% of the time (which
2339 // is roughly 63% of the remaining cases). Test to see if the loaded
2340 // check-offset points into the subklass display list or the 1-element
2341 // cache. If it points to the display (and NOT the cache) and the display
2342 // missed then it's not a subtype.
2343 Node *cacheoff = _gvn.intcon(cacheoff_con);
2344 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2345 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2346 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2347 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2348 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2350 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2351 // No performance impact (too rare) but allows sharing of secondary arrays
2352 // which has some footprint reduction.
2353 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2354 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2355 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2356 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2357 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2359 // -- Roads not taken here: --
2360 // We could also have chosen to perform the self-check at the beginning
2361 // of this code sequence, as the assembler does. This would not pay off
2362 // the same way, since the optimizer, unlike the assembler, can perform
2363 // static type analysis to fold away many successful self-checks.
2364 // Non-foldable self checks work better here in second position, because
2365 // the initial primary superclass check subsumes a self-check for most
2366 // types. An exception would be a secondary type like array-of-interface,
2367 // which does not appear in its own primary supertype display.
2368 // Finally, we could have chosen to move the self-check into the
2369 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2370 // dependent manner. But it is worthwhile to have the check here,
2371 // where it can be perhaps be optimized. The cost in code space is
2372 // small (register compare, branch).
2374 // Now do a linear scan of the secondary super-klass array. Again, no real
2375 // performance impact (too rare) but it's gotta be done.
2376 // Since the code is rarely used, there is no penalty for moving it
2377 // out of line, and it can only improve I-cache density.
2378 // The decision to inline or out-of-line this final check is platform
2379 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2380 Node* psc = _gvn.transform(
2381 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2383 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2384 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2385 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2386 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2387 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2389 // Return false path; set default control to true path.
2390 set_control( _gvn.transform(r_ok_subtype) );
2391 return _gvn.transform(r_not_subtype);
2392 }
2394 //----------------------------static_subtype_check-----------------------------
2395 // Shortcut important common cases when superklass is exact:
2396 // (0) superklass is java.lang.Object (can occur in reflective code)
2397 // (1) subklass is already limited to a subtype of superklass => always ok
2398 // (2) subklass does not overlap with superklass => always fail
2399 // (3) superklass has NO subtypes and we can check with a simple compare.
2400 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2401 if (StressReflectiveCode) {
2402 return SSC_full_test; // Let caller generate the general case.
2403 }
2405 if (superk == env()->Object_klass()) {
2406 return SSC_always_true; // (0) this test cannot fail
2407 }
2409 ciType* superelem = superk;
2410 if (superelem->is_array_klass())
2411 superelem = superelem->as_array_klass()->base_element_type();
2413 if (!subk->is_interface()) { // cannot trust static interface types yet
2414 if (subk->is_subtype_of(superk)) {
2415 return SSC_always_true; // (1) false path dead; no dynamic test needed
2416 }
2417 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2418 !superk->is_subtype_of(subk)) {
2419 return SSC_always_false;
2420 }
2421 }
2423 // If casting to an instance klass, it must have no subtypes
2424 if (superk->is_interface()) {
2425 // Cannot trust interfaces yet.
2426 // %%% S.B. superk->nof_implementors() == 1
2427 } else if (superelem->is_instance_klass()) {
2428 ciInstanceKlass* ik = superelem->as_instance_klass();
2429 if (!ik->has_subklass() && !ik->is_interface()) {
2430 if (!ik->is_final()) {
2431 // Add a dependency if there is a chance of a later subclass.
2432 C->dependencies()->assert_leaf_type(ik);
2433 }
2434 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2435 }
2436 } else {
2437 // A primitive array type has no subtypes.
2438 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2439 }
2441 return SSC_full_test;
2442 }
2444 // Profile-driven exact type check:
2445 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2446 float prob,
2447 Node* *casted_receiver) {
2448 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2449 Node* recv_klass = load_object_klass(receiver);
2450 Node* want_klass = makecon(tklass);
2451 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2452 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2453 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2454 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2455 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2457 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2458 assert(recv_xtype->klass_is_exact(), "");
2460 // Subsume downstream occurrences of receiver with a cast to
2461 // recv_xtype, since now we know what the type will be.
2462 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2463 (*casted_receiver) = _gvn.transform(cast);
2464 // (User must make the replace_in_map call.)
2466 return fail;
2467 }
2470 //------------------------------seems_never_null-------------------------------
2471 // Use null_seen information if it is available from the profile.
2472 // If we see an unexpected null at a type check we record it and force a
2473 // recompile; the offending check will be recompiled to handle NULLs.
2474 // If we see several offending BCIs, then all checks in the
2475 // method will be recompiled.
2476 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2477 if (UncommonNullCast // Cutout for this technique
2478 && obj != null() // And not the -Xcomp stupid case?
2479 && !too_many_traps(Deoptimization::Reason_null_check)
2480 ) {
2481 if (data == NULL)
2482 // Edge case: no mature data. Be optimistic here.
2483 return true;
2484 // If the profile has not seen a null, assume it won't happen.
2485 assert(java_bc() == Bytecodes::_checkcast ||
2486 java_bc() == Bytecodes::_instanceof ||
2487 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2488 return !data->as_BitData()->null_seen();
2489 }
2490 return false;
2491 }
2493 //------------------------maybe_cast_profiled_receiver-------------------------
2494 // If the profile has seen exactly one type, narrow to exactly that type.
2495 // Subsequent type checks will always fold up.
2496 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2497 ciProfileData* data,
2498 ciKlass* require_klass) {
2499 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2500 if (data == NULL) return NULL;
2502 // Make sure we haven't already deoptimized from this tactic.
2503 if (too_many_traps(Deoptimization::Reason_class_check))
2504 return NULL;
2506 // (No, this isn't a call, but it's enough like a virtual call
2507 // to use the same ciMethod accessor to get the profile info...)
2508 ciCallProfile profile = method()->call_profile_at_bci(bci());
2509 if (profile.count() >= 0 && // no cast failures here
2510 profile.has_receiver(0) &&
2511 profile.morphism() == 1) {
2512 ciKlass* exact_kls = profile.receiver(0);
2513 if (require_klass == NULL ||
2514 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2515 // If we narrow the type to match what the type profile sees,
2516 // we can then remove the rest of the cast.
2517 // This is a win, even if the exact_kls is very specific,
2518 // because downstream operations, such as method calls,
2519 // will often benefit from the sharper type.
2520 Node* exact_obj = not_null_obj; // will get updated in place...
2521 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2522 &exact_obj);
2523 { PreserveJVMState pjvms(this);
2524 set_control(slow_ctl);
2525 uncommon_trap(Deoptimization::Reason_class_check,
2526 Deoptimization::Action_maybe_recompile);
2527 }
2528 replace_in_map(not_null_obj, exact_obj);
2529 return exact_obj;
2530 }
2531 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2532 }
2534 return NULL;
2535 }
2538 //-------------------------------gen_instanceof--------------------------------
2539 // Generate an instance-of idiom. Used by both the instance-of bytecode
2540 // and the reflective instance-of call.
2541 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) {
2542 kill_dead_locals(); // Benefit all the uncommon traps
2543 assert( !stopped(), "dead parse path should be checked in callers" );
2544 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2545 "must check for not-null not-dead klass in callers");
2547 // Make the merge point
2548 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2549 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2550 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2551 C->set_has_split_ifs(true); // Has chance for split-if optimization
2553 ciProfileData* data = NULL;
2554 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2555 data = method()->method_data()->bci_to_data(bci());
2556 }
2557 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2558 && seems_never_null(obj, data));
2560 // Null check; get casted pointer; set region slot 3
2561 Node* null_ctl = top();
2562 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2564 // If not_null_obj is dead, only null-path is taken
2565 if (stopped()) { // Doing instance-of on a NULL?
2566 set_control(null_ctl);
2567 return intcon(0);
2568 }
2569 region->init_req(_null_path, null_ctl);
2570 phi ->init_req(_null_path, intcon(0)); // Set null path value
2571 if (null_ctl == top()) {
2572 // Do this eagerly, so that pattern matches like is_diamond_phi
2573 // will work even during parsing.
2574 assert(_null_path == PATH_LIMIT-1, "delete last");
2575 region->del_req(_null_path);
2576 phi ->del_req(_null_path);
2577 }
2579 if (ProfileDynamicTypes && data != NULL) {
2580 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL);
2581 if (stopped()) { // Profile disagrees with this path.
2582 set_control(null_ctl); // Null is the only remaining possibility.
2583 return intcon(0);
2584 }
2585 if (cast_obj != NULL)
2586 not_null_obj = cast_obj;
2587 }
2589 // Load the object's klass
2590 Node* obj_klass = load_object_klass(not_null_obj);
2592 // Generate the subtype check
2593 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2595 // Plug in the success path to the general merge in slot 1.
2596 region->init_req(_obj_path, control());
2597 phi ->init_req(_obj_path, intcon(1));
2599 // Plug in the failing path to the general merge in slot 2.
2600 region->init_req(_fail_path, not_subtype_ctrl);
2601 phi ->init_req(_fail_path, intcon(0));
2603 // Return final merged results
2604 set_control( _gvn.transform(region) );
2605 record_for_igvn(region);
2606 return _gvn.transform(phi);
2607 }
2609 //-------------------------------gen_checkcast---------------------------------
2610 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2611 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2612 // uncommon-trap paths work. Adjust stack after this call.
2613 // If failure_control is supplied and not null, it is filled in with
2614 // the control edge for the cast failure. Otherwise, an appropriate
2615 // uncommon trap or exception is thrown.
2616 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2617 Node* *failure_control) {
2618 kill_dead_locals(); // Benefit all the uncommon traps
2619 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2620 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2622 // Fast cutout: Check the case that the cast is vacuously true.
2623 // This detects the common cases where the test will short-circuit
2624 // away completely. We do this before we perform the null check,
2625 // because if the test is going to turn into zero code, we don't
2626 // want a residual null check left around. (Causes a slowdown,
2627 // for example, in some objArray manipulations, such as a[i]=a[j].)
2628 if (tk->singleton()) {
2629 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2630 if (objtp != NULL && objtp->klass() != NULL) {
2631 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2632 case SSC_always_true:
2633 return obj;
2634 case SSC_always_false:
2635 // It needs a null check because a null will *pass* the cast check.
2636 // A non-null value will always produce an exception.
2637 return do_null_assert(obj, T_OBJECT);
2638 }
2639 }
2640 }
2642 ciProfileData* data = NULL;
2643 if (failure_control == NULL) { // use MDO in regular case only
2644 assert(java_bc() == Bytecodes::_aastore ||
2645 java_bc() == Bytecodes::_checkcast,
2646 "interpreter profiles type checks only for these BCs");
2647 data = method()->method_data()->bci_to_data(bci());
2648 }
2650 // Make the merge point
2651 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2652 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2653 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
2654 C->set_has_split_ifs(true); // Has chance for split-if optimization
2656 // Use null-cast information if it is available
2657 bool never_see_null = ((failure_control == NULL) // regular case only
2658 && seems_never_null(obj, data));
2660 // Null check; get casted pointer; set region slot 3
2661 Node* null_ctl = top();
2662 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2664 // If not_null_obj is dead, only null-path is taken
2665 if (stopped()) { // Doing instance-of on a NULL?
2666 set_control(null_ctl);
2667 return null();
2668 }
2669 region->init_req(_null_path, null_ctl);
2670 phi ->init_req(_null_path, null()); // Set null path value
2671 if (null_ctl == top()) {
2672 // Do this eagerly, so that pattern matches like is_diamond_phi
2673 // will work even during parsing.
2674 assert(_null_path == PATH_LIMIT-1, "delete last");
2675 region->del_req(_null_path);
2676 phi ->del_req(_null_path);
2677 }
2679 Node* cast_obj = NULL;
2680 if (data != NULL &&
2681 // Counter has never been decremented (due to cast failure).
2682 // ...This is a reasonable thing to expect. It is true of
2683 // all casts inserted by javac to implement generic types.
2684 data->as_CounterData()->count() >= 0) {
2685 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass());
2686 if (cast_obj != NULL) {
2687 if (failure_control != NULL) // failure is now impossible
2688 (*failure_control) = top();
2689 // adjust the type of the phi to the exact klass:
2690 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR));
2691 }
2692 }
2694 if (cast_obj == NULL) {
2695 // Load the object's klass
2696 Node* obj_klass = load_object_klass(not_null_obj);
2698 // Generate the subtype check
2699 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2701 // Plug in success path into the merge
2702 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2703 not_null_obj, toop));
2704 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2705 if (failure_control == NULL) {
2706 if (not_subtype_ctrl != top()) { // If failure is possible
2707 PreserveJVMState pjvms(this);
2708 set_control(not_subtype_ctrl);
2709 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2710 }
2711 } else {
2712 (*failure_control) = not_subtype_ctrl;
2713 }
2714 }
2716 region->init_req(_obj_path, control());
2717 phi ->init_req(_obj_path, cast_obj);
2719 // A merge of NULL or Casted-NotNull obj
2720 Node* res = _gvn.transform(phi);
2722 // Note I do NOT always 'replace_in_map(obj,result)' here.
2723 // if( tk->klass()->can_be_primary_super() )
2724 // This means that if I successfully store an Object into an array-of-String
2725 // I 'forget' that the Object is really now known to be a String. I have to
2726 // do this because we don't have true union types for interfaces - if I store
2727 // a Baz into an array-of-Interface and then tell the optimizer it's an
2728 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2729 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2730 // replace_in_map( obj, res );
2732 // Return final merged results
2733 set_control( _gvn.transform(region) );
2734 record_for_igvn(region);
2735 return res;
2736 }
2738 //------------------------------next_monitor-----------------------------------
2739 // What number should be given to the next monitor?
2740 int GraphKit::next_monitor() {
2741 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2742 int next = current + C->sync_stack_slots();
2743 // Keep the toplevel high water mark current:
2744 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2745 return current;
2746 }
2748 //------------------------------insert_mem_bar---------------------------------
2749 // Memory barrier to avoid floating things around
2750 // The membar serves as a pinch point between both control and all memory slices.
2751 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2752 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2753 mb->init_req(TypeFunc::Control, control());
2754 mb->init_req(TypeFunc::Memory, reset_memory());
2755 Node* membar = _gvn.transform(mb);
2756 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2757 set_all_memory_call(membar);
2758 return membar;
2759 }
2761 //-------------------------insert_mem_bar_volatile----------------------------
2762 // Memory barrier to avoid floating things around
2763 // The membar serves as a pinch point between both control and memory(alias_idx).
2764 // If you want to make a pinch point on all memory slices, do not use this
2765 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2766 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2767 // When Parse::do_put_xxx updates a volatile field, it appends a series
2768 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2769 // The first membar is on the same memory slice as the field store opcode.
2770 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2771 // All the other membars (for other volatile slices, including AliasIdxBot,
2772 // which stands for all unknown volatile slices) are control-dependent
2773 // on the first membar. This prevents later volatile loads or stores
2774 // from sliding up past the just-emitted store.
2776 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2777 mb->set_req(TypeFunc::Control,control());
2778 if (alias_idx == Compile::AliasIdxBot) {
2779 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2780 } else {
2781 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2782 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2783 }
2784 Node* membar = _gvn.transform(mb);
2785 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2786 if (alias_idx == Compile::AliasIdxBot) {
2787 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2788 } else {
2789 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2790 }
2791 return membar;
2792 }
2794 //------------------------------shared_lock------------------------------------
2795 // Emit locking code.
2796 FastLockNode* GraphKit::shared_lock(Node* obj) {
2797 // bci is either a monitorenter bc or InvocationEntryBci
2798 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2799 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2801 if( !GenerateSynchronizationCode )
2802 return NULL; // Not locking things?
2803 if (stopped()) // Dead monitor?
2804 return NULL;
2806 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2808 // Box the stack location
2809 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2810 Node* mem = reset_memory();
2812 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2813 if (PrintPreciseBiasedLockingStatistics) {
2814 // Create the counters for this fast lock.
2815 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2816 }
2817 // Add monitor to debug info for the slow path. If we block inside the
2818 // slow path and de-opt, we need the monitor hanging around
2819 map()->push_monitor( flock );
2821 const TypeFunc *tf = LockNode::lock_type();
2822 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2824 lock->init_req( TypeFunc::Control, control() );
2825 lock->init_req( TypeFunc::Memory , mem );
2826 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2827 lock->init_req( TypeFunc::FramePtr, frameptr() );
2828 lock->init_req( TypeFunc::ReturnAdr, top() );
2830 lock->init_req(TypeFunc::Parms + 0, obj);
2831 lock->init_req(TypeFunc::Parms + 1, box);
2832 lock->init_req(TypeFunc::Parms + 2, flock);
2833 add_safepoint_edges(lock);
2835 lock = _gvn.transform( lock )->as_Lock();
2837 // lock has no side-effects, sets few values
2838 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2840 insert_mem_bar(Op_MemBarAcquire);
2842 // Add this to the worklist so that the lock can be eliminated
2843 record_for_igvn(lock);
2845 #ifndef PRODUCT
2846 if (PrintLockStatistics) {
2847 // Update the counter for this lock. Don't bother using an atomic
2848 // operation since we don't require absolute accuracy.
2849 lock->create_lock_counter(map()->jvms());
2850 increment_counter(lock->counter()->addr());
2851 }
2852 #endif
2854 return flock;
2855 }
2858 //------------------------------shared_unlock----------------------------------
2859 // Emit unlocking code.
2860 void GraphKit::shared_unlock(Node* box, Node* obj) {
2861 // bci is either a monitorenter bc or InvocationEntryBci
2862 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2863 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2865 if( !GenerateSynchronizationCode )
2866 return;
2867 if (stopped()) { // Dead monitor?
2868 map()->pop_monitor(); // Kill monitor from debug info
2869 return;
2870 }
2872 // Memory barrier to avoid floating things down past the locked region
2873 insert_mem_bar(Op_MemBarRelease);
2875 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2876 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2877 uint raw_idx = Compile::AliasIdxRaw;
2878 unlock->init_req( TypeFunc::Control, control() );
2879 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2880 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2881 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2882 unlock->init_req( TypeFunc::ReturnAdr, top() );
2884 unlock->init_req(TypeFunc::Parms + 0, obj);
2885 unlock->init_req(TypeFunc::Parms + 1, box);
2886 unlock = _gvn.transform(unlock)->as_Unlock();
2888 Node* mem = reset_memory();
2890 // unlock has no side-effects, sets few values
2891 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2893 // Kill monitor from debug info
2894 map()->pop_monitor( );
2895 }
2897 //-------------------------------get_layout_helper-----------------------------
2898 // If the given klass is a constant or known to be an array,
2899 // fetch the constant layout helper value into constant_value
2900 // and return (Node*)NULL. Otherwise, load the non-constant
2901 // layout helper value, and return the node which represents it.
2902 // This two-faced routine is useful because allocation sites
2903 // almost always feature constant types.
2904 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2905 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2906 if (!StressReflectiveCode && inst_klass != NULL) {
2907 ciKlass* klass = inst_klass->klass();
2908 bool xklass = inst_klass->klass_is_exact();
2909 if (xklass || klass->is_array_klass()) {
2910 jint lhelper = klass->layout_helper();
2911 if (lhelper != Klass::_lh_neutral_value) {
2912 constant_value = lhelper;
2913 return (Node*) NULL;
2914 }
2915 }
2916 }
2917 constant_value = Klass::_lh_neutral_value; // put in a known value
2918 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
2919 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2920 }
2922 // We just put in an allocate/initialize with a big raw-memory effect.
2923 // Hook selected additional alias categories on the initialization.
2924 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2925 MergeMemNode* init_in_merge,
2926 Node* init_out_raw) {
2927 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2928 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2930 Node* prevmem = kit.memory(alias_idx);
2931 init_in_merge->set_memory_at(alias_idx, prevmem);
2932 kit.set_memory(init_out_raw, alias_idx);
2933 }
2935 //---------------------------set_output_for_allocation-------------------------
2936 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2937 const TypeOopPtr* oop_type,
2938 bool raw_mem_only) {
2939 int rawidx = Compile::AliasIdxRaw;
2940 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2941 add_safepoint_edges(alloc);
2942 Node* allocx = _gvn.transform(alloc);
2943 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2944 // create memory projection for i_o
2945 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2946 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2948 // create a memory projection as for the normal control path
2949 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2950 set_memory(malloc, rawidx);
2952 // a normal slow-call doesn't change i_o, but an allocation does
2953 // we create a separate i_o projection for the normal control path
2954 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2955 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2957 // put in an initialization barrier
2958 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2959 rawoop)->as_Initialize();
2960 assert(alloc->initialization() == init, "2-way macro link must work");
2961 assert(init ->allocation() == alloc, "2-way macro link must work");
2962 if (ReduceFieldZeroing && !raw_mem_only) {
2963 // Extract memory strands which may participate in the new object's
2964 // initialization, and source them from the new InitializeNode.
2965 // This will allow us to observe initializations when they occur,
2966 // and link them properly (as a group) to the InitializeNode.
2967 assert(init->in(InitializeNode::Memory) == malloc, "");
2968 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
2969 init->set_req(InitializeNode::Memory, minit_in);
2970 record_for_igvn(minit_in); // fold it up later, if possible
2971 Node* minit_out = memory(rawidx);
2972 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
2973 if (oop_type->isa_aryptr()) {
2974 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
2975 int elemidx = C->get_alias_index(telemref);
2976 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
2977 } else if (oop_type->isa_instptr()) {
2978 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
2979 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
2980 ciField* field = ik->nonstatic_field_at(i);
2981 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
2982 continue; // do not bother to track really large numbers of fields
2983 // Find (or create) the alias category for this field:
2984 int fieldidx = C->alias_type(field)->index();
2985 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
2986 }
2987 }
2988 }
2990 // Cast raw oop to the real thing...
2991 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
2992 javaoop = _gvn.transform(javaoop);
2993 C->set_recent_alloc(control(), javaoop);
2994 assert(just_allocated_object(control()) == javaoop, "just allocated");
2996 #ifdef ASSERT
2997 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
2998 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
2999 "Ideal_allocation works");
3000 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3001 "Ideal_allocation works");
3002 if (alloc->is_AllocateArray()) {
3003 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3004 "Ideal_allocation works");
3005 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3006 "Ideal_allocation works");
3007 } else {
3008 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3009 }
3010 }
3011 #endif //ASSERT
3013 return javaoop;
3014 }
3016 //---------------------------new_instance--------------------------------------
3017 // This routine takes a klass_node which may be constant (for a static type)
3018 // or may be non-constant (for reflective code). It will work equally well
3019 // for either, and the graph will fold nicely if the optimizer later reduces
3020 // the type to a constant.
3021 // The optional arguments are for specialized use by intrinsics:
3022 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3023 // - If 'raw_mem_only', do not cast the result to an oop.
3024 // - If 'return_size_val', report the the total object size to the caller.
3025 Node* GraphKit::new_instance(Node* klass_node,
3026 Node* extra_slow_test,
3027 bool raw_mem_only, // affect only raw memory
3028 Node* *return_size_val) {
3029 // Compute size in doublewords
3030 // The size is always an integral number of doublewords, represented
3031 // as a positive bytewise size stored in the klass's layout_helper.
3032 // The layout_helper also encodes (in a low bit) the need for a slow path.
3033 jint layout_con = Klass::_lh_neutral_value;
3034 Node* layout_val = get_layout_helper(klass_node, layout_con);
3035 int layout_is_con = (layout_val == NULL);
3037 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3038 // Generate the initial go-slow test. It's either ALWAYS (return a
3039 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3040 // case) a computed value derived from the layout_helper.
3041 Node* initial_slow_test = NULL;
3042 if (layout_is_con) {
3043 assert(!StressReflectiveCode, "stress mode does not use these paths");
3044 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3045 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3047 } else { // reflective case
3048 // This reflective path is used by Unsafe.allocateInstance.
3049 // (It may be stress-tested by specifying StressReflectiveCode.)
3050 // Basically, we want to get into the VM is there's an illegal argument.
3051 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3052 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
3053 if (extra_slow_test != intcon(0)) {
3054 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
3055 }
3056 // (Macro-expander will further convert this to a Bool, if necessary.)
3057 }
3059 // Find the size in bytes. This is easy; it's the layout_helper.
3060 // The size value must be valid even if the slow path is taken.
3061 Node* size = NULL;
3062 if (layout_is_con) {
3063 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3064 } else { // reflective case
3065 // This reflective path is used by clone and Unsafe.allocateInstance.
3066 size = ConvI2X(layout_val);
3068 // Clear the low bits to extract layout_helper_size_in_bytes:
3069 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3070 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3071 size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
3072 }
3073 if (return_size_val != NULL) {
3074 (*return_size_val) = size;
3075 }
3077 // This is a precise notnull oop of the klass.
3078 // (Actually, it need not be precise if this is a reflective allocation.)
3079 // It's what we cast the result to.
3080 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3081 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3082 const TypeOopPtr* oop_type = tklass->as_instance_type();
3084 // Now generate allocation code
3086 // The entire memory state is needed for slow path of the allocation
3087 // since GC and deoptimization can happened.
3088 Node *mem = reset_memory();
3089 set_all_memory(mem); // Create new memory state
3091 AllocateNode* alloc
3092 = new (C, AllocateNode::ParmLimit)
3093 AllocateNode(C, AllocateNode::alloc_type(),
3094 control(), mem, i_o(),
3095 size, klass_node,
3096 initial_slow_test);
3098 return set_output_for_allocation(alloc, oop_type, raw_mem_only);
3099 }
3101 //-------------------------------new_array-------------------------------------
3102 // helper for both newarray and anewarray
3103 // The 'length' parameter is (obviously) the length of the array.
3104 // See comments on new_instance for the meaning of the other arguments.
3105 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3106 Node* length, // number of array elements
3107 int nargs, // number of arguments to push back for uncommon trap
3108 bool raw_mem_only, // affect only raw memory
3109 Node* *return_size_val) {
3110 jint layout_con = Klass::_lh_neutral_value;
3111 Node* layout_val = get_layout_helper(klass_node, layout_con);
3112 int layout_is_con = (layout_val == NULL);
3114 if (!layout_is_con && !StressReflectiveCode &&
3115 !too_many_traps(Deoptimization::Reason_class_check)) {
3116 // This is a reflective array creation site.
3117 // Optimistically assume that it is a subtype of Object[],
3118 // so that we can fold up all the address arithmetic.
3119 layout_con = Klass::array_layout_helper(T_OBJECT);
3120 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
3121 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
3122 { BuildCutout unless(this, bol_lh, PROB_MAX);
3123 _sp += nargs;
3124 uncommon_trap(Deoptimization::Reason_class_check,
3125 Deoptimization::Action_maybe_recompile);
3126 }
3127 layout_val = NULL;
3128 layout_is_con = true;
3129 }
3131 // Generate the initial go-slow test. Make sure we do not overflow
3132 // if length is huge (near 2Gig) or negative! We do not need
3133 // exact double-words here, just a close approximation of needed
3134 // double-words. We can't add any offset or rounding bits, lest we
3135 // take a size -1 of bytes and make it positive. Use an unsigned
3136 // compare, so negative sizes look hugely positive.
3137 int fast_size_limit = FastAllocateSizeLimit;
3138 if (layout_is_con) {
3139 assert(!StressReflectiveCode, "stress mode does not use these paths");
3140 // Increase the size limit if we have exact knowledge of array type.
3141 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3142 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3143 }
3145 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3146 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3147 if (initial_slow_test->is_Bool()) {
3148 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3149 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3150 }
3152 // --- Size Computation ---
3153 // array_size = round_to_heap(array_header + (length << elem_shift));
3154 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3155 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3156 // The rounding mask is strength-reduced, if possible.
3157 int round_mask = MinObjAlignmentInBytes - 1;
3158 Node* header_size = NULL;
3159 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3160 // (T_BYTE has the weakest alignment and size restrictions...)
3161 if (layout_is_con) {
3162 int hsize = Klass::layout_helper_header_size(layout_con);
3163 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3164 BasicType etype = Klass::layout_helper_element_type(layout_con);
3165 if ((round_mask & ~right_n_bits(eshift)) == 0)
3166 round_mask = 0; // strength-reduce it if it goes away completely
3167 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3168 assert(header_size_min <= hsize, "generic minimum is smallest");
3169 header_size_min = hsize;
3170 header_size = intcon(hsize + round_mask);
3171 } else {
3172 Node* hss = intcon(Klass::_lh_header_size_shift);
3173 Node* hsm = intcon(Klass::_lh_header_size_mask);
3174 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3175 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3176 Node* mask = intcon(round_mask);
3177 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3178 }
3180 Node* elem_shift = NULL;
3181 if (layout_is_con) {
3182 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3183 if (eshift != 0)
3184 elem_shift = intcon(eshift);
3185 } else {
3186 // There is no need to mask or shift this value.
3187 // The semantics of LShiftINode include an implicit mask to 0x1F.
3188 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3189 elem_shift = layout_val;
3190 }
3192 // Transition to native address size for all offset calculations:
3193 Node* lengthx = ConvI2X(length);
3194 Node* headerx = ConvI2X(header_size);
3195 #ifdef _LP64
3196 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3197 if (tllen != NULL && tllen->_lo < 0) {
3198 // Add a manual constraint to a positive range. Cf. array_element_address.
3199 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3200 if (size_max > tllen->_hi) size_max = tllen->_hi;
3201 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3202 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3203 }
3204 }
3205 #endif
3207 // Combine header size (plus rounding) and body size. Then round down.
3208 // This computation cannot overflow, because it is used only in two
3209 // places, one where the length is sharply limited, and the other
3210 // after a successful allocation.
3211 Node* abody = lengthx;
3212 if (elem_shift != NULL)
3213 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3214 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3215 if (round_mask != 0) {
3216 Node* mask = MakeConX(~round_mask);
3217 size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3218 }
3219 // else if round_mask == 0, the size computation is self-rounding
3221 if (return_size_val != NULL) {
3222 // This is the size
3223 (*return_size_val) = size;
3224 }
3226 // Now generate allocation code
3228 // The entire memory state is needed for slow path of the allocation
3229 // since GC and deoptimization can happened.
3230 Node *mem = reset_memory();
3231 set_all_memory(mem); // Create new memory state
3233 // Create the AllocateArrayNode and its result projections
3234 AllocateArrayNode* alloc
3235 = new (C, AllocateArrayNode::ParmLimit)
3236 AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3237 control(), mem, i_o(),
3238 size, klass_node,
3239 initial_slow_test,
3240 length);
3242 // Cast to correct type. Note that the klass_node may be constant or not,
3243 // and in the latter case the actual array type will be inexact also.
3244 // (This happens via a non-constant argument to inline_native_newArray.)
3245 // In any case, the value of klass_node provides the desired array type.
3246 const TypeInt* length_type = _gvn.find_int_type(length);
3247 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3248 if (ary_type->isa_aryptr() && length_type != NULL) {
3249 // Try to get a better type than POS for the size
3250 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3251 }
3253 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
3255 // Cast length on remaining path to be as narrow as possible
3256 if (map()->find_edge(length) >= 0) {
3257 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3258 if (ccast != length) {
3259 _gvn.set_type_bottom(ccast);
3260 record_for_igvn(ccast);
3261 replace_in_map(length, ccast);
3262 }
3263 }
3265 return javaoop;
3266 }
3268 // The following "Ideal_foo" functions are placed here because they recognize
3269 // the graph shapes created by the functions immediately above.
3271 //---------------------------Ideal_allocation----------------------------------
3272 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3273 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3274 if (ptr == NULL) { // reduce dumb test in callers
3275 return NULL;
3276 }
3277 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3278 ptr = ptr->in(1);
3279 if (ptr == NULL) return NULL;
3280 }
3281 if (ptr->is_Proj()) {
3282 Node* allo = ptr->in(0);
3283 if (allo != NULL && allo->is_Allocate()) {
3284 return allo->as_Allocate();
3285 }
3286 }
3287 // Report failure to match.
3288 return NULL;
3289 }
3291 // Fancy version which also strips off an offset (and reports it to caller).
3292 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3293 intptr_t& offset) {
3294 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3295 if (base == NULL) return NULL;
3296 return Ideal_allocation(base, phase);
3297 }
3299 // Trace Initialize <- Proj[Parm] <- Allocate
3300 AllocateNode* InitializeNode::allocation() {
3301 Node* rawoop = in(InitializeNode::RawAddress);
3302 if (rawoop->is_Proj()) {
3303 Node* alloc = rawoop->in(0);
3304 if (alloc->is_Allocate()) {
3305 return alloc->as_Allocate();
3306 }
3307 }
3308 return NULL;
3309 }
3311 // Trace Allocate -> Proj[Parm] -> Initialize
3312 InitializeNode* AllocateNode::initialization() {
3313 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3314 if (rawoop == NULL) return NULL;
3315 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3316 Node* init = rawoop->fast_out(i);
3317 if (init->is_Initialize()) {
3318 assert(init->as_Initialize()->allocation() == this, "2-way link");
3319 return init->as_Initialize();
3320 }
3321 }
3322 return NULL;
3323 }
3325 //----------------------------- store barriers ----------------------------
3326 #define __ ideal.
3328 void GraphKit::sync_kit(IdealKit& ideal) {
3329 // Final sync IdealKit and graphKit.
3330 __ drain_delay_transform();
3331 set_all_memory(__ merged_memory());
3332 set_control(__ ctrl());
3333 }
3335 // vanilla/CMS post barrier
3336 // Insert a write-barrier store. This is to let generational GC work; we have
3337 // to flag all oop-stores before the next GC point.
3338 void GraphKit::write_barrier_post(Node* oop_store,
3339 Node* obj,
3340 Node* adr,
3341 uint adr_idx,
3342 Node* val,
3343 bool use_precise) {
3344 // No store check needed if we're storing a NULL or an old object
3345 // (latter case is probably a string constant). The concurrent
3346 // mark sweep garbage collector, however, needs to have all nonNull
3347 // oop updates flagged via card-marks.
3348 if (val != NULL && val->is_Con()) {
3349 // must be either an oop or NULL
3350 const Type* t = val->bottom_type();
3351 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3352 // stores of null never (?) need barriers
3353 return;
3354 ciObject* con = t->is_oopptr()->const_oop();
3355 if (con != NULL
3356 && con->is_perm()
3357 && Universe::heap()->can_elide_permanent_oop_store_barriers())
3358 // no store barrier needed, because no old-to-new ref created
3359 return;
3360 }
3362 if (use_ReduceInitialCardMarks()
3363 && obj == just_allocated_object(control())) {
3364 // We can skip marks on a freshly-allocated object in Eden.
3365 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3366 // That routine informs GC to take appropriate compensating steps,
3367 // upon a slow-path allocation, so as to make this card-mark
3368 // elision safe.
3369 return;
3370 }
3372 if (!use_precise) {
3373 // All card marks for a (non-array) instance are in one place:
3374 adr = obj;
3375 }
3376 // (Else it's an array (or unknown), and we want more precise card marks.)
3377 assert(adr != NULL, "");
3379 IdealKit ideal(gvn(), control(), merged_memory(), true);
3381 // Convert the pointer to an int prior to doing math on it
3382 Node* cast = __ CastPX(__ ctrl(), adr);
3384 // Divide by card size
3385 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3386 "Only one we handle so far.");
3387 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3389 // Combine card table base and card offset
3390 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3392 // Get the alias_index for raw card-mark memory
3393 int adr_type = Compile::AliasIdxRaw;
3394 // Smash zero into card
3395 Node* zero = __ ConI(0);
3396 BasicType bt = T_BYTE;
3397 if( !UseConcMarkSweepGC ) {
3398 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3399 } else {
3400 // Specialized path for CM store barrier
3401 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3402 }
3404 // Final sync IdealKit and GraphKit.
3405 sync_kit(ideal);
3406 }
3408 // G1 pre/post barriers
3409 void GraphKit::g1_write_barrier_pre(Node* obj,
3410 Node* adr,
3411 uint alias_idx,
3412 Node* val,
3413 const TypeOopPtr* val_type,
3414 BasicType bt) {
3415 IdealKit ideal(gvn(), control(), merged_memory(), true);
3417 Node* tls = __ thread(); // ThreadLocalStorage
3419 Node* no_ctrl = NULL;
3420 Node* no_base = __ top();
3421 Node* zero = __ ConI(0);
3423 float likely = PROB_LIKELY(0.999);
3424 float unlikely = PROB_UNLIKELY(0.999);
3426 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3427 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3429 // Offsets into the thread
3430 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3431 PtrQueue::byte_offset_of_active());
3432 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3433 PtrQueue::byte_offset_of_index());
3434 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3435 PtrQueue::byte_offset_of_buf());
3436 // Now the actual pointers into the thread
3438 // set_control( ctl);
3440 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3441 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3442 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3444 // Now some of the values
3446 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3448 // if (!marking)
3449 __ if_then(marking, BoolTest::ne, zero); {
3450 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3452 const Type* t1 = adr->bottom_type();
3453 const Type* t2 = val->bottom_type();
3455 Node* orig = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3456 // if (orig != NULL)
3457 __ if_then(orig, BoolTest::ne, null()); {
3458 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3460 // load original value
3461 // alias_idx correct??
3463 // is the queue for this thread full?
3464 __ if_then(index, BoolTest::ne, zero, likely); {
3466 // decrement the index
3467 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3468 Node* next_indexX = next_index;
3469 #ifdef _LP64
3470 // We could refine the type for what it's worth
3471 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3472 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3473 #endif
3475 // Now get the buffer location we will log the original value into and store it
3476 Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3477 __ store(__ ctrl(), log_addr, orig, T_OBJECT, Compile::AliasIdxRaw);
3479 // update the index
3480 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3482 } __ else_(); {
3484 // logging buffer is full, call the runtime
3485 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3486 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", orig, tls);
3487 } __ end_if(); // (!index)
3488 } __ end_if(); // (orig != NULL)
3489 } __ end_if(); // (!marking)
3491 // Final sync IdealKit and GraphKit.
3492 sync_kit(ideal);
3493 }
3495 //
3496 // Update the card table and add card address to the queue
3497 //
3498 void GraphKit::g1_mark_card(IdealKit& ideal,
3499 Node* card_adr,
3500 Node* oop_store,
3501 uint oop_alias_idx,
3502 Node* index,
3503 Node* index_adr,
3504 Node* buffer,
3505 const TypeFunc* tf) {
3507 Node* zero = __ ConI(0);
3508 Node* no_base = __ top();
3509 BasicType card_bt = T_BYTE;
3510 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3511 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3513 // Now do the queue work
3514 __ if_then(index, BoolTest::ne, zero); {
3516 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3517 Node* next_indexX = next_index;
3518 #ifdef _LP64
3519 // We could refine the type for what it's worth
3520 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3521 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3522 #endif // _LP64
3523 Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3525 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3526 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3528 } __ else_(); {
3529 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3530 } __ end_if();
3532 }
3534 void GraphKit::g1_write_barrier_post(Node* oop_store,
3535 Node* obj,
3536 Node* adr,
3537 uint alias_idx,
3538 Node* val,
3539 BasicType bt,
3540 bool use_precise) {
3541 // If we are writing a NULL then we need no post barrier
3543 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3544 // Must be NULL
3545 const Type* t = val->bottom_type();
3546 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3547 // No post barrier if writing NULLx
3548 return;
3549 }
3551 if (!use_precise) {
3552 // All card marks for a (non-array) instance are in one place:
3553 adr = obj;
3554 }
3555 // (Else it's an array (or unknown), and we want more precise card marks.)
3556 assert(adr != NULL, "");
3558 IdealKit ideal(gvn(), control(), merged_memory(), true);
3560 Node* tls = __ thread(); // ThreadLocalStorage
3562 Node* no_base = __ top();
3563 float likely = PROB_LIKELY(0.999);
3564 float unlikely = PROB_UNLIKELY(0.999);
3565 Node* zero = __ ConI(0);
3566 Node* zeroX = __ ConX(0);
3568 // Get the alias_index for raw card-mark memory
3569 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3571 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3573 // Offsets into the thread
3574 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3575 PtrQueue::byte_offset_of_index());
3576 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3577 PtrQueue::byte_offset_of_buf());
3579 // Pointers into the thread
3581 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3582 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3584 // Now some values
3585 // Use ctrl to avoid hoisting these values past a safepoint, which could
3586 // potentially reset these fields in the JavaThread.
3587 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3588 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3590 // Convert the store obj pointer to an int prior to doing math on it
3591 // Must use ctrl to prevent "integerized oop" existing across safepoint
3592 Node* cast = __ CastPX(__ ctrl(), adr);
3594 // Divide pointer by card size
3595 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3597 // Combine card table base and card offset
3598 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3600 // If we know the value being stored does it cross regions?
3602 if (val != NULL) {
3603 // Does the store cause us to cross regions?
3605 // Should be able to do an unsigned compare of region_size instead of
3606 // and extra shift. Do we have an unsigned compare??
3607 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3608 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3610 // if (xor_res == 0) same region so skip
3611 __ if_then(xor_res, BoolTest::ne, zeroX); {
3613 // No barrier if we are storing a NULL
3614 __ if_then(val, BoolTest::ne, null(), unlikely); {
3616 // Ok must mark the card if not already dirty
3618 // load the original value of the card
3619 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3621 __ if_then(card_val, BoolTest::ne, zero); {
3622 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3623 } __ end_if();
3624 } __ end_if();
3625 } __ end_if();
3626 } else {
3627 // Object.clone() instrinsic uses this path.
3628 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3629 }
3631 // Final sync IdealKit and GraphKit.
3632 sync_kit(ideal);
3633 }
3634 #undef __