1 /*
2 * Copyright (c) 2001, 2011, 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 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc_implementation/g1/heapRegion.hpp"
29 #include "gc_interface/collectedHeap.hpp"
30 #include "memory/barrierSet.hpp"
31 #include "memory/cardTableModRefBS.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/graphKit.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/locknode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/rootnode.hpp"
39 #include "opto/runtime.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/sharedRuntime.hpp"
42
43 //----------------------------GraphKit-----------------------------------------
44 // Main utility constructor.
45 GraphKit::GraphKit(JVMState* jvms)
46 : Phase(Phase::Parser),
47 _env(C->env()),
48 _gvn(*C->initial_gvn())
49 {
50 _exceptions = jvms->map()->next_exception();
51 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
52 set_jvms(jvms);
53 }
54
55 // Private constructor for parser.
56 GraphKit::GraphKit()
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn())
60 {
61 _exceptions = NULL;
62 set_map(NULL);
63 debug_only(_sp = -99);
64 debug_only(set_bci(-99));
65 }
66
67
68
69 //---------------------------clean_stack---------------------------------------
70 // Clear away rubbish from the stack area of the JVM state.
71 // This destroys any arguments that may be waiting on the stack.
72 void GraphKit::clean_stack(int from_sp) {
73 SafePointNode* map = this->map();
74 JVMState* jvms = this->jvms();
75 int stk_size = jvms->stk_size();
76 int stkoff = jvms->stkoff();
77 Node* top = this->top();
78 for (int i = from_sp; i < stk_size; i++) {
79 if (map->in(stkoff + i) != top) {
80 map->set_req(stkoff + i, top);
81 }
82 }
83 }
84
85
86 //--------------------------------sync_jvms-----------------------------------
87 // Make sure our current jvms agrees with our parse state.
88 JVMState* GraphKit::sync_jvms() const {
89 JVMState* jvms = this->jvms();
90 jvms->set_bci(bci()); // Record the new bci in the JVMState
91 jvms->set_sp(sp()); // Record the new sp in the JVMState
92 assert(jvms_in_sync(), "jvms is now in sync");
93 return jvms;
94 }
95
96 #ifdef ASSERT
97 bool GraphKit::jvms_in_sync() const {
98 Parse* parse = is_Parse();
99 if (parse == NULL) {
100 if (bci() != jvms()->bci()) return false;
101 if (sp() != (int)jvms()->sp()) return false;
102 return true;
103 }
104 if (jvms()->method() != parse->method()) return false;
105 if (jvms()->bci() != parse->bci()) return false;
106 int jvms_sp = jvms()->sp();
107 if (jvms_sp != parse->sp()) return false;
108 int jvms_depth = jvms()->depth();
109 if (jvms_depth != parse->depth()) return false;
110 return true;
111 }
112
113 // Local helper checks for special internal merge points
114 // used to accumulate and merge exception states.
115 // They are marked by the region's in(0) edge being the map itself.
116 // Such merge points must never "escape" into the parser at large,
117 // until they have been handed to gvn.transform.
118 static bool is_hidden_merge(Node* reg) {
119 if (reg == NULL) return false;
120 if (reg->is_Phi()) {
121 reg = reg->in(0);
122 if (reg == NULL) return false;
123 }
124 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
125 }
126
127 void GraphKit::verify_map() const {
128 if (map() == NULL) return; // null map is OK
129 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
130 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
131 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
132 }
133
134 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
135 assert(ex_map->next_exception() == NULL, "not already part of a chain");
136 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
137 }
138 #endif
139
140 //---------------------------stop_and_kill_map---------------------------------
141 // Set _map to NULL, signalling a stop to further bytecode execution.
142 // First smash the current map's control to a constant, to mark it dead.
143 void GraphKit::stop_and_kill_map() {
144 SafePointNode* dead_map = stop();
145 if (dead_map != NULL) {
146 dead_map->disconnect_inputs(NULL); // Mark the map as killed.
147 assert(dead_map->is_killed(), "must be so marked");
148 }
149 }
150
151
152 //--------------------------------stopped--------------------------------------
153 // Tell if _map is NULL, or control is top.
154 bool GraphKit::stopped() {
155 if (map() == NULL) return true;
156 else if (control() == top()) return true;
157 else return false;
158 }
159
160
161 //-----------------------------has_ex_handler----------------------------------
162 // Tell if this method or any caller method has exception handlers.
163 bool GraphKit::has_ex_handler() {
164 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
165 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
166 return true;
167 }
168 }
169 return false;
170 }
171
172 //------------------------------save_ex_oop------------------------------------
173 // Save an exception without blowing stack contents or other JVM state.
174 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
175 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
176 ex_map->add_req(ex_oop);
177 debug_only(verify_exception_state(ex_map));
178 }
179
180 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
181 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
182 Node* ex_oop = ex_map->in(ex_map->req()-1);
183 if (clear_it) ex_map->del_req(ex_map->req()-1);
184 return ex_oop;
185 }
186
187 //-----------------------------saved_ex_oop------------------------------------
188 // Recover a saved exception from its map.
189 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
190 return common_saved_ex_oop(ex_map, false);
191 }
192
193 //--------------------------clear_saved_ex_oop---------------------------------
194 // Erase a previously saved exception from its map.
195 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
196 return common_saved_ex_oop(ex_map, true);
197 }
198
199 #ifdef ASSERT
200 //---------------------------has_saved_ex_oop----------------------------------
201 // Erase a previously saved exception from its map.
202 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
203 return ex_map->req() == ex_map->jvms()->endoff()+1;
204 }
205 #endif
206
207 //-------------------------make_exception_state--------------------------------
208 // Turn the current JVM state into an exception state, appending the ex_oop.
209 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
210 sync_jvms();
211 SafePointNode* ex_map = stop(); // do not manipulate this map any more
212 set_saved_ex_oop(ex_map, ex_oop);
213 return ex_map;
214 }
215
216
217 //--------------------------add_exception_state--------------------------------
218 // Add an exception to my list of exceptions.
219 void GraphKit::add_exception_state(SafePointNode* ex_map) {
220 if (ex_map == NULL || ex_map->control() == top()) {
221 return;
222 }
223 #ifdef ASSERT
224 verify_exception_state(ex_map);
225 if (has_exceptions()) {
226 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
227 }
228 #endif
229
230 // If there is already an exception of exactly this type, merge with it.
231 // In particular, null-checks and other low-level exceptions common up here.
232 Node* ex_oop = saved_ex_oop(ex_map);
233 const Type* ex_type = _gvn.type(ex_oop);
234 if (ex_oop == top()) {
235 // No action needed.
236 return;
237 }
238 assert(ex_type->isa_instptr(), "exception must be an instance");
239 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
240 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
241 // We check sp also because call bytecodes can generate exceptions
242 // both before and after arguments are popped!
243 if (ex_type2 == ex_type
244 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
245 combine_exception_states(ex_map, e2);
246 return;
247 }
248 }
249
250 // No pre-existing exception of the same type. Chain it on the list.
251 push_exception_state(ex_map);
252 }
253
254 //-----------------------add_exception_states_from-----------------------------
255 void GraphKit::add_exception_states_from(JVMState* jvms) {
256 SafePointNode* ex_map = jvms->map()->next_exception();
257 if (ex_map != NULL) {
258 jvms->map()->set_next_exception(NULL);
259 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
260 next_map = ex_map->next_exception();
261 ex_map->set_next_exception(NULL);
262 add_exception_state(ex_map);
263 }
264 }
265 }
266
267 //-----------------------transfer_exceptions_into_jvms-------------------------
268 JVMState* GraphKit::transfer_exceptions_into_jvms() {
269 if (map() == NULL) {
270 // We need a JVMS to carry the exceptions, but the map has gone away.
271 // Create a scratch JVMS, cloned from any of the exception states...
272 if (has_exceptions()) {
273 _map = _exceptions;
274 _map = clone_map();
275 _map->set_next_exception(NULL);
276 clear_saved_ex_oop(_map);
277 debug_only(verify_map());
278 } else {
279 // ...or created from scratch
280 JVMState* jvms = new (C) JVMState(_method, NULL);
281 jvms->set_bci(_bci);
282 jvms->set_sp(_sp);
283 jvms->set_map(new (C, TypeFunc::Parms) SafePointNode(TypeFunc::Parms, jvms));
284 set_jvms(jvms);
285 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
286 set_all_memory(top());
287 while (map()->req() < jvms->endoff()) map()->add_req(top());
288 }
289 // (This is a kludge, in case you didn't notice.)
290 set_control(top());
291 }
292 JVMState* jvms = sync_jvms();
293 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
294 jvms->map()->set_next_exception(_exceptions);
295 _exceptions = NULL; // done with this set of exceptions
296 return jvms;
297 }
298
299 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
300 assert(is_hidden_merge(dstphi), "must be a special merge node");
301 assert(is_hidden_merge(srcphi), "must be a special merge node");
302 uint limit = srcphi->req();
303 for (uint i = PhiNode::Input; i < limit; i++) {
304 dstphi->add_req(srcphi->in(i));
305 }
306 }
307 static inline void add_one_req(Node* dstphi, Node* src) {
308 assert(is_hidden_merge(dstphi), "must be a special merge node");
309 assert(!is_hidden_merge(src), "must not be a special merge node");
310 dstphi->add_req(src);
311 }
312
313 //-----------------------combine_exception_states------------------------------
314 // This helper function combines exception states by building phis on a
315 // specially marked state-merging region. These regions and phis are
316 // untransformed, and can build up gradually. The region is marked by
317 // having a control input of its exception map, rather than NULL. Such
318 // regions do not appear except in this function, and in use_exception_state.
319 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
320 if (failing()) return; // dying anyway...
321 JVMState* ex_jvms = ex_map->_jvms;
322 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
323 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
324 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
325 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
326 assert(ex_map->req() == phi_map->req(), "matching maps");
327 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
328 Node* hidden_merge_mark = root();
329 Node* region = phi_map->control();
330 MergeMemNode* phi_mem = phi_map->merged_memory();
331 MergeMemNode* ex_mem = ex_map->merged_memory();
332 if (region->in(0) != hidden_merge_mark) {
333 // The control input is not (yet) a specially-marked region in phi_map.
334 // Make it so, and build some phis.
335 region = new (C, 2) RegionNode(2);
336 _gvn.set_type(region, Type::CONTROL);
337 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
338 region->init_req(1, phi_map->control());
339 phi_map->set_control(region);
340 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
341 record_for_igvn(io_phi);
342 _gvn.set_type(io_phi, Type::ABIO);
343 phi_map->set_i_o(io_phi);
344 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
345 Node* m = mms.memory();
346 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
347 record_for_igvn(m_phi);
348 _gvn.set_type(m_phi, Type::MEMORY);
349 mms.set_memory(m_phi);
350 }
351 }
352
353 // Either or both of phi_map and ex_map might already be converted into phis.
354 Node* ex_control = ex_map->control();
355 // if there is special marking on ex_map also, we add multiple edges from src
356 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
357 // how wide was the destination phi_map, originally?
358 uint orig_width = region->req();
359
360 if (add_multiple) {
361 add_n_reqs(region, ex_control);
362 add_n_reqs(phi_map->i_o(), ex_map->i_o());
363 } else {
364 // ex_map has no merges, so we just add single edges everywhere
365 add_one_req(region, ex_control);
366 add_one_req(phi_map->i_o(), ex_map->i_o());
367 }
368 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
369 if (mms.is_empty()) {
370 // get a copy of the base memory, and patch some inputs into it
371 const TypePtr* adr_type = mms.adr_type(C);
372 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
373 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
374 mms.set_memory(phi);
375 // Prepare to append interesting stuff onto the newly sliced phi:
376 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
377 }
378 // Append stuff from ex_map:
379 if (add_multiple) {
380 add_n_reqs(mms.memory(), mms.memory2());
381 } else {
382 add_one_req(mms.memory(), mms.memory2());
383 }
384 }
385 uint limit = ex_map->req();
386 for (uint i = TypeFunc::Parms; i < limit; i++) {
387 // Skip everything in the JVMS after tos. (The ex_oop follows.)
388 if (i == tos) i = ex_jvms->monoff();
389 Node* src = ex_map->in(i);
390 Node* dst = phi_map->in(i);
391 if (src != dst) {
392 PhiNode* phi;
393 if (dst->in(0) != region) {
394 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
395 record_for_igvn(phi);
396 _gvn.set_type(phi, phi->type());
397 phi_map->set_req(i, dst);
398 // Prepare to append interesting stuff onto the new phi:
399 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
400 } else {
401 assert(dst->is_Phi(), "nobody else uses a hidden region");
402 phi = (PhiNode*)dst;
403 }
404 if (add_multiple && src->in(0) == ex_control) {
405 // Both are phis.
406 add_n_reqs(dst, src);
407 } else {
408 while (dst->req() < region->req()) add_one_req(dst, src);
409 }
410 const Type* srctype = _gvn.type(src);
411 if (phi->type() != srctype) {
412 const Type* dsttype = phi->type()->meet(srctype);
413 if (phi->type() != dsttype) {
414 phi->set_type(dsttype);
415 _gvn.set_type(phi, dsttype);
416 }
417 }
418 }
419 }
420 }
421
422 //--------------------------use_exception_state--------------------------------
423 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
424 if (failing()) { stop(); return top(); }
425 Node* region = phi_map->control();
426 Node* hidden_merge_mark = root();
427 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
428 Node* ex_oop = clear_saved_ex_oop(phi_map);
429 if (region->in(0) == hidden_merge_mark) {
430 // Special marking for internal ex-states. Process the phis now.
431 region->set_req(0, region); // now it's an ordinary region
432 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
433 // Note: Setting the jvms also sets the bci and sp.
434 set_control(_gvn.transform(region));
435 uint tos = jvms()->stkoff() + sp();
436 for (uint i = 1; i < tos; i++) {
437 Node* x = phi_map->in(i);
438 if (x->in(0) == region) {
439 assert(x->is_Phi(), "expected a special phi");
440 phi_map->set_req(i, _gvn.transform(x));
441 }
442 }
443 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
444 Node* x = mms.memory();
445 if (x->in(0) == region) {
446 assert(x->is_Phi(), "nobody else uses a hidden region");
447 mms.set_memory(_gvn.transform(x));
448 }
449 }
450 if (ex_oop->in(0) == region) {
451 assert(ex_oop->is_Phi(), "expected a special phi");
452 ex_oop = _gvn.transform(ex_oop);
453 }
454 } else {
455 set_jvms(phi_map->jvms());
456 }
457
458 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
459 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
460 return ex_oop;
461 }
462
463 //---------------------------------java_bc-------------------------------------
464 Bytecodes::Code GraphKit::java_bc() const {
465 ciMethod* method = this->method();
466 int bci = this->bci();
467 if (method != NULL && bci != InvocationEntryBci)
468 return method->java_code_at_bci(bci);
469 else
470 return Bytecodes::_illegal;
471 }
472
473 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
474 bool must_throw) {
475 // if the exception capability is set, then we will generate code
476 // to check the JavaThread.should_post_on_exceptions flag to see
477 // if we actually need to report exception events (for this
478 // thread). If we don't need to report exception events, we will
479 // take the normal fast path provided by add_exception_events. If
480 // exception event reporting is enabled for this thread, we will
481 // take the uncommon_trap in the BuildCutout below.
482
483 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
484 Node* jthread = _gvn.transform(new (C, 1) ThreadLocalNode());
485 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
486 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, false);
487
488 // Test the should_post_on_exceptions_flag vs. 0
489 Node* chk = _gvn.transform( new (C, 3) CmpINode(should_post_flag, intcon(0)) );
490 Node* tst = _gvn.transform( new (C, 2) BoolNode(chk, BoolTest::eq) );
491
492 // Branch to slow_path if should_post_on_exceptions_flag was true
493 { BuildCutout unless(this, tst, PROB_MAX);
494 // Do not try anything fancy if we're notifying the VM on every throw.
495 // Cf. case Bytecodes::_athrow in parse2.cpp.
496 uncommon_trap(reason, Deoptimization::Action_none,
497 (ciKlass*)NULL, (char*)NULL, must_throw);
498 }
499
500 }
501
502 //------------------------------builtin_throw----------------------------------
503 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
504 bool must_throw = true;
505
506 if (env()->jvmti_can_post_on_exceptions()) {
507 // check if we must post exception events, take uncommon trap if so
508 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
509 // here if should_post_on_exceptions is false
510 // continue on with the normal codegen
511 }
512
513 // If this particular condition has not yet happened at this
514 // bytecode, then use the uncommon trap mechanism, and allow for
515 // a future recompilation if several traps occur here.
516 // If the throw is hot, try to use a more complicated inline mechanism
517 // which keeps execution inside the compiled code.
518 bool treat_throw_as_hot = false;
519 ciMethodData* md = method()->method_data();
520
521 if (ProfileTraps) {
522 if (too_many_traps(reason)) {
523 treat_throw_as_hot = true;
524 }
525 // (If there is no MDO at all, assume it is early in
526 // execution, and that any deopts are part of the
527 // startup transient, and don't need to be remembered.)
528
529 // Also, if there is a local exception handler, treat all throws
530 // as hot if there has been at least one in this method.
531 if (C->trap_count(reason) != 0
532 && method()->method_data()->trap_count(reason) != 0
533 && has_ex_handler()) {
534 treat_throw_as_hot = true;
535 }
536 }
537
538 // If this throw happens frequently, an uncommon trap might cause
539 // a performance pothole. If there is a local exception handler,
540 // and if this particular bytecode appears to be deoptimizing often,
541 // let us handle the throw inline, with a preconstructed instance.
542 // Note: If the deopt count has blown up, the uncommon trap
543 // runtime is going to flush this nmethod, not matter what.
544 if (treat_throw_as_hot
545 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
546 // If the throw is local, we use a pre-existing instance and
547 // punt on the backtrace. This would lead to a missing backtrace
548 // (a repeat of 4292742) if the backtrace object is ever asked
549 // for its backtrace.
550 // Fixing this remaining case of 4292742 requires some flavor of
551 // escape analysis. Leave that for the future.
552 ciInstance* ex_obj = NULL;
553 switch (reason) {
554 case Deoptimization::Reason_null_check:
555 ex_obj = env()->NullPointerException_instance();
556 break;
557 case Deoptimization::Reason_div0_check:
558 ex_obj = env()->ArithmeticException_instance();
559 break;
560 case Deoptimization::Reason_range_check:
561 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
562 break;
563 case Deoptimization::Reason_class_check:
564 if (java_bc() == Bytecodes::_aastore) {
565 ex_obj = env()->ArrayStoreException_instance();
566 } else {
567 ex_obj = env()->ClassCastException_instance();
568 }
569 break;
570 }
571 if (failing()) { stop(); return; } // exception allocation might fail
572 if (ex_obj != NULL) {
573 // Cheat with a preallocated exception object.
574 if (C->log() != NULL)
575 C->log()->elem("hot_throw preallocated='1' reason='%s'",
576 Deoptimization::trap_reason_name(reason));
577 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
578 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
579
580 // Clear the detail message of the preallocated exception object.
581 // Weblogic sometimes mutates the detail message of exceptions
582 // using reflection.
583 int offset = java_lang_Throwable::get_detailMessage_offset();
584 const TypePtr* adr_typ = ex_con->add_offset(offset);
585
586 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
587 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
588 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT);
589
590 add_exception_state(make_exception_state(ex_node));
591 return;
592 }
593 }
594
595 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
596 // It won't be much cheaper than bailing to the interp., since we'll
597 // have to pass up all the debug-info, and the runtime will have to
598 // create the stack trace.
599
600 // Usual case: Bail to interpreter.
601 // Reserve the right to recompile if we haven't seen anything yet.
602
603 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
604 if (treat_throw_as_hot
605 && (method()->method_data()->trap_recompiled_at(bci())
606 || C->too_many_traps(reason))) {
607 // We cannot afford to take more traps here. Suffer in the interpreter.
608 if (C->log() != NULL)
609 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
610 Deoptimization::trap_reason_name(reason),
611 C->trap_count(reason));
612 action = Deoptimization::Action_none;
613 }
614
615 // "must_throw" prunes the JVM state to include only the stack, if there
616 // are no local exception handlers. This should cut down on register
617 // allocation time and code size, by drastically reducing the number
618 // of in-edges on the call to the uncommon trap.
619
620 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
621 }
622
623
624 //----------------------------PreserveJVMState---------------------------------
625 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
626 debug_only(kit->verify_map());
627 _kit = kit;
628 _map = kit->map(); // preserve the map
629 _sp = kit->sp();
630 kit->set_map(clone_map ? kit->clone_map() : NULL);
631 #ifdef ASSERT
632 _bci = kit->bci();
633 Parse* parser = kit->is_Parse();
634 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
635 _block = block;
636 #endif
637 }
638 PreserveJVMState::~PreserveJVMState() {
639 GraphKit* kit = _kit;
640 #ifdef ASSERT
641 assert(kit->bci() == _bci, "bci must not shift");
642 Parse* parser = kit->is_Parse();
643 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
644 assert(block == _block, "block must not shift");
645 #endif
646 kit->set_map(_map);
647 kit->set_sp(_sp);
648 }
649
650
651 //-----------------------------BuildCutout-------------------------------------
652 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
653 : PreserveJVMState(kit)
654 {
655 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
656 SafePointNode* outer_map = _map; // preserved map is caller's
657 SafePointNode* inner_map = kit->map();
658 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
659 outer_map->set_control(kit->gvn().transform( new (kit->C, 1) IfTrueNode(iff) ));
660 inner_map->set_control(kit->gvn().transform( new (kit->C, 1) IfFalseNode(iff) ));
661 }
662 BuildCutout::~BuildCutout() {
663 GraphKit* kit = _kit;
664 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
665 }
666
667 //---------------------------PreserveReexecuteState----------------------------
668 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
669 assert(!kit->stopped(), "must call stopped() before");
670 _kit = kit;
671 _sp = kit->sp();
672 _reexecute = kit->jvms()->_reexecute;
673 }
674 PreserveReexecuteState::~PreserveReexecuteState() {
675 if (_kit->stopped()) return;
676 _kit->jvms()->_reexecute = _reexecute;
677 _kit->set_sp(_sp);
678 }
679
680 //------------------------------clone_map--------------------------------------
681 // Implementation of PreserveJVMState
682 //
683 // Only clone_map(...) here. If this function is only used in the
684 // PreserveJVMState class we may want to get rid of this extra
685 // function eventually and do it all there.
686
687 SafePointNode* GraphKit::clone_map() {
688 if (map() == NULL) return NULL;
689
690 // Clone the memory edge first
691 Node* mem = MergeMemNode::make(C, map()->memory());
692 gvn().set_type_bottom(mem);
693
694 SafePointNode *clonemap = (SafePointNode*)map()->clone();
695 JVMState* jvms = this->jvms();
696 JVMState* clonejvms = jvms->clone_shallow(C);
697 clonemap->set_memory(mem);
698 clonemap->set_jvms(clonejvms);
699 clonejvms->set_map(clonemap);
700 record_for_igvn(clonemap);
701 gvn().set_type_bottom(clonemap);
702 return clonemap;
703 }
704
705
706 //-----------------------------set_map_clone-----------------------------------
707 void GraphKit::set_map_clone(SafePointNode* m) {
708 _map = m;
709 _map = clone_map();
710 _map->set_next_exception(NULL);
711 debug_only(verify_map());
712 }
713
714
715 //----------------------------kill_dead_locals---------------------------------
716 // Detect any locals which are known to be dead, and force them to top.
717 void GraphKit::kill_dead_locals() {
718 // Consult the liveness information for the locals. If any
719 // of them are unused, then they can be replaced by top(). This
720 // should help register allocation time and cut down on the size
721 // of the deoptimization information.
722
723 // This call is made from many of the bytecode handling
724 // subroutines called from the Big Switch in do_one_bytecode.
725 // Every bytecode which might include a slow path is responsible
726 // for killing its dead locals. The more consistent we
727 // are about killing deads, the fewer useless phis will be
728 // constructed for them at various merge points.
729
730 // bci can be -1 (InvocationEntryBci). We return the entry
731 // liveness for the method.
732
733 if (method() == NULL || method()->code_size() == 0) {
734 // We are building a graph for a call to a native method.
735 // All locals are live.
736 return;
737 }
738
739 ResourceMark rm;
740
741 // Consult the liveness information for the locals. If any
742 // of them are unused, then they can be replaced by top(). This
743 // should help register allocation time and cut down on the size
744 // of the deoptimization information.
745 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
746
747 int len = (int)live_locals.size();
748 assert(len <= jvms()->loc_size(), "too many live locals");
749 for (int local = 0; local < len; local++) {
750 if (!live_locals.at(local)) {
751 set_local(local, top());
752 }
753 }
754 }
755
756 #ifdef ASSERT
757 //-------------------------dead_locals_are_killed------------------------------
758 // Return true if all dead locals are set to top in the map.
759 // Used to assert "clean" debug info at various points.
760 bool GraphKit::dead_locals_are_killed() {
761 if (method() == NULL || method()->code_size() == 0) {
762 // No locals need to be dead, so all is as it should be.
763 return true;
764 }
765
766 // Make sure somebody called kill_dead_locals upstream.
767 ResourceMark rm;
768 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
769 if (jvms->loc_size() == 0) continue; // no locals to consult
770 SafePointNode* map = jvms->map();
771 ciMethod* method = jvms->method();
772 int bci = jvms->bci();
773 if (jvms == this->jvms()) {
774 bci = this->bci(); // it might not yet be synched
775 }
776 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
777 int len = (int)live_locals.size();
778 if (!live_locals.is_valid() || len == 0)
779 // This method is trivial, or is poisoned by a breakpoint.
780 return true;
781 assert(len == jvms->loc_size(), "live map consistent with locals map");
782 for (int local = 0; local < len; local++) {
783 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
784 if (PrintMiscellaneous && (Verbose || WizardMode)) {
785 tty->print_cr("Zombie local %d: ", local);
786 jvms->dump();
787 }
788 return false;
789 }
790 }
791 }
792 return true;
793 }
794
795 #endif //ASSERT
796
797 // Helper function for enforcing certain bytecodes to reexecute if
798 // deoptimization happens
799 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
800 ciMethod* cur_method = jvms->method();
801 int cur_bci = jvms->bci();
802 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
803 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
804 return Interpreter::bytecode_should_reexecute(code) ||
805 is_anewarray && code == Bytecodes::_multianewarray;
806 // Reexecute _multianewarray bytecode which was replaced with
807 // sequence of [a]newarray. See Parse::do_multianewarray().
808 //
809 // Note: interpreter should not have it set since this optimization
810 // is limited by dimensions and guarded by flag so in some cases
811 // multianewarray() runtime calls will be generated and
812 // the bytecode should not be reexecutes (stack will not be reset).
813 } else
814 return false;
815 }
816
817 // Helper function for adding JVMState and debug information to node
818 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
819 // Add the safepoint edges to the call (or other safepoint).
820
821 // Make sure dead locals are set to top. This
822 // should help register allocation time and cut down on the size
823 // of the deoptimization information.
824 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
825
826 // Walk the inline list to fill in the correct set of JVMState's
827 // Also fill in the associated edges for each JVMState.
828
829 JVMState* youngest_jvms = sync_jvms();
830
831 // If we are guaranteed to throw, we can prune everything but the
832 // input to the current bytecode.
833 bool can_prune_locals = false;
834 uint stack_slots_not_pruned = 0;
835 int inputs = 0, depth = 0;
836 if (must_throw) {
837 assert(method() == youngest_jvms->method(), "sanity");
838 if (compute_stack_effects(inputs, depth)) {
839 can_prune_locals = true;
840 stack_slots_not_pruned = inputs;
841 }
842 }
843
844 if (env()->jvmti_can_access_local_variables()) {
845 // At any safepoint, this method can get breakpointed, which would
846 // then require an immediate deoptimization.
847 can_prune_locals = false; // do not prune locals
848 stack_slots_not_pruned = 0;
849 }
850
851 // do not scribble on the input jvms
852 JVMState* out_jvms = youngest_jvms->clone_deep(C);
853 call->set_jvms(out_jvms); // Start jvms list for call node
854
855 // For a known set of bytecodes, the interpreter should reexecute them if
856 // deoptimization happens. We set the reexecute state for them here
857 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
858 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
859 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
860 }
861
862 // Presize the call:
863 debug_only(uint non_debug_edges = call->req());
864 call->add_req_batch(top(), youngest_jvms->debug_depth());
865 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
866
867 // Set up edges so that the call looks like this:
868 // Call [state:] ctl io mem fptr retadr
869 // [parms:] parm0 ... parmN
870 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
871 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
872 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
873 // Note that caller debug info precedes callee debug info.
874
875 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
876 uint debug_ptr = call->req();
877
878 // Loop over the map input edges associated with jvms, add them
879 // to the call node, & reset all offsets to match call node array.
880 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
881 uint debug_end = debug_ptr;
882 uint debug_start = debug_ptr - in_jvms->debug_size();
883 debug_ptr = debug_start; // back up the ptr
884
885 uint p = debug_start; // walks forward in [debug_start, debug_end)
886 uint j, k, l;
887 SafePointNode* in_map = in_jvms->map();
888 out_jvms->set_map(call);
889
890 if (can_prune_locals) {
891 assert(in_jvms->method() == out_jvms->method(), "sanity");
892 // If the current throw can reach an exception handler in this JVMS,
893 // then we must keep everything live that can reach that handler.
894 // As a quick and dirty approximation, we look for any handlers at all.
895 if (in_jvms->method()->has_exception_handlers()) {
896 can_prune_locals = false;
897 }
898 }
899
900 // Add the Locals
901 k = in_jvms->locoff();
902 l = in_jvms->loc_size();
903 out_jvms->set_locoff(p);
904 if (!can_prune_locals) {
905 for (j = 0; j < l; j++)
906 call->set_req(p++, in_map->in(k+j));
907 } else {
908 p += l; // already set to top above by add_req_batch
909 }
910
911 // Add the Expression Stack
912 k = in_jvms->stkoff();
913 l = in_jvms->sp();
914 out_jvms->set_stkoff(p);
915 if (!can_prune_locals) {
916 for (j = 0; j < l; j++)
917 call->set_req(p++, in_map->in(k+j));
918 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
919 // Divide stack into {S0,...,S1}, where S0 is set to top.
920 uint s1 = stack_slots_not_pruned;
921 stack_slots_not_pruned = 0; // for next iteration
922 if (s1 > l) s1 = l;
923 uint s0 = l - s1;
924 p += s0; // skip the tops preinstalled by add_req_batch
925 for (j = s0; j < l; j++)
926 call->set_req(p++, in_map->in(k+j));
927 } else {
928 p += l; // already set to top above by add_req_batch
929 }
930
931 // Add the Monitors
932 k = in_jvms->monoff();
933 l = in_jvms->mon_size();
934 out_jvms->set_monoff(p);
935 for (j = 0; j < l; j++)
936 call->set_req(p++, in_map->in(k+j));
937
938 // Copy any scalar object fields.
939 k = in_jvms->scloff();
940 l = in_jvms->scl_size();
941 out_jvms->set_scloff(p);
942 for (j = 0; j < l; j++)
943 call->set_req(p++, in_map->in(k+j));
944
945 // Finish the new jvms.
946 out_jvms->set_endoff(p);
947
948 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
949 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
950 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
951 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
952 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
953 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
954
955 // Update the two tail pointers in parallel.
956 out_jvms = out_jvms->caller();
957 in_jvms = in_jvms->caller();
958 }
959
960 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
961
962 // Test the correctness of JVMState::debug_xxx accessors:
963 assert(call->jvms()->debug_start() == non_debug_edges, "");
964 assert(call->jvms()->debug_end() == call->req(), "");
965 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
966 }
967
968 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
969 Bytecodes::Code code = java_bc();
970 if (code == Bytecodes::_wide) {
971 code = method()->java_code_at_bci(bci() + 1);
972 }
973
974 BasicType rtype = T_ILLEGAL;
975 int rsize = 0;
976
977 if (code != Bytecodes::_illegal) {
978 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
979 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
980 if (rtype < T_CONFLICT)
981 rsize = type2size[rtype];
982 }
983
984 switch (code) {
985 case Bytecodes::_illegal:
986 return false;
987
988 case Bytecodes::_ldc:
989 case Bytecodes::_ldc_w:
990 case Bytecodes::_ldc2_w:
991 inputs = 0;
992 break;
993
994 case Bytecodes::_dup: inputs = 1; break;
995 case Bytecodes::_dup_x1: inputs = 2; break;
996 case Bytecodes::_dup_x2: inputs = 3; break;
997 case Bytecodes::_dup2: inputs = 2; break;
998 case Bytecodes::_dup2_x1: inputs = 3; break;
999 case Bytecodes::_dup2_x2: inputs = 4; break;
1000 case Bytecodes::_swap: inputs = 2; break;
1001 case Bytecodes::_arraylength: inputs = 1; break;
1002
1003 case Bytecodes::_getstatic:
1004 case Bytecodes::_putstatic:
1005 case Bytecodes::_getfield:
1006 case Bytecodes::_putfield:
1007 {
1008 bool is_get = (depth >= 0), is_static = (depth & 1);
1009 bool ignore;
1010 ciBytecodeStream iter(method());
1011 iter.reset_to_bci(bci());
1012 iter.next();
1013 ciField* field = iter.get_field(ignore);
1014 int size = field->type()->size();
1015 inputs = (is_static ? 0 : 1);
1016 if (is_get) {
1017 depth = size - inputs;
1018 } else {
1019 inputs += size; // putxxx pops the value from the stack
1020 depth = - inputs;
1021 }
1022 }
1023 break;
1024
1025 case Bytecodes::_invokevirtual:
1026 case Bytecodes::_invokespecial:
1027 case Bytecodes::_invokestatic:
1028 case Bytecodes::_invokedynamic:
1029 case Bytecodes::_invokeinterface:
1030 {
1031 bool ignore;
1032 ciBytecodeStream iter(method());
1033 iter.reset_to_bci(bci());
1034 iter.next();
1035 ciMethod* method = iter.get_method(ignore);
1036 inputs = method->arg_size_no_receiver();
1037 // Add a receiver argument, maybe:
1038 if (code != Bytecodes::_invokestatic &&
1039 code != Bytecodes::_invokedynamic)
1040 inputs += 1;
1041 // (Do not use ciMethod::arg_size(), because
1042 // it might be an unloaded method, which doesn't
1043 // know whether it is static or not.)
1044 int size = method->return_type()->size();
1045 depth = size - inputs;
1046 }
1047 break;
1048
1049 case Bytecodes::_multianewarray:
1050 {
1051 ciBytecodeStream iter(method());
1052 iter.reset_to_bci(bci());
1053 iter.next();
1054 inputs = iter.get_dimensions();
1055 assert(rsize == 1, "");
1056 depth = rsize - inputs;
1057 }
1058 break;
1059
1060 case Bytecodes::_ireturn:
1061 case Bytecodes::_lreturn:
1062 case Bytecodes::_freturn:
1063 case Bytecodes::_dreturn:
1064 case Bytecodes::_areturn:
1065 assert(rsize = -depth, "");
1066 inputs = rsize;
1067 break;
1068
1069 case Bytecodes::_jsr:
1070 case Bytecodes::_jsr_w:
1071 inputs = 0;
1072 depth = 1; // S.B. depth=1, not zero
1073 break;
1074
1075 default:
1076 // bytecode produces a typed result
1077 inputs = rsize - depth;
1078 assert(inputs >= 0, "");
1079 break;
1080 }
1081
1082 #ifdef ASSERT
1083 // spot check
1084 int outputs = depth + inputs;
1085 assert(outputs >= 0, "sanity");
1086 switch (code) {
1087 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1088 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1089 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1090 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1091 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1092 }
1093 #endif //ASSERT
1094
1095 return true;
1096 }
1097
1098
1099
1100 //------------------------------basic_plus_adr---------------------------------
1101 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1102 // short-circuit a common case
1103 if (offset == intcon(0)) return ptr;
1104 return _gvn.transform( new (C, 4) AddPNode(base, ptr, offset) );
1105 }
1106
1107 Node* GraphKit::ConvI2L(Node* offset) {
1108 // short-circuit a common case
1109 jint offset_con = find_int_con(offset, Type::OffsetBot);
1110 if (offset_con != Type::OffsetBot) {
1111 return longcon((long) offset_con);
1112 }
1113 return _gvn.transform( new (C, 2) ConvI2LNode(offset));
1114 }
1115 Node* GraphKit::ConvL2I(Node* offset) {
1116 // short-circuit a common case
1117 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1118 if (offset_con != (jlong)Type::OffsetBot) {
1119 return intcon((int) offset_con);
1120 }
1121 return _gvn.transform( new (C, 2) ConvL2INode(offset));
1122 }
1123
1124 //-------------------------load_object_klass-----------------------------------
1125 Node* GraphKit::load_object_klass(Node* obj) {
1126 // Special-case a fresh allocation to avoid building nodes:
1127 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1128 if (akls != NULL) return akls;
1129 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1130 return _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS) );
1131 }
1132
1133 //-------------------------load_array_length-----------------------------------
1134 Node* GraphKit::load_array_length(Node* array) {
1135 // Special-case a fresh allocation to avoid building nodes:
1136 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1137 Node *alen;
1138 if (alloc == NULL) {
1139 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1140 alen = _gvn.transform( new (C, 3) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1141 } else {
1142 alen = alloc->Ideal_length();
1143 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1144 if (ccast != alen) {
1145 alen = _gvn.transform(ccast);
1146 }
1147 }
1148 return alen;
1149 }
1150
1151 //------------------------------do_null_check----------------------------------
1152 // Helper function to do a NULL pointer check. Returned value is
1153 // the incoming address with NULL casted away. You are allowed to use the
1154 // not-null value only if you are control dependent on the test.
1155 extern int explicit_null_checks_inserted,
1156 explicit_null_checks_elided;
1157 Node* GraphKit::null_check_common(Node* value, BasicType type,
1158 // optional arguments for variations:
1159 bool assert_null,
1160 Node* *null_control) {
1161 assert(!assert_null || null_control == NULL, "not both at once");
1162 if (stopped()) return top();
1163 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1164 // For some performance testing, we may wish to suppress null checking.
1165 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1166 return value;
1167 }
1168 explicit_null_checks_inserted++;
1169
1170 // Construct NULL check
1171 Node *chk = NULL;
1172 switch(type) {
1173 case T_LONG : chk = new (C, 3) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1174 case T_INT : chk = new (C, 3) CmpINode( value, _gvn.intcon(0)); break;
1175 case T_ARRAY : // fall through
1176 type = T_OBJECT; // simplify further tests
1177 case T_OBJECT : {
1178 const Type *t = _gvn.type( value );
1179
1180 const TypeOopPtr* tp = t->isa_oopptr();
1181 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1182 // Only for do_null_check, not any of its siblings:
1183 && !assert_null && null_control == NULL) {
1184 // Usually, any field access or invocation on an unloaded oop type
1185 // will simply fail to link, since the statically linked class is
1186 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1187 // the static class is loaded but the sharper oop type is not.
1188 // Rather than checking for this obscure case in lots of places,
1189 // we simply observe that a null check on an unloaded class
1190 // will always be followed by a nonsense operation, so we
1191 // can just issue the uncommon trap here.
1192 // Our access to the unloaded class will only be correct
1193 // after it has been loaded and initialized, which requires
1194 // a trip through the interpreter.
1195 #ifndef PRODUCT
1196 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1197 #endif
1198 uncommon_trap(Deoptimization::Reason_unloaded,
1199 Deoptimization::Action_reinterpret,
1200 tp->klass(), "!loaded");
1201 return top();
1202 }
1203
1204 if (assert_null) {
1205 // See if the type is contained in NULL_PTR.
1206 // If so, then the value is already null.
1207 if (t->higher_equal(TypePtr::NULL_PTR)) {
1208 explicit_null_checks_elided++;
1209 return value; // Elided null assert quickly!
1210 }
1211 } else {
1212 // See if mixing in the NULL pointer changes type.
1213 // If so, then the NULL pointer was not allowed in the original
1214 // type. In other words, "value" was not-null.
1215 if (t->meet(TypePtr::NULL_PTR) != t) {
1216 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1217 explicit_null_checks_elided++;
1218 return value; // Elided null check quickly!
1219 }
1220 }
1221 chk = new (C, 3) CmpPNode( value, null() );
1222 break;
1223 }
1224
1225 default : ShouldNotReachHere();
1226 }
1227 assert(chk != NULL, "sanity check");
1228 chk = _gvn.transform(chk);
1229
1230 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1231 BoolNode *btst = new (C, 2) BoolNode( chk, btest);
1232 Node *tst = _gvn.transform( btst );
1233
1234 //-----------
1235 // if peephole optimizations occurred, a prior test existed.
1236 // If a prior test existed, maybe it dominates as we can avoid this test.
1237 if (tst != btst && type == T_OBJECT) {
1238 // At this point we want to scan up the CFG to see if we can
1239 // find an identical test (and so avoid this test altogether).
1240 Node *cfg = control();
1241 int depth = 0;
1242 while( depth < 16 ) { // Limit search depth for speed
1243 if( cfg->Opcode() == Op_IfTrue &&
1244 cfg->in(0)->in(1) == tst ) {
1245 // Found prior test. Use "cast_not_null" to construct an identical
1246 // CastPP (and hence hash to) as already exists for the prior test.
1247 // Return that casted value.
1248 if (assert_null) {
1249 replace_in_map(value, null());
1250 return null(); // do not issue the redundant test
1251 }
1252 Node *oldcontrol = control();
1253 set_control(cfg);
1254 Node *res = cast_not_null(value);
1255 set_control(oldcontrol);
1256 explicit_null_checks_elided++;
1257 return res;
1258 }
1259 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1260 if (cfg == NULL) break; // Quit at region nodes
1261 depth++;
1262 }
1263 }
1264
1265 //-----------
1266 // Branch to failure if null
1267 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1268 Deoptimization::DeoptReason reason;
1269 if (assert_null)
1270 reason = Deoptimization::Reason_null_assert;
1271 else if (type == T_OBJECT)
1272 reason = Deoptimization::Reason_null_check;
1273 else
1274 reason = Deoptimization::Reason_div0_check;
1275
1276 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1277 // ciMethodData::has_trap_at will return a conservative -1 if any
1278 // must-be-null assertion has failed. This could cause performance
1279 // problems for a method after its first do_null_assert failure.
1280 // Consider using 'Reason_class_check' instead?
1281
1282 // To cause an implicit null check, we set the not-null probability
1283 // to the maximum (PROB_MAX). For an explicit check the probability
1284 // is set to a smaller value.
1285 if (null_control != NULL || too_many_traps(reason)) {
1286 // probability is less likely
1287 ok_prob = PROB_LIKELY_MAG(3);
1288 } else if (!assert_null &&
1289 (ImplicitNullCheckThreshold > 0) &&
1290 method() != NULL &&
1291 (method()->method_data()->trap_count(reason)
1292 >= (uint)ImplicitNullCheckThreshold)) {
1293 ok_prob = PROB_LIKELY_MAG(3);
1294 }
1295
1296 if (null_control != NULL) {
1297 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1298 Node* null_true = _gvn.transform( new (C, 1) IfFalseNode(iff));
1299 set_control( _gvn.transform( new (C, 1) IfTrueNode(iff)));
1300 if (null_true == top())
1301 explicit_null_checks_elided++;
1302 (*null_control) = null_true;
1303 } else {
1304 BuildCutout unless(this, tst, ok_prob);
1305 // Check for optimizer eliding test at parse time
1306 if (stopped()) {
1307 // Failure not possible; do not bother making uncommon trap.
1308 explicit_null_checks_elided++;
1309 } else if (assert_null) {
1310 uncommon_trap(reason,
1311 Deoptimization::Action_make_not_entrant,
1312 NULL, "assert_null");
1313 } else {
1314 replace_in_map(value, zerocon(type));
1315 builtin_throw(reason);
1316 }
1317 }
1318
1319 // Must throw exception, fall-thru not possible?
1320 if (stopped()) {
1321 return top(); // No result
1322 }
1323
1324 if (assert_null) {
1325 // Cast obj to null on this path.
1326 replace_in_map(value, zerocon(type));
1327 return zerocon(type);
1328 }
1329
1330 // Cast obj to not-null on this path, if there is no null_control.
1331 // (If there is a null_control, a non-null value may come back to haunt us.)
1332 if (type == T_OBJECT) {
1333 Node* cast = cast_not_null(value, false);
1334 if (null_control == NULL || (*null_control) == top())
1335 replace_in_map(value, cast);
1336 value = cast;
1337 }
1338
1339 return value;
1340 }
1341
1342
1343 //------------------------------cast_not_null----------------------------------
1344 // Cast obj to not-null on this path
1345 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1346 const Type *t = _gvn.type(obj);
1347 const Type *t_not_null = t->join(TypePtr::NOTNULL);
1348 // Object is already not-null?
1349 if( t == t_not_null ) return obj;
1350
1351 Node *cast = new (C, 2) CastPPNode(obj,t_not_null);
1352 cast->init_req(0, control());
1353 cast = _gvn.transform( cast );
1354
1355 // Scan for instances of 'obj' in the current JVM mapping.
1356 // These instances are known to be not-null after the test.
1357 if (do_replace_in_map)
1358 replace_in_map(obj, cast);
1359
1360 return cast; // Return casted value
1361 }
1362
1363
1364 //--------------------------replace_in_map-------------------------------------
1365 void GraphKit::replace_in_map(Node* old, Node* neww) {
1366 this->map()->replace_edge(old, neww);
1367
1368 // Note: This operation potentially replaces any edge
1369 // on the map. This includes locals, stack, and monitors
1370 // of the current (innermost) JVM state.
1371
1372 // We can consider replacing in caller maps.
1373 // The idea would be that an inlined function's null checks
1374 // can be shared with the entire inlining tree.
1375 // The expense of doing this is that the PreserveJVMState class
1376 // would have to preserve caller states too, with a deep copy.
1377 }
1378
1379
1380
1381 //=============================================================================
1382 //--------------------------------memory---------------------------------------
1383 Node* GraphKit::memory(uint alias_idx) {
1384 MergeMemNode* mem = merged_memory();
1385 Node* p = mem->memory_at(alias_idx);
1386 _gvn.set_type(p, Type::MEMORY); // must be mapped
1387 return p;
1388 }
1389
1390 //-----------------------------reset_memory------------------------------------
1391 Node* GraphKit::reset_memory() {
1392 Node* mem = map()->memory();
1393 // do not use this node for any more parsing!
1394 debug_only( map()->set_memory((Node*)NULL) );
1395 return _gvn.transform( mem );
1396 }
1397
1398 //------------------------------set_all_memory---------------------------------
1399 void GraphKit::set_all_memory(Node* newmem) {
1400 Node* mergemem = MergeMemNode::make(C, newmem);
1401 gvn().set_type_bottom(mergemem);
1402 map()->set_memory(mergemem);
1403 }
1404
1405 //------------------------------set_all_memory_call----------------------------
1406 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1407 Node* newmem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1408 set_all_memory(newmem);
1409 }
1410
1411 //=============================================================================
1412 //
1413 // parser factory methods for MemNodes
1414 //
1415 // These are layered on top of the factory methods in LoadNode and StoreNode,
1416 // and integrate with the parser's memory state and _gvn engine.
1417 //
1418
1419 // factory methods in "int adr_idx"
1420 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1421 int adr_idx,
1422 bool require_atomic_access) {
1423 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1424 const TypePtr* adr_type = NULL; // debug-mode-only argument
1425 debug_only(adr_type = C->get_adr_type(adr_idx));
1426 Node* mem = memory(adr_idx);
1427 Node* ld;
1428 if (require_atomic_access && bt == T_LONG) {
1429 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t);
1430 } else {
1431 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt);
1432 }
1433 return _gvn.transform(ld);
1434 }
1435
1436 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1437 int adr_idx,
1438 bool require_atomic_access) {
1439 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1440 const TypePtr* adr_type = NULL;
1441 debug_only(adr_type = C->get_adr_type(adr_idx));
1442 Node *mem = memory(adr_idx);
1443 Node* st;
1444 if (require_atomic_access && bt == T_LONG) {
1445 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val);
1446 } else {
1447 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt);
1448 }
1449 st = _gvn.transform(st);
1450 set_memory(st, adr_idx);
1451 // Back-to-back stores can only remove intermediate store with DU info
1452 // so push on worklist for optimizer.
1453 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1454 record_for_igvn(st);
1455
1456 return st;
1457 }
1458
1459
1460 void GraphKit::pre_barrier(bool do_load,
1461 Node* ctl,
1462 Node* obj,
1463 Node* adr,
1464 uint adr_idx,
1465 Node* val,
1466 const TypeOopPtr* val_type,
1467 Node* pre_val,
1468 BasicType bt) {
1469
1470 BarrierSet* bs = Universe::heap()->barrier_set();
1471 set_control(ctl);
1472 switch (bs->kind()) {
1473 case BarrierSet::G1SATBCT:
1474 case BarrierSet::G1SATBCTLogging:
1475 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1476 break;
1477
1478 case BarrierSet::CardTableModRef:
1479 case BarrierSet::CardTableExtension:
1480 case BarrierSet::ModRef:
1481 break;
1482
1483 case BarrierSet::Other:
1484 default :
1485 ShouldNotReachHere();
1486
1487 }
1488 }
1489
1490 void GraphKit::post_barrier(Node* ctl,
1491 Node* store,
1492 Node* obj,
1493 Node* adr,
1494 uint adr_idx,
1495 Node* val,
1496 BasicType bt,
1497 bool use_precise) {
1498 BarrierSet* bs = Universe::heap()->barrier_set();
1499 set_control(ctl);
1500 switch (bs->kind()) {
1501 case BarrierSet::G1SATBCT:
1502 case BarrierSet::G1SATBCTLogging:
1503 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1504 break;
1505
1506 case BarrierSet::CardTableModRef:
1507 case BarrierSet::CardTableExtension:
1508 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1509 break;
1510
1511 case BarrierSet::ModRef:
1512 break;
1513
1514 case BarrierSet::Other:
1515 default :
1516 ShouldNotReachHere();
1517
1518 }
1519 }
1520
1521 Node* GraphKit::store_oop(Node* ctl,
1522 Node* obj,
1523 Node* adr,
1524 const TypePtr* adr_type,
1525 Node* val,
1526 const TypeOopPtr* val_type,
1527 BasicType bt,
1528 bool use_precise) {
1529
1530 set_control(ctl);
1531 if (stopped()) return top(); // Dead path ?
1532
1533 assert(bt == T_OBJECT, "sanity");
1534 assert(val != NULL, "not dead path");
1535 uint adr_idx = C->get_alias_index(adr_type);
1536 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1537
1538 pre_barrier(true /* do_load */,
1539 control(), obj, adr, adr_idx, val, val_type,
1540 NULL /* pre_val */,
1541 bt);
1542
1543 Node* store = store_to_memory(control(), adr, val, bt, adr_idx);
1544 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1545 return store;
1546 }
1547
1548 // Could be an array or object we don't know at compile time (unsafe ref.)
1549 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1550 Node* obj, // containing obj
1551 Node* adr, // actual adress to store val at
1552 const TypePtr* adr_type,
1553 Node* val,
1554 BasicType bt) {
1555 Compile::AliasType* at = C->alias_type(adr_type);
1556 const TypeOopPtr* val_type = NULL;
1557 if (adr_type->isa_instptr()) {
1558 if (at->field() != NULL) {
1559 // known field. This code is a copy of the do_put_xxx logic.
1560 ciField* field = at->field();
1561 if (!field->type()->is_loaded()) {
1562 val_type = TypeInstPtr::BOTTOM;
1563 } else {
1564 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1565 }
1566 }
1567 } else if (adr_type->isa_aryptr()) {
1568 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1569 }
1570 if (val_type == NULL) {
1571 val_type = TypeInstPtr::BOTTOM;
1572 }
1573 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true);
1574 }
1575
1576
1577 //-------------------------array_element_address-------------------------
1578 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1579 const TypeInt* sizetype) {
1580 uint shift = exact_log2(type2aelembytes(elembt));
1581 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1582
1583 // short-circuit a common case (saves lots of confusing waste motion)
1584 jint idx_con = find_int_con(idx, -1);
1585 if (idx_con >= 0) {
1586 intptr_t offset = header + ((intptr_t)idx_con << shift);
1587 return basic_plus_adr(ary, offset);
1588 }
1589
1590 // must be correct type for alignment purposes
1591 Node* base = basic_plus_adr(ary, header);
1592 #ifdef _LP64
1593 // The scaled index operand to AddP must be a clean 64-bit value.
1594 // Java allows a 32-bit int to be incremented to a negative
1595 // value, which appears in a 64-bit register as a large
1596 // positive number. Using that large positive number as an
1597 // operand in pointer arithmetic has bad consequences.
1598 // On the other hand, 32-bit overflow is rare, and the possibility
1599 // can often be excluded, if we annotate the ConvI2L node with
1600 // a type assertion that its value is known to be a small positive
1601 // number. (The prior range check has ensured this.)
1602 // This assertion is used by ConvI2LNode::Ideal.
1603 int index_max = max_jint - 1; // array size is max_jint, index is one less
1604 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1605 const TypeLong* lidxtype = TypeLong::make(CONST64(0), index_max, Type::WidenMax);
1606 idx = _gvn.transform( new (C, 2) ConvI2LNode(idx, lidxtype) );
1607 #endif
1608 Node* scale = _gvn.transform( new (C, 3) LShiftXNode(idx, intcon(shift)) );
1609 return basic_plus_adr(ary, base, scale);
1610 }
1611
1612 //-------------------------load_array_element-------------------------
1613 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1614 const Type* elemtype = arytype->elem();
1615 BasicType elembt = elemtype->array_element_basic_type();
1616 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1617 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype);
1618 return ld;
1619 }
1620
1621 //-------------------------set_arguments_for_java_call-------------------------
1622 // Arguments (pre-popped from the stack) are taken from the JVMS.
1623 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1624 // Add the call arguments:
1625 uint nargs = call->method()->arg_size();
1626 for (uint i = 0; i < nargs; i++) {
1627 Node* arg = argument(i);
1628 call->init_req(i + TypeFunc::Parms, arg);
1629 }
1630 }
1631
1632 //---------------------------set_edges_for_java_call---------------------------
1633 // Connect a newly created call into the current JVMS.
1634 // A return value node (if any) is returned from set_edges_for_java_call.
1635 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1636
1637 // Add the predefined inputs:
1638 call->init_req( TypeFunc::Control, control() );
1639 call->init_req( TypeFunc::I_O , i_o() );
1640 call->init_req( TypeFunc::Memory , reset_memory() );
1641 call->init_req( TypeFunc::FramePtr, frameptr() );
1642 call->init_req( TypeFunc::ReturnAdr, top() );
1643
1644 add_safepoint_edges(call, must_throw);
1645
1646 Node* xcall = _gvn.transform(call);
1647
1648 if (xcall == top()) {
1649 set_control(top());
1650 return;
1651 }
1652 assert(xcall == call, "call identity is stable");
1653
1654 // Re-use the current map to produce the result.
1655
1656 set_control(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Control)));
1657 set_i_o( _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1658 set_all_memory_call(xcall, separate_io_proj);
1659
1660 //return xcall; // no need, caller already has it
1661 }
1662
1663 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1664 if (stopped()) return top(); // maybe the call folded up?
1665
1666 // Capture the return value, if any.
1667 Node* ret;
1668 if (call->method() == NULL ||
1669 call->method()->return_type()->basic_type() == T_VOID)
1670 ret = top();
1671 else ret = _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
1672
1673 // Note: Since any out-of-line call can produce an exception,
1674 // we always insert an I_O projection from the call into the result.
1675
1676 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1677
1678 if (separate_io_proj) {
1679 // The caller requested separate projections be used by the fall
1680 // through and exceptional paths, so replace the projections for
1681 // the fall through path.
1682 set_i_o(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O) ));
1683 set_all_memory(_gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) ));
1684 }
1685 return ret;
1686 }
1687
1688 //--------------------set_predefined_input_for_runtime_call--------------------
1689 // Reading and setting the memory state is way conservative here.
1690 // The real problem is that I am not doing real Type analysis on memory,
1691 // so I cannot distinguish card mark stores from other stores. Across a GC
1692 // point the Store Barrier and the card mark memory has to agree. I cannot
1693 // have a card mark store and its barrier split across the GC point from
1694 // either above or below. Here I get that to happen by reading ALL of memory.
1695 // A better answer would be to separate out card marks from other memory.
1696 // For now, return the input memory state, so that it can be reused
1697 // after the call, if this call has restricted memory effects.
1698 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1699 // Set fixed predefined input arguments
1700 Node* memory = reset_memory();
1701 call->init_req( TypeFunc::Control, control() );
1702 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1703 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1704 call->init_req( TypeFunc::FramePtr, frameptr() );
1705 call->init_req( TypeFunc::ReturnAdr, top() );
1706 return memory;
1707 }
1708
1709 //-------------------set_predefined_output_for_runtime_call--------------------
1710 // Set control and memory (not i_o) from the call.
1711 // If keep_mem is not NULL, use it for the output state,
1712 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1713 // If hook_mem is NULL, this call produces no memory effects at all.
1714 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1715 // then only that memory slice is taken from the call.
1716 // In the last case, we must put an appropriate memory barrier before
1717 // the call, so as to create the correct anti-dependencies on loads
1718 // preceding the call.
1719 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1720 Node* keep_mem,
1721 const TypePtr* hook_mem) {
1722 // no i/o
1723 set_control(_gvn.transform( new (C, 1) ProjNode(call,TypeFunc::Control) ));
1724 if (keep_mem) {
1725 // First clone the existing memory state
1726 set_all_memory(keep_mem);
1727 if (hook_mem != NULL) {
1728 // Make memory for the call
1729 Node* mem = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::Memory) );
1730 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1731 // We also use hook_mem to extract specific effects from arraycopy stubs.
1732 set_memory(mem, hook_mem);
1733 }
1734 // ...else the call has NO memory effects.
1735
1736 // Make sure the call advertises its memory effects precisely.
1737 // This lets us build accurate anti-dependences in gcm.cpp.
1738 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1739 "call node must be constructed correctly");
1740 } else {
1741 assert(hook_mem == NULL, "");
1742 // This is not a "slow path" call; all memory comes from the call.
1743 set_all_memory_call(call);
1744 }
1745 }
1746
1747
1748 // Replace the call with the current state of the kit.
1749 void GraphKit::replace_call(CallNode* call, Node* result) {
1750 JVMState* ejvms = NULL;
1751 if (has_exceptions()) {
1752 ejvms = transfer_exceptions_into_jvms();
1753 }
1754
1755 SafePointNode* final_state = stop();
1756
1757 // Find all the needed outputs of this call
1758 CallProjections callprojs;
1759 call->extract_projections(&callprojs, true);
1760
1761 // Replace all the old call edges with the edges from the inlining result
1762 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_state->in(TypeFunc::Control));
1763 C->gvn_replace_by(callprojs.fallthrough_memproj, final_state->in(TypeFunc::Memory));
1764 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_state->in(TypeFunc::I_O));
1765 Node* final_mem = final_state->in(TypeFunc::Memory);
1766
1767 // Replace the result with the new result if it exists and is used
1768 if (callprojs.resproj != NULL && result != NULL) {
1769 C->gvn_replace_by(callprojs.resproj, result);
1770 }
1771
1772 if (ejvms == NULL) {
1773 // No exception edges to simply kill off those paths
1774 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1775 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1776 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1777
1778 // Replace the old exception object with top
1779 if (callprojs.exobj != NULL) {
1780 C->gvn_replace_by(callprojs.exobj, C->top());
1781 }
1782 } else {
1783 GraphKit ekit(ejvms);
1784
1785 // Load my combined exception state into the kit, with all phis transformed:
1786 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1787
1788 Node* ex_oop = ekit.use_exception_state(ex_map);
1789
1790 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1791 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1792 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1793
1794 // Replace the old exception object with the newly created one
1795 if (callprojs.exobj != NULL) {
1796 C->gvn_replace_by(callprojs.exobj, ex_oop);
1797 }
1798 }
1799
1800 // Disconnect the call from the graph
1801 call->disconnect_inputs(NULL);
1802 C->gvn_replace_by(call, C->top());
1803
1804 // Clean up any MergeMems that feed other MergeMems since the
1805 // optimizer doesn't like that.
1806 if (final_mem->is_MergeMem()) {
1807 Node_List wl;
1808 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1809 Node* m = i.get();
1810 if (m->is_MergeMem() && !wl.contains(m)) {
1811 wl.push(m);
1812 }
1813 }
1814 while (wl.size() > 0) {
1815 _gvn.transform(wl.pop());
1816 }
1817 }
1818 }
1819
1820
1821 //------------------------------increment_counter------------------------------
1822 // for statistics: increment a VM counter by 1
1823
1824 void GraphKit::increment_counter(address counter_addr) {
1825 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1826 increment_counter(adr1);
1827 }
1828
1829 void GraphKit::increment_counter(Node* counter_addr) {
1830 int adr_type = Compile::AliasIdxRaw;
1831 Node* ctrl = control();
1832 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type);
1833 Node* incr = _gvn.transform(new (C, 3) AddINode(cnt, _gvn.intcon(1)));
1834 store_to_memory( ctrl, counter_addr, incr, T_INT, adr_type );
1835 }
1836
1837
1838 //------------------------------uncommon_trap----------------------------------
1839 // Bail out to the interpreter in mid-method. Implemented by calling the
1840 // uncommon_trap blob. This helper function inserts a runtime call with the
1841 // right debug info.
1842 void GraphKit::uncommon_trap(int trap_request,
1843 ciKlass* klass, const char* comment,
1844 bool must_throw,
1845 bool keep_exact_action) {
1846 if (failing()) stop();
1847 if (stopped()) return; // trap reachable?
1848
1849 // Note: If ProfileTraps is true, and if a deopt. actually
1850 // occurs here, the runtime will make sure an MDO exists. There is
1851 // no need to call method()->ensure_method_data() at this point.
1852
1853 #ifdef ASSERT
1854 if (!must_throw) {
1855 // Make sure the stack has at least enough depth to execute
1856 // the current bytecode.
1857 int inputs, ignore;
1858 if (compute_stack_effects(inputs, ignore)) {
1859 assert(sp() >= inputs, "must have enough JVMS stack to execute");
1860 // It is a frequent error in library_call.cpp to issue an
1861 // uncommon trap with the _sp value already popped.
1862 }
1863 }
1864 #endif
1865
1866 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
1867 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
1868
1869 switch (action) {
1870 case Deoptimization::Action_maybe_recompile:
1871 case Deoptimization::Action_reinterpret:
1872 // Temporary fix for 6529811 to allow virtual calls to be sure they
1873 // get the chance to go from mono->bi->mega
1874 if (!keep_exact_action &&
1875 Deoptimization::trap_request_index(trap_request) < 0 &&
1876 too_many_recompiles(reason)) {
1877 // This BCI is causing too many recompilations.
1878 action = Deoptimization::Action_none;
1879 trap_request = Deoptimization::make_trap_request(reason, action);
1880 } else {
1881 C->set_trap_can_recompile(true);
1882 }
1883 break;
1884 case Deoptimization::Action_make_not_entrant:
1885 C->set_trap_can_recompile(true);
1886 break;
1887 #ifdef ASSERT
1888 case Deoptimization::Action_none:
1889 case Deoptimization::Action_make_not_compilable:
1890 break;
1891 default:
1892 assert(false, "bad action");
1893 #endif
1894 }
1895
1896 if (TraceOptoParse) {
1897 char buf[100];
1898 tty->print_cr("Uncommon trap %s at bci:%d",
1899 Deoptimization::format_trap_request(buf, sizeof(buf),
1900 trap_request), bci());
1901 }
1902
1903 CompileLog* log = C->log();
1904 if (log != NULL) {
1905 int kid = (klass == NULL)? -1: log->identify(klass);
1906 log->begin_elem("uncommon_trap bci='%d'", bci());
1907 char buf[100];
1908 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
1909 trap_request));
1910 if (kid >= 0) log->print(" klass='%d'", kid);
1911 if (comment != NULL) log->print(" comment='%s'", comment);
1912 log->end_elem();
1913 }
1914
1915 // Make sure any guarding test views this path as very unlikely
1916 Node *i0 = control()->in(0);
1917 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
1918 IfNode *iff = i0->as_If();
1919 float f = iff->_prob; // Get prob
1920 if (control()->Opcode() == Op_IfTrue) {
1921 if (f > PROB_UNLIKELY_MAG(4))
1922 iff->_prob = PROB_MIN;
1923 } else {
1924 if (f < PROB_LIKELY_MAG(4))
1925 iff->_prob = PROB_MAX;
1926 }
1927 }
1928
1929 // Clear out dead values from the debug info.
1930 kill_dead_locals();
1931
1932 // Now insert the uncommon trap subroutine call
1933 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
1934 const TypePtr* no_memory_effects = NULL;
1935 // Pass the index of the class to be loaded
1936 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
1937 (must_throw ? RC_MUST_THROW : 0),
1938 OptoRuntime::uncommon_trap_Type(),
1939 call_addr, "uncommon_trap", no_memory_effects,
1940 intcon(trap_request));
1941 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
1942 "must extract request correctly from the graph");
1943 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
1944
1945 call->set_req(TypeFunc::ReturnAdr, returnadr());
1946 // The debug info is the only real input to this call.
1947
1948 // Halt-and-catch fire here. The above call should never return!
1949 HaltNode* halt = new(C, TypeFunc::Parms) HaltNode(control(), frameptr());
1950 _gvn.set_type_bottom(halt);
1951 root()->add_req(halt);
1952
1953 stop_and_kill_map();
1954 }
1955
1956
1957 //--------------------------just_allocated_object------------------------------
1958 // Report the object that was just allocated.
1959 // It must be the case that there are no intervening safepoints.
1960 // We use this to determine if an object is so "fresh" that
1961 // it does not require card marks.
1962 Node* GraphKit::just_allocated_object(Node* current_control) {
1963 if (C->recent_alloc_ctl() == current_control)
1964 return C->recent_alloc_obj();
1965 return NULL;
1966 }
1967
1968
1969 void GraphKit::round_double_arguments(ciMethod* dest_method) {
1970 // (Note: TypeFunc::make has a cache that makes this fast.)
1971 const TypeFunc* tf = TypeFunc::make(dest_method);
1972 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
1973 for (int j = 0; j < nargs; j++) {
1974 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
1975 if( targ->basic_type() == T_DOUBLE ) {
1976 // If any parameters are doubles, they must be rounded before
1977 // the call, dstore_rounding does gvn.transform
1978 Node *arg = argument(j);
1979 arg = dstore_rounding(arg);
1980 set_argument(j, arg);
1981 }
1982 }
1983 }
1984
1985 void GraphKit::round_double_result(ciMethod* dest_method) {
1986 // A non-strict method may return a double value which has an extended
1987 // exponent, but this must not be visible in a caller which is 'strict'
1988 // If a strict caller invokes a non-strict callee, round a double result
1989
1990 BasicType result_type = dest_method->return_type()->basic_type();
1991 assert( method() != NULL, "must have caller context");
1992 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
1993 // Destination method's return value is on top of stack
1994 // dstore_rounding() does gvn.transform
1995 Node *result = pop_pair();
1996 result = dstore_rounding(result);
1997 push_pair(result);
1998 }
1999 }
2000
2001 // rounding for strict float precision conformance
2002 Node* GraphKit::precision_rounding(Node* n) {
2003 return UseStrictFP && _method->flags().is_strict()
2004 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2005 ? _gvn.transform( new (C, 2) RoundFloatNode(0, n) )
2006 : n;
2007 }
2008
2009 // rounding for strict double precision conformance
2010 Node* GraphKit::dprecision_rounding(Node *n) {
2011 return UseStrictFP && _method->flags().is_strict()
2012 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2013 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
2014 : n;
2015 }
2016
2017 // rounding for non-strict double stores
2018 Node* GraphKit::dstore_rounding(Node* n) {
2019 return Matcher::strict_fp_requires_explicit_rounding
2020 && UseSSE <= 1
2021 ? _gvn.transform( new (C, 2) RoundDoubleNode(0, n) )
2022 : n;
2023 }
2024
2025 //=============================================================================
2026 // Generate a fast path/slow path idiom. Graph looks like:
2027 // [foo] indicates that 'foo' is a parameter
2028 //
2029 // [in] NULL
2030 // \ /
2031 // CmpP
2032 // Bool ne
2033 // If
2034 // / \
2035 // True False-<2>
2036 // / |
2037 // / cast_not_null
2038 // Load | | ^
2039 // [fast_test] | |
2040 // gvn to opt_test | |
2041 // / \ | <1>
2042 // True False |
2043 // | \\ |
2044 // [slow_call] \[fast_result]
2045 // Ctl Val \ \
2046 // | \ \
2047 // Catch <1> \ \
2048 // / \ ^ \ \
2049 // Ex No_Ex | \ \
2050 // | \ \ | \ <2> \
2051 // ... \ [slow_res] | | \ [null_result]
2052 // \ \--+--+--- | |
2053 // \ | / \ | /
2054 // --------Region Phi
2055 //
2056 //=============================================================================
2057 // Code is structured as a series of driver functions all called 'do_XXX' that
2058 // call a set of helper functions. Helper functions first, then drivers.
2059
2060 //------------------------------null_check_oop---------------------------------
2061 // Null check oop. Set null-path control into Region in slot 3.
2062 // Make a cast-not-nullness use the other not-null control. Return cast.
2063 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2064 bool never_see_null) {
2065 // Initial NULL check taken path
2066 (*null_control) = top();
2067 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2068
2069 // Generate uncommon_trap:
2070 if (never_see_null && (*null_control) != top()) {
2071 // If we see an unexpected null at a check-cast we record it and force a
2072 // recompile; the offending check-cast will be compiled to handle NULLs.
2073 // If we see more than one offending BCI, then all checkcasts in the
2074 // method will be compiled to handle NULLs.
2075 PreserveJVMState pjvms(this);
2076 set_control(*null_control);
2077 replace_in_map(value, null());
2078 uncommon_trap(Deoptimization::Reason_null_check,
2079 Deoptimization::Action_make_not_entrant);
2080 (*null_control) = top(); // NULL path is dead
2081 }
2082
2083 // Cast away null-ness on the result
2084 return cast;
2085 }
2086
2087 //------------------------------opt_iff----------------------------------------
2088 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2089 // Return slow-path control.
2090 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2091 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2092
2093 // Fast path taken; set region slot 2
2094 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_iff) );
2095 region->init_req(2,fast_taken); // Capture fast-control
2096
2097 // Fast path not-taken, i.e. slow path
2098 Node *slow_taken = _gvn.transform( new (C, 1) IfTrueNode(opt_iff) );
2099 return slow_taken;
2100 }
2101
2102 //-----------------------------make_runtime_call-------------------------------
2103 Node* GraphKit::make_runtime_call(int flags,
2104 const TypeFunc* call_type, address call_addr,
2105 const char* call_name,
2106 const TypePtr* adr_type,
2107 // The following parms are all optional.
2108 // The first NULL ends the list.
2109 Node* parm0, Node* parm1,
2110 Node* parm2, Node* parm3,
2111 Node* parm4, Node* parm5,
2112 Node* parm6, Node* parm7) {
2113 // Slow-path call
2114 int size = call_type->domain()->cnt();
2115 bool is_leaf = !(flags & RC_NO_LEAF);
2116 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2117 if (call_name == NULL) {
2118 assert(!is_leaf, "must supply name for leaf");
2119 call_name = OptoRuntime::stub_name(call_addr);
2120 }
2121 CallNode* call;
2122 if (!is_leaf) {
2123 call = new(C, size) CallStaticJavaNode(call_type, call_addr, call_name,
2124 bci(), adr_type);
2125 } else if (flags & RC_NO_FP) {
2126 call = new(C, size) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2127 } else {
2128 call = new(C, size) CallLeafNode(call_type, call_addr, call_name, adr_type);
2129 }
2130
2131 // The following is similar to set_edges_for_java_call,
2132 // except that the memory effects of the call are restricted to AliasIdxRaw.
2133
2134 // Slow path call has no side-effects, uses few values
2135 bool wide_in = !(flags & RC_NARROW_MEM);
2136 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2137
2138 Node* prev_mem = NULL;
2139 if (wide_in) {
2140 prev_mem = set_predefined_input_for_runtime_call(call);
2141 } else {
2142 assert(!wide_out, "narrow in => narrow out");
2143 Node* narrow_mem = memory(adr_type);
2144 prev_mem = reset_memory();
2145 map()->set_memory(narrow_mem);
2146 set_predefined_input_for_runtime_call(call);
2147 }
2148
2149 // Hook each parm in order. Stop looking at the first NULL.
2150 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2151 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2152 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2153 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2154 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2155 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2156 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2157 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2158 /* close each nested if ===> */ } } } } } } } }
2159 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2160
2161 if (!is_leaf) {
2162 // Non-leaves can block and take safepoints:
2163 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2164 }
2165 // Non-leaves can throw exceptions:
2166 if (has_io) {
2167 call->set_req(TypeFunc::I_O, i_o());
2168 }
2169
2170 if (flags & RC_UNCOMMON) {
2171 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2172 // (An "if" probability corresponds roughly to an unconditional count.
2173 // Sort of.)
2174 call->set_cnt(PROB_UNLIKELY_MAG(4));
2175 }
2176
2177 Node* c = _gvn.transform(call);
2178 assert(c == call, "cannot disappear");
2179
2180 if (wide_out) {
2181 // Slow path call has full side-effects.
2182 set_predefined_output_for_runtime_call(call);
2183 } else {
2184 // Slow path call has few side-effects, and/or sets few values.
2185 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2186 }
2187
2188 if (has_io) {
2189 set_i_o(_gvn.transform(new (C, 1) ProjNode(call, TypeFunc::I_O)));
2190 }
2191 return call;
2192
2193 }
2194
2195 //------------------------------merge_memory-----------------------------------
2196 // Merge memory from one path into the current memory state.
2197 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2198 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2199 Node* old_slice = mms.force_memory();
2200 Node* new_slice = mms.memory2();
2201 if (old_slice != new_slice) {
2202 PhiNode* phi;
2203 if (new_slice->is_Phi() && new_slice->as_Phi()->region() == region) {
2204 phi = new_slice->as_Phi();
2205 #ifdef ASSERT
2206 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region)
2207 old_slice = old_slice->in(new_path);
2208 // Caller is responsible for ensuring that any pre-existing
2209 // phis are already aware of old memory.
2210 int old_path = (new_path > 1) ? 1 : 2; // choose old_path != new_path
2211 assert(phi->in(old_path) == old_slice, "pre-existing phis OK");
2212 #endif
2213 mms.set_memory(phi);
2214 } else {
2215 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2216 _gvn.set_type(phi, Type::MEMORY);
2217 phi->set_req(new_path, new_slice);
2218 mms.set_memory(_gvn.transform(phi)); // assume it is complete
2219 }
2220 }
2221 }
2222 }
2223
2224 //------------------------------make_slow_call_ex------------------------------
2225 // Make the exception handler hookups for the slow call
2226 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj) {
2227 if (stopped()) return;
2228
2229 // Make a catch node with just two handlers: fall-through and catch-all
2230 Node* i_o = _gvn.transform( new (C, 1) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2231 Node* catc = _gvn.transform( new (C, 2) CatchNode(control(), i_o, 2) );
2232 Node* norm = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2233 Node* excp = _gvn.transform( new (C, 1) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2234
2235 { PreserveJVMState pjvms(this);
2236 set_control(excp);
2237 set_i_o(i_o);
2238
2239 if (excp != top()) {
2240 // Create an exception state also.
2241 // Use an exact type if the caller has specified a specific exception.
2242 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2243 Node* ex_oop = new (C, 2) CreateExNode(ex_type, control(), i_o);
2244 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2245 }
2246 }
2247
2248 // Get the no-exception control from the CatchNode.
2249 set_control(norm);
2250 }
2251
2252
2253 //-------------------------------gen_subtype_check-----------------------------
2254 // Generate a subtyping check. Takes as input the subtype and supertype.
2255 // Returns 2 values: sets the default control() to the true path and returns
2256 // the false path. Only reads invariant memory; sets no (visible) memory.
2257 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2258 // but that's not exposed to the optimizer. This call also doesn't take in an
2259 // Object; if you wish to check an Object you need to load the Object's class
2260 // prior to coming here.
2261 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2262 // Fast check for identical types, perhaps identical constants.
2263 // The types can even be identical non-constants, in cases
2264 // involving Array.newInstance, Object.clone, etc.
2265 if (subklass == superklass)
2266 return top(); // false path is dead; no test needed.
2267
2268 if (_gvn.type(superklass)->singleton()) {
2269 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2270 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2271
2272 // In the common case of an exact superklass, try to fold up the
2273 // test before generating code. You may ask, why not just generate
2274 // the code and then let it fold up? The answer is that the generated
2275 // code will necessarily include null checks, which do not always
2276 // completely fold away. If they are also needless, then they turn
2277 // into a performance loss. Example:
2278 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2279 // Here, the type of 'fa' is often exact, so the store check
2280 // of fa[1]=x will fold up, without testing the nullness of x.
2281 switch (static_subtype_check(superk, subk)) {
2282 case SSC_always_false:
2283 {
2284 Node* always_fail = control();
2285 set_control(top());
2286 return always_fail;
2287 }
2288 case SSC_always_true:
2289 return top();
2290 case SSC_easy_test:
2291 {
2292 // Just do a direct pointer compare and be done.
2293 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(subklass, superklass) );
2294 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2295 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2296 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ) );
2297 return _gvn.transform( new(C, 1) IfFalseNode(iff) );
2298 }
2299 case SSC_full_test:
2300 break;
2301 default:
2302 ShouldNotReachHere();
2303 }
2304 }
2305
2306 // %%% Possible further optimization: Even if the superklass is not exact,
2307 // if the subklass is the unique subtype of the superklass, the check
2308 // will always succeed. We could leave a dependency behind to ensure this.
2309
2310 // First load the super-klass's check-offset
2311 Node *p1 = basic_plus_adr( superklass, superklass, sizeof(oopDesc) + Klass::super_check_offset_offset_in_bytes() );
2312 Node *chk_off = _gvn.transform( new (C, 3) LoadINode( NULL, memory(p1), p1, _gvn.type(p1)->is_ptr() ) );
2313 int cacheoff_con = sizeof(oopDesc) + Klass::secondary_super_cache_offset_in_bytes();
2314 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2315
2316 // Load from the sub-klass's super-class display list, or a 1-word cache of
2317 // the secondary superclass list, or a failing value with a sentinel offset
2318 // if the super-klass is an interface or exceptionally deep in the Java
2319 // hierarchy and we have to scan the secondary superclass list the hard way.
2320 // Worst-case type is a little odd: NULL is allowed as a result (usually
2321 // klass loads can never produce a NULL).
2322 Node *chk_off_X = ConvI2X(chk_off);
2323 Node *p2 = _gvn.transform( new (C, 4) AddPNode(subklass,subklass,chk_off_X) );
2324 // For some types like interfaces the following loadKlass is from a 1-word
2325 // cache which is mutable so can't use immutable memory. Other
2326 // types load from the super-class display table which is immutable.
2327 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2328 Node *nkls = _gvn.transform( LoadKlassNode::make( _gvn, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL ) );
2329
2330 // Compile speed common case: ARE a subtype and we canNOT fail
2331 if( superklass == nkls )
2332 return top(); // false path is dead; no test needed.
2333
2334 // See if we get an immediate positive hit. Happens roughly 83% of the
2335 // time. Test to see if the value loaded just previously from the subklass
2336 // is exactly the superklass.
2337 Node *cmp1 = _gvn.transform( new (C, 3) CmpPNode( superklass, nkls ) );
2338 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp1, BoolTest::eq ) );
2339 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2340 Node *iftrue1 = _gvn.transform( new (C, 1) IfTrueNode ( iff1 ) );
2341 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff1 ) ) );
2342
2343 // Compile speed common case: Check for being deterministic right now. If
2344 // chk_off is a constant and not equal to cacheoff then we are NOT a
2345 // subklass. In this case we need exactly the 1 test above and we can
2346 // return those results immediately.
2347 if (!might_be_cache) {
2348 Node* not_subtype_ctrl = control();
2349 set_control(iftrue1); // We need exactly the 1 test above
2350 return not_subtype_ctrl;
2351 }
2352
2353 // Gather the various success & failures here
2354 RegionNode *r_ok_subtype = new (C, 4) RegionNode(4);
2355 record_for_igvn(r_ok_subtype);
2356 RegionNode *r_not_subtype = new (C, 3) RegionNode(3);
2357 record_for_igvn(r_not_subtype);
2358
2359 r_ok_subtype->init_req(1, iftrue1);
2360
2361 // Check for immediate negative hit. Happens roughly 11% of the time (which
2362 // is roughly 63% of the remaining cases). Test to see if the loaded
2363 // check-offset points into the subklass display list or the 1-element
2364 // cache. If it points to the display (and NOT the cache) and the display
2365 // missed then it's not a subtype.
2366 Node *cacheoff = _gvn.intcon(cacheoff_con);
2367 Node *cmp2 = _gvn.transform( new (C, 3) CmpINode( chk_off, cacheoff ) );
2368 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmp2, BoolTest::ne ) );
2369 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2370 r_not_subtype->init_req(1, _gvn.transform( new (C, 1) IfTrueNode (iff2) ) );
2371 set_control( _gvn.transform( new (C, 1) IfFalseNode(iff2) ) );
2372
2373 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2374 // No performance impact (too rare) but allows sharing of secondary arrays
2375 // which has some footprint reduction.
2376 Node *cmp3 = _gvn.transform( new (C, 3) CmpPNode( subklass, superklass ) );
2377 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmp3, BoolTest::eq ) );
2378 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2379 r_ok_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode ( iff3 ) ) );
2380 set_control( _gvn.transform( new (C, 1) IfFalseNode( iff3 ) ) );
2381
2382 // -- Roads not taken here: --
2383 // We could also have chosen to perform the self-check at the beginning
2384 // of this code sequence, as the assembler does. This would not pay off
2385 // the same way, since the optimizer, unlike the assembler, can perform
2386 // static type analysis to fold away many successful self-checks.
2387 // Non-foldable self checks work better here in second position, because
2388 // the initial primary superclass check subsumes a self-check for most
2389 // types. An exception would be a secondary type like array-of-interface,
2390 // which does not appear in its own primary supertype display.
2391 // Finally, we could have chosen to move the self-check into the
2392 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2393 // dependent manner. But it is worthwhile to have the check here,
2394 // where it can be perhaps be optimized. The cost in code space is
2395 // small (register compare, branch).
2396
2397 // Now do a linear scan of the secondary super-klass array. Again, no real
2398 // performance impact (too rare) but it's gotta be done.
2399 // Since the code is rarely used, there is no penalty for moving it
2400 // out of line, and it can only improve I-cache density.
2401 // The decision to inline or out-of-line this final check is platform
2402 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2403 Node* psc = _gvn.transform(
2404 new (C, 3) PartialSubtypeCheckNode(control(), subklass, superklass) );
2405
2406 Node *cmp4 = _gvn.transform( new (C, 3) CmpPNode( psc, null() ) );
2407 Node *bol4 = _gvn.transform( new (C, 2) BoolNode( cmp4, BoolTest::ne ) );
2408 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2409 r_not_subtype->init_req(2, _gvn.transform( new (C, 1) IfTrueNode (iff4) ) );
2410 r_ok_subtype ->init_req(3, _gvn.transform( new (C, 1) IfFalseNode(iff4) ) );
2411
2412 // Return false path; set default control to true path.
2413 set_control( _gvn.transform(r_ok_subtype) );
2414 return _gvn.transform(r_not_subtype);
2415 }
2416
2417 //----------------------------static_subtype_check-----------------------------
2418 // Shortcut important common cases when superklass is exact:
2419 // (0) superklass is java.lang.Object (can occur in reflective code)
2420 // (1) subklass is already limited to a subtype of superklass => always ok
2421 // (2) subklass does not overlap with superklass => always fail
2422 // (3) superklass has NO subtypes and we can check with a simple compare.
2423 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2424 if (StressReflectiveCode) {
2425 return SSC_full_test; // Let caller generate the general case.
2426 }
2427
2428 if (superk == env()->Object_klass()) {
2429 return SSC_always_true; // (0) this test cannot fail
2430 }
2431
2432 ciType* superelem = superk;
2433 if (superelem->is_array_klass())
2434 superelem = superelem->as_array_klass()->base_element_type();
2435
2436 if (!subk->is_interface()) { // cannot trust static interface types yet
2437 if (subk->is_subtype_of(superk)) {
2438 return SSC_always_true; // (1) false path dead; no dynamic test needed
2439 }
2440 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2441 !superk->is_subtype_of(subk)) {
2442 return SSC_always_false;
2443 }
2444 }
2445
2446 // If casting to an instance klass, it must have no subtypes
2447 if (superk->is_interface()) {
2448 // Cannot trust interfaces yet.
2449 // %%% S.B. superk->nof_implementors() == 1
2450 } else if (superelem->is_instance_klass()) {
2451 ciInstanceKlass* ik = superelem->as_instance_klass();
2452 if (!ik->has_subklass() && !ik->is_interface()) {
2453 if (!ik->is_final()) {
2454 // Add a dependency if there is a chance of a later subclass.
2455 C->dependencies()->assert_leaf_type(ik);
2456 }
2457 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2458 }
2459 } else {
2460 // A primitive array type has no subtypes.
2461 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2462 }
2463
2464 return SSC_full_test;
2465 }
2466
2467 // Profile-driven exact type check:
2468 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2469 float prob,
2470 Node* *casted_receiver) {
2471 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2472 Node* recv_klass = load_object_klass(receiver);
2473 Node* want_klass = makecon(tklass);
2474 Node* cmp = _gvn.transform( new(C, 3) CmpPNode(recv_klass, want_klass) );
2475 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
2476 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2477 set_control( _gvn.transform( new(C, 1) IfTrueNode (iff) ));
2478 Node* fail = _gvn.transform( new(C, 1) IfFalseNode(iff) );
2479
2480 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2481 assert(recv_xtype->klass_is_exact(), "");
2482
2483 // Subsume downstream occurrences of receiver with a cast to
2484 // recv_xtype, since now we know what the type will be.
2485 Node* cast = new(C, 2) CheckCastPPNode(control(), receiver, recv_xtype);
2486 (*casted_receiver) = _gvn.transform(cast);
2487 // (User must make the replace_in_map call.)
2488
2489 return fail;
2490 }
2491
2492
2493 //------------------------------seems_never_null-------------------------------
2494 // Use null_seen information if it is available from the profile.
2495 // If we see an unexpected null at a type check we record it and force a
2496 // recompile; the offending check will be recompiled to handle NULLs.
2497 // If we see several offending BCIs, then all checks in the
2498 // method will be recompiled.
2499 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2500 if (UncommonNullCast // Cutout for this technique
2501 && obj != null() // And not the -Xcomp stupid case?
2502 && !too_many_traps(Deoptimization::Reason_null_check)
2503 ) {
2504 if (data == NULL)
2505 // Edge case: no mature data. Be optimistic here.
2506 return true;
2507 // If the profile has not seen a null, assume it won't happen.
2508 assert(java_bc() == Bytecodes::_checkcast ||
2509 java_bc() == Bytecodes::_instanceof ||
2510 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2511 return !data->as_BitData()->null_seen();
2512 }
2513 return false;
2514 }
2515
2516 //------------------------maybe_cast_profiled_receiver-------------------------
2517 // If the profile has seen exactly one type, narrow to exactly that type.
2518 // Subsequent type checks will always fold up.
2519 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2520 ciProfileData* data,
2521 ciKlass* require_klass) {
2522 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2523 if (data == NULL) return NULL;
2524
2525 // Make sure we haven't already deoptimized from this tactic.
2526 if (too_many_traps(Deoptimization::Reason_class_check))
2527 return NULL;
2528
2529 // (No, this isn't a call, but it's enough like a virtual call
2530 // to use the same ciMethod accessor to get the profile info...)
2531 ciCallProfile profile = method()->call_profile_at_bci(bci());
2532 if (profile.count() >= 0 && // no cast failures here
2533 profile.has_receiver(0) &&
2534 profile.morphism() == 1) {
2535 ciKlass* exact_kls = profile.receiver(0);
2536 if (require_klass == NULL ||
2537 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2538 // If we narrow the type to match what the type profile sees,
2539 // we can then remove the rest of the cast.
2540 // This is a win, even if the exact_kls is very specific,
2541 // because downstream operations, such as method calls,
2542 // will often benefit from the sharper type.
2543 Node* exact_obj = not_null_obj; // will get updated in place...
2544 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2545 &exact_obj);
2546 { PreserveJVMState pjvms(this);
2547 set_control(slow_ctl);
2548 uncommon_trap(Deoptimization::Reason_class_check,
2549 Deoptimization::Action_maybe_recompile);
2550 }
2551 replace_in_map(not_null_obj, exact_obj);
2552 return exact_obj;
2553 }
2554 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2555 }
2556
2557 return NULL;
2558 }
2559
2560
2561 //-------------------------------gen_instanceof--------------------------------
2562 // Generate an instance-of idiom. Used by both the instance-of bytecode
2563 // and the reflective instance-of call.
2564 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass) {
2565 kill_dead_locals(); // Benefit all the uncommon traps
2566 assert( !stopped(), "dead parse path should be checked in callers" );
2567 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2568 "must check for not-null not-dead klass in callers");
2569
2570 // Make the merge point
2571 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2572 RegionNode* region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2573 Node* phi = new(C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
2574 C->set_has_split_ifs(true); // Has chance for split-if optimization
2575
2576 ciProfileData* data = NULL;
2577 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2578 data = method()->method_data()->bci_to_data(bci());
2579 }
2580 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2581 && seems_never_null(obj, data));
2582
2583 // Null check; get casted pointer; set region slot 3
2584 Node* null_ctl = top();
2585 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2586
2587 // If not_null_obj is dead, only null-path is taken
2588 if (stopped()) { // Doing instance-of on a NULL?
2589 set_control(null_ctl);
2590 return intcon(0);
2591 }
2592 region->init_req(_null_path, null_ctl);
2593 phi ->init_req(_null_path, intcon(0)); // Set null path value
2594 if (null_ctl == top()) {
2595 // Do this eagerly, so that pattern matches like is_diamond_phi
2596 // will work even during parsing.
2597 assert(_null_path == PATH_LIMIT-1, "delete last");
2598 region->del_req(_null_path);
2599 phi ->del_req(_null_path);
2600 }
2601
2602 if (ProfileDynamicTypes && data != NULL) {
2603 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, NULL);
2604 if (stopped()) { // Profile disagrees with this path.
2605 set_control(null_ctl); // Null is the only remaining possibility.
2606 return intcon(0);
2607 }
2608 if (cast_obj != NULL)
2609 not_null_obj = cast_obj;
2610 }
2611
2612 // Load the object's klass
2613 Node* obj_klass = load_object_klass(not_null_obj);
2614
2615 // Generate the subtype check
2616 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2617
2618 // Plug in the success path to the general merge in slot 1.
2619 region->init_req(_obj_path, control());
2620 phi ->init_req(_obj_path, intcon(1));
2621
2622 // Plug in the failing path to the general merge in slot 2.
2623 region->init_req(_fail_path, not_subtype_ctrl);
2624 phi ->init_req(_fail_path, intcon(0));
2625
2626 // Return final merged results
2627 set_control( _gvn.transform(region) );
2628 record_for_igvn(region);
2629 return _gvn.transform(phi);
2630 }
2631
2632 //-------------------------------gen_checkcast---------------------------------
2633 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2634 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2635 // uncommon-trap paths work. Adjust stack after this call.
2636 // If failure_control is supplied and not null, it is filled in with
2637 // the control edge for the cast failure. Otherwise, an appropriate
2638 // uncommon trap or exception is thrown.
2639 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2640 Node* *failure_control) {
2641 kill_dead_locals(); // Benefit all the uncommon traps
2642 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2643 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2644
2645 // Fast cutout: Check the case that the cast is vacuously true.
2646 // This detects the common cases where the test will short-circuit
2647 // away completely. We do this before we perform the null check,
2648 // because if the test is going to turn into zero code, we don't
2649 // want a residual null check left around. (Causes a slowdown,
2650 // for example, in some objArray manipulations, such as a[i]=a[j].)
2651 if (tk->singleton()) {
2652 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2653 if (objtp != NULL && objtp->klass() != NULL) {
2654 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2655 case SSC_always_true:
2656 return obj;
2657 case SSC_always_false:
2658 // It needs a null check because a null will *pass* the cast check.
2659 // A non-null value will always produce an exception.
2660 return do_null_assert(obj, T_OBJECT);
2661 }
2662 }
2663 }
2664
2665 ciProfileData* data = NULL;
2666 if (failure_control == NULL) { // use MDO in regular case only
2667 assert(java_bc() == Bytecodes::_aastore ||
2668 java_bc() == Bytecodes::_checkcast,
2669 "interpreter profiles type checks only for these BCs");
2670 data = method()->method_data()->bci_to_data(bci());
2671 }
2672
2673 // Make the merge point
2674 enum { _obj_path = 1, _null_path, PATH_LIMIT };
2675 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2676 Node* phi = new (C, PATH_LIMIT) PhiNode(region, toop);
2677 C->set_has_split_ifs(true); // Has chance for split-if optimization
2678
2679 // Use null-cast information if it is available
2680 bool never_see_null = ((failure_control == NULL) // regular case only
2681 && seems_never_null(obj, data));
2682
2683 // Null check; get casted pointer; set region slot 3
2684 Node* null_ctl = top();
2685 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null);
2686
2687 // If not_null_obj is dead, only null-path is taken
2688 if (stopped()) { // Doing instance-of on a NULL?
2689 set_control(null_ctl);
2690 return null();
2691 }
2692 region->init_req(_null_path, null_ctl);
2693 phi ->init_req(_null_path, null()); // Set null path value
2694 if (null_ctl == top()) {
2695 // Do this eagerly, so that pattern matches like is_diamond_phi
2696 // will work even during parsing.
2697 assert(_null_path == PATH_LIMIT-1, "delete last");
2698 region->del_req(_null_path);
2699 phi ->del_req(_null_path);
2700 }
2701
2702 Node* cast_obj = NULL;
2703 if (data != NULL &&
2704 // Counter has never been decremented (due to cast failure).
2705 // ...This is a reasonable thing to expect. It is true of
2706 // all casts inserted by javac to implement generic types.
2707 data->as_CounterData()->count() >= 0) {
2708 cast_obj = maybe_cast_profiled_receiver(not_null_obj, data, tk->klass());
2709 if (cast_obj != NULL) {
2710 if (failure_control != NULL) // failure is now impossible
2711 (*failure_control) = top();
2712 // adjust the type of the phi to the exact klass:
2713 phi->raise_bottom_type(_gvn.type(cast_obj)->meet(TypePtr::NULL_PTR));
2714 }
2715 }
2716
2717 if (cast_obj == NULL) {
2718 // Load the object's klass
2719 Node* obj_klass = load_object_klass(not_null_obj);
2720
2721 // Generate the subtype check
2722 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
2723
2724 // Plug in success path into the merge
2725 cast_obj = _gvn.transform(new (C, 2) CheckCastPPNode(control(),
2726 not_null_obj, toop));
2727 // Failure path ends in uncommon trap (or may be dead - failure impossible)
2728 if (failure_control == NULL) {
2729 if (not_subtype_ctrl != top()) { // If failure is possible
2730 PreserveJVMState pjvms(this);
2731 set_control(not_subtype_ctrl);
2732 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
2733 }
2734 } else {
2735 (*failure_control) = not_subtype_ctrl;
2736 }
2737 }
2738
2739 region->init_req(_obj_path, control());
2740 phi ->init_req(_obj_path, cast_obj);
2741
2742 // A merge of NULL or Casted-NotNull obj
2743 Node* res = _gvn.transform(phi);
2744
2745 // Note I do NOT always 'replace_in_map(obj,result)' here.
2746 // if( tk->klass()->can_be_primary_super() )
2747 // This means that if I successfully store an Object into an array-of-String
2748 // I 'forget' that the Object is really now known to be a String. I have to
2749 // do this because we don't have true union types for interfaces - if I store
2750 // a Baz into an array-of-Interface and then tell the optimizer it's an
2751 // Interface, I forget that it's also a Baz and cannot do Baz-like field
2752 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
2753 // replace_in_map( obj, res );
2754
2755 // Return final merged results
2756 set_control( _gvn.transform(region) );
2757 record_for_igvn(region);
2758 return res;
2759 }
2760
2761 //------------------------------next_monitor-----------------------------------
2762 // What number should be given to the next monitor?
2763 int GraphKit::next_monitor() {
2764 int current = jvms()->monitor_depth()* C->sync_stack_slots();
2765 int next = current + C->sync_stack_slots();
2766 // Keep the toplevel high water mark current:
2767 if (C->fixed_slots() < next) C->set_fixed_slots(next);
2768 return current;
2769 }
2770
2771 //------------------------------insert_mem_bar---------------------------------
2772 // Memory barrier to avoid floating things around
2773 // The membar serves as a pinch point between both control and all memory slices.
2774 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
2775 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
2776 mb->init_req(TypeFunc::Control, control());
2777 mb->init_req(TypeFunc::Memory, reset_memory());
2778 Node* membar = _gvn.transform(mb);
2779 set_control(_gvn.transform(new (C, 1) ProjNode(membar,TypeFunc::Control) ));
2780 set_all_memory_call(membar);
2781 return membar;
2782 }
2783
2784 //-------------------------insert_mem_bar_volatile----------------------------
2785 // Memory barrier to avoid floating things around
2786 // The membar serves as a pinch point between both control and memory(alias_idx).
2787 // If you want to make a pinch point on all memory slices, do not use this
2788 // function (even with AliasIdxBot); use insert_mem_bar() instead.
2789 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
2790 // When Parse::do_put_xxx updates a volatile field, it appends a series
2791 // of MemBarVolatile nodes, one for *each* volatile field alias category.
2792 // The first membar is on the same memory slice as the field store opcode.
2793 // This forces the membar to follow the store. (Bug 6500685 broke this.)
2794 // All the other membars (for other volatile slices, including AliasIdxBot,
2795 // which stands for all unknown volatile slices) are control-dependent
2796 // on the first membar. This prevents later volatile loads or stores
2797 // from sliding up past the just-emitted store.
2798
2799 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
2800 mb->set_req(TypeFunc::Control,control());
2801 if (alias_idx == Compile::AliasIdxBot) {
2802 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
2803 } else {
2804 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
2805 mb->set_req(TypeFunc::Memory, memory(alias_idx));
2806 }
2807 Node* membar = _gvn.transform(mb);
2808 set_control(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Control)));
2809 if (alias_idx == Compile::AliasIdxBot) {
2810 merged_memory()->set_base_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)));
2811 } else {
2812 set_memory(_gvn.transform(new (C, 1) ProjNode(membar, TypeFunc::Memory)),alias_idx);
2813 }
2814 return membar;
2815 }
2816
2817 //------------------------------shared_lock------------------------------------
2818 // Emit locking code.
2819 FastLockNode* GraphKit::shared_lock(Node* obj) {
2820 // bci is either a monitorenter bc or InvocationEntryBci
2821 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2822 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2823
2824 if( !GenerateSynchronizationCode )
2825 return NULL; // Not locking things?
2826 if (stopped()) // Dead monitor?
2827 return NULL;
2828
2829 assert(dead_locals_are_killed(), "should kill locals before sync. point");
2830
2831 // Box the stack location
2832 Node* box = _gvn.transform(new (C, 1) BoxLockNode(next_monitor()));
2833 Node* mem = reset_memory();
2834
2835 FastLockNode * flock = _gvn.transform(new (C, 3) FastLockNode(0, obj, box) )->as_FastLock();
2836 if (PrintPreciseBiasedLockingStatistics) {
2837 // Create the counters for this fast lock.
2838 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
2839 }
2840 // Add monitor to debug info for the slow path. If we block inside the
2841 // slow path and de-opt, we need the monitor hanging around
2842 map()->push_monitor( flock );
2843
2844 const TypeFunc *tf = LockNode::lock_type();
2845 LockNode *lock = new (C, tf->domain()->cnt()) LockNode(C, tf);
2846
2847 lock->init_req( TypeFunc::Control, control() );
2848 lock->init_req( TypeFunc::Memory , mem );
2849 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2850 lock->init_req( TypeFunc::FramePtr, frameptr() );
2851 lock->init_req( TypeFunc::ReturnAdr, top() );
2852
2853 lock->init_req(TypeFunc::Parms + 0, obj);
2854 lock->init_req(TypeFunc::Parms + 1, box);
2855 lock->init_req(TypeFunc::Parms + 2, flock);
2856 add_safepoint_edges(lock);
2857
2858 lock = _gvn.transform( lock )->as_Lock();
2859
2860 // lock has no side-effects, sets few values
2861 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
2862
2863 insert_mem_bar(Op_MemBarAcquire);
2864
2865 // Add this to the worklist so that the lock can be eliminated
2866 record_for_igvn(lock);
2867
2868 #ifndef PRODUCT
2869 if (PrintLockStatistics) {
2870 // Update the counter for this lock. Don't bother using an atomic
2871 // operation since we don't require absolute accuracy.
2872 lock->create_lock_counter(map()->jvms());
2873 increment_counter(lock->counter()->addr());
2874 }
2875 #endif
2876
2877 return flock;
2878 }
2879
2880
2881 //------------------------------shared_unlock----------------------------------
2882 // Emit unlocking code.
2883 void GraphKit::shared_unlock(Node* box, Node* obj) {
2884 // bci is either a monitorenter bc or InvocationEntryBci
2885 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
2886 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
2887
2888 if( !GenerateSynchronizationCode )
2889 return;
2890 if (stopped()) { // Dead monitor?
2891 map()->pop_monitor(); // Kill monitor from debug info
2892 return;
2893 }
2894
2895 // Memory barrier to avoid floating things down past the locked region
2896 insert_mem_bar(Op_MemBarRelease);
2897
2898 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
2899 UnlockNode *unlock = new (C, tf->domain()->cnt()) UnlockNode(C, tf);
2900 uint raw_idx = Compile::AliasIdxRaw;
2901 unlock->init_req( TypeFunc::Control, control() );
2902 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
2903 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
2904 unlock->init_req( TypeFunc::FramePtr, frameptr() );
2905 unlock->init_req( TypeFunc::ReturnAdr, top() );
2906
2907 unlock->init_req(TypeFunc::Parms + 0, obj);
2908 unlock->init_req(TypeFunc::Parms + 1, box);
2909 unlock = _gvn.transform(unlock)->as_Unlock();
2910
2911 Node* mem = reset_memory();
2912
2913 // unlock has no side-effects, sets few values
2914 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
2915
2916 // Kill monitor from debug info
2917 map()->pop_monitor( );
2918 }
2919
2920 //-------------------------------get_layout_helper-----------------------------
2921 // If the given klass is a constant or known to be an array,
2922 // fetch the constant layout helper value into constant_value
2923 // and return (Node*)NULL. Otherwise, load the non-constant
2924 // layout helper value, and return the node which represents it.
2925 // This two-faced routine is useful because allocation sites
2926 // almost always feature constant types.
2927 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
2928 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
2929 if (!StressReflectiveCode && inst_klass != NULL) {
2930 ciKlass* klass = inst_klass->klass();
2931 bool xklass = inst_klass->klass_is_exact();
2932 if (xklass || klass->is_array_klass()) {
2933 jint lhelper = klass->layout_helper();
2934 if (lhelper != Klass::_lh_neutral_value) {
2935 constant_value = lhelper;
2936 return (Node*) NULL;
2937 }
2938 }
2939 }
2940 constant_value = Klass::_lh_neutral_value; // put in a known value
2941 Node* lhp = basic_plus_adr(klass_node, klass_node, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc));
2942 return make_load(NULL, lhp, TypeInt::INT, T_INT);
2943 }
2944
2945 // We just put in an allocate/initialize with a big raw-memory effect.
2946 // Hook selected additional alias categories on the initialization.
2947 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
2948 MergeMemNode* init_in_merge,
2949 Node* init_out_raw) {
2950 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
2951 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
2952
2953 Node* prevmem = kit.memory(alias_idx);
2954 init_in_merge->set_memory_at(alias_idx, prevmem);
2955 kit.set_memory(init_out_raw, alias_idx);
2956 }
2957
2958 //---------------------------set_output_for_allocation-------------------------
2959 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
2960 const TypeOopPtr* oop_type,
2961 bool raw_mem_only) {
2962 int rawidx = Compile::AliasIdxRaw;
2963 alloc->set_req( TypeFunc::FramePtr, frameptr() );
2964 add_safepoint_edges(alloc);
2965 Node* allocx = _gvn.transform(alloc);
2966 set_control( _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Control) ) );
2967 // create memory projection for i_o
2968 set_memory ( _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
2969 make_slow_call_ex(allocx, env()->OutOfMemoryError_klass(), true);
2970
2971 // create a memory projection as for the normal control path
2972 Node* malloc = _gvn.transform(new (C, 1) ProjNode(allocx, TypeFunc::Memory));
2973 set_memory(malloc, rawidx);
2974
2975 // a normal slow-call doesn't change i_o, but an allocation does
2976 // we create a separate i_o projection for the normal control path
2977 set_i_o(_gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::I_O, false) ) );
2978 Node* rawoop = _gvn.transform( new (C, 1) ProjNode(allocx, TypeFunc::Parms) );
2979
2980 // put in an initialization barrier
2981 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
2982 rawoop)->as_Initialize();
2983 assert(alloc->initialization() == init, "2-way macro link must work");
2984 assert(init ->allocation() == alloc, "2-way macro link must work");
2985 if (ReduceFieldZeroing && !raw_mem_only) {
2986 // Extract memory strands which may participate in the new object's
2987 // initialization, and source them from the new InitializeNode.
2988 // This will allow us to observe initializations when they occur,
2989 // and link them properly (as a group) to the InitializeNode.
2990 assert(init->in(InitializeNode::Memory) == malloc, "");
2991 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
2992 init->set_req(InitializeNode::Memory, minit_in);
2993 record_for_igvn(minit_in); // fold it up later, if possible
2994 Node* minit_out = memory(rawidx);
2995 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
2996 if (oop_type->isa_aryptr()) {
2997 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
2998 int elemidx = C->get_alias_index(telemref);
2999 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3000 } else if (oop_type->isa_instptr()) {
3001 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3002 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3003 ciField* field = ik->nonstatic_field_at(i);
3004 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3005 continue; // do not bother to track really large numbers of fields
3006 // Find (or create) the alias category for this field:
3007 int fieldidx = C->alias_type(field)->index();
3008 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3009 }
3010 }
3011 }
3012
3013 // Cast raw oop to the real thing...
3014 Node* javaoop = new (C, 2) CheckCastPPNode(control(), rawoop, oop_type);
3015 javaoop = _gvn.transform(javaoop);
3016 C->set_recent_alloc(control(), javaoop);
3017 assert(just_allocated_object(control()) == javaoop, "just allocated");
3018
3019 #ifdef ASSERT
3020 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3021 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3022 "Ideal_allocation works");
3023 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3024 "Ideal_allocation works");
3025 if (alloc->is_AllocateArray()) {
3026 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3027 "Ideal_allocation works");
3028 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3029 "Ideal_allocation works");
3030 } else {
3031 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3032 }
3033 }
3034 #endif //ASSERT
3035
3036 return javaoop;
3037 }
3038
3039 //---------------------------new_instance--------------------------------------
3040 // This routine takes a klass_node which may be constant (for a static type)
3041 // or may be non-constant (for reflective code). It will work equally well
3042 // for either, and the graph will fold nicely if the optimizer later reduces
3043 // the type to a constant.
3044 // The optional arguments are for specialized use by intrinsics:
3045 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3046 // - If 'raw_mem_only', do not cast the result to an oop.
3047 // - If 'return_size_val', report the the total object size to the caller.
3048 Node* GraphKit::new_instance(Node* klass_node,
3049 Node* extra_slow_test,
3050 bool raw_mem_only, // affect only raw memory
3051 Node* *return_size_val) {
3052 // Compute size in doublewords
3053 // The size is always an integral number of doublewords, represented
3054 // as a positive bytewise size stored in the klass's layout_helper.
3055 // The layout_helper also encodes (in a low bit) the need for a slow path.
3056 jint layout_con = Klass::_lh_neutral_value;
3057 Node* layout_val = get_layout_helper(klass_node, layout_con);
3058 int layout_is_con = (layout_val == NULL);
3059
3060 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3061 // Generate the initial go-slow test. It's either ALWAYS (return a
3062 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3063 // case) a computed value derived from the layout_helper.
3064 Node* initial_slow_test = NULL;
3065 if (layout_is_con) {
3066 assert(!StressReflectiveCode, "stress mode does not use these paths");
3067 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3068 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3069
3070 } else { // reflective case
3071 // This reflective path is used by Unsafe.allocateInstance.
3072 // (It may be stress-tested by specifying StressReflectiveCode.)
3073 // Basically, we want to get into the VM is there's an illegal argument.
3074 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3075 initial_slow_test = _gvn.transform( new (C, 3) AndINode(layout_val, bit) );
3076 if (extra_slow_test != intcon(0)) {
3077 initial_slow_test = _gvn.transform( new (C, 3) OrINode(initial_slow_test, extra_slow_test) );
3078 }
3079 // (Macro-expander will further convert this to a Bool, if necessary.)
3080 }
3081
3082 // Find the size in bytes. This is easy; it's the layout_helper.
3083 // The size value must be valid even if the slow path is taken.
3084 Node* size = NULL;
3085 if (layout_is_con) {
3086 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3087 } else { // reflective case
3088 // This reflective path is used by clone and Unsafe.allocateInstance.
3089 size = ConvI2X(layout_val);
3090
3091 // Clear the low bits to extract layout_helper_size_in_bytes:
3092 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3093 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3094 size = _gvn.transform( new (C, 3) AndXNode(size, mask) );
3095 }
3096 if (return_size_val != NULL) {
3097 (*return_size_val) = size;
3098 }
3099
3100 // This is a precise notnull oop of the klass.
3101 // (Actually, it need not be precise if this is a reflective allocation.)
3102 // It's what we cast the result to.
3103 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3104 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3105 const TypeOopPtr* oop_type = tklass->as_instance_type();
3106
3107 // Now generate allocation code
3108
3109 // The entire memory state is needed for slow path of the allocation
3110 // since GC and deoptimization can happened.
3111 Node *mem = reset_memory();
3112 set_all_memory(mem); // Create new memory state
3113
3114 AllocateNode* alloc
3115 = new (C, AllocateNode::ParmLimit)
3116 AllocateNode(C, AllocateNode::alloc_type(),
3117 control(), mem, i_o(),
3118 size, klass_node,
3119 initial_slow_test);
3120
3121 return set_output_for_allocation(alloc, oop_type, raw_mem_only);
3122 }
3123
3124 //-------------------------------new_array-------------------------------------
3125 // helper for both newarray and anewarray
3126 // The 'length' parameter is (obviously) the length of the array.
3127 // See comments on new_instance for the meaning of the other arguments.
3128 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3129 Node* length, // number of array elements
3130 int nargs, // number of arguments to push back for uncommon trap
3131 bool raw_mem_only, // affect only raw memory
3132 Node* *return_size_val) {
3133 jint layout_con = Klass::_lh_neutral_value;
3134 Node* layout_val = get_layout_helper(klass_node, layout_con);
3135 int layout_is_con = (layout_val == NULL);
3136
3137 if (!layout_is_con && !StressReflectiveCode &&
3138 !too_many_traps(Deoptimization::Reason_class_check)) {
3139 // This is a reflective array creation site.
3140 // Optimistically assume that it is a subtype of Object[],
3141 // so that we can fold up all the address arithmetic.
3142 layout_con = Klass::array_layout_helper(T_OBJECT);
3143 Node* cmp_lh = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(layout_con)) );
3144 Node* bol_lh = _gvn.transform( new(C, 2) BoolNode(cmp_lh, BoolTest::eq) );
3145 { BuildCutout unless(this, bol_lh, PROB_MAX);
3146 _sp += nargs;
3147 uncommon_trap(Deoptimization::Reason_class_check,
3148 Deoptimization::Action_maybe_recompile);
3149 }
3150 layout_val = NULL;
3151 layout_is_con = true;
3152 }
3153
3154 // Generate the initial go-slow test. Make sure we do not overflow
3155 // if length is huge (near 2Gig) or negative! We do not need
3156 // exact double-words here, just a close approximation of needed
3157 // double-words. We can't add any offset or rounding bits, lest we
3158 // take a size -1 of bytes and make it positive. Use an unsigned
3159 // compare, so negative sizes look hugely positive.
3160 int fast_size_limit = FastAllocateSizeLimit;
3161 if (layout_is_con) {
3162 assert(!StressReflectiveCode, "stress mode does not use these paths");
3163 // Increase the size limit if we have exact knowledge of array type.
3164 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3165 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3166 }
3167
3168 Node* initial_slow_cmp = _gvn.transform( new (C, 3) CmpUNode( length, intcon( fast_size_limit ) ) );
3169 Node* initial_slow_test = _gvn.transform( new (C, 2) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3170 if (initial_slow_test->is_Bool()) {
3171 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3172 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3173 }
3174
3175 // --- Size Computation ---
3176 // array_size = round_to_heap(array_header + (length << elem_shift));
3177 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3178 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3179 // The rounding mask is strength-reduced, if possible.
3180 int round_mask = MinObjAlignmentInBytes - 1;
3181 Node* header_size = NULL;
3182 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3183 // (T_BYTE has the weakest alignment and size restrictions...)
3184 if (layout_is_con) {
3185 int hsize = Klass::layout_helper_header_size(layout_con);
3186 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3187 BasicType etype = Klass::layout_helper_element_type(layout_con);
3188 if ((round_mask & ~right_n_bits(eshift)) == 0)
3189 round_mask = 0; // strength-reduce it if it goes away completely
3190 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3191 assert(header_size_min <= hsize, "generic minimum is smallest");
3192 header_size_min = hsize;
3193 header_size = intcon(hsize + round_mask);
3194 } else {
3195 Node* hss = intcon(Klass::_lh_header_size_shift);
3196 Node* hsm = intcon(Klass::_lh_header_size_mask);
3197 Node* hsize = _gvn.transform( new(C, 3) URShiftINode(layout_val, hss) );
3198 hsize = _gvn.transform( new(C, 3) AndINode(hsize, hsm) );
3199 Node* mask = intcon(round_mask);
3200 header_size = _gvn.transform( new(C, 3) AddINode(hsize, mask) );
3201 }
3202
3203 Node* elem_shift = NULL;
3204 if (layout_is_con) {
3205 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3206 if (eshift != 0)
3207 elem_shift = intcon(eshift);
3208 } else {
3209 // There is no need to mask or shift this value.
3210 // The semantics of LShiftINode include an implicit mask to 0x1F.
3211 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3212 elem_shift = layout_val;
3213 }
3214
3215 // Transition to native address size for all offset calculations:
3216 Node* lengthx = ConvI2X(length);
3217 Node* headerx = ConvI2X(header_size);
3218 #ifdef _LP64
3219 { const TypeLong* tllen = _gvn.find_long_type(lengthx);
3220 if (tllen != NULL && tllen->_lo < 0) {
3221 // Add a manual constraint to a positive range. Cf. array_element_address.
3222 jlong size_max = arrayOopDesc::max_array_length(T_BYTE);
3223 if (size_max > tllen->_hi) size_max = tllen->_hi;
3224 const TypeLong* tlcon = TypeLong::make(CONST64(0), size_max, Type::WidenMin);
3225 lengthx = _gvn.transform( new (C, 2) ConvI2LNode(length, tlcon));
3226 }
3227 }
3228 #endif
3229
3230 // Combine header size (plus rounding) and body size. Then round down.
3231 // This computation cannot overflow, because it is used only in two
3232 // places, one where the length is sharply limited, and the other
3233 // after a successful allocation.
3234 Node* abody = lengthx;
3235 if (elem_shift != NULL)
3236 abody = _gvn.transform( new(C, 3) LShiftXNode(lengthx, elem_shift) );
3237 Node* size = _gvn.transform( new(C, 3) AddXNode(headerx, abody) );
3238 if (round_mask != 0) {
3239 Node* mask = MakeConX(~round_mask);
3240 size = _gvn.transform( new(C, 3) AndXNode(size, mask) );
3241 }
3242 // else if round_mask == 0, the size computation is self-rounding
3243
3244 if (return_size_val != NULL) {
3245 // This is the size
3246 (*return_size_val) = size;
3247 }
3248
3249 // Now generate allocation code
3250
3251 // The entire memory state is needed for slow path of the allocation
3252 // since GC and deoptimization can happened.
3253 Node *mem = reset_memory();
3254 set_all_memory(mem); // Create new memory state
3255
3256 // Create the AllocateArrayNode and its result projections
3257 AllocateArrayNode* alloc
3258 = new (C, AllocateArrayNode::ParmLimit)
3259 AllocateArrayNode(C, AllocateArrayNode::alloc_type(),
3260 control(), mem, i_o(),
3261 size, klass_node,
3262 initial_slow_test,
3263 length);
3264
3265 // Cast to correct type. Note that the klass_node may be constant or not,
3266 // and in the latter case the actual array type will be inexact also.
3267 // (This happens via a non-constant argument to inline_native_newArray.)
3268 // In any case, the value of klass_node provides the desired array type.
3269 const TypeInt* length_type = _gvn.find_int_type(length);
3270 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3271 if (ary_type->isa_aryptr() && length_type != NULL) {
3272 // Try to get a better type than POS for the size
3273 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3274 }
3275
3276 Node* javaoop = set_output_for_allocation(alloc, ary_type, raw_mem_only);
3277
3278 // Cast length on remaining path to be as narrow as possible
3279 if (map()->find_edge(length) >= 0) {
3280 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3281 if (ccast != length) {
3282 _gvn.set_type_bottom(ccast);
3283 record_for_igvn(ccast);
3284 replace_in_map(length, ccast);
3285 }
3286 }
3287
3288 return javaoop;
3289 }
3290
3291 // The following "Ideal_foo" functions are placed here because they recognize
3292 // the graph shapes created by the functions immediately above.
3293
3294 //---------------------------Ideal_allocation----------------------------------
3295 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3296 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3297 if (ptr == NULL) { // reduce dumb test in callers
3298 return NULL;
3299 }
3300 if (ptr->is_CheckCastPP()) { // strip a raw-to-oop cast
3301 ptr = ptr->in(1);
3302 if (ptr == NULL) return NULL;
3303 }
3304 if (ptr->is_Proj()) {
3305 Node* allo = ptr->in(0);
3306 if (allo != NULL && allo->is_Allocate()) {
3307 return allo->as_Allocate();
3308 }
3309 }
3310 // Report failure to match.
3311 return NULL;
3312 }
3313
3314 // Fancy version which also strips off an offset (and reports it to caller).
3315 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3316 intptr_t& offset) {
3317 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3318 if (base == NULL) return NULL;
3319 return Ideal_allocation(base, phase);
3320 }
3321
3322 // Trace Initialize <- Proj[Parm] <- Allocate
3323 AllocateNode* InitializeNode::allocation() {
3324 Node* rawoop = in(InitializeNode::RawAddress);
3325 if (rawoop->is_Proj()) {
3326 Node* alloc = rawoop->in(0);
3327 if (alloc->is_Allocate()) {
3328 return alloc->as_Allocate();
3329 }
3330 }
3331 return NULL;
3332 }
3333
3334 // Trace Allocate -> Proj[Parm] -> Initialize
3335 InitializeNode* AllocateNode::initialization() {
3336 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3337 if (rawoop == NULL) return NULL;
3338 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3339 Node* init = rawoop->fast_out(i);
3340 if (init->is_Initialize()) {
3341 assert(init->as_Initialize()->allocation() == this, "2-way link");
3342 return init->as_Initialize();
3343 }
3344 }
3345 return NULL;
3346 }
3347
3348 //----------------------------- store barriers ----------------------------
3349 #define __ ideal.
3350
3351 void GraphKit::sync_kit(IdealKit& ideal) {
3352 // Final sync IdealKit and graphKit.
3353 __ drain_delay_transform();
3354 set_all_memory(__ merged_memory());
3355 set_control(__ ctrl());
3356 }
3357
3358 // vanilla/CMS post barrier
3359 // Insert a write-barrier store. This is to let generational GC work; we have
3360 // to flag all oop-stores before the next GC point.
3361 void GraphKit::write_barrier_post(Node* oop_store,
3362 Node* obj,
3363 Node* adr,
3364 uint adr_idx,
3365 Node* val,
3366 bool use_precise) {
3367 // No store check needed if we're storing a NULL or an old object
3368 // (latter case is probably a string constant). The concurrent
3369 // mark sweep garbage collector, however, needs to have all nonNull
3370 // oop updates flagged via card-marks.
3371 if (val != NULL && val->is_Con()) {
3372 // must be either an oop or NULL
3373 const Type* t = val->bottom_type();
3374 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3375 // stores of null never (?) need barriers
3376 return;
3377 ciObject* con = t->is_oopptr()->const_oop();
3378 if (con != NULL
3379 && con->is_perm()
3380 && Universe::heap()->can_elide_permanent_oop_store_barriers())
3381 // no store barrier needed, because no old-to-new ref created
3382 return;
3383 }
3384
3385 if (use_ReduceInitialCardMarks()
3386 && obj == just_allocated_object(control())) {
3387 // We can skip marks on a freshly-allocated object in Eden.
3388 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3389 // That routine informs GC to take appropriate compensating steps,
3390 // upon a slow-path allocation, so as to make this card-mark
3391 // elision safe.
3392 return;
3393 }
3394
3395 if (!use_precise) {
3396 // All card marks for a (non-array) instance are in one place:
3397 adr = obj;
3398 }
3399 // (Else it's an array (or unknown), and we want more precise card marks.)
3400 assert(adr != NULL, "");
3401
3402 IdealKit ideal(gvn(), control(), merged_memory(), true);
3403
3404 // Convert the pointer to an int prior to doing math on it
3405 Node* cast = __ CastPX(__ ctrl(), adr);
3406
3407 // Divide by card size
3408 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3409 "Only one we handle so far.");
3410 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3411
3412 // Combine card table base and card offset
3413 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3414
3415 // Get the alias_index for raw card-mark memory
3416 int adr_type = Compile::AliasIdxRaw;
3417 // Smash zero into card
3418 Node* zero = __ ConI(0);
3419 BasicType bt = T_BYTE;
3420 if( !UseConcMarkSweepGC ) {
3421 __ store(__ ctrl(), card_adr, zero, bt, adr_type);
3422 } else {
3423 // Specialized path for CM store barrier
3424 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3425 }
3426
3427 // Final sync IdealKit and GraphKit.
3428 sync_kit(ideal);
3429 }
3430
3431 // G1 pre/post barriers
3432 void GraphKit::g1_write_barrier_pre(bool do_load,
3433 Node* obj,
3434 Node* adr,
3435 uint alias_idx,
3436 Node* val,
3437 const TypeOopPtr* val_type,
3438 Node* pre_val,
3439 BasicType bt) {
3440
3441 // Some sanity checks
3442 // Note: val is unused in this routine.
3443
3444 if (do_load) {
3445 // We need to generate the load of the previous value
3446 assert(obj != NULL, "must have a base");
3447 assert(adr != NULL, "where are loading from?");
3448 assert(pre_val == NULL, "loaded already?");
3449 assert(val_type != NULL, "need a type");
3450 } else {
3451 // In this case both val_type and alias_idx are unused.
3452 assert(pre_val != NULL, "must be loaded already");
3453 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3454 }
3455 assert(bt == T_OBJECT, "or we shouldn't be here");
3456
3457 IdealKit ideal(gvn(), control(), merged_memory(), true);
3458
3459 Node* tls = __ thread(); // ThreadLocalStorage
3460
3461 Node* no_ctrl = NULL;
3462 Node* no_base = __ top();
3463 Node* zero = __ ConI(0);
3464
3465 float likely = PROB_LIKELY(0.999);
3466 float unlikely = PROB_UNLIKELY(0.999);
3467
3468 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3469 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3470
3471 // Offsets into the thread
3472 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3473 PtrQueue::byte_offset_of_active());
3474 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3475 PtrQueue::byte_offset_of_index());
3476 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3477 PtrQueue::byte_offset_of_buf());
3478
3479 // Now the actual pointers into the thread
3480 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3481 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3482 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3483
3484 // Now some of the values
3485 Node* marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3486
3487 // if (!marking)
3488 __ if_then(marking, BoolTest::ne, zero); {
3489 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3490
3491 if (do_load) {
3492 // load original value
3493 // alias_idx correct??
3494 pre_val = __ load(no_ctrl, adr, val_type, bt, alias_idx);
3495 }
3496
3497 // if (pre_val != NULL)
3498 __ if_then(pre_val, BoolTest::ne, null()); {
3499 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3500
3501 // is the queue for this thread full?
3502 __ if_then(index, BoolTest::ne, zero, likely); {
3503
3504 // decrement the index
3505 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3506 Node* next_indexX = next_index;
3507 #ifdef _LP64
3508 // We could refine the type for what it's worth
3509 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3510 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3511 #endif
3512
3513 // Now get the buffer location we will log the previous value into and store it
3514 Node *log_addr = __ AddP(no_base, buffer, next_indexX);
3515 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw);
3516 // update the index
3517 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3518
3519 } __ else_(); {
3520
3521 // logging buffer is full, call the runtime
3522 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3523 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3524 } __ end_if(); // (!index)
3525 } __ end_if(); // (pre_val != NULL)
3526 } __ end_if(); // (!marking)
3527
3528 // Final sync IdealKit and GraphKit.
3529 sync_kit(ideal);
3530 }
3531
3532 //
3533 // Update the card table and add card address to the queue
3534 //
3535 void GraphKit::g1_mark_card(IdealKit& ideal,
3536 Node* card_adr,
3537 Node* oop_store,
3538 uint oop_alias_idx,
3539 Node* index,
3540 Node* index_adr,
3541 Node* buffer,
3542 const TypeFunc* tf) {
3543
3544 Node* zero = __ ConI(0);
3545 Node* no_base = __ top();
3546 BasicType card_bt = T_BYTE;
3547 // Smash zero into card. MUST BE ORDERED WRT TO STORE
3548 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
3549
3550 // Now do the queue work
3551 __ if_then(index, BoolTest::ne, zero); {
3552
3553 Node* next_index = __ SubI(index, __ ConI(sizeof(intptr_t)));
3554 Node* next_indexX = next_index;
3555 #ifdef _LP64
3556 // We could refine the type for what it's worth
3557 // const TypeLong* lidxtype = TypeLong::make(CONST64(0), get_size_from_queue);
3558 next_indexX = _gvn.transform( new (C, 2) ConvI2LNode(next_index, TypeLong::make(0, max_jlong, Type::WidenMax)) );
3559 #endif // _LP64
3560 Node* log_addr = __ AddP(no_base, buffer, next_indexX);
3561
3562 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw);
3563 __ store(__ ctrl(), index_adr, next_index, T_INT, Compile::AliasIdxRaw);
3564
3565 } __ else_(); {
3566 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
3567 } __ end_if();
3568
3569 }
3570
3571 void GraphKit::g1_write_barrier_post(Node* oop_store,
3572 Node* obj,
3573 Node* adr,
3574 uint alias_idx,
3575 Node* val,
3576 BasicType bt,
3577 bool use_precise) {
3578 // If we are writing a NULL then we need no post barrier
3579
3580 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
3581 // Must be NULL
3582 const Type* t = val->bottom_type();
3583 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
3584 // No post barrier if writing NULLx
3585 return;
3586 }
3587
3588 if (!use_precise) {
3589 // All card marks for a (non-array) instance are in one place:
3590 adr = obj;
3591 }
3592 // (Else it's an array (or unknown), and we want more precise card marks.)
3593 assert(adr != NULL, "");
3594
3595 IdealKit ideal(gvn(), control(), merged_memory(), true);
3596
3597 Node* tls = __ thread(); // ThreadLocalStorage
3598
3599 Node* no_base = __ top();
3600 float likely = PROB_LIKELY(0.999);
3601 float unlikely = PROB_UNLIKELY(0.999);
3602 Node* zero = __ ConI(0);
3603 Node* zeroX = __ ConX(0);
3604
3605 // Get the alias_index for raw card-mark memory
3606 const TypePtr* card_type = TypeRawPtr::BOTTOM;
3607
3608 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
3609
3610 // Offsets into the thread
3611 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3612 PtrQueue::byte_offset_of_index());
3613 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
3614 PtrQueue::byte_offset_of_buf());
3615
3616 // Pointers into the thread
3617
3618 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3619 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3620
3621 // Now some values
3622 // Use ctrl to avoid hoisting these values past a safepoint, which could
3623 // potentially reset these fields in the JavaThread.
3624 Node* index = __ load(__ ctrl(), index_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
3625 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3626
3627 // Convert the store obj pointer to an int prior to doing math on it
3628 // Must use ctrl to prevent "integerized oop" existing across safepoint
3629 Node* cast = __ CastPX(__ ctrl(), adr);
3630
3631 // Divide pointer by card size
3632 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3633
3634 // Combine card table base and card offset
3635 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
3636
3637 // If we know the value being stored does it cross regions?
3638
3639 if (val != NULL) {
3640 // Does the store cause us to cross regions?
3641
3642 // Should be able to do an unsigned compare of region_size instead of
3643 // and extra shift. Do we have an unsigned compare??
3644 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
3645 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
3646
3647 // if (xor_res == 0) same region so skip
3648 __ if_then(xor_res, BoolTest::ne, zeroX); {
3649
3650 // No barrier if we are storing a NULL
3651 __ if_then(val, BoolTest::ne, null(), unlikely); {
3652
3653 // Ok must mark the card if not already dirty
3654
3655 // load the original value of the card
3656 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
3657
3658 __ if_then(card_val, BoolTest::ne, zero); {
3659 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3660 } __ end_if();
3661 } __ end_if();
3662 } __ end_if();
3663 } else {
3664 // Object.clone() instrinsic uses this path.
3665 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
3666 }
3667
3668 // Final sync IdealKit and GraphKit.
3669 sync_kit(ideal);
3670 }
3671 #undef __