1 /*
2 * Copyright (c) 2001, 2018, 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 "ci/ciUtilities.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/g1/g1BarrierSet.hpp"
29 #include "gc/g1/g1CardTable.hpp"
30 #include "gc/g1/heapRegion.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/cardTable.hpp"
33 #include "gc/shared/cardTableBarrierSet.hpp"
34 #include "gc/shared/collectedHeap.hpp"
35 #include "gc/shenandoah/brooksPointer.hpp"
36 #include "gc/shenandoah/shenandoahConnectionMatrix.hpp"
37 #include "gc/shenandoah/shenandoahHeap.hpp"
38 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
39 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
40 #include "interpreter/interpreter.hpp"
41 #include "memory/resourceArea.hpp"
42 #include "opto/addnode.hpp"
43 #include "opto/castnode.hpp"
44 #include "opto/convertnode.hpp"
45 #include "opto/graphKit.hpp"
46 #include "opto/idealKit.hpp"
47 #include "opto/intrinsicnode.hpp"
48 #include "opto/locknode.hpp"
49 #include "opto/machnode.hpp"
50 #include "opto/opaquenode.hpp"
51 #include "opto/parse.hpp"
52 #include "opto/rootnode.hpp"
53 #include "opto/runtime.hpp"
54 #include "opto/shenandoahSupport.hpp"
55 #include "runtime/deoptimization.hpp"
56 #include "runtime/sharedRuntime.hpp"
57 #if INCLUDE_ALL_GCS
58 #include "gc/g1/g1ThreadLocalData.hpp"
59 #endif // INCLUDE_ALL_GCS
60
61 //----------------------------GraphKit-----------------------------------------
62 // Main utility constructor.
63 GraphKit::GraphKit(JVMState* jvms)
64 : Phase(Phase::Parser),
65 _env(C->env()),
66 _gvn(*C->initial_gvn())
67 {
68 _exceptions = jvms->map()->next_exception();
69 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
70 set_jvms(jvms);
71 }
72
73 // Private constructor for parser.
74 GraphKit::GraphKit()
75 : Phase(Phase::Parser),
76 _env(C->env()),
77 _gvn(*C->initial_gvn())
78 {
79 _exceptions = NULL;
80 set_map(NULL);
81 debug_only(_sp = -99);
82 debug_only(set_bci(-99));
83 }
84
85
86
87 //---------------------------clean_stack---------------------------------------
88 // Clear away rubbish from the stack area of the JVM state.
89 // This destroys any arguments that may be waiting on the stack.
90 void GraphKit::clean_stack(int from_sp) {
91 SafePointNode* map = this->map();
92 JVMState* jvms = this->jvms();
93 int stk_size = jvms->stk_size();
94 int stkoff = jvms->stkoff();
95 Node* top = this->top();
96 for (int i = from_sp; i < stk_size; i++) {
97 if (map->in(stkoff + i) != top) {
98 map->set_req(stkoff + i, top);
99 }
100 }
101 }
102
103
104 //--------------------------------sync_jvms-----------------------------------
105 // Make sure our current jvms agrees with our parse state.
106 JVMState* GraphKit::sync_jvms() const {
107 JVMState* jvms = this->jvms();
108 jvms->set_bci(bci()); // Record the new bci in the JVMState
109 jvms->set_sp(sp()); // Record the new sp in the JVMState
110 assert(jvms_in_sync(), "jvms is now in sync");
111 return jvms;
112 }
113
114 //--------------------------------sync_jvms_for_reexecute---------------------
115 // Make sure our current jvms agrees with our parse state. This version
116 // uses the reexecute_sp for reexecuting bytecodes.
117 JVMState* GraphKit::sync_jvms_for_reexecute() {
118 JVMState* jvms = this->jvms();
119 jvms->set_bci(bci()); // Record the new bci in the JVMState
120 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
121 return jvms;
122 }
123
124 #ifdef ASSERT
125 bool GraphKit::jvms_in_sync() const {
126 Parse* parse = is_Parse();
127 if (parse == NULL) {
128 if (bci() != jvms()->bci()) return false;
129 if (sp() != (int)jvms()->sp()) return false;
130 return true;
131 }
132 if (jvms()->method() != parse->method()) return false;
133 if (jvms()->bci() != parse->bci()) return false;
134 int jvms_sp = jvms()->sp();
135 if (jvms_sp != parse->sp()) return false;
136 int jvms_depth = jvms()->depth();
137 if (jvms_depth != parse->depth()) return false;
138 return true;
139 }
140
141 // Local helper checks for special internal merge points
142 // used to accumulate and merge exception states.
143 // They are marked by the region's in(0) edge being the map itself.
144 // Such merge points must never "escape" into the parser at large,
145 // until they have been handed to gvn.transform.
146 static bool is_hidden_merge(Node* reg) {
147 if (reg == NULL) return false;
148 if (reg->is_Phi()) {
149 reg = reg->in(0);
150 if (reg == NULL) return false;
151 }
152 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
153 }
154
155 void GraphKit::verify_map() const {
156 if (map() == NULL) return; // null map is OK
157 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
158 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
159 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
160 }
161
162 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
163 assert(ex_map->next_exception() == NULL, "not already part of a chain");
164 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
165 }
166 #endif
167
168 //---------------------------stop_and_kill_map---------------------------------
169 // Set _map to NULL, signalling a stop to further bytecode execution.
170 // First smash the current map's control to a constant, to mark it dead.
171 void GraphKit::stop_and_kill_map() {
172 SafePointNode* dead_map = stop();
173 if (dead_map != NULL) {
174 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
175 assert(dead_map->is_killed(), "must be so marked");
176 }
177 }
178
179
180 //--------------------------------stopped--------------------------------------
181 // Tell if _map is NULL, or control is top.
182 bool GraphKit::stopped() {
183 if (map() == NULL) return true;
184 else if (control() == top()) return true;
185 else return false;
186 }
187
188
189 //-----------------------------has_ex_handler----------------------------------
190 // Tell if this method or any caller method has exception handlers.
191 bool GraphKit::has_ex_handler() {
192 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
193 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
194 return true;
195 }
196 }
197 return false;
198 }
199
200 //------------------------------save_ex_oop------------------------------------
201 // Save an exception without blowing stack contents or other JVM state.
202 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
203 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
204 ex_map->add_req(ex_oop);
205 debug_only(verify_exception_state(ex_map));
206 }
207
208 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
209 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
210 Node* ex_oop = ex_map->in(ex_map->req()-1);
211 if (clear_it) ex_map->del_req(ex_map->req()-1);
212 return ex_oop;
213 }
214
215 //-----------------------------saved_ex_oop------------------------------------
216 // Recover a saved exception from its map.
217 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
218 return common_saved_ex_oop(ex_map, false);
219 }
220
221 //--------------------------clear_saved_ex_oop---------------------------------
222 // Erase a previously saved exception from its map.
223 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
224 return common_saved_ex_oop(ex_map, true);
225 }
226
227 #ifdef ASSERT
228 //---------------------------has_saved_ex_oop----------------------------------
229 // Erase a previously saved exception from its map.
230 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
231 return ex_map->req() == ex_map->jvms()->endoff()+1;
232 }
233 #endif
234
235 //-------------------------make_exception_state--------------------------------
236 // Turn the current JVM state into an exception state, appending the ex_oop.
237 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
238 sync_jvms();
239 SafePointNode* ex_map = stop(); // do not manipulate this map any more
240 set_saved_ex_oop(ex_map, ex_oop);
241 return ex_map;
242 }
243
244
245 //--------------------------add_exception_state--------------------------------
246 // Add an exception to my list of exceptions.
247 void GraphKit::add_exception_state(SafePointNode* ex_map) {
248 if (ex_map == NULL || ex_map->control() == top()) {
249 return;
250 }
251 #ifdef ASSERT
252 verify_exception_state(ex_map);
253 if (has_exceptions()) {
254 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
255 }
256 #endif
257
258 // If there is already an exception of exactly this type, merge with it.
259 // In particular, null-checks and other low-level exceptions common up here.
260 Node* ex_oop = saved_ex_oop(ex_map);
261 const Type* ex_type = _gvn.type(ex_oop);
262 if (ex_oop == top()) {
263 // No action needed.
264 return;
265 }
266 assert(ex_type->isa_instptr(), "exception must be an instance");
267 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
268 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
269 // We check sp also because call bytecodes can generate exceptions
270 // both before and after arguments are popped!
271 if (ex_type2 == ex_type
272 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
273 combine_exception_states(ex_map, e2);
274 return;
275 }
276 }
277
278 // No pre-existing exception of the same type. Chain it on the list.
279 push_exception_state(ex_map);
280 }
281
282 //-----------------------add_exception_states_from-----------------------------
283 void GraphKit::add_exception_states_from(JVMState* jvms) {
284 SafePointNode* ex_map = jvms->map()->next_exception();
285 if (ex_map != NULL) {
286 jvms->map()->set_next_exception(NULL);
287 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
288 next_map = ex_map->next_exception();
289 ex_map->set_next_exception(NULL);
290 add_exception_state(ex_map);
291 }
292 }
293 }
294
295 //-----------------------transfer_exceptions_into_jvms-------------------------
296 JVMState* GraphKit::transfer_exceptions_into_jvms() {
297 if (map() == NULL) {
298 // We need a JVMS to carry the exceptions, but the map has gone away.
299 // Create a scratch JVMS, cloned from any of the exception states...
300 if (has_exceptions()) {
301 _map = _exceptions;
302 _map = clone_map();
303 _map->set_next_exception(NULL);
304 clear_saved_ex_oop(_map);
305 debug_only(verify_map());
306 } else {
307 // ...or created from scratch
308 JVMState* jvms = new (C) JVMState(_method, NULL);
309 jvms->set_bci(_bci);
310 jvms->set_sp(_sp);
311 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
312 set_jvms(jvms);
313 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
314 set_all_memory(top());
315 while (map()->req() < jvms->endoff()) map()->add_req(top());
316 }
317 // (This is a kludge, in case you didn't notice.)
318 set_control(top());
319 }
320 JVMState* jvms = sync_jvms();
321 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
322 jvms->map()->set_next_exception(_exceptions);
323 _exceptions = NULL; // done with this set of exceptions
324 return jvms;
325 }
326
327 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
328 assert(is_hidden_merge(dstphi), "must be a special merge node");
329 assert(is_hidden_merge(srcphi), "must be a special merge node");
330 uint limit = srcphi->req();
331 for (uint i = PhiNode::Input; i < limit; i++) {
332 dstphi->add_req(srcphi->in(i));
333 }
334 }
335 static inline void add_one_req(Node* dstphi, Node* src) {
336 assert(is_hidden_merge(dstphi), "must be a special merge node");
337 assert(!is_hidden_merge(src), "must not be a special merge node");
338 dstphi->add_req(src);
339 }
340
341 //-----------------------combine_exception_states------------------------------
342 // This helper function combines exception states by building phis on a
343 // specially marked state-merging region. These regions and phis are
344 // untransformed, and can build up gradually. The region is marked by
345 // having a control input of its exception map, rather than NULL. Such
346 // regions do not appear except in this function, and in use_exception_state.
347 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
348 if (failing()) return; // dying anyway...
349 JVMState* ex_jvms = ex_map->_jvms;
350 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
351 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
352 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
353 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
354 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
355 assert(ex_map->req() == phi_map->req(), "matching maps");
356 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
357 Node* hidden_merge_mark = root();
358 Node* region = phi_map->control();
359 MergeMemNode* phi_mem = phi_map->merged_memory();
360 MergeMemNode* ex_mem = ex_map->merged_memory();
361 if (region->in(0) != hidden_merge_mark) {
362 // The control input is not (yet) a specially-marked region in phi_map.
363 // Make it so, and build some phis.
364 region = new RegionNode(2);
365 _gvn.set_type(region, Type::CONTROL);
366 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
367 region->init_req(1, phi_map->control());
368 phi_map->set_control(region);
369 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
370 record_for_igvn(io_phi);
371 _gvn.set_type(io_phi, Type::ABIO);
372 phi_map->set_i_o(io_phi);
373 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
374 Node* m = mms.memory();
375 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
376 record_for_igvn(m_phi);
377 _gvn.set_type(m_phi, Type::MEMORY);
378 mms.set_memory(m_phi);
379 }
380 }
381
382 // Either or both of phi_map and ex_map might already be converted into phis.
383 Node* ex_control = ex_map->control();
384 // if there is special marking on ex_map also, we add multiple edges from src
385 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
386 // how wide was the destination phi_map, originally?
387 uint orig_width = region->req();
388
389 if (add_multiple) {
390 add_n_reqs(region, ex_control);
391 add_n_reqs(phi_map->i_o(), ex_map->i_o());
392 } else {
393 // ex_map has no merges, so we just add single edges everywhere
394 add_one_req(region, ex_control);
395 add_one_req(phi_map->i_o(), ex_map->i_o());
396 }
397 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
398 if (mms.is_empty()) {
399 // get a copy of the base memory, and patch some inputs into it
400 const TypePtr* adr_type = mms.adr_type(C);
401 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
402 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
403 mms.set_memory(phi);
404 // Prepare to append interesting stuff onto the newly sliced phi:
405 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
406 }
407 // Append stuff from ex_map:
408 if (add_multiple) {
409 add_n_reqs(mms.memory(), mms.memory2());
410 } else {
411 add_one_req(mms.memory(), mms.memory2());
412 }
413 }
414 uint limit = ex_map->req();
415 for (uint i = TypeFunc::Parms; i < limit; i++) {
416 // Skip everything in the JVMS after tos. (The ex_oop follows.)
417 if (i == tos) i = ex_jvms->monoff();
418 Node* src = ex_map->in(i);
419 Node* dst = phi_map->in(i);
420 if (src != dst) {
421 PhiNode* phi;
422 if (dst->in(0) != region) {
423 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
424 record_for_igvn(phi);
425 _gvn.set_type(phi, phi->type());
426 phi_map->set_req(i, dst);
427 // Prepare to append interesting stuff onto the new phi:
428 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
429 } else {
430 assert(dst->is_Phi(), "nobody else uses a hidden region");
431 phi = dst->as_Phi();
432 }
433 if (add_multiple && src->in(0) == ex_control) {
434 // Both are phis.
435 add_n_reqs(dst, src);
436 } else {
437 while (dst->req() < region->req()) add_one_req(dst, src);
438 }
439 const Type* srctype = _gvn.type(src);
440 if (phi->type() != srctype) {
441 const Type* dsttype = phi->type()->meet_speculative(srctype);
442 if (phi->type() != dsttype) {
443 phi->set_type(dsttype);
444 _gvn.set_type(phi, dsttype);
445 }
446 }
447 }
448 }
449 phi_map->merge_replaced_nodes_with(ex_map);
450 }
451
452 //--------------------------use_exception_state--------------------------------
453 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
454 if (failing()) { stop(); return top(); }
455 Node* region = phi_map->control();
456 Node* hidden_merge_mark = root();
457 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
458 Node* ex_oop = clear_saved_ex_oop(phi_map);
459 if (region->in(0) == hidden_merge_mark) {
460 // Special marking for internal ex-states. Process the phis now.
461 region->set_req(0, region); // now it's an ordinary region
462 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
463 // Note: Setting the jvms also sets the bci and sp.
464 set_control(_gvn.transform(region));
465 uint tos = jvms()->stkoff() + sp();
466 for (uint i = 1; i < tos; i++) {
467 Node* x = phi_map->in(i);
468 if (x->in(0) == region) {
469 assert(x->is_Phi(), "expected a special phi");
470 phi_map->set_req(i, _gvn.transform(x));
471 }
472 }
473 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
474 Node* x = mms.memory();
475 if (x->in(0) == region) {
476 assert(x->is_Phi(), "nobody else uses a hidden region");
477 mms.set_memory(_gvn.transform(x));
478 }
479 }
480 if (ex_oop->in(0) == region) {
481 assert(ex_oop->is_Phi(), "expected a special phi");
482 ex_oop = _gvn.transform(ex_oop);
483 }
484 } else {
485 set_jvms(phi_map->jvms());
486 }
487
488 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
489 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
490 return ex_oop;
491 }
492
493 //---------------------------------java_bc-------------------------------------
494 Bytecodes::Code GraphKit::java_bc() const {
495 ciMethod* method = this->method();
496 int bci = this->bci();
497 if (method != NULL && bci != InvocationEntryBci)
498 return method->java_code_at_bci(bci);
499 else
500 return Bytecodes::_illegal;
501 }
502
503 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
504 bool must_throw) {
505 // if the exception capability is set, then we will generate code
506 // to check the JavaThread.should_post_on_exceptions flag to see
507 // if we actually need to report exception events (for this
508 // thread). If we don't need to report exception events, we will
509 // take the normal fast path provided by add_exception_events. If
510 // exception event reporting is enabled for this thread, we will
511 // take the uncommon_trap in the BuildCutout below.
512
513 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
514 Node* jthread = _gvn.transform(new ThreadLocalNode());
515 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
516 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
517
518 // Test the should_post_on_exceptions_flag vs. 0
519 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
520 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
521
522 // Branch to slow_path if should_post_on_exceptions_flag was true
523 { BuildCutout unless(this, tst, PROB_MAX);
524 // Do not try anything fancy if we're notifying the VM on every throw.
525 // Cf. case Bytecodes::_athrow in parse2.cpp.
526 uncommon_trap(reason, Deoptimization::Action_none,
527 (ciKlass*)NULL, (char*)NULL, must_throw);
528 }
529
530 }
531
532 //------------------------------builtin_throw----------------------------------
533 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
534 bool must_throw = true;
535
536 if (env()->jvmti_can_post_on_exceptions()) {
537 // check if we must post exception events, take uncommon trap if so
538 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
539 // here if should_post_on_exceptions is false
540 // continue on with the normal codegen
541 }
542
543 // If this particular condition has not yet happened at this
544 // bytecode, then use the uncommon trap mechanism, and allow for
545 // a future recompilation if several traps occur here.
546 // If the throw is hot, try to use a more complicated inline mechanism
547 // which keeps execution inside the compiled code.
548 bool treat_throw_as_hot = false;
549 ciMethodData* md = method()->method_data();
550
551 if (ProfileTraps) {
552 if (too_many_traps(reason)) {
553 treat_throw_as_hot = true;
554 }
555 // (If there is no MDO at all, assume it is early in
556 // execution, and that any deopts are part of the
557 // startup transient, and don't need to be remembered.)
558
559 // Also, if there is a local exception handler, treat all throws
560 // as hot if there has been at least one in this method.
561 if (C->trap_count(reason) != 0
562 && method()->method_data()->trap_count(reason) != 0
563 && has_ex_handler()) {
564 treat_throw_as_hot = true;
565 }
566 }
567
568 // If this throw happens frequently, an uncommon trap might cause
569 // a performance pothole. If there is a local exception handler,
570 // and if this particular bytecode appears to be deoptimizing often,
571 // let us handle the throw inline, with a preconstructed instance.
572 // Note: If the deopt count has blown up, the uncommon trap
573 // runtime is going to flush this nmethod, not matter what.
574 if (treat_throw_as_hot
575 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
576 // If the throw is local, we use a pre-existing instance and
577 // punt on the backtrace. This would lead to a missing backtrace
578 // (a repeat of 4292742) if the backtrace object is ever asked
579 // for its backtrace.
580 // Fixing this remaining case of 4292742 requires some flavor of
581 // escape analysis. Leave that for the future.
582 ciInstance* ex_obj = NULL;
583 switch (reason) {
584 case Deoptimization::Reason_null_check:
585 ex_obj = env()->NullPointerException_instance();
586 break;
587 case Deoptimization::Reason_div0_check:
588 ex_obj = env()->ArithmeticException_instance();
589 break;
590 case Deoptimization::Reason_range_check:
591 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
592 break;
593 case Deoptimization::Reason_class_check:
594 if (java_bc() == Bytecodes::_aastore) {
595 ex_obj = env()->ArrayStoreException_instance();
596 } else {
597 ex_obj = env()->ClassCastException_instance();
598 }
599 break;
600 default:
601 break;
602 }
603 if (failing()) { stop(); return; } // exception allocation might fail
604 if (ex_obj != NULL) {
605 // Cheat with a preallocated exception object.
606 if (C->log() != NULL)
607 C->log()->elem("hot_throw preallocated='1' reason='%s'",
608 Deoptimization::trap_reason_name(reason));
609 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
610 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
611
612 ex_node = shenandoah_write_barrier(ex_node);
613
614 // Clear the detail message of the preallocated exception object.
615 // Weblogic sometimes mutates the detail message of exceptions
616 // using reflection.
617 int offset = java_lang_Throwable::get_detailMessage_offset();
618 const TypePtr* adr_typ = ex_con->add_offset(offset);
619
620 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
621 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
622 // Conservatively release stores of object references.
623 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
624
625 add_exception_state(make_exception_state(ex_node));
626 return;
627 }
628 }
629
630 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
631 // It won't be much cheaper than bailing to the interp., since we'll
632 // have to pass up all the debug-info, and the runtime will have to
633 // create the stack trace.
634
635 // Usual case: Bail to interpreter.
636 // Reserve the right to recompile if we haven't seen anything yet.
637
638 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
639 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
640 if (treat_throw_as_hot
641 && (method()->method_data()->trap_recompiled_at(bci(), m)
642 || C->too_many_traps(reason))) {
643 // We cannot afford to take more traps here. Suffer in the interpreter.
644 if (C->log() != NULL)
645 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
646 Deoptimization::trap_reason_name(reason),
647 C->trap_count(reason));
648 action = Deoptimization::Action_none;
649 }
650
651 // "must_throw" prunes the JVM state to include only the stack, if there
652 // are no local exception handlers. This should cut down on register
653 // allocation time and code size, by drastically reducing the number
654 // of in-edges on the call to the uncommon trap.
655
656 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
657 }
658
659
660 //----------------------------PreserveJVMState---------------------------------
661 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
662 debug_only(kit->verify_map());
663 _kit = kit;
664 _map = kit->map(); // preserve the map
665 _sp = kit->sp();
666 kit->set_map(clone_map ? kit->clone_map() : NULL);
667 #ifdef ASSERT
668 _bci = kit->bci();
669 Parse* parser = kit->is_Parse();
670 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
671 _block = block;
672 #endif
673 }
674 PreserveJVMState::~PreserveJVMState() {
675 GraphKit* kit = _kit;
676 #ifdef ASSERT
677 assert(kit->bci() == _bci, "bci must not shift");
678 Parse* parser = kit->is_Parse();
679 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
680 assert(block == _block, "block must not shift");
681 #endif
682 kit->set_map(_map);
683 kit->set_sp(_sp);
684 }
685
686
687 //-----------------------------BuildCutout-------------------------------------
688 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
689 : PreserveJVMState(kit)
690 {
691 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
692 SafePointNode* outer_map = _map; // preserved map is caller's
693 SafePointNode* inner_map = kit->map();
694 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
695 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
696 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
697 }
698 BuildCutout::~BuildCutout() {
699 GraphKit* kit = _kit;
700 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
701 }
702
703 //---------------------------PreserveReexecuteState----------------------------
704 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
705 assert(!kit->stopped(), "must call stopped() before");
706 _kit = kit;
707 _sp = kit->sp();
708 _reexecute = kit->jvms()->_reexecute;
709 }
710 PreserveReexecuteState::~PreserveReexecuteState() {
711 if (_kit->stopped()) return;
712 _kit->jvms()->_reexecute = _reexecute;
713 _kit->set_sp(_sp);
714 }
715
716 //------------------------------clone_map--------------------------------------
717 // Implementation of PreserveJVMState
718 //
719 // Only clone_map(...) here. If this function is only used in the
720 // PreserveJVMState class we may want to get rid of this extra
721 // function eventually and do it all there.
722
723 SafePointNode* GraphKit::clone_map() {
724 if (map() == NULL) return NULL;
725
726 // Clone the memory edge first
727 Node* mem = MergeMemNode::make(map()->memory());
728 gvn().set_type_bottom(mem);
729
730 SafePointNode *clonemap = (SafePointNode*)map()->clone();
731 JVMState* jvms = this->jvms();
732 JVMState* clonejvms = jvms->clone_shallow(C);
733 clonemap->set_memory(mem);
734 clonemap->set_jvms(clonejvms);
735 clonejvms->set_map(clonemap);
736 record_for_igvn(clonemap);
737 gvn().set_type_bottom(clonemap);
738 return clonemap;
739 }
740
741
742 //-----------------------------set_map_clone-----------------------------------
743 void GraphKit::set_map_clone(SafePointNode* m) {
744 _map = m;
745 _map = clone_map();
746 _map->set_next_exception(NULL);
747 debug_only(verify_map());
748 }
749
750
751 //----------------------------kill_dead_locals---------------------------------
752 // Detect any locals which are known to be dead, and force them to top.
753 void GraphKit::kill_dead_locals() {
754 // Consult the liveness information for the locals. If any
755 // of them are unused, then they can be replaced by top(). This
756 // should help register allocation time and cut down on the size
757 // of the deoptimization information.
758
759 // This call is made from many of the bytecode handling
760 // subroutines called from the Big Switch in do_one_bytecode.
761 // Every bytecode which might include a slow path is responsible
762 // for killing its dead locals. The more consistent we
763 // are about killing deads, the fewer useless phis will be
764 // constructed for them at various merge points.
765
766 // bci can be -1 (InvocationEntryBci). We return the entry
767 // liveness for the method.
768
769 if (method() == NULL || method()->code_size() == 0) {
770 // We are building a graph for a call to a native method.
771 // All locals are live.
772 return;
773 }
774
775 ResourceMark rm;
776
777 // Consult the liveness information for the locals. If any
778 // of them are unused, then they can be replaced by top(). This
779 // should help register allocation time and cut down on the size
780 // of the deoptimization information.
781 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
782
783 int len = (int)live_locals.size();
784 assert(len <= jvms()->loc_size(), "too many live locals");
785 for (int local = 0; local < len; local++) {
786 if (!live_locals.at(local)) {
787 set_local(local, top());
788 }
789 }
790 }
791
792 #ifdef ASSERT
793 //-------------------------dead_locals_are_killed------------------------------
794 // Return true if all dead locals are set to top in the map.
795 // Used to assert "clean" debug info at various points.
796 bool GraphKit::dead_locals_are_killed() {
797 if (method() == NULL || method()->code_size() == 0) {
798 // No locals need to be dead, so all is as it should be.
799 return true;
800 }
801
802 // Make sure somebody called kill_dead_locals upstream.
803 ResourceMark rm;
804 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
805 if (jvms->loc_size() == 0) continue; // no locals to consult
806 SafePointNode* map = jvms->map();
807 ciMethod* method = jvms->method();
808 int bci = jvms->bci();
809 if (jvms == this->jvms()) {
810 bci = this->bci(); // it might not yet be synched
811 }
812 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
813 int len = (int)live_locals.size();
814 if (!live_locals.is_valid() || len == 0)
815 // This method is trivial, or is poisoned by a breakpoint.
816 return true;
817 assert(len == jvms->loc_size(), "live map consistent with locals map");
818 for (int local = 0; local < len; local++) {
819 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
820 if (PrintMiscellaneous && (Verbose || WizardMode)) {
821 tty->print_cr("Zombie local %d: ", local);
822 jvms->dump();
823 }
824 return false;
825 }
826 }
827 }
828 return true;
829 }
830
831 #endif //ASSERT
832
833 // Helper function for enforcing certain bytecodes to reexecute if
834 // deoptimization happens
835 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
836 ciMethod* cur_method = jvms->method();
837 int cur_bci = jvms->bci();
838 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
839 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
840 return Interpreter::bytecode_should_reexecute(code) ||
841 (is_anewarray && code == Bytecodes::_multianewarray);
842 // Reexecute _multianewarray bytecode which was replaced with
843 // sequence of [a]newarray. See Parse::do_multianewarray().
844 //
845 // Note: interpreter should not have it set since this optimization
846 // is limited by dimensions and guarded by flag so in some cases
847 // multianewarray() runtime calls will be generated and
848 // the bytecode should not be reexecutes (stack will not be reset).
849 } else
850 return false;
851 }
852
853 // Helper function for adding JVMState and debug information to node
854 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
855 // Add the safepoint edges to the call (or other safepoint).
856
857 // Make sure dead locals are set to top. This
858 // should help register allocation time and cut down on the size
859 // of the deoptimization information.
860 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
861
862 // Walk the inline list to fill in the correct set of JVMState's
863 // Also fill in the associated edges for each JVMState.
864
865 // If the bytecode needs to be reexecuted we need to put
866 // the arguments back on the stack.
867 const bool should_reexecute = jvms()->should_reexecute();
868 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
869
870 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
871 // undefined if the bci is different. This is normal for Parse but it
872 // should not happen for LibraryCallKit because only one bci is processed.
873 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
874 "in LibraryCallKit the reexecute bit should not change");
875
876 // If we are guaranteed to throw, we can prune everything but the
877 // input to the current bytecode.
878 bool can_prune_locals = false;
879 uint stack_slots_not_pruned = 0;
880 int inputs = 0, depth = 0;
881 if (must_throw) {
882 assert(method() == youngest_jvms->method(), "sanity");
883 if (compute_stack_effects(inputs, depth)) {
884 can_prune_locals = true;
885 stack_slots_not_pruned = inputs;
886 }
887 }
888
889 if (env()->should_retain_local_variables()) {
890 // At any safepoint, this method can get breakpointed, which would
891 // then require an immediate deoptimization.
892 can_prune_locals = false; // do not prune locals
893 stack_slots_not_pruned = 0;
894 }
895
896 // do not scribble on the input jvms
897 JVMState* out_jvms = youngest_jvms->clone_deep(C);
898 call->set_jvms(out_jvms); // Start jvms list for call node
899
900 // For a known set of bytecodes, the interpreter should reexecute them if
901 // deoptimization happens. We set the reexecute state for them here
902 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
903 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
904 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
905 }
906
907 // Presize the call:
908 DEBUG_ONLY(uint non_debug_edges = call->req());
909 call->add_req_batch(top(), youngest_jvms->debug_depth());
910 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
911
912 // Set up edges so that the call looks like this:
913 // Call [state:] ctl io mem fptr retadr
914 // [parms:] parm0 ... parmN
915 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
916 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
917 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
918 // Note that caller debug info precedes callee debug info.
919
920 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
921 uint debug_ptr = call->req();
922
923 // Loop over the map input edges associated with jvms, add them
924 // to the call node, & reset all offsets to match call node array.
925 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
926 uint debug_end = debug_ptr;
927 uint debug_start = debug_ptr - in_jvms->debug_size();
928 debug_ptr = debug_start; // back up the ptr
929
930 uint p = debug_start; // walks forward in [debug_start, debug_end)
931 uint j, k, l;
932 SafePointNode* in_map = in_jvms->map();
933 out_jvms->set_map(call);
934
935 if (can_prune_locals) {
936 assert(in_jvms->method() == out_jvms->method(), "sanity");
937 // If the current throw can reach an exception handler in this JVMS,
938 // then we must keep everything live that can reach that handler.
939 // As a quick and dirty approximation, we look for any handlers at all.
940 if (in_jvms->method()->has_exception_handlers()) {
941 can_prune_locals = false;
942 }
943 }
944
945 // Add the Locals
946 k = in_jvms->locoff();
947 l = in_jvms->loc_size();
948 out_jvms->set_locoff(p);
949 if (!can_prune_locals) {
950 for (j = 0; j < l; j++)
951 call->set_req(p++, in_map->in(k+j));
952 } else {
953 p += l; // already set to top above by add_req_batch
954 }
955
956 // Add the Expression Stack
957 k = in_jvms->stkoff();
958 l = in_jvms->sp();
959 out_jvms->set_stkoff(p);
960 if (!can_prune_locals) {
961 for (j = 0; j < l; j++)
962 call->set_req(p++, in_map->in(k+j));
963 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
964 // Divide stack into {S0,...,S1}, where S0 is set to top.
965 uint s1 = stack_slots_not_pruned;
966 stack_slots_not_pruned = 0; // for next iteration
967 if (s1 > l) s1 = l;
968 uint s0 = l - s1;
969 p += s0; // skip the tops preinstalled by add_req_batch
970 for (j = s0; j < l; j++)
971 call->set_req(p++, in_map->in(k+j));
972 } else {
973 p += l; // already set to top above by add_req_batch
974 }
975
976 // Add the Monitors
977 k = in_jvms->monoff();
978 l = in_jvms->mon_size();
979 out_jvms->set_monoff(p);
980 for (j = 0; j < l; j++)
981 call->set_req(p++, in_map->in(k+j));
982
983 // Copy any scalar object fields.
984 k = in_jvms->scloff();
985 l = in_jvms->scl_size();
986 out_jvms->set_scloff(p);
987 for (j = 0; j < l; j++)
988 call->set_req(p++, in_map->in(k+j));
989
990 // Finish the new jvms.
991 out_jvms->set_endoff(p);
992
993 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
994 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
995 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
996 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
997 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
998 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
999
1000 // Update the two tail pointers in parallel.
1001 out_jvms = out_jvms->caller();
1002 in_jvms = in_jvms->caller();
1003 }
1004
1005 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1006
1007 // Test the correctness of JVMState::debug_xxx accessors:
1008 assert(call->jvms()->debug_start() == non_debug_edges, "");
1009 assert(call->jvms()->debug_end() == call->req(), "");
1010 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1011 }
1012
1013 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1014 Bytecodes::Code code = java_bc();
1015 if (code == Bytecodes::_wide) {
1016 code = method()->java_code_at_bci(bci() + 1);
1017 }
1018
1019 BasicType rtype = T_ILLEGAL;
1020 int rsize = 0;
1021
1022 if (code != Bytecodes::_illegal) {
1023 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1024 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1025 if (rtype < T_CONFLICT)
1026 rsize = type2size[rtype];
1027 }
1028
1029 switch (code) {
1030 case Bytecodes::_illegal:
1031 return false;
1032
1033 case Bytecodes::_ldc:
1034 case Bytecodes::_ldc_w:
1035 case Bytecodes::_ldc2_w:
1036 inputs = 0;
1037 break;
1038
1039 case Bytecodes::_dup: inputs = 1; break;
1040 case Bytecodes::_dup_x1: inputs = 2; break;
1041 case Bytecodes::_dup_x2: inputs = 3; break;
1042 case Bytecodes::_dup2: inputs = 2; break;
1043 case Bytecodes::_dup2_x1: inputs = 3; break;
1044 case Bytecodes::_dup2_x2: inputs = 4; break;
1045 case Bytecodes::_swap: inputs = 2; break;
1046 case Bytecodes::_arraylength: inputs = 1; break;
1047
1048 case Bytecodes::_getstatic:
1049 case Bytecodes::_putstatic:
1050 case Bytecodes::_getfield:
1051 case Bytecodes::_putfield:
1052 {
1053 bool ignored_will_link;
1054 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1055 int size = field->type()->size();
1056 bool is_get = (depth >= 0), is_static = (depth & 1);
1057 inputs = (is_static ? 0 : 1);
1058 if (is_get) {
1059 depth = size - inputs;
1060 } else {
1061 inputs += size; // putxxx pops the value from the stack
1062 depth = - inputs;
1063 }
1064 }
1065 break;
1066
1067 case Bytecodes::_invokevirtual:
1068 case Bytecodes::_invokespecial:
1069 case Bytecodes::_invokestatic:
1070 case Bytecodes::_invokedynamic:
1071 case Bytecodes::_invokeinterface:
1072 {
1073 bool ignored_will_link;
1074 ciSignature* declared_signature = NULL;
1075 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1076 assert(declared_signature != NULL, "cannot be null");
1077 inputs = declared_signature->arg_size_for_bc(code);
1078 int size = declared_signature->return_type()->size();
1079 depth = size - inputs;
1080 }
1081 break;
1082
1083 case Bytecodes::_multianewarray:
1084 {
1085 ciBytecodeStream iter(method());
1086 iter.reset_to_bci(bci());
1087 iter.next();
1088 inputs = iter.get_dimensions();
1089 assert(rsize == 1, "");
1090 depth = rsize - inputs;
1091 }
1092 break;
1093
1094 case Bytecodes::_ireturn:
1095 case Bytecodes::_lreturn:
1096 case Bytecodes::_freturn:
1097 case Bytecodes::_dreturn:
1098 case Bytecodes::_areturn:
1099 assert(rsize == -depth, "");
1100 inputs = rsize;
1101 break;
1102
1103 case Bytecodes::_jsr:
1104 case Bytecodes::_jsr_w:
1105 inputs = 0;
1106 depth = 1; // S.B. depth=1, not zero
1107 break;
1108
1109 default:
1110 // bytecode produces a typed result
1111 inputs = rsize - depth;
1112 assert(inputs >= 0, "");
1113 break;
1114 }
1115
1116 #ifdef ASSERT
1117 // spot check
1118 int outputs = depth + inputs;
1119 assert(outputs >= 0, "sanity");
1120 switch (code) {
1121 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1122 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1123 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1124 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1125 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1126 default: break;
1127 }
1128 #endif //ASSERT
1129
1130 return true;
1131 }
1132
1133
1134
1135 //------------------------------basic_plus_adr---------------------------------
1136 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1137 // short-circuit a common case
1138 if (offset == intcon(0)) return ptr;
1139 return _gvn.transform( new AddPNode(base, ptr, offset) );
1140 }
1141
1142 Node* GraphKit::ConvI2L(Node* offset) {
1143 // short-circuit a common case
1144 jint offset_con = find_int_con(offset, Type::OffsetBot);
1145 if (offset_con != Type::OffsetBot) {
1146 return longcon((jlong) offset_con);
1147 }
1148 return _gvn.transform( new ConvI2LNode(offset));
1149 }
1150
1151 Node* GraphKit::ConvI2UL(Node* offset) {
1152 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1153 if (offset_con != (juint) Type::OffsetBot) {
1154 return longcon((julong) offset_con);
1155 }
1156 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1157 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1158 return _gvn.transform( new AndLNode(conv, mask) );
1159 }
1160
1161 Node* GraphKit::ConvL2I(Node* offset) {
1162 // short-circuit a common case
1163 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1164 if (offset_con != (jlong)Type::OffsetBot) {
1165 return intcon((int) offset_con);
1166 }
1167 return _gvn.transform( new ConvL2INode(offset));
1168 }
1169
1170 //-------------------------load_object_klass-----------------------------------
1171 Node* GraphKit::load_object_klass(Node* obj) {
1172 // Special-case a fresh allocation to avoid building nodes:
1173 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1174 if (akls != NULL) return akls;
1175 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1176 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1177 }
1178
1179 //-------------------------load_array_length-----------------------------------
1180 Node* GraphKit::load_array_length(Node* array) {
1181 // Special-case a fresh allocation to avoid building nodes:
1182 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1183 Node *alen;
1184 if (alloc == NULL) {
1185 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1186 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1187 } else {
1188 alen = alloc->Ideal_length();
1189 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1190 if (ccast != alen) {
1191 alen = _gvn.transform(ccast);
1192 }
1193 }
1194 return alen;
1195 }
1196
1197 //------------------------------do_null_check----------------------------------
1198 // Helper function to do a NULL pointer check. Returned value is
1199 // the incoming address with NULL casted away. You are allowed to use the
1200 // not-null value only if you are control dependent on the test.
1201 #ifndef PRODUCT
1202 extern int explicit_null_checks_inserted,
1203 explicit_null_checks_elided;
1204 #endif
1205 Node* GraphKit::null_check_common(Node* value, BasicType type,
1206 // optional arguments for variations:
1207 bool assert_null,
1208 Node* *null_control,
1209 bool speculative) {
1210 assert(!assert_null || null_control == NULL, "not both at once");
1211 if (stopped()) return top();
1212 NOT_PRODUCT(explicit_null_checks_inserted++);
1213
1214 // Construct NULL check
1215 Node *chk = NULL;
1216 switch(type) {
1217 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1218 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1219 case T_ARRAY : // fall through
1220 type = T_OBJECT; // simplify further tests
1221 case T_OBJECT : {
1222 const Type *t = _gvn.type( value );
1223
1224 const TypeOopPtr* tp = t->isa_oopptr();
1225 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1226 // Only for do_null_check, not any of its siblings:
1227 && !assert_null && null_control == NULL) {
1228 // Usually, any field access or invocation on an unloaded oop type
1229 // will simply fail to link, since the statically linked class is
1230 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1231 // the static class is loaded but the sharper oop type is not.
1232 // Rather than checking for this obscure case in lots of places,
1233 // we simply observe that a null check on an unloaded class
1234 // will always be followed by a nonsense operation, so we
1235 // can just issue the uncommon trap here.
1236 // Our access to the unloaded class will only be correct
1237 // after it has been loaded and initialized, which requires
1238 // a trip through the interpreter.
1239 #ifndef PRODUCT
1240 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1241 #endif
1242 uncommon_trap(Deoptimization::Reason_unloaded,
1243 Deoptimization::Action_reinterpret,
1244 tp->klass(), "!loaded");
1245 return top();
1246 }
1247
1248 if (assert_null) {
1249 // See if the type is contained in NULL_PTR.
1250 // If so, then the value is already null.
1251 if (t->higher_equal(TypePtr::NULL_PTR)) {
1252 NOT_PRODUCT(explicit_null_checks_elided++);
1253 return value; // Elided null assert quickly!
1254 }
1255 } else {
1256 // See if mixing in the NULL pointer changes type.
1257 // If so, then the NULL pointer was not allowed in the original
1258 // type. In other words, "value" was not-null.
1259 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1260 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1261 NOT_PRODUCT(explicit_null_checks_elided++);
1262 return value; // Elided null check quickly!
1263 }
1264 }
1265 chk = new CmpPNode( value, null() );
1266 break;
1267 }
1268
1269 default:
1270 fatal("unexpected type: %s", type2name(type));
1271 }
1272 assert(chk != NULL, "sanity check");
1273 chk = _gvn.transform(chk);
1274
1275 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1276 BoolNode *btst = new BoolNode( chk, btest);
1277 Node *tst = _gvn.transform( btst );
1278
1279 //-----------
1280 // if peephole optimizations occurred, a prior test existed.
1281 // If a prior test existed, maybe it dominates as we can avoid this test.
1282 if (tst != btst && type == T_OBJECT) {
1283 // At this point we want to scan up the CFG to see if we can
1284 // find an identical test (and so avoid this test altogether).
1285 Node *cfg = control();
1286 int depth = 0;
1287 while( depth < 16 ) { // Limit search depth for speed
1288 if( cfg->Opcode() == Op_IfTrue &&
1289 cfg->in(0)->in(1) == tst ) {
1290 // Found prior test. Use "cast_not_null" to construct an identical
1291 // CastPP (and hence hash to) as already exists for the prior test.
1292 // Return that casted value.
1293 if (assert_null) {
1294 replace_in_map(value, null());
1295 return null(); // do not issue the redundant test
1296 }
1297 Node *oldcontrol = control();
1298 set_control(cfg);
1299 Node *res = cast_not_null(value);
1300 set_control(oldcontrol);
1301 NOT_PRODUCT(explicit_null_checks_elided++);
1302 return res;
1303 }
1304 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1305 if (cfg == NULL) break; // Quit at region nodes
1306 depth++;
1307 }
1308 }
1309
1310 //-----------
1311 // Branch to failure if null
1312 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1313 Deoptimization::DeoptReason reason;
1314 if (assert_null) {
1315 reason = Deoptimization::reason_null_assert(speculative);
1316 } else if (type == T_OBJECT) {
1317 reason = Deoptimization::reason_null_check(speculative);
1318 } else {
1319 reason = Deoptimization::Reason_div0_check;
1320 }
1321 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1322 // ciMethodData::has_trap_at will return a conservative -1 if any
1323 // must-be-null assertion has failed. This could cause performance
1324 // problems for a method after its first do_null_assert failure.
1325 // Consider using 'Reason_class_check' instead?
1326
1327 // To cause an implicit null check, we set the not-null probability
1328 // to the maximum (PROB_MAX). For an explicit check the probability
1329 // is set to a smaller value.
1330 if (null_control != NULL || too_many_traps(reason)) {
1331 // probability is less likely
1332 ok_prob = PROB_LIKELY_MAG(3);
1333 } else if (!assert_null &&
1334 (ImplicitNullCheckThreshold > 0) &&
1335 method() != NULL &&
1336 (method()->method_data()->trap_count(reason)
1337 >= (uint)ImplicitNullCheckThreshold)) {
1338 ok_prob = PROB_LIKELY_MAG(3);
1339 }
1340
1341 if (null_control != NULL) {
1342 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1343 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1344 set_control( _gvn.transform( new IfTrueNode(iff)));
1345 #ifndef PRODUCT
1346 if (null_true == top()) {
1347 explicit_null_checks_elided++;
1348 }
1349 #endif
1350 (*null_control) = null_true;
1351 } else {
1352 BuildCutout unless(this, tst, ok_prob);
1353 // Check for optimizer eliding test at parse time
1354 if (stopped()) {
1355 // Failure not possible; do not bother making uncommon trap.
1356 NOT_PRODUCT(explicit_null_checks_elided++);
1357 } else if (assert_null) {
1358 uncommon_trap(reason,
1359 Deoptimization::Action_make_not_entrant,
1360 NULL, "assert_null");
1361 } else {
1362 replace_in_map(value, zerocon(type));
1363 builtin_throw(reason);
1364 }
1365 }
1366
1367 // Must throw exception, fall-thru not possible?
1368 if (stopped()) {
1369 return top(); // No result
1370 }
1371
1372 if (assert_null) {
1373 // Cast obj to null on this path.
1374 replace_in_map(value, zerocon(type));
1375 return zerocon(type);
1376 }
1377
1378 // Cast obj to not-null on this path, if there is no null_control.
1379 // (If there is a null_control, a non-null value may come back to haunt us.)
1380 if (type == T_OBJECT) {
1381 Node* cast = cast_not_null(value, false);
1382 if (null_control == NULL || (*null_control) == top())
1383 replace_in_map(value, cast);
1384 value = cast;
1385 }
1386
1387 return value;
1388 }
1389
1390
1391 //------------------------------cast_not_null----------------------------------
1392 // Cast obj to not-null on this path
1393 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1394 const Type *t = _gvn.type(obj);
1395 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1396 // Object is already not-null?
1397 if( t == t_not_null ) return obj;
1398
1399 Node *cast = new CastPPNode(obj,t_not_null);
1400 cast->init_req(0, control());
1401 cast = _gvn.transform( cast );
1402
1403 // Scan for instances of 'obj' in the current JVM mapping.
1404 // These instances are known to be not-null after the test.
1405 if (do_replace_in_map)
1406 replace_in_map(obj, cast);
1407
1408 return cast; // Return casted value
1409 }
1410
1411 // Sometimes in intrinsics, we implicitly know an object is not null
1412 // (there's no actual null check) so we can cast it to not null. In
1413 // the course of optimizations, the input to the cast can become null.
1414 // In that case that data path will die and we need the control path
1415 // to become dead as well to keep the graph consistent. So we have to
1416 // add a check for null for which one branch can't be taken. It uses
1417 // an Opaque4 node that will cause the check to be removed after loop
1418 // opts so the test goes away and the compiled code doesn't execute a
1419 // useless check.
1420 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1421 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1422 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1423 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1424 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1425 _gvn.set_type(iff, iff->Value(&_gvn));
1426 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1427 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1428 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1429 C->root()->add_req(halt);
1430 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1431 set_control(if_t);
1432 return cast_not_null(value, do_replace_in_map);
1433 }
1434
1435
1436 //--------------------------replace_in_map-------------------------------------
1437 void GraphKit::replace_in_map(Node* old, Node* neww) {
1438 if (old == neww) {
1439 return;
1440 }
1441
1442 map()->replace_edge(old, neww);
1443
1444 // Note: This operation potentially replaces any edge
1445 // on the map. This includes locals, stack, and monitors
1446 // of the current (innermost) JVM state.
1447
1448 // don't let inconsistent types from profiling escape this
1449 // method
1450
1451 const Type* told = _gvn.type(old);
1452 const Type* tnew = _gvn.type(neww);
1453
1454 if (!tnew->higher_equal(told)) {
1455 return;
1456 }
1457
1458 map()->record_replaced_node(old, neww);
1459 }
1460
1461
1462 //=============================================================================
1463 //--------------------------------memory---------------------------------------
1464 Node* GraphKit::memory(uint alias_idx) {
1465 MergeMemNode* mem = merged_memory();
1466 Node* p = mem->memory_at(alias_idx);
1467 _gvn.set_type(p, Type::MEMORY); // must be mapped
1468 return p;
1469 }
1470
1471 //-----------------------------reset_memory------------------------------------
1472 Node* GraphKit::reset_memory() {
1473 Node* mem = map()->memory();
1474 // do not use this node for any more parsing!
1475 debug_only( map()->set_memory((Node*)NULL) );
1476 return _gvn.transform( mem );
1477 }
1478
1479 //------------------------------set_all_memory---------------------------------
1480 void GraphKit::set_all_memory(Node* newmem) {
1481 Node* mergemem = MergeMemNode::make(newmem);
1482 gvn().set_type_bottom(mergemem);
1483 map()->set_memory(mergemem);
1484 }
1485
1486 //------------------------------set_all_memory_call----------------------------
1487 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1488 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1489 set_all_memory(newmem);
1490 }
1491
1492 //=============================================================================
1493 //
1494 // parser factory methods for MemNodes
1495 //
1496 // These are layered on top of the factory methods in LoadNode and StoreNode,
1497 // and integrate with the parser's memory state and _gvn engine.
1498 //
1499
1500 // factory methods in "int adr_idx"
1501 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1502 int adr_idx,
1503 MemNode::MemOrd mo,
1504 LoadNode::ControlDependency control_dependency,
1505 bool require_atomic_access,
1506 bool unaligned,
1507 bool mismatched) {
1508 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1509 const TypePtr* adr_type = NULL; // debug-mode-only argument
1510 debug_only(adr_type = C->get_adr_type(adr_idx));
1511 Node* mem = memory(adr_idx);
1512 Node* ld;
1513 if (require_atomic_access && bt == T_LONG) {
1514 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1515 } else if (require_atomic_access && bt == T_DOUBLE) {
1516 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1517 } else {
1518 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1519 }
1520 ld = _gvn.transform(ld);
1521 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1522 // Improve graph before escape analysis and boxing elimination.
1523 record_for_igvn(ld);
1524 }
1525 return ld;
1526 }
1527
1528 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1529 int adr_idx,
1530 MemNode::MemOrd mo,
1531 bool require_atomic_access,
1532 bool unaligned,
1533 bool mismatched) {
1534 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1535 const TypePtr* adr_type = NULL;
1536 debug_only(adr_type = C->get_adr_type(adr_idx));
1537 Node *mem = memory(adr_idx);
1538 Node* st;
1539 if (require_atomic_access && bt == T_LONG) {
1540 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1541 } else if (require_atomic_access && bt == T_DOUBLE) {
1542 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1543 } else {
1544 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1545 }
1546 if (unaligned) {
1547 st->as_Store()->set_unaligned_access();
1548 }
1549 if (mismatched) {
1550 st->as_Store()->set_mismatched_access();
1551 }
1552 st = _gvn.transform(st);
1553 set_memory(st, adr_idx);
1554 // Back-to-back stores can only remove intermediate store with DU info
1555 // so push on worklist for optimizer.
1556 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1557 record_for_igvn(st);
1558
1559 return st;
1560 }
1561
1562
1563 void GraphKit::pre_barrier(bool do_load,
1564 Node* ctl,
1565 Node* obj,
1566 Node* adr,
1567 uint adr_idx,
1568 Node* val,
1569 const TypeOopPtr* val_type,
1570 Node* pre_val,
1571 BasicType bt) {
1572
1573 BarrierSet* bs = BarrierSet::barrier_set();
1574 set_control(ctl);
1575 switch (bs->kind()) {
1576 case BarrierSet::G1BarrierSet:
1577 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1578 break;
1579 case BarrierSet::Shenandoah:
1580 shenandoah_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1581 break;
1582
1583 case BarrierSet::CardTableBarrierSet:
1584 break;
1585
1586 default :
1587 ShouldNotReachHere();
1588
1589 }
1590 }
1591
1592 bool GraphKit::can_move_pre_barrier() const {
1593 BarrierSet* bs = BarrierSet::barrier_set();
1594 switch (bs->kind()) {
1595 case BarrierSet::G1BarrierSet:
1596 case BarrierSet::Shenandoah:
1597 return true; // Can move it if no safepoint
1598
1599 case BarrierSet::CardTableBarrierSet:
1600 return true; // There is no pre-barrier
1601
1602 default :
1603 ShouldNotReachHere();
1604 }
1605 return false;
1606 }
1607
1608 void GraphKit::post_barrier(Node* ctl,
1609 Node* store,
1610 Node* obj,
1611 Node* adr,
1612 uint adr_idx,
1613 Node* val,
1614 BasicType bt,
1615 bool use_precise) {
1616 BarrierSet* bs = BarrierSet::barrier_set();
1617 set_control(ctl);
1618 switch (bs->kind()) {
1619 case BarrierSet::G1BarrierSet:
1620 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1621 break;
1622
1623 case BarrierSet::CardTableBarrierSet:
1624 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1625 break;
1626
1627 case BarrierSet::Shenandoah:
1628 break;
1629
1630 default :
1631 ShouldNotReachHere();
1632
1633 }
1634 }
1635
1636 void GraphKit::keep_alive_barrier(Node* ctl, Node* obj) {
1637 BarrierSet* bs = BarrierSet::barrier_set();
1638 switch (bs->kind()) {
1639 case BarrierSet::G1BarrierSet:
1640 pre_barrier(false /* do_load */,
1641 ctl,
1642 NULL /* obj */,
1643 NULL /* adr */,
1644 max_juint /* alias_idx */,
1645 NULL /* val */,
1646 NULL /* val_type */,
1647 obj /* pre_val */,
1648 T_OBJECT);
1649 break;
1650 case BarrierSet::Shenandoah:
1651 if (ShenandoahKeepAliveBarrier) {
1652 pre_barrier(false /* do_load */,
1653 ctl,
1654 NULL /* obj */,
1655 NULL /* adr */,
1656 max_juint /* alias_idx */,
1657 NULL /* val */,
1658 NULL /* val_type */,
1659 obj /* pre_val */,
1660 T_OBJECT);
1661 }
1662 break;
1663 case BarrierSet::CardTableBarrierSet:
1664 break;
1665 default :
1666 ShouldNotReachHere();
1667
1668 }
1669 }
1670
1671 Node* GraphKit::store_oop(Node* ctl,
1672 Node* obj,
1673 Node* adr,
1674 const TypePtr* adr_type,
1675 Node* val,
1676 const TypeOopPtr* val_type,
1677 BasicType bt,
1678 bool use_precise,
1679 MemNode::MemOrd mo,
1680 bool mismatched) {
1681 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1682 // could be delayed during Parse (for example, in adjust_map_after_if()).
1683 // Execute transformation here to avoid barrier generation in such case.
1684 if (_gvn.type(val) == TypePtr::NULL_PTR)
1685 val = _gvn.makecon(TypePtr::NULL_PTR);
1686
1687 set_control(ctl);
1688 if (stopped()) return top(); // Dead path ?
1689
1690 assert(bt == T_OBJECT, "sanity");
1691 assert(val != NULL, "not dead path");
1692 uint adr_idx = C->get_alias_index(adr_type);
1693 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1694
1695 val = shenandoah_storeval_barrier(val);
1696
1697 pre_barrier(true /* do_load */,
1698 control(), obj, adr, adr_idx, val, val_type,
1699 NULL /* pre_val */,
1700 bt);
1701
1702 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1703 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1704 return store;
1705 }
1706
1707 // Could be an array or object we don't know at compile time (unsafe ref.)
1708 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1709 Node* obj, // containing obj
1710 Node* adr, // actual adress to store val at
1711 const TypePtr* adr_type,
1712 Node* val,
1713 BasicType bt,
1714 MemNode::MemOrd mo,
1715 bool mismatched) {
1716 Compile::AliasType* at = C->alias_type(adr_type);
1717 const TypeOopPtr* val_type = NULL;
1718 if (adr_type->isa_instptr()) {
1719 if (at->field() != NULL) {
1720 // known field. This code is a copy of the do_put_xxx logic.
1721 ciField* field = at->field();
1722 if (!field->type()->is_loaded()) {
1723 val_type = TypeInstPtr::BOTTOM;
1724 } else {
1725 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1726 }
1727 }
1728 } else if (adr_type->isa_aryptr()) {
1729 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1730 }
1731 if (val_type == NULL) {
1732 val_type = TypeInstPtr::BOTTOM;
1733 }
1734 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1735 }
1736
1737
1738 //-------------------------array_element_address-------------------------
1739 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1740 const TypeInt* sizetype, Node* ctrl) {
1741 uint shift = exact_log2(type2aelembytes(elembt));
1742 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1743
1744 // short-circuit a common case (saves lots of confusing waste motion)
1745 jint idx_con = find_int_con(idx, -1);
1746 if (idx_con >= 0) {
1747 intptr_t offset = header + ((intptr_t)idx_con << shift);
1748 return basic_plus_adr(ary, offset);
1749 }
1750
1751 // must be correct type for alignment purposes
1752 Node* base = basic_plus_adr(ary, header);
1753 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1754 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1755 return basic_plus_adr(ary, base, scale);
1756 }
1757
1758 //-------------------------load_array_element-------------------------
1759 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1760 const Type* elemtype = arytype->elem();
1761 BasicType elembt = elemtype->array_element_basic_type();
1762 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1763 if (elembt == T_NARROWOOP) {
1764 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1765 }
1766 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1767 return ld;
1768 }
1769
1770 //-------------------------set_arguments_for_java_call-------------------------
1771 // Arguments (pre-popped from the stack) are taken from the JVMS.
1772 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1773 // Add the call arguments:
1774 uint nargs = call->method()->arg_size();
1775 for (uint i = 0; i < nargs; i++) {
1776 Node* arg = argument(i);
1777 call->init_req(i + TypeFunc::Parms, arg);
1778 }
1779 }
1780
1781 //---------------------------set_edges_for_java_call---------------------------
1782 // Connect a newly created call into the current JVMS.
1783 // A return value node (if any) is returned from set_edges_for_java_call.
1784 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1785
1786 // Add the predefined inputs:
1787 call->init_req( TypeFunc::Control, control() );
1788 call->init_req( TypeFunc::I_O , i_o() );
1789 call->init_req( TypeFunc::Memory , reset_memory() );
1790 call->init_req( TypeFunc::FramePtr, frameptr() );
1791 call->init_req( TypeFunc::ReturnAdr, top() );
1792
1793 add_safepoint_edges(call, must_throw);
1794
1795 Node* xcall = _gvn.transform(call);
1796
1797 if (xcall == top()) {
1798 set_control(top());
1799 return;
1800 }
1801 assert(xcall == call, "call identity is stable");
1802
1803 // Re-use the current map to produce the result.
1804
1805 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1806 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1807 set_all_memory_call(xcall, separate_io_proj);
1808
1809 //return xcall; // no need, caller already has it
1810 }
1811
1812 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1813 if (stopped()) return top(); // maybe the call folded up?
1814
1815 // Capture the return value, if any.
1816 Node* ret;
1817 if (call->method() == NULL ||
1818 call->method()->return_type()->basic_type() == T_VOID)
1819 ret = top();
1820 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1821
1822 // Note: Since any out-of-line call can produce an exception,
1823 // we always insert an I_O projection from the call into the result.
1824
1825 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1826
1827 if (separate_io_proj) {
1828 // The caller requested separate projections be used by the fall
1829 // through and exceptional paths, so replace the projections for
1830 // the fall through path.
1831 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1832 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1833 }
1834 return ret;
1835 }
1836
1837 //--------------------set_predefined_input_for_runtime_call--------------------
1838 // Reading and setting the memory state is way conservative here.
1839 // The real problem is that I am not doing real Type analysis on memory,
1840 // so I cannot distinguish card mark stores from other stores. Across a GC
1841 // point the Store Barrier and the card mark memory has to agree. I cannot
1842 // have a card mark store and its barrier split across the GC point from
1843 // either above or below. Here I get that to happen by reading ALL of memory.
1844 // A better answer would be to separate out card marks from other memory.
1845 // For now, return the input memory state, so that it can be reused
1846 // after the call, if this call has restricted memory effects.
1847 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1848 // Set fixed predefined input arguments
1849 Node* memory = reset_memory();
1850 call->init_req( TypeFunc::Control, control() );
1851 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1852 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1853 call->init_req( TypeFunc::FramePtr, frameptr() );
1854 call->init_req( TypeFunc::ReturnAdr, top() );
1855 return memory;
1856 }
1857
1858 //-------------------set_predefined_output_for_runtime_call--------------------
1859 // Set control and memory (not i_o) from the call.
1860 // If keep_mem is not NULL, use it for the output state,
1861 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1862 // If hook_mem is NULL, this call produces no memory effects at all.
1863 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1864 // then only that memory slice is taken from the call.
1865 // In the last case, we must put an appropriate memory barrier before
1866 // the call, so as to create the correct anti-dependencies on loads
1867 // preceding the call.
1868 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1869 Node* keep_mem,
1870 const TypePtr* hook_mem) {
1871 // no i/o
1872 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1873 if (keep_mem) {
1874 // First clone the existing memory state
1875 set_all_memory(keep_mem);
1876 if (hook_mem != NULL) {
1877 // Make memory for the call
1878 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1879 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1880 // We also use hook_mem to extract specific effects from arraycopy stubs.
1881 set_memory(mem, hook_mem);
1882 }
1883 // ...else the call has NO memory effects.
1884
1885 // Make sure the call advertises its memory effects precisely.
1886 // This lets us build accurate anti-dependences in gcm.cpp.
1887 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1888 "call node must be constructed correctly");
1889 } else {
1890 assert(hook_mem == NULL, "");
1891 // This is not a "slow path" call; all memory comes from the call.
1892 set_all_memory_call(call);
1893 }
1894 }
1895
1896
1897 // Replace the call with the current state of the kit.
1898 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1899 JVMState* ejvms = NULL;
1900 if (has_exceptions()) {
1901 ejvms = transfer_exceptions_into_jvms();
1902 }
1903
1904 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1905 ReplacedNodes replaced_nodes_exception;
1906 Node* ex_ctl = top();
1907
1908 SafePointNode* final_state = stop();
1909
1910 // Find all the needed outputs of this call
1911 CallProjections callprojs;
1912 call->extract_projections(&callprojs, true);
1913
1914 Node* init_mem = call->in(TypeFunc::Memory);
1915 Node* final_mem = final_state->in(TypeFunc::Memory);
1916 Node* final_ctl = final_state->in(TypeFunc::Control);
1917 Node* final_io = final_state->in(TypeFunc::I_O);
1918
1919 // Replace all the old call edges with the edges from the inlining result
1920 if (callprojs.fallthrough_catchproj != NULL) {
1921 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1922 }
1923 if (callprojs.fallthrough_memproj != NULL) {
1924 if (final_mem->is_MergeMem()) {
1925 // Parser's exits MergeMem was not transformed but may be optimized
1926 final_mem = _gvn.transform(final_mem);
1927 }
1928 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1929 }
1930 if (callprojs.fallthrough_ioproj != NULL) {
1931 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1932 }
1933
1934 // Replace the result with the new result if it exists and is used
1935 if (callprojs.resproj != NULL && result != NULL) {
1936 C->gvn_replace_by(callprojs.resproj, result);
1937 }
1938
1939 if (ejvms == NULL) {
1940 // No exception edges to simply kill off those paths
1941 if (callprojs.catchall_catchproj != NULL) {
1942 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1943 }
1944 if (callprojs.catchall_memproj != NULL) {
1945 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1946 }
1947 if (callprojs.catchall_ioproj != NULL) {
1948 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1949 }
1950 // Replace the old exception object with top
1951 if (callprojs.exobj != NULL) {
1952 C->gvn_replace_by(callprojs.exobj, C->top());
1953 }
1954 } else {
1955 GraphKit ekit(ejvms);
1956
1957 // Load my combined exception state into the kit, with all phis transformed:
1958 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1959 replaced_nodes_exception = ex_map->replaced_nodes();
1960
1961 Node* ex_oop = ekit.use_exception_state(ex_map);
1962
1963 if (callprojs.catchall_catchproj != NULL) {
1964 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1965 ex_ctl = ekit.control();
1966 }
1967 if (callprojs.catchall_memproj != NULL) {
1968 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1969 }
1970 if (callprojs.catchall_ioproj != NULL) {
1971 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1972 }
1973
1974 // Replace the old exception object with the newly created one
1975 if (callprojs.exobj != NULL) {
1976 C->gvn_replace_by(callprojs.exobj, ex_oop);
1977 }
1978 }
1979
1980 // Disconnect the call from the graph
1981 call->disconnect_inputs(NULL, C);
1982 C->gvn_replace_by(call, C->top());
1983
1984 // Clean up any MergeMems that feed other MergeMems since the
1985 // optimizer doesn't like that.
1986 if (final_mem->is_MergeMem()) {
1987 Node_List wl;
1988 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1989 Node* m = i.get();
1990 if (m->is_MergeMem() && !wl.contains(m)) {
1991 wl.push(m);
1992 }
1993 }
1994 while (wl.size() > 0) {
1995 _gvn.transform(wl.pop());
1996 }
1997 }
1998
1999 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2000 replaced_nodes.apply(C, final_ctl);
2001 }
2002 if (!ex_ctl->is_top() && do_replaced_nodes) {
2003 replaced_nodes_exception.apply(C, ex_ctl);
2004 }
2005 }
2006
2007
2008 //------------------------------increment_counter------------------------------
2009 // for statistics: increment a VM counter by 1
2010
2011 void GraphKit::increment_counter(address counter_addr) {
2012 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2013 increment_counter(adr1);
2014 }
2015
2016 void GraphKit::increment_counter(Node* counter_addr) {
2017 int adr_type = Compile::AliasIdxRaw;
2018 Node* ctrl = control();
2019 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2020 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2021 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2022 }
2023
2024
2025 //------------------------------uncommon_trap----------------------------------
2026 // Bail out to the interpreter in mid-method. Implemented by calling the
2027 // uncommon_trap blob. This helper function inserts a runtime call with the
2028 // right debug info.
2029 void GraphKit::uncommon_trap(int trap_request,
2030 ciKlass* klass, const char* comment,
2031 bool must_throw,
2032 bool keep_exact_action) {
2033 if (failing()) stop();
2034 if (stopped()) return; // trap reachable?
2035
2036 // Note: If ProfileTraps is true, and if a deopt. actually
2037 // occurs here, the runtime will make sure an MDO exists. There is
2038 // no need to call method()->ensure_method_data() at this point.
2039
2040 // Set the stack pointer to the right value for reexecution:
2041 set_sp(reexecute_sp());
2042
2043 #ifdef ASSERT
2044 if (!must_throw) {
2045 // Make sure the stack has at least enough depth to execute
2046 // the current bytecode.
2047 int inputs, ignored_depth;
2048 if (compute_stack_effects(inputs, ignored_depth)) {
2049 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2050 Bytecodes::name(java_bc()), sp(), inputs);
2051 }
2052 }
2053 #endif
2054
2055 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2056 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2057
2058 switch (action) {
2059 case Deoptimization::Action_maybe_recompile:
2060 case Deoptimization::Action_reinterpret:
2061 // Temporary fix for 6529811 to allow virtual calls to be sure they
2062 // get the chance to go from mono->bi->mega
2063 if (!keep_exact_action &&
2064 Deoptimization::trap_request_index(trap_request) < 0 &&
2065 too_many_recompiles(reason)) {
2066 // This BCI is causing too many recompilations.
2067 if (C->log() != NULL) {
2068 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2069 Deoptimization::trap_reason_name(reason),
2070 Deoptimization::trap_action_name(action));
2071 }
2072 action = Deoptimization::Action_none;
2073 trap_request = Deoptimization::make_trap_request(reason, action);
2074 } else {
2075 C->set_trap_can_recompile(true);
2076 }
2077 break;
2078 case Deoptimization::Action_make_not_entrant:
2079 C->set_trap_can_recompile(true);
2080 break;
2081 case Deoptimization::Action_none:
2082 case Deoptimization::Action_make_not_compilable:
2083 break;
2084 default:
2085 #ifdef ASSERT
2086 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2087 #endif
2088 break;
2089 }
2090
2091 if (TraceOptoParse) {
2092 char buf[100];
2093 tty->print_cr("Uncommon trap %s at bci:%d",
2094 Deoptimization::format_trap_request(buf, sizeof(buf),
2095 trap_request), bci());
2096 }
2097
2098 CompileLog* log = C->log();
2099 if (log != NULL) {
2100 int kid = (klass == NULL)? -1: log->identify(klass);
2101 log->begin_elem("uncommon_trap bci='%d'", bci());
2102 char buf[100];
2103 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2104 trap_request));
2105 if (kid >= 0) log->print(" klass='%d'", kid);
2106 if (comment != NULL) log->print(" comment='%s'", comment);
2107 log->end_elem();
2108 }
2109
2110 // Make sure any guarding test views this path as very unlikely
2111 Node *i0 = control()->in(0);
2112 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2113 IfNode *iff = i0->as_If();
2114 float f = iff->_prob; // Get prob
2115 if (control()->Opcode() == Op_IfTrue) {
2116 if (f > PROB_UNLIKELY_MAG(4))
2117 iff->_prob = PROB_MIN;
2118 } else {
2119 if (f < PROB_LIKELY_MAG(4))
2120 iff->_prob = PROB_MAX;
2121 }
2122 }
2123
2124 // Clear out dead values from the debug info.
2125 kill_dead_locals();
2126
2127 // Now insert the uncommon trap subroutine call
2128 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2129 const TypePtr* no_memory_effects = NULL;
2130 // Pass the index of the class to be loaded
2131 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2132 (must_throw ? RC_MUST_THROW : 0),
2133 OptoRuntime::uncommon_trap_Type(),
2134 call_addr, "uncommon_trap", no_memory_effects,
2135 intcon(trap_request));
2136 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2137 "must extract request correctly from the graph");
2138 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2139
2140 call->set_req(TypeFunc::ReturnAdr, returnadr());
2141 // The debug info is the only real input to this call.
2142
2143 // Halt-and-catch fire here. The above call should never return!
2144 HaltNode* halt = new HaltNode(control(), frameptr());
2145 _gvn.set_type_bottom(halt);
2146 root()->add_req(halt);
2147
2148 stop_and_kill_map();
2149 }
2150
2151
2152 //--------------------------just_allocated_object------------------------------
2153 // Report the object that was just allocated.
2154 // It must be the case that there are no intervening safepoints.
2155 // We use this to determine if an object is so "fresh" that
2156 // it does not require card marks.
2157 Node* GraphKit::just_allocated_object(Node* current_control) {
2158 if (C->recent_alloc_ctl() == current_control)
2159 return C->recent_alloc_obj();
2160 return NULL;
2161 }
2162
2163
2164 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2165 // (Note: TypeFunc::make has a cache that makes this fast.)
2166 const TypeFunc* tf = TypeFunc::make(dest_method);
2167 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2168 for (int j = 0; j < nargs; j++) {
2169 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2170 if( targ->basic_type() == T_DOUBLE ) {
2171 // If any parameters are doubles, they must be rounded before
2172 // the call, dstore_rounding does gvn.transform
2173 Node *arg = argument(j);
2174 arg = dstore_rounding(arg);
2175 set_argument(j, arg);
2176 }
2177 }
2178 }
2179
2180 /**
2181 * Record profiling data exact_kls for Node n with the type system so
2182 * that it can propagate it (speculation)
2183 *
2184 * @param n node that the type applies to
2185 * @param exact_kls type from profiling
2186 * @param maybe_null did profiling see null?
2187 *
2188 * @return node with improved type
2189 */
2190 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2191 const Type* current_type = _gvn.type(n);
2192 assert(UseTypeSpeculation, "type speculation must be on");
2193
2194 const TypePtr* speculative = current_type->speculative();
2195
2196 // Should the klass from the profile be recorded in the speculative type?
2197 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2198 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2199 const TypeOopPtr* xtype = tklass->as_instance_type();
2200 assert(xtype->klass_is_exact(), "Should be exact");
2201 // Any reason to believe n is not null (from this profiling or a previous one)?
2202 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2203 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2204 // record the new speculative type's depth
2205 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2206 speculative = speculative->with_inline_depth(jvms()->depth());
2207 } else if (current_type->would_improve_ptr(ptr_kind)) {
2208 // Profiling report that null was never seen so we can change the
2209 // speculative type to non null ptr.
2210 if (ptr_kind == ProfileAlwaysNull) {
2211 speculative = TypePtr::NULL_PTR;
2212 } else {
2213 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2214 const TypePtr* ptr = TypePtr::NOTNULL;
2215 if (speculative != NULL) {
2216 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2217 } else {
2218 speculative = ptr;
2219 }
2220 }
2221 }
2222
2223 if (speculative != current_type->speculative()) {
2224 // Build a type with a speculative type (what we think we know
2225 // about the type but will need a guard when we use it)
2226 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2227 // We're changing the type, we need a new CheckCast node to carry
2228 // the new type. The new type depends on the control: what
2229 // profiling tells us is only valid from here as far as we can
2230 // tell.
2231 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2232 cast = _gvn.transform(cast);
2233 replace_in_map(n, cast);
2234 n = cast;
2235 }
2236
2237 return n;
2238 }
2239
2240 /**
2241 * Record profiling data from receiver profiling at an invoke with the
2242 * type system so that it can propagate it (speculation)
2243 *
2244 * @param n receiver node
2245 *
2246 * @return node with improved type
2247 */
2248 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2249 if (!UseTypeSpeculation) {
2250 return n;
2251 }
2252 ciKlass* exact_kls = profile_has_unique_klass();
2253 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2254 if ((java_bc() == Bytecodes::_checkcast ||
2255 java_bc() == Bytecodes::_instanceof ||
2256 java_bc() == Bytecodes::_aastore) &&
2257 method()->method_data()->is_mature()) {
2258 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2259 if (data != NULL) {
2260 if (!data->as_BitData()->null_seen()) {
2261 ptr_kind = ProfileNeverNull;
2262 } else {
2263 assert(data->is_ReceiverTypeData(), "bad profile data type");
2264 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2265 uint i = 0;
2266 for (; i < call->row_limit(); i++) {
2267 ciKlass* receiver = call->receiver(i);
2268 if (receiver != NULL) {
2269 break;
2270 }
2271 }
2272 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2273 }
2274 }
2275 }
2276 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2277 }
2278
2279 /**
2280 * Record profiling data from argument profiling at an invoke with the
2281 * type system so that it can propagate it (speculation)
2282 *
2283 * @param dest_method target method for the call
2284 * @param bc what invoke bytecode is this?
2285 */
2286 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2287 if (!UseTypeSpeculation) {
2288 return;
2289 }
2290 const TypeFunc* tf = TypeFunc::make(dest_method);
2291 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2292 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2293 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2294 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2295 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2296 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2297 ciKlass* better_type = NULL;
2298 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2299 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2300 }
2301 i++;
2302 }
2303 }
2304 }
2305
2306 /**
2307 * Record profiling data from parameter profiling at an invoke with
2308 * the type system so that it can propagate it (speculation)
2309 */
2310 void GraphKit::record_profiled_parameters_for_speculation() {
2311 if (!UseTypeSpeculation) {
2312 return;
2313 }
2314 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2315 if (_gvn.type(local(i))->isa_oopptr()) {
2316 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2317 ciKlass* better_type = NULL;
2318 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2319 record_profile_for_speculation(local(i), better_type, ptr_kind);
2320 }
2321 j++;
2322 }
2323 }
2324 }
2325
2326 /**
2327 * Record profiling data from return value profiling at an invoke with
2328 * the type system so that it can propagate it (speculation)
2329 */
2330 void GraphKit::record_profiled_return_for_speculation() {
2331 if (!UseTypeSpeculation) {
2332 return;
2333 }
2334 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2335 ciKlass* better_type = NULL;
2336 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2337 // If profiling reports a single type for the return value,
2338 // feed it to the type system so it can propagate it as a
2339 // speculative type
2340 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2341 }
2342 }
2343
2344 void GraphKit::round_double_result(ciMethod* dest_method) {
2345 // A non-strict method may return a double value which has an extended
2346 // exponent, but this must not be visible in a caller which is 'strict'
2347 // If a strict caller invokes a non-strict callee, round a double result
2348
2349 BasicType result_type = dest_method->return_type()->basic_type();
2350 assert( method() != NULL, "must have caller context");
2351 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2352 // Destination method's return value is on top of stack
2353 // dstore_rounding() does gvn.transform
2354 Node *result = pop_pair();
2355 result = dstore_rounding(result);
2356 push_pair(result);
2357 }
2358 }
2359
2360 // rounding for strict float precision conformance
2361 Node* GraphKit::precision_rounding(Node* n) {
2362 return UseStrictFP && _method->flags().is_strict()
2363 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2364 ? _gvn.transform( new RoundFloatNode(0, n) )
2365 : n;
2366 }
2367
2368 // rounding for strict double precision conformance
2369 Node* GraphKit::dprecision_rounding(Node *n) {
2370 return UseStrictFP && _method->flags().is_strict()
2371 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2372 ? _gvn.transform( new RoundDoubleNode(0, n) )
2373 : n;
2374 }
2375
2376 // rounding for non-strict double stores
2377 Node* GraphKit::dstore_rounding(Node* n) {
2378 return Matcher::strict_fp_requires_explicit_rounding
2379 && UseSSE <= 1
2380 ? _gvn.transform( new RoundDoubleNode(0, n) )
2381 : n;
2382 }
2383
2384 //=============================================================================
2385 // Generate a fast path/slow path idiom. Graph looks like:
2386 // [foo] indicates that 'foo' is a parameter
2387 //
2388 // [in] NULL
2389 // \ /
2390 // CmpP
2391 // Bool ne
2392 // If
2393 // / \
2394 // True False-<2>
2395 // / |
2396 // / cast_not_null
2397 // Load | | ^
2398 // [fast_test] | |
2399 // gvn to opt_test | |
2400 // / \ | <1>
2401 // True False |
2402 // | \\ |
2403 // [slow_call] \[fast_result]
2404 // Ctl Val \ \
2405 // | \ \
2406 // Catch <1> \ \
2407 // / \ ^ \ \
2408 // Ex No_Ex | \ \
2409 // | \ \ | \ <2> \
2410 // ... \ [slow_res] | | \ [null_result]
2411 // \ \--+--+--- | |
2412 // \ | / \ | /
2413 // --------Region Phi
2414 //
2415 //=============================================================================
2416 // Code is structured as a series of driver functions all called 'do_XXX' that
2417 // call a set of helper functions. Helper functions first, then drivers.
2418
2419 //------------------------------null_check_oop---------------------------------
2420 // Null check oop. Set null-path control into Region in slot 3.
2421 // Make a cast-not-nullness use the other not-null control. Return cast.
2422 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2423 bool never_see_null,
2424 bool safe_for_replace,
2425 bool speculative) {
2426 // Initial NULL check taken path
2427 (*null_control) = top();
2428 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2429
2430 // Generate uncommon_trap:
2431 if (never_see_null && (*null_control) != top()) {
2432 // If we see an unexpected null at a check-cast we record it and force a
2433 // recompile; the offending check-cast will be compiled to handle NULLs.
2434 // If we see more than one offending BCI, then all checkcasts in the
2435 // method will be compiled to handle NULLs.
2436 PreserveJVMState pjvms(this);
2437 set_control(*null_control);
2438 replace_in_map(value, null());
2439 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2440 uncommon_trap(reason,
2441 Deoptimization::Action_make_not_entrant);
2442 (*null_control) = top(); // NULL path is dead
2443 }
2444 if ((*null_control) == top() && safe_for_replace) {
2445 replace_in_map(value, cast);
2446 }
2447
2448 // Cast away null-ness on the result
2449 return cast;
2450 }
2451
2452 //------------------------------opt_iff----------------------------------------
2453 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2454 // Return slow-path control.
2455 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2456 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2457
2458 // Fast path taken; set region slot 2
2459 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2460 region->init_req(2,fast_taken); // Capture fast-control
2461
2462 // Fast path not-taken, i.e. slow path
2463 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2464 return slow_taken;
2465 }
2466
2467 //-----------------------------make_runtime_call-------------------------------
2468 Node* GraphKit::make_runtime_call(int flags,
2469 const TypeFunc* call_type, address call_addr,
2470 const char* call_name,
2471 const TypePtr* adr_type,
2472 // The following parms are all optional.
2473 // The first NULL ends the list.
2474 Node* parm0, Node* parm1,
2475 Node* parm2, Node* parm3,
2476 Node* parm4, Node* parm5,
2477 Node* parm6, Node* parm7) {
2478 // Slow-path call
2479 bool is_leaf = !(flags & RC_NO_LEAF);
2480 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2481 if (call_name == NULL) {
2482 assert(!is_leaf, "must supply name for leaf");
2483 call_name = OptoRuntime::stub_name(call_addr);
2484 }
2485 CallNode* call;
2486 if (!is_leaf) {
2487 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2488 bci(), adr_type);
2489 } else if (flags & RC_NO_FP) {
2490 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2491 } else {
2492 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2493 }
2494
2495 // The following is similar to set_edges_for_java_call,
2496 // except that the memory effects of the call are restricted to AliasIdxRaw.
2497
2498 // Slow path call has no side-effects, uses few values
2499 bool wide_in = !(flags & RC_NARROW_MEM);
2500 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2501
2502 Node* prev_mem = NULL;
2503 if (wide_in) {
2504 prev_mem = set_predefined_input_for_runtime_call(call);
2505 } else {
2506 assert(!wide_out, "narrow in => narrow out");
2507 Node* narrow_mem = memory(adr_type);
2508 prev_mem = reset_memory();
2509 map()->set_memory(narrow_mem);
2510 set_predefined_input_for_runtime_call(call);
2511 }
2512
2513 // Hook each parm in order. Stop looking at the first NULL.
2514 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2515 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2516 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2517 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2518 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2519 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2520 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2521 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2522 /* close each nested if ===> */ } } } } } } } }
2523 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2524
2525 if (!is_leaf) {
2526 // Non-leaves can block and take safepoints:
2527 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2528 }
2529 // Non-leaves can throw exceptions:
2530 if (has_io) {
2531 call->set_req(TypeFunc::I_O, i_o());
2532 }
2533
2534 if (flags & RC_UNCOMMON) {
2535 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2536 // (An "if" probability corresponds roughly to an unconditional count.
2537 // Sort of.)
2538 call->set_cnt(PROB_UNLIKELY_MAG(4));
2539 }
2540
2541 Node* c = _gvn.transform(call);
2542 assert(c == call, "cannot disappear");
2543
2544 if (wide_out) {
2545 // Slow path call has full side-effects.
2546 set_predefined_output_for_runtime_call(call);
2547 } else {
2548 // Slow path call has few side-effects, and/or sets few values.
2549 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2550 }
2551
2552 if (has_io) {
2553 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2554 }
2555 return call;
2556
2557 }
2558
2559 //------------------------------merge_memory-----------------------------------
2560 // Merge memory from one path into the current memory state.
2561 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2562 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2563 Node* old_slice = mms.force_memory();
2564 Node* new_slice = mms.memory2();
2565 if (old_slice != new_slice) {
2566 PhiNode* phi;
2567 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2568 if (mms.is_empty()) {
2569 // clone base memory Phi's inputs for this memory slice
2570 assert(old_slice == mms.base_memory(), "sanity");
2571 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2572 _gvn.set_type(phi, Type::MEMORY);
2573 for (uint i = 1; i < phi->req(); i++) {
2574 phi->init_req(i, old_slice->in(i));
2575 }
2576 } else {
2577 phi = old_slice->as_Phi(); // Phi was generated already
2578 }
2579 } else {
2580 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2581 _gvn.set_type(phi, Type::MEMORY);
2582 }
2583 phi->set_req(new_path, new_slice);
2584 mms.set_memory(phi);
2585 }
2586 }
2587 }
2588
2589 //------------------------------make_slow_call_ex------------------------------
2590 // Make the exception handler hookups for the slow call
2591 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2592 if (stopped()) return;
2593
2594 // Make a catch node with just two handlers: fall-through and catch-all
2595 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2596 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2597 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2598 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2599
2600 { PreserveJVMState pjvms(this);
2601 set_control(excp);
2602 set_i_o(i_o);
2603
2604 if (excp != top()) {
2605 if (deoptimize) {
2606 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2607 uncommon_trap(Deoptimization::Reason_unhandled,
2608 Deoptimization::Action_none);
2609 } else {
2610 // Create an exception state also.
2611 // Use an exact type if the caller has specified a specific exception.
2612 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2613 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2614 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2615 }
2616 }
2617 }
2618
2619 // Get the no-exception control from the CatchNode.
2620 set_control(norm);
2621 }
2622
2623 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2624 Node* cmp = NULL;
2625 switch(bt) {
2626 case T_INT: cmp = new CmpINode(in1, in2); break;
2627 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2628 default: fatal("unexpected comparison type %s", type2name(bt));
2629 }
2630 gvn->transform(cmp);
2631 Node* bol = gvn->transform(new BoolNode(cmp, test));
2632 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2633 gvn->transform(iff);
2634 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2635 return iff;
2636 }
2637
2638
2639 //-------------------------------gen_subtype_check-----------------------------
2640 // Generate a subtyping check. Takes as input the subtype and supertype.
2641 // Returns 2 values: sets the default control() to the true path and returns
2642 // the false path. Only reads invariant memory; sets no (visible) memory.
2643 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2644 // but that's not exposed to the optimizer. This call also doesn't take in an
2645 // Object; if you wish to check an Object you need to load the Object's class
2646 // prior to coming here.
2647 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2648 Compile* C = gvn->C;
2649
2650 if ((*ctrl)->is_top()) {
2651 return C->top();
2652 }
2653
2654 // Fast check for identical types, perhaps identical constants.
2655 // The types can even be identical non-constants, in cases
2656 // involving Array.newInstance, Object.clone, etc.
2657 if (subklass == superklass)
2658 return C->top(); // false path is dead; no test needed.
2659
2660 if (gvn->type(superklass)->singleton()) {
2661 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2662 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2663
2664 // In the common case of an exact superklass, try to fold up the
2665 // test before generating code. You may ask, why not just generate
2666 // the code and then let it fold up? The answer is that the generated
2667 // code will necessarily include null checks, which do not always
2668 // completely fold away. If they are also needless, then they turn
2669 // into a performance loss. Example:
2670 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2671 // Here, the type of 'fa' is often exact, so the store check
2672 // of fa[1]=x will fold up, without testing the nullness of x.
2673 switch (C->static_subtype_check(superk, subk)) {
2674 case Compile::SSC_always_false:
2675 {
2676 Node* always_fail = *ctrl;
2677 *ctrl = gvn->C->top();
2678 return always_fail;
2679 }
2680 case Compile::SSC_always_true:
2681 return C->top();
2682 case Compile::SSC_easy_test:
2683 {
2684 // Just do a direct pointer compare and be done.
2685 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2686 *ctrl = gvn->transform(new IfTrueNode(iff));
2687 return gvn->transform(new IfFalseNode(iff));
2688 }
2689 case Compile::SSC_full_test:
2690 break;
2691 default:
2692 ShouldNotReachHere();
2693 }
2694 }
2695
2696 // %%% Possible further optimization: Even if the superklass is not exact,
2697 // if the subklass is the unique subtype of the superklass, the check
2698 // will always succeed. We could leave a dependency behind to ensure this.
2699
2700 // First load the super-klass's check-offset
2701 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2702 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2703 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2704 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2705 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2706
2707 // Load from the sub-klass's super-class display list, or a 1-word cache of
2708 // the secondary superclass list, or a failing value with a sentinel offset
2709 // if the super-klass is an interface or exceptionally deep in the Java
2710 // hierarchy and we have to scan the secondary superclass list the hard way.
2711 // Worst-case type is a little odd: NULL is allowed as a result (usually
2712 // klass loads can never produce a NULL).
2713 Node *chk_off_X = chk_off;
2714 #ifdef _LP64
2715 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2716 #endif
2717 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2718 // For some types like interfaces the following loadKlass is from a 1-word
2719 // cache which is mutable so can't use immutable memory. Other
2720 // types load from the super-class display table which is immutable.
2721 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2722 Node *kmem = might_be_cache ? m : C->immutable_memory();
2723 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2724
2725 // Compile speed common case: ARE a subtype and we canNOT fail
2726 if( superklass == nkls )
2727 return C->top(); // false path is dead; no test needed.
2728
2729 // See if we get an immediate positive hit. Happens roughly 83% of the
2730 // time. Test to see if the value loaded just previously from the subklass
2731 // is exactly the superklass.
2732 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2733 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2734 *ctrl = gvn->transform(new IfFalseNode(iff1));
2735
2736 // Compile speed common case: Check for being deterministic right now. If
2737 // chk_off is a constant and not equal to cacheoff then we are NOT a
2738 // subklass. In this case we need exactly the 1 test above and we can
2739 // return those results immediately.
2740 if (!might_be_cache) {
2741 Node* not_subtype_ctrl = *ctrl;
2742 *ctrl = iftrue1; // We need exactly the 1 test above
2743 return not_subtype_ctrl;
2744 }
2745
2746 // Gather the various success & failures here
2747 RegionNode *r_ok_subtype = new RegionNode(4);
2748 gvn->record_for_igvn(r_ok_subtype);
2749 RegionNode *r_not_subtype = new RegionNode(3);
2750 gvn->record_for_igvn(r_not_subtype);
2751
2752 r_ok_subtype->init_req(1, iftrue1);
2753
2754 // Check for immediate negative hit. Happens roughly 11% of the time (which
2755 // is roughly 63% of the remaining cases). Test to see if the loaded
2756 // check-offset points into the subklass display list or the 1-element
2757 // cache. If it points to the display (and NOT the cache) and the display
2758 // missed then it's not a subtype.
2759 Node *cacheoff = gvn->intcon(cacheoff_con);
2760 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2761 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2762 *ctrl = gvn->transform(new IfFalseNode(iff2));
2763
2764 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2765 // No performance impact (too rare) but allows sharing of secondary arrays
2766 // which has some footprint reduction.
2767 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2768 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2769 *ctrl = gvn->transform(new IfFalseNode(iff3));
2770
2771 // -- Roads not taken here: --
2772 // We could also have chosen to perform the self-check at the beginning
2773 // of this code sequence, as the assembler does. This would not pay off
2774 // the same way, since the optimizer, unlike the assembler, can perform
2775 // static type analysis to fold away many successful self-checks.
2776 // Non-foldable self checks work better here in second position, because
2777 // the initial primary superclass check subsumes a self-check for most
2778 // types. An exception would be a secondary type like array-of-interface,
2779 // which does not appear in its own primary supertype display.
2780 // Finally, we could have chosen to move the self-check into the
2781 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2782 // dependent manner. But it is worthwhile to have the check here,
2783 // where it can be perhaps be optimized. The cost in code space is
2784 // small (register compare, branch).
2785
2786 // Now do a linear scan of the secondary super-klass array. Again, no real
2787 // performance impact (too rare) but it's gotta be done.
2788 // Since the code is rarely used, there is no penalty for moving it
2789 // out of line, and it can only improve I-cache density.
2790 // The decision to inline or out-of-line this final check is platform
2791 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2792 Node* psc = gvn->transform(
2793 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2794
2795 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2796 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2797 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2798
2799 // Return false path; set default control to true path.
2800 *ctrl = gvn->transform(r_ok_subtype);
2801 return gvn->transform(r_not_subtype);
2802 }
2803
2804 // Profile-driven exact type check:
2805 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2806 float prob,
2807 Node* *casted_receiver) {
2808 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2809 Node* recv_klass = load_object_klass(receiver);
2810 Node* want_klass = makecon(tklass);
2811 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2812 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2813 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2814 set_control( _gvn.transform( new IfTrueNode (iff) ));
2815 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2816
2817 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2818 assert(recv_xtype->klass_is_exact(), "");
2819
2820 // Subsume downstream occurrences of receiver with a cast to
2821 // recv_xtype, since now we know what the type will be.
2822 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2823 (*casted_receiver) = _gvn.transform(cast);
2824 // (User must make the replace_in_map call.)
2825
2826 return fail;
2827 }
2828
2829
2830 //------------------------------seems_never_null-------------------------------
2831 // Use null_seen information if it is available from the profile.
2832 // If we see an unexpected null at a type check we record it and force a
2833 // recompile; the offending check will be recompiled to handle NULLs.
2834 // If we see several offending BCIs, then all checks in the
2835 // method will be recompiled.
2836 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2837 speculating = !_gvn.type(obj)->speculative_maybe_null();
2838 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2839 if (UncommonNullCast // Cutout for this technique
2840 && obj != null() // And not the -Xcomp stupid case?
2841 && !too_many_traps(reason)
2842 ) {
2843 if (speculating) {
2844 return true;
2845 }
2846 if (data == NULL)
2847 // Edge case: no mature data. Be optimistic here.
2848 return true;
2849 // If the profile has not seen a null, assume it won't happen.
2850 assert(java_bc() == Bytecodes::_checkcast ||
2851 java_bc() == Bytecodes::_instanceof ||
2852 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2853 return !data->as_BitData()->null_seen();
2854 }
2855 speculating = false;
2856 return false;
2857 }
2858
2859 //------------------------maybe_cast_profiled_receiver-------------------------
2860 // If the profile has seen exactly one type, narrow to exactly that type.
2861 // Subsequent type checks will always fold up.
2862 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2863 ciKlass* require_klass,
2864 ciKlass* spec_klass,
2865 bool safe_for_replace) {
2866 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2867
2868 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2869
2870 // Make sure we haven't already deoptimized from this tactic.
2871 if (too_many_traps(reason) || too_many_recompiles(reason))
2872 return NULL;
2873
2874 // (No, this isn't a call, but it's enough like a virtual call
2875 // to use the same ciMethod accessor to get the profile info...)
2876 // If we have a speculative type use it instead of profiling (which
2877 // may not help us)
2878 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2879 if (exact_kls != NULL) {// no cast failures here
2880 if (require_klass == NULL ||
2881 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2882 // If we narrow the type to match what the type profile sees or
2883 // the speculative type, we can then remove the rest of the
2884 // cast.
2885 // This is a win, even if the exact_kls is very specific,
2886 // because downstream operations, such as method calls,
2887 // will often benefit from the sharper type.
2888 Node* exact_obj = not_null_obj; // will get updated in place...
2889 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2890 &exact_obj);
2891 { PreserveJVMState pjvms(this);
2892 set_control(slow_ctl);
2893 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2894 }
2895 if (safe_for_replace) {
2896 replace_in_map(not_null_obj, exact_obj);
2897 }
2898 return exact_obj;
2899 }
2900 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2901 }
2902
2903 return NULL;
2904 }
2905
2906 /**
2907 * Cast obj to type and emit guard unless we had too many traps here
2908 * already
2909 *
2910 * @param obj node being casted
2911 * @param type type to cast the node to
2912 * @param not_null true if we know node cannot be null
2913 */
2914 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2915 ciKlass* type,
2916 bool not_null) {
2917 if (stopped()) {
2918 return obj;
2919 }
2920
2921 // type == NULL if profiling tells us this object is always null
2922 if (type != NULL) {
2923 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2924 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2925
2926 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2927 !too_many_traps(class_reason) &&
2928 !too_many_recompiles(class_reason)) {
2929 Node* not_null_obj = NULL;
2930 // not_null is true if we know the object is not null and
2931 // there's no need for a null check
2932 if (!not_null) {
2933 Node* null_ctl = top();
2934 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2935 assert(null_ctl->is_top(), "no null control here");
2936 } else {
2937 not_null_obj = obj;
2938 }
2939
2940 Node* exact_obj = not_null_obj;
2941 ciKlass* exact_kls = type;
2942 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2943 &exact_obj);
2944 {
2945 PreserveJVMState pjvms(this);
2946 set_control(slow_ctl);
2947 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2948 }
2949 replace_in_map(not_null_obj, exact_obj);
2950 obj = exact_obj;
2951 }
2952 } else {
2953 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2954 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2955 Node* exact_obj = null_assert(obj);
2956 replace_in_map(obj, exact_obj);
2957 obj = exact_obj;
2958 }
2959 }
2960 return obj;
2961 }
2962
2963 //-------------------------------gen_instanceof--------------------------------
2964 // Generate an instance-of idiom. Used by both the instance-of bytecode
2965 // and the reflective instance-of call.
2966 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2967 kill_dead_locals(); // Benefit all the uncommon traps
2968 assert( !stopped(), "dead parse path should be checked in callers" );
2969 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2970 "must check for not-null not-dead klass in callers");
2971
2972 // Make the merge point
2973 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2974 RegionNode* region = new RegionNode(PATH_LIMIT);
2975 Node* phi = new PhiNode(region, TypeInt::BOOL);
2976 C->set_has_split_ifs(true); // Has chance for split-if optimization
2977
2978 ciProfileData* data = NULL;
2979 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2980 data = method()->method_data()->bci_to_data(bci());
2981 }
2982 bool speculative_not_null = false;
2983 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2984 && seems_never_null(obj, data, speculative_not_null));
2985
2986 // Null check; get casted pointer; set region slot 3
2987 Node* null_ctl = top();
2988 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2989
2990 // If not_null_obj is dead, only null-path is taken
2991 if (stopped()) { // Doing instance-of on a NULL?
2992 set_control(null_ctl);
2993 return intcon(0);
2994 }
2995 region->init_req(_null_path, null_ctl);
2996 phi ->init_req(_null_path, intcon(0)); // Set null path value
2997 if (null_ctl == top()) {
2998 // Do this eagerly, so that pattern matches like is_diamond_phi
2999 // will work even during parsing.
3000 assert(_null_path == PATH_LIMIT-1, "delete last");
3001 region->del_req(_null_path);
3002 phi ->del_req(_null_path);
3003 }
3004
3005 // Do we know the type check always succeed?
3006 bool known_statically = false;
3007 if (_gvn.type(superklass)->singleton()) {
3008 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3009 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3010 if (subk != NULL && subk->is_loaded()) {
3011 int static_res = C->static_subtype_check(superk, subk);
3012 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3013 }
3014 }
3015
3016 if (!known_statically) {
3017 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3018 // We may not have profiling here or it may not help us. If we
3019 // have a speculative type use it to perform an exact cast.
3020 ciKlass* spec_obj_type = obj_type->speculative_type();
3021 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3022 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3023 if (stopped()) { // Profile disagrees with this path.
3024 set_control(null_ctl); // Null is the only remaining possibility.
3025 return intcon(0);
3026 }
3027 if (cast_obj != NULL) {
3028 not_null_obj = cast_obj;
3029 }
3030 }
3031 }
3032
3033 // Load the object's klass
3034 Node* obj_klass = load_object_klass(not_null_obj);
3035
3036 // Generate the subtype check
3037 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3038
3039 // Plug in the success path to the general merge in slot 1.
3040 region->init_req(_obj_path, control());
3041 phi ->init_req(_obj_path, intcon(1));
3042
3043 // Plug in the failing path to the general merge in slot 2.
3044 region->init_req(_fail_path, not_subtype_ctrl);
3045 phi ->init_req(_fail_path, intcon(0));
3046
3047 // Return final merged results
3048 set_control( _gvn.transform(region) );
3049 record_for_igvn(region);
3050
3051 // If we know the type check always succeeds then we don't use the
3052 // profiling data at this bytecode. Don't lose it, feed it to the
3053 // type system as a speculative type.
3054 if (safe_for_replace) {
3055 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3056 replace_in_map(obj, casted_obj);
3057 }
3058
3059 return _gvn.transform(phi);
3060 }
3061
3062 //-------------------------------gen_checkcast---------------------------------
3063 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3064 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3065 // uncommon-trap paths work. Adjust stack after this call.
3066 // If failure_control is supplied and not null, it is filled in with
3067 // the control edge for the cast failure. Otherwise, an appropriate
3068 // uncommon trap or exception is thrown.
3069 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3070 Node* *failure_control) {
3071 kill_dead_locals(); // Benefit all the uncommon traps
3072 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3073 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3074
3075 // Fast cutout: Check the case that the cast is vacuously true.
3076 // This detects the common cases where the test will short-circuit
3077 // away completely. We do this before we perform the null check,
3078 // because if the test is going to turn into zero code, we don't
3079 // want a residual null check left around. (Causes a slowdown,
3080 // for example, in some objArray manipulations, such as a[i]=a[j].)
3081 if (tk->singleton()) {
3082 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3083 if (objtp != NULL && objtp->klass() != NULL) {
3084 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3085 case Compile::SSC_always_true:
3086 // If we know the type check always succeed then we don't use
3087 // the profiling data at this bytecode. Don't lose it, feed it
3088 // to the type system as a speculative type.
3089 return record_profiled_receiver_for_speculation(obj);
3090 case Compile::SSC_always_false:
3091 // It needs a null check because a null will *pass* the cast check.
3092 // A non-null value will always produce an exception.
3093 return null_assert(obj);
3094 }
3095 }
3096 }
3097
3098 ciProfileData* data = NULL;
3099 bool safe_for_replace = false;
3100 if (failure_control == NULL) { // use MDO in regular case only
3101 assert(java_bc() == Bytecodes::_aastore ||
3102 java_bc() == Bytecodes::_checkcast,
3103 "interpreter profiles type checks only for these BCs");
3104 data = method()->method_data()->bci_to_data(bci());
3105 safe_for_replace = true;
3106 }
3107
3108 // Make the merge point
3109 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3110 RegionNode* region = new RegionNode(PATH_LIMIT);
3111 Node* phi = new PhiNode(region, toop);
3112 C->set_has_split_ifs(true); // Has chance for split-if optimization
3113
3114 // Use null-cast information if it is available
3115 bool speculative_not_null = false;
3116 bool never_see_null = ((failure_control == NULL) // regular case only
3117 && seems_never_null(obj, data, speculative_not_null));
3118
3119 // Null check; get casted pointer; set region slot 3
3120 Node* null_ctl = top();
3121 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3122
3123 // If not_null_obj is dead, only null-path is taken
3124 if (stopped()) { // Doing instance-of on a NULL?
3125 set_control(null_ctl);
3126 return null();
3127 }
3128 region->init_req(_null_path, null_ctl);
3129 phi ->init_req(_null_path, null()); // Set null path value
3130 if (null_ctl == top()) {
3131 // Do this eagerly, so that pattern matches like is_diamond_phi
3132 // will work even during parsing.
3133 assert(_null_path == PATH_LIMIT-1, "delete last");
3134 region->del_req(_null_path);
3135 phi ->del_req(_null_path);
3136 }
3137
3138 Node* cast_obj = NULL;
3139 if (tk->klass_is_exact()) {
3140 // The following optimization tries to statically cast the speculative type of the object
3141 // (for example obtained during profiling) to the type of the superklass and then do a
3142 // dynamic check that the type of the object is what we expect. To work correctly
3143 // for checkcast and aastore the type of superklass should be exact.
3144 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3145 // We may not have profiling here or it may not help us. If we have
3146 // a speculative type use it to perform an exact cast.
3147 ciKlass* spec_obj_type = obj_type->speculative_type();
3148 if (spec_obj_type != NULL || data != NULL) {
3149 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3150 if (cast_obj != NULL) {
3151 if (failure_control != NULL) // failure is now impossible
3152 (*failure_control) = top();
3153 // adjust the type of the phi to the exact klass:
3154 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3155 }
3156 }
3157 }
3158
3159 if (cast_obj == NULL) {
3160 // Load the object's klass
3161 Node* obj_klass = load_object_klass(not_null_obj);
3162
3163 // Generate the subtype check
3164 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3165
3166 // Plug in success path into the merge
3167 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3168 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3169 if (failure_control == NULL) {
3170 if (not_subtype_ctrl != top()) { // If failure is possible
3171 PreserveJVMState pjvms(this);
3172 set_control(not_subtype_ctrl);
3173 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3174 }
3175 } else {
3176 (*failure_control) = not_subtype_ctrl;
3177 }
3178 }
3179
3180 region->init_req(_obj_path, control());
3181 phi ->init_req(_obj_path, cast_obj);
3182
3183 // A merge of NULL or Casted-NotNull obj
3184 Node* res = _gvn.transform(phi);
3185
3186 // Note I do NOT always 'replace_in_map(obj,result)' here.
3187 // if( tk->klass()->can_be_primary_super() )
3188 // This means that if I successfully store an Object into an array-of-String
3189 // I 'forget' that the Object is really now known to be a String. I have to
3190 // do this because we don't have true union types for interfaces - if I store
3191 // a Baz into an array-of-Interface and then tell the optimizer it's an
3192 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3193 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3194 // replace_in_map( obj, res );
3195
3196 // Return final merged results
3197 set_control( _gvn.transform(region) );
3198 record_for_igvn(region);
3199
3200 return record_profiled_receiver_for_speculation(res);
3201 }
3202
3203 //------------------------------next_monitor-----------------------------------
3204 // What number should be given to the next monitor?
3205 int GraphKit::next_monitor() {
3206 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3207 int next = current + C->sync_stack_slots();
3208 // Keep the toplevel high water mark current:
3209 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3210 return current;
3211 }
3212
3213 //------------------------------insert_mem_bar---------------------------------
3214 // Memory barrier to avoid floating things around
3215 // The membar serves as a pinch point between both control and all memory slices.
3216 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3217 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3218 mb->init_req(TypeFunc::Control, control());
3219 mb->init_req(TypeFunc::Memory, reset_memory());
3220 Node* membar = _gvn.transform(mb);
3221 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3222 set_all_memory_call(membar);
3223 return membar;
3224 }
3225
3226 //-------------------------insert_mem_bar_volatile----------------------------
3227 // Memory barrier to avoid floating things around
3228 // The membar serves as a pinch point between both control and memory(alias_idx).
3229 // If you want to make a pinch point on all memory slices, do not use this
3230 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3231 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3232 // When Parse::do_put_xxx updates a volatile field, it appends a series
3233 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3234 // The first membar is on the same memory slice as the field store opcode.
3235 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3236 // All the other membars (for other volatile slices, including AliasIdxBot,
3237 // which stands for all unknown volatile slices) are control-dependent
3238 // on the first membar. This prevents later volatile loads or stores
3239 // from sliding up past the just-emitted store.
3240
3241 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3242 mb->set_req(TypeFunc::Control,control());
3243 if (alias_idx == Compile::AliasIdxBot) {
3244 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3245 } else {
3246 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3247 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3248 }
3249 Node* membar = _gvn.transform(mb);
3250 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3251 if (alias_idx == Compile::AliasIdxBot) {
3252 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3253 } else {
3254 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3255 }
3256 return membar;
3257 }
3258
3259 //------------------------------shared_lock------------------------------------
3260 // Emit locking code.
3261 FastLockNode* GraphKit::shared_lock(Node* obj) {
3262 // bci is either a monitorenter bc or InvocationEntryBci
3263 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3264 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3265
3266 if( !GenerateSynchronizationCode )
3267 return NULL; // Not locking things?
3268 if (stopped()) // Dead monitor?
3269 return NULL;
3270
3271 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3272
3273 obj = shenandoah_write_barrier(obj);
3274
3275 // Box the stack location
3276 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3277 Node* mem = reset_memory();
3278
3279 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3280 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3281 // Create the counters for this fast lock.
3282 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3283 }
3284
3285 // Create the rtm counters for this fast lock if needed.
3286 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3287
3288 // Add monitor to debug info for the slow path. If we block inside the
3289 // slow path and de-opt, we need the monitor hanging around
3290 map()->push_monitor( flock );
3291
3292 const TypeFunc *tf = LockNode::lock_type();
3293 LockNode *lock = new LockNode(C, tf);
3294
3295 lock->init_req( TypeFunc::Control, control() );
3296 lock->init_req( TypeFunc::Memory , mem );
3297 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3298 lock->init_req( TypeFunc::FramePtr, frameptr() );
3299 lock->init_req( TypeFunc::ReturnAdr, top() );
3300
3301 lock->init_req(TypeFunc::Parms + 0, obj);
3302 lock->init_req(TypeFunc::Parms + 1, box);
3303 lock->init_req(TypeFunc::Parms + 2, flock);
3304 add_safepoint_edges(lock);
3305
3306 lock = _gvn.transform( lock )->as_Lock();
3307
3308 // lock has no side-effects, sets few values
3309 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3310
3311 insert_mem_bar(Op_MemBarAcquireLock);
3312
3313 // Add this to the worklist so that the lock can be eliminated
3314 record_for_igvn(lock);
3315
3316 #ifndef PRODUCT
3317 if (PrintLockStatistics) {
3318 // Update the counter for this lock. Don't bother using an atomic
3319 // operation since we don't require absolute accuracy.
3320 lock->create_lock_counter(map()->jvms());
3321 increment_counter(lock->counter()->addr());
3322 }
3323 #endif
3324
3325 return flock;
3326 }
3327
3328
3329 //------------------------------shared_unlock----------------------------------
3330 // Emit unlocking code.
3331 void GraphKit::shared_unlock(Node* box, Node* obj) {
3332 // bci is either a monitorenter bc or InvocationEntryBci
3333 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3334 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3335
3336 if( !GenerateSynchronizationCode )
3337 return;
3338 if (stopped()) { // Dead monitor?
3339 map()->pop_monitor(); // Kill monitor from debug info
3340 return;
3341 }
3342
3343 // Memory barrier to avoid floating things down past the locked region
3344 insert_mem_bar(Op_MemBarReleaseLock);
3345
3346 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3347 UnlockNode *unlock = new UnlockNode(C, tf);
3348 #ifdef ASSERT
3349 unlock->set_dbg_jvms(sync_jvms());
3350 #endif
3351 uint raw_idx = Compile::AliasIdxRaw;
3352 unlock->init_req( TypeFunc::Control, control() );
3353 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3354 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3355 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3356 unlock->init_req( TypeFunc::ReturnAdr, top() );
3357
3358 unlock->init_req(TypeFunc::Parms + 0, obj);
3359 unlock->init_req(TypeFunc::Parms + 1, box);
3360 unlock = _gvn.transform(unlock)->as_Unlock();
3361
3362 Node* mem = reset_memory();
3363
3364 // unlock has no side-effects, sets few values
3365 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3366
3367 // Kill monitor from debug info
3368 map()->pop_monitor( );
3369 }
3370
3371 //-------------------------------get_layout_helper-----------------------------
3372 // If the given klass is a constant or known to be an array,
3373 // fetch the constant layout helper value into constant_value
3374 // and return (Node*)NULL. Otherwise, load the non-constant
3375 // layout helper value, and return the node which represents it.
3376 // This two-faced routine is useful because allocation sites
3377 // almost always feature constant types.
3378 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3379 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3380 if (!StressReflectiveCode && inst_klass != NULL) {
3381 ciKlass* klass = inst_klass->klass();
3382 bool xklass = inst_klass->klass_is_exact();
3383 if (xklass || klass->is_array_klass()) {
3384 jint lhelper = klass->layout_helper();
3385 if (lhelper != Klass::_lh_neutral_value) {
3386 constant_value = lhelper;
3387 return (Node*) NULL;
3388 }
3389 }
3390 }
3391 constant_value = Klass::_lh_neutral_value; // put in a known value
3392 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3393 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3394 }
3395
3396 // We just put in an allocate/initialize with a big raw-memory effect.
3397 // Hook selected additional alias categories on the initialization.
3398 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3399 MergeMemNode* init_in_merge,
3400 Node* init_out_raw) {
3401 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3402 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3403
3404 Node* prevmem = kit.memory(alias_idx);
3405 init_in_merge->set_memory_at(alias_idx, prevmem);
3406 kit.set_memory(init_out_raw, alias_idx);
3407 }
3408
3409 //---------------------------set_output_for_allocation-------------------------
3410 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3411 const TypeOopPtr* oop_type,
3412 bool deoptimize_on_exception) {
3413 int rawidx = Compile::AliasIdxRaw;
3414 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3415 add_safepoint_edges(alloc);
3416 Node* allocx = _gvn.transform(alloc);
3417 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3418 // create memory projection for i_o
3419 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3420 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3421
3422 // create a memory projection as for the normal control path
3423 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3424 set_memory(malloc, rawidx);
3425
3426 // a normal slow-call doesn't change i_o, but an allocation does
3427 // we create a separate i_o projection for the normal control path
3428 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3429 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3430
3431 // put in an initialization barrier
3432 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3433 rawoop)->as_Initialize();
3434 assert(alloc->initialization() == init, "2-way macro link must work");
3435 assert(init ->allocation() == alloc, "2-way macro link must work");
3436 {
3437 // Extract memory strands which may participate in the new object's
3438 // initialization, and source them from the new InitializeNode.
3439 // This will allow us to observe initializations when they occur,
3440 // and link them properly (as a group) to the InitializeNode.
3441 assert(init->in(InitializeNode::Memory) == malloc, "");
3442 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3443 init->set_req(InitializeNode::Memory, minit_in);
3444 record_for_igvn(minit_in); // fold it up later, if possible
3445 Node* minit_out = memory(rawidx);
3446 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3447 if (oop_type->isa_aryptr()) {
3448 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3449 int elemidx = C->get_alias_index(telemref);
3450 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3451 } else if (oop_type->isa_instptr()) {
3452 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3453 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3454 ciField* field = ik->nonstatic_field_at(i);
3455 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3456 continue; // do not bother to track really large numbers of fields
3457 // Find (or create) the alias category for this field:
3458 int fieldidx = C->alias_type(field)->index();
3459 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3460 }
3461 }
3462 }
3463
3464 // Cast raw oop to the real thing...
3465 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3466 javaoop = _gvn.transform(javaoop);
3467 C->set_recent_alloc(control(), javaoop);
3468 assert(just_allocated_object(control()) == javaoop, "just allocated");
3469
3470 #ifdef ASSERT
3471 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3472 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3473 "Ideal_allocation works");
3474 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3475 "Ideal_allocation works");
3476 if (alloc->is_AllocateArray()) {
3477 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3478 "Ideal_allocation works");
3479 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3480 "Ideal_allocation works");
3481 } else {
3482 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3483 }
3484 }
3485 #endif //ASSERT
3486
3487 return javaoop;
3488 }
3489
3490 //---------------------------new_instance--------------------------------------
3491 // This routine takes a klass_node which may be constant (for a static type)
3492 // or may be non-constant (for reflective code). It will work equally well
3493 // for either, and the graph will fold nicely if the optimizer later reduces
3494 // the type to a constant.
3495 // The optional arguments are for specialized use by intrinsics:
3496 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3497 // - If 'return_size_val', report the the total object size to the caller.
3498 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3499 Node* GraphKit::new_instance(Node* klass_node,
3500 Node* extra_slow_test,
3501 Node* *return_size_val,
3502 bool deoptimize_on_exception) {
3503 // Compute size in doublewords
3504 // The size is always an integral number of doublewords, represented
3505 // as a positive bytewise size stored in the klass's layout_helper.
3506 // The layout_helper also encodes (in a low bit) the need for a slow path.
3507 jint layout_con = Klass::_lh_neutral_value;
3508 Node* layout_val = get_layout_helper(klass_node, layout_con);
3509 int layout_is_con = (layout_val == NULL);
3510
3511 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3512 // Generate the initial go-slow test. It's either ALWAYS (return a
3513 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3514 // case) a computed value derived from the layout_helper.
3515 Node* initial_slow_test = NULL;
3516 if (layout_is_con) {
3517 assert(!StressReflectiveCode, "stress mode does not use these paths");
3518 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3519 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3520 } else { // reflective case
3521 // This reflective path is used by Unsafe.allocateInstance.
3522 // (It may be stress-tested by specifying StressReflectiveCode.)
3523 // Basically, we want to get into the VM is there's an illegal argument.
3524 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3525 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3526 if (extra_slow_test != intcon(0)) {
3527 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3528 }
3529 // (Macro-expander will further convert this to a Bool, if necessary.)
3530 }
3531
3532 // Find the size in bytes. This is easy; it's the layout_helper.
3533 // The size value must be valid even if the slow path is taken.
3534 Node* size = NULL;
3535 if (layout_is_con) {
3536 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3537 } else { // reflective case
3538 // This reflective path is used by clone and Unsafe.allocateInstance.
3539 size = ConvI2X(layout_val);
3540
3541 // Clear the low bits to extract layout_helper_size_in_bytes:
3542 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3543 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3544 size = _gvn.transform( new AndXNode(size, mask) );
3545 }
3546 if (return_size_val != NULL) {
3547 (*return_size_val) = size;
3548 }
3549
3550 // This is a precise notnull oop of the klass.
3551 // (Actually, it need not be precise if this is a reflective allocation.)
3552 // It's what we cast the result to.
3553 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3554 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3555 const TypeOopPtr* oop_type = tklass->as_instance_type();
3556
3557 // Now generate allocation code
3558
3559 // The entire memory state is needed for slow path of the allocation
3560 // since GC and deoptimization can happened.
3561 Node *mem = reset_memory();
3562 set_all_memory(mem); // Create new memory state
3563
3564 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3565 control(), mem, i_o(),
3566 size, klass_node,
3567 initial_slow_test);
3568
3569 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3570 }
3571
3572 //-------------------------------new_array-------------------------------------
3573 // helper for both newarray and anewarray
3574 // The 'length' parameter is (obviously) the length of the array.
3575 // See comments on new_instance for the meaning of the other arguments.
3576 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3577 Node* length, // number of array elements
3578 int nargs, // number of arguments to push back for uncommon trap
3579 Node* *return_size_val,
3580 bool deoptimize_on_exception) {
3581 jint layout_con = Klass::_lh_neutral_value;
3582 Node* layout_val = get_layout_helper(klass_node, layout_con);
3583 int layout_is_con = (layout_val == NULL);
3584
3585 if (!layout_is_con && !StressReflectiveCode &&
3586 !too_many_traps(Deoptimization::Reason_class_check)) {
3587 // This is a reflective array creation site.
3588 // Optimistically assume that it is a subtype of Object[],
3589 // so that we can fold up all the address arithmetic.
3590 layout_con = Klass::array_layout_helper(T_OBJECT);
3591 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3592 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3593 { BuildCutout unless(this, bol_lh, PROB_MAX);
3594 inc_sp(nargs);
3595 uncommon_trap(Deoptimization::Reason_class_check,
3596 Deoptimization::Action_maybe_recompile);
3597 }
3598 layout_val = NULL;
3599 layout_is_con = true;
3600 }
3601
3602 // Generate the initial go-slow test. Make sure we do not overflow
3603 // if length is huge (near 2Gig) or negative! We do not need
3604 // exact double-words here, just a close approximation of needed
3605 // double-words. We can't add any offset or rounding bits, lest we
3606 // take a size -1 of bytes and make it positive. Use an unsigned
3607 // compare, so negative sizes look hugely positive.
3608 int fast_size_limit = FastAllocateSizeLimit;
3609 if (layout_is_con) {
3610 assert(!StressReflectiveCode, "stress mode does not use these paths");
3611 // Increase the size limit if we have exact knowledge of array type.
3612 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3613 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3614 }
3615
3616 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3617 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3618
3619 // --- Size Computation ---
3620 // array_size = round_to_heap(array_header + (length << elem_shift));
3621 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3622 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3623 // The rounding mask is strength-reduced, if possible.
3624 int round_mask = MinObjAlignmentInBytes - 1;
3625 Node* header_size = NULL;
3626 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3627 // (T_BYTE has the weakest alignment and size restrictions...)
3628 if (layout_is_con) {
3629 int hsize = Klass::layout_helper_header_size(layout_con);
3630 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3631 BasicType etype = Klass::layout_helper_element_type(layout_con);
3632 if ((round_mask & ~right_n_bits(eshift)) == 0)
3633 round_mask = 0; // strength-reduce it if it goes away completely
3634 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3635 assert(header_size_min <= hsize, "generic minimum is smallest");
3636 header_size_min = hsize;
3637 header_size = intcon(hsize + round_mask);
3638 } else {
3639 Node* hss = intcon(Klass::_lh_header_size_shift);
3640 Node* hsm = intcon(Klass::_lh_header_size_mask);
3641 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3642 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3643 Node* mask = intcon(round_mask);
3644 header_size = _gvn.transform( new AddINode(hsize, mask) );
3645 }
3646
3647 Node* elem_shift = NULL;
3648 if (layout_is_con) {
3649 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3650 if (eshift != 0)
3651 elem_shift = intcon(eshift);
3652 } else {
3653 // There is no need to mask or shift this value.
3654 // The semantics of LShiftINode include an implicit mask to 0x1F.
3655 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3656 elem_shift = layout_val;
3657 }
3658
3659 // Transition to native address size for all offset calculations:
3660 Node* lengthx = ConvI2X(length);
3661 Node* headerx = ConvI2X(header_size);
3662 #ifdef _LP64
3663 { const TypeInt* tilen = _gvn.find_int_type(length);
3664 if (tilen != NULL && tilen->_lo < 0) {
3665 // Add a manual constraint to a positive range. Cf. array_element_address.
3666 jint size_max = fast_size_limit;
3667 if (size_max > tilen->_hi) size_max = tilen->_hi;
3668 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3669
3670 // Only do a narrow I2L conversion if the range check passed.
3671 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3672 _gvn.transform(iff);
3673 RegionNode* region = new RegionNode(3);
3674 _gvn.set_type(region, Type::CONTROL);
3675 lengthx = new PhiNode(region, TypeLong::LONG);
3676 _gvn.set_type(lengthx, TypeLong::LONG);
3677
3678 // Range check passed. Use ConvI2L node with narrow type.
3679 Node* passed = IfFalse(iff);
3680 region->init_req(1, passed);
3681 // Make I2L conversion control dependent to prevent it from
3682 // floating above the range check during loop optimizations.
3683 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3684
3685 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3686 region->init_req(2, IfTrue(iff));
3687 lengthx->init_req(2, ConvI2X(length));
3688
3689 set_control(region);
3690 record_for_igvn(region);
3691 record_for_igvn(lengthx);
3692 }
3693 }
3694 #endif
3695
3696 // Combine header size (plus rounding) and body size. Then round down.
3697 // This computation cannot overflow, because it is used only in two
3698 // places, one where the length is sharply limited, and the other
3699 // after a successful allocation.
3700 Node* abody = lengthx;
3701 if (elem_shift != NULL)
3702 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3703 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3704 if (round_mask != 0) {
3705 Node* mask = MakeConX(~round_mask);
3706 size = _gvn.transform( new AndXNode(size, mask) );
3707 }
3708 // else if round_mask == 0, the size computation is self-rounding
3709
3710 if (return_size_val != NULL) {
3711 // This is the size
3712 (*return_size_val) = size;
3713 }
3714
3715 // Now generate allocation code
3716
3717 // The entire memory state is needed for slow path of the allocation
3718 // since GC and deoptimization can happened.
3719 Node *mem = reset_memory();
3720 set_all_memory(mem); // Create new memory state
3721
3722 if (initial_slow_test->is_Bool()) {
3723 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3724 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3725 }
3726
3727 // Create the AllocateArrayNode and its result projections
3728 AllocateArrayNode* alloc
3729 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3730 control(), mem, i_o(),
3731 size, klass_node,
3732 initial_slow_test,
3733 length);
3734
3735 // Cast to correct type. Note that the klass_node may be constant or not,
3736 // and in the latter case the actual array type will be inexact also.
3737 // (This happens via a non-constant argument to inline_native_newArray.)
3738 // In any case, the value of klass_node provides the desired array type.
3739 const TypeInt* length_type = _gvn.find_int_type(length);
3740 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3741 if (ary_type->isa_aryptr() && length_type != NULL) {
3742 // Try to get a better type than POS for the size
3743 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3744 }
3745
3746 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3747
3748 // Cast length on remaining path to be as narrow as possible
3749 if (map()->find_edge(length) >= 0) {
3750 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3751 if (ccast != length) {
3752 _gvn.set_type_bottom(ccast);
3753 record_for_igvn(ccast);
3754 replace_in_map(length, ccast);
3755 }
3756 }
3757
3758 return javaoop;
3759 }
3760
3761 // The following "Ideal_foo" functions are placed here because they recognize
3762 // the graph shapes created by the functions immediately above.
3763
3764 //---------------------------Ideal_allocation----------------------------------
3765 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3766 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3767 if (ptr == NULL) { // reduce dumb test in callers
3768 return NULL;
3769 }
3770
3771 // Attempt to see through Shenandoah barriers.
3772 ptr = ShenandoahBarrierNode::skip_through_barrier(ptr);
3773
3774 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3775 ptr = ptr->in(1);
3776 if (ptr == NULL) return NULL;
3777 }
3778 // Return NULL for allocations with several casts:
3779 // j.l.reflect.Array.newInstance(jobject, jint)
3780 // Object.clone()
3781 // to keep more precise type from last cast.
3782 if (ptr->is_Proj()) {
3783 Node* allo = ptr->in(0);
3784 if (allo != NULL && allo->is_Allocate()) {
3785 return allo->as_Allocate();
3786 }
3787 }
3788 // Report failure to match.
3789 return NULL;
3790 }
3791
3792 // Fancy version which also strips off an offset (and reports it to caller).
3793 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3794 intptr_t& offset) {
3795 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3796 if (base == NULL) return NULL;
3797 return Ideal_allocation(base, phase);
3798 }
3799
3800 // Trace Initialize <- Proj[Parm] <- Allocate
3801 AllocateNode* InitializeNode::allocation() {
3802 Node* rawoop = in(InitializeNode::RawAddress);
3803 if (rawoop->is_Proj()) {
3804 Node* alloc = rawoop->in(0);
3805 if (alloc->is_Allocate()) {
3806 return alloc->as_Allocate();
3807 }
3808 }
3809 return NULL;
3810 }
3811
3812 // Trace Allocate -> Proj[Parm] -> Initialize
3813 InitializeNode* AllocateNode::initialization() {
3814 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
3815 if (rawoop == NULL) return NULL;
3816 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3817 Node* init = rawoop->fast_out(i);
3818 if (init->is_Initialize()) {
3819 assert(init->as_Initialize()->allocation() == this, "2-way link");
3820 return init->as_Initialize();
3821 }
3822 }
3823 return NULL;
3824 }
3825
3826 //----------------------------- loop predicates ---------------------------
3827
3828 //------------------------------add_predicate_impl----------------------------
3829 bool GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3830 // Too many traps seen?
3831 if (too_many_traps(reason)) {
3832 #ifdef ASSERT
3833 if (TraceLoopPredicate) {
3834 int tc = C->trap_count(reason);
3835 tty->print("too many traps=%s tcount=%d in ",
3836 Deoptimization::trap_reason_name(reason), tc);
3837 method()->print(); // which method has too many predicate traps
3838 tty->cr();
3839 }
3840 #endif
3841 // We cannot afford to take more traps here,
3842 // do not generate predicate.
3843 return false;
3844 }
3845
3846 Node *cont = _gvn.intcon(1);
3847 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3848 Node *bol = _gvn.transform(new Conv2BNode(opq));
3849 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3850 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3851 C->add_predicate_opaq(opq);
3852 {
3853 PreserveJVMState pjvms(this);
3854 set_control(iffalse);
3855 inc_sp(nargs);
3856 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3857 }
3858 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3859 set_control(iftrue);
3860 return true;
3861 }
3862
3863 //------------------------------add_predicate---------------------------------
3864 void GraphKit::add_predicate(int nargs) {
3865 bool added = false;
3866 if (UseProfiledLoopPredicate) {
3867 added = add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
3868 }
3869 if (!added && UseLoopPredicate) {
3870 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3871 }
3872 // loop's limit check predicate should be near the loop.
3873 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3874 }
3875
3876 //----------------------------- store barriers ----------------------------
3877 #define __ ideal.
3878
3879 bool GraphKit::use_ReduceInitialCardMarks() {
3880 BarrierSet *bs = BarrierSet::barrier_set();
3881 return bs->is_a(BarrierSet::CardTableBarrierSet)
3882 && barrier_set_cast<CardTableBarrierSet>(bs)->can_elide_tlab_store_barriers()
3883 && ReduceInitialCardMarks;
3884 }
3885
3886 void GraphKit::sync_kit(IdealKit& ideal) {
3887 set_all_memory(__ merged_memory());
3888 set_i_o(__ i_o());
3889 set_control(__ ctrl());
3890 }
3891
3892 void GraphKit::final_sync(IdealKit& ideal) {
3893 // Final sync IdealKit and graphKit.
3894 sync_kit(ideal);
3895 }
3896
3897 Node* GraphKit::byte_map_base_node() {
3898 // Get base of card map
3899 jbyte* card_table_base = ci_card_table_address();
3900 if (card_table_base != NULL) {
3901 return makecon(TypeRawPtr::make((address)card_table_base));
3902 } else {
3903 return null();
3904 }
3905 }
3906
3907 // vanilla/CMS post barrier
3908 // Insert a write-barrier store. This is to let generational GC work; we have
3909 // to flag all oop-stores before the next GC point.
3910 void GraphKit::write_barrier_post(Node* oop_store,
3911 Node* obj,
3912 Node* adr,
3913 uint adr_idx,
3914 Node* val,
3915 bool use_precise) {
3916 // No store check needed if we're storing a NULL or an old object
3917 // (latter case is probably a string constant). The concurrent
3918 // mark sweep garbage collector, however, needs to have all nonNull
3919 // oop updates flagged via card-marks.
3920 if (val != NULL && val->is_Con()) {
3921 // must be either an oop or NULL
3922 const Type* t = val->bottom_type();
3923 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3924 // stores of null never (?) need barriers
3925 return;
3926 }
3927
3928 if (use_ReduceInitialCardMarks()
3929 && obj == just_allocated_object(control())) {
3930 // We can skip marks on a freshly-allocated object in Eden.
3931 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
3932 // That routine informs GC to take appropriate compensating steps,
3933 // upon a slow-path allocation, so as to make this card-mark
3934 // elision safe.
3935 return;
3936 }
3937
3938 if (!use_precise) {
3939 // All card marks for a (non-array) instance are in one place:
3940 adr = obj;
3941 }
3942 // (Else it's an array (or unknown), and we want more precise card marks.)
3943 assert(adr != NULL, "");
3944
3945 IdealKit ideal(this, true);
3946
3947 // Convert the pointer to an int prior to doing math on it
3948 Node* cast = __ CastPX(__ ctrl(), adr);
3949
3950 // Divide by card size
3951 assert(BarrierSet::barrier_set()->is_a(BarrierSet::CardTableBarrierSet),
3952 "Only one we handle so far.");
3953 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
3954
3955 // Combine card table base and card offset
3956 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3957
3958 // Get the alias_index for raw card-mark memory
3959 int adr_type = Compile::AliasIdxRaw;
3960 Node* zero = __ ConI(0); // Dirty card value
3961 BasicType bt = T_BYTE;
3962
3963 if (UseConcMarkSweepGC && UseCondCardMark) {
3964 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
3965 __ sync_kit(this);
3966 }
3967
3968 if (UseCondCardMark) {
3969 // The classic GC reference write barrier is typically implemented
3970 // as a store into the global card mark table. Unfortunately
3971 // unconditional stores can result in false sharing and excessive
3972 // coherence traffic as well as false transactional aborts.
3973 // UseCondCardMark enables MP "polite" conditional card mark
3974 // stores. In theory we could relax the load from ctrl() to
3975 // no_ctrl, but that doesn't buy much latitude.
3976 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3977 __ if_then(card_val, BoolTest::ne, zero);
3978 }
3979
3980 // Smash zero into card
3981 if( !UseConcMarkSweepGC ) {
3982 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3983 } else {
3984 // Specialized path for CM store barrier
3985 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3986 }
3987
3988 if (UseCondCardMark) {
3989 __ end_if();
3990 }
3991
3992 // Final sync IdealKit and GraphKit.
3993 final_sync(ideal);
3994 }
3995 /*
3996 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3997 * required by SATB to make sure all objects live at the start of the
3998 * marking are kept alive, all reference updates need to any previous
3999 * reference stored before writing.
4000 *
4001 * If the previous value is NULL there is no need to save the old value.
4002 * References that are NULL are filtered during runtime by the barrier
4003 * code to avoid unnecessary queuing.
4004 *
4005 * However in the case of newly allocated objects it might be possible to
4006 * prove that the reference about to be overwritten is NULL during compile
4007 * time and avoid adding the barrier code completely.
4008 *
4009 * The compiler needs to determine that the object in which a field is about
4010 * to be written is newly allocated, and that no prior store to the same field
4011 * has happened since the allocation.
4012 *
4013 * Returns true if the pre-barrier can be removed
4014 */
4015 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
4016 BasicType bt, uint adr_idx) {
4017 intptr_t offset = 0;
4018 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4019 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4020
4021 if (offset == Type::OffsetBot) {
4022 return false; // cannot unalias unless there are precise offsets
4023 }
4024
4025 if (alloc == NULL) {
4026 return false; // No allocation found
4027 }
4028
4029 intptr_t size_in_bytes = type2aelembytes(bt);
4030
4031 Node* mem = memory(adr_idx); // start searching here...
4032
4033 for (int cnt = 0; cnt < 50; cnt++) {
4034
4035 if (mem->is_Store()) {
4036
4037 Node* st_adr = mem->in(MemNode::Address);
4038 intptr_t st_offset = 0;
4039 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
4040
4041 if (st_base == NULL) {
4042 break; // inscrutable pointer
4043 }
4044
4045 // Break we have found a store with same base and offset as ours so break
4046 if (st_base == base && st_offset == offset) {
4047 break;
4048 }
4049
4050 if (st_offset != offset && st_offset != Type::OffsetBot) {
4051 const int MAX_STORE = BytesPerLong;
4052 if (st_offset >= offset + size_in_bytes ||
4053 st_offset <= offset - MAX_STORE ||
4054 st_offset <= offset - mem->as_Store()->memory_size()) {
4055 // Success: The offsets are provably independent.
4056 // (You may ask, why not just test st_offset != offset and be done?
4057 // The answer is that stores of different sizes can co-exist
4058 // in the same sequence of RawMem effects. We sometimes initialize
4059 // a whole 'tile' of array elements with a single jint or jlong.)
4060 mem = mem->in(MemNode::Memory);
4061 continue; // advance through independent store memory
4062 }
4063 }
4064
4065 if (st_base != base
4066 && MemNode::detect_ptr_independence(base, alloc, st_base,
4067 AllocateNode::Ideal_allocation(st_base, phase),
4068 phase)) {
4069 // Success: The bases are provably independent.
4070 mem = mem->in(MemNode::Memory);
4071 continue; // advance through independent store memory
4072 }
4073 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4074
4075 InitializeNode* st_init = mem->in(0)->as_Initialize();
4076 AllocateNode* st_alloc = st_init->allocation();
4077
4078 // Make sure that we are looking at the same allocation site.
4079 // The alloc variable is guaranteed to not be null here from earlier check.
4080 if (alloc == st_alloc) {
4081 // Check that the initialization is storing NULL so that no previous store
4082 // has been moved up and directly write a reference
4083 Node* captured_store = st_init->find_captured_store(offset,
4084 type2aelembytes(T_OBJECT),
4085 phase);
4086 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4087 return true;
4088 }
4089 }
4090 }
4091
4092 // Unless there is an explicit 'continue', we must bail out here,
4093 // because 'mem' is an inscrutable memory state (e.g., a call).
4094 break;
4095 }
4096
4097 return false;
4098 }
4099
4100 static void g1_write_barrier_pre_helper(const GraphKit& kit, Node* adr) {
4101 if (UseShenandoahGC && ShenandoahSATBBarrier && adr != NULL) {
4102 Node* c = kit.control();
4103 Node* call = c->in(1)->in(1)->in(1)->in(0);
4104 assert(call->is_g1_wb_pre_call(), "g1_wb_pre call expected");
4105 call->add_req(adr);
4106 }
4107 }
4108
4109 // G1 pre/post barriers
4110 void GraphKit::g1_write_barrier_pre(bool do_load,
4111 Node* obj,
4112 Node* adr,
4113 uint alias_idx,
4114 Node* val,
4115 const TypeOopPtr* val_type,
4116 Node* pre_val,
4117 BasicType bt) {
4118
4119 // Some sanity checks
4120 // Note: val is unused in this routine.
4121
4122 if (do_load) {
4123 // We need to generate the load of the previous value
4124 assert(obj != NULL, "must have a base");
4125 assert(adr != NULL, "where are loading from?");
4126 assert(pre_val == NULL, "loaded already?");
4127 assert(val_type != NULL, "need a type");
4128
4129 if (use_ReduceInitialCardMarks()
4130 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4131 return;
4132 }
4133
4134 } else {
4135 // In this case both val_type and alias_idx are unused.
4136 assert(pre_val != NULL, "must be loaded already");
4137 // Nothing to be done if pre_val is null.
4138 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4139 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4140 }
4141 assert(bt == T_OBJECT, "or we shouldn't be here");
4142
4143 IdealKit ideal(this, true);
4144
4145 Node* tls = __ thread(); // ThreadLocalStorage
4146
4147 Node* no_ctrl = NULL;
4148 Node* no_base = __ top();
4149 Node* zero = __ ConI(0);
4150 Node* zeroX = __ ConX(0);
4151
4152 float likely = PROB_LIKELY(0.999);
4153 float unlikely = PROB_UNLIKELY(0.999);
4154
4155 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4156 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4157
4158 // Offsets into the thread
4159 const int marking_offset = in_bytes(UseG1GC ? G1ThreadLocalData::satb_mark_queue_active_offset()
4160 : ShenandoahThreadLocalData::satb_mark_queue_active_offset());
4161 const int index_offset = in_bytes(UseG1GC ? G1ThreadLocalData::satb_mark_queue_index_offset()
4162 : ShenandoahThreadLocalData::satb_mark_queue_index_offset());
4163 const int buffer_offset = in_bytes(UseG1GC ? G1ThreadLocalData::satb_mark_queue_buffer_offset()
4164 : ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
4165
4166 // Now the actual pointers into the thread
4167 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4168 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4169 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4170
4171 // Now some of the values
4172 Node* marking;
4173 if (UseShenandoahGC) {
4174 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
4175 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
4176 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
4177 assert(ShenandoahWriteBarrierNode::is_gc_state_load(ld), "Should match the shape");
4178 } else {
4179 assert(UseG1GC, "should be");
4180 marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4181 }
4182
4183 // if (!marking)
4184 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4185 BasicType index_bt = TypeX_X->basic_type();
4186 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4187 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4188
4189 if (do_load) {
4190 // load original value
4191 // alias_idx correct??
4192 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4193 }
4194
4195 // if (pre_val != NULL)
4196 __ if_then(pre_val, BoolTest::ne, null()); {
4197 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4198
4199 // is the queue for this thread full?
4200 __ if_then(index, BoolTest::ne, zeroX, likely); {
4201
4202 // decrement the index
4203 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4204
4205 // Now get the buffer location we will log the previous value into and store it
4206 Node *log_addr = __ AddP(no_base, buffer, next_index);
4207 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4208 // update the index
4209 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4210
4211 } __ else_(); {
4212
4213 // logging buffer is full, call the runtime
4214 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4215 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4216 } __ end_if(); // (!index)
4217 } __ end_if(); // (pre_val != NULL)
4218 } __ end_if(); // (!marking)
4219
4220 // Final sync IdealKit and GraphKit.
4221 final_sync(ideal);
4222 g1_write_barrier_pre_helper(*this, adr);
4223 }
4224
4225 void GraphKit::shenandoah_enqueue_barrier(Node* pre_val) {
4226
4227 // Some sanity checks
4228 assert(pre_val != NULL, "must be loaded already");
4229 // Nothing to be done if pre_val is null.
4230 if (pre_val->bottom_type()->higher_equal(TypePtr::NULL_PTR)) return;
4231 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4232
4233 IdealKit ideal(this, true);
4234
4235 Node* tls = __ thread(); // ThreadLocalStorage
4236
4237 Node* no_base = __ top();
4238 Node* zero = __ ConI(0);
4239 Node* zeroX = __ ConX(0);
4240
4241 float likely = PROB_LIKELY(0.999);
4242 float unlikely = PROB_UNLIKELY(0.999);
4243
4244 // Offsets into the thread
4245 const int gc_state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
4246 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
4247 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
4248
4249 // Now the actual pointers into the thread
4250 Node* gc_state_adr = __ AddP(no_base, tls, __ ConX(gc_state_offset));
4251 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4252 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4253
4254 const Type* obj_type = pre_val->bottom_type();
4255 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) {
4256 // dunno if it's NULL or not.
4257 // if (pre_val != NULL)
4258 __ if_then(pre_val, BoolTest::ne, null()); {
4259
4260 // Now some of the values
4261 Node* marking = __ load(__ ctrl(), gc_state_adr, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
4262 // if (!marking)
4263 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4264 BasicType index_bt = TypeX_X->basic_type();
4265 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4266 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4267
4268 // is the queue for this thread full?
4269 __ if_then(index, BoolTest::ne, zeroX, likely); {
4270
4271 // decrement the index
4272 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4273
4274 // Now get the buffer location we will log the previous value into and store it
4275 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4276 Node *log_addr = __ AddP(no_base, buffer, next_index);
4277 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4278 // update the index
4279 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4280
4281 } __ else_(); {
4282 // logging buffer is full, call the runtime
4283 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4284 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4285 } __ end_if(); // (!index)
4286 } __ end_if(); // (!marking)
4287 } __ end_if(); // (pre_val != NULL)
4288 } else {
4289 // We know it is not null.
4290 // Now some of the values
4291 Node* marking = __ load(__ ctrl(), gc_state_adr, TypeInt::BOOL, T_BOOLEAN, Compile::AliasIdxRaw);
4292
4293 // if (!marking)
4294 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4295 BasicType index_bt = TypeX_X->basic_type();
4296 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4297 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4298
4299 // is the queue for this thread full?
4300 __ if_then(index, BoolTest::ne, zeroX, likely); {
4301
4302 // decrement the index
4303 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4304
4305 // Now get the buffer location we will log the previous value into and store it
4306 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4307 Node *log_addr = __ AddP(no_base, buffer, next_index);
4308 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4309 // update the index
4310 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4311
4312 } __ else_(); {
4313 // logging buffer is full, call the runtime
4314 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4315 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4316 } __ end_if(); // (!index)
4317 } __ end_if(); // (!marking)
4318 }
4319 // Final sync IdealKit and GraphKit.
4320 final_sync(ideal);
4321 }
4322
4323 void GraphKit::shenandoah_write_barrier_pre(bool do_load,
4324 Node* obj,
4325 Node* adr,
4326 uint alias_idx,
4327 Node* val,
4328 const TypeOopPtr* val_type,
4329 Node* pre_val,
4330 BasicType bt) {
4331
4332 // Some sanity checks
4333 // Note: val is unused in this routine.
4334
4335 if (val != NULL) {
4336 shenandoah_update_matrix(adr, val);
4337 }
4338
4339 if (ShenandoahSATBBarrier) {
4340 g1_write_barrier_pre(do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
4341 }
4342 }
4343
4344 void GraphKit::shenandoah_update_matrix(Node* adr, Node* val) {
4345 if (!UseShenandoahMatrix) {
4346 return;
4347 }
4348
4349 assert(val != NULL, "checked before");
4350 if (adr == NULL) {
4351 return; // Nothing to do
4352 }
4353 assert(adr != NULL, "must not happen");
4354 if (val->bottom_type()->higher_equal(TypePtr::NULL_PTR)) {
4355 // Nothing to do.
4356 return;
4357 }
4358
4359 ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix();
4360
4361 enum { _set_path = 1, _already_set_path, _val_null_path, PATH_LIMIT };
4362 RegionNode* region = new RegionNode(PATH_LIMIT);
4363 Node* prev_mem = memory(Compile::AliasIdxRaw);
4364 Node* memphi = PhiNode::make(region, prev_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
4365 Node* null_ctrl = top();
4366 Node* not_null_val = null_check_oop(val, &null_ctrl);
4367
4368 // Null path: nothing to do.
4369 region->init_req(_val_null_path, null_ctrl);
4370 memphi->init_req(_val_null_path, prev_mem);
4371
4372 // Not null path. Update the matrix.
4373
4374 // This uses a fast calculation for the matrix address. For a description,
4375 // see src/share/vm/gc/shenandoah/shenandoahConnectionMatrix.inline.hpp,
4376 // ShenandoahConnectionMatrix::compute_address(const void* from, const void* to).
4377 address heap_base = ShenandoahHeap::heap()->base();
4378 jint stride = matrix->stride_jint();
4379 jint rs = ShenandoahHeapRegion::region_size_bytes_shift_jint();
4380
4381 guarantee(stride < ShenandoahHeapRegion::region_size_bytes_jint(), "sanity");
4382 guarantee(is_aligned(heap_base, ShenandoahHeapRegion::region_size_bytes()), "sanity");
4383
4384 Node* magic_con = MakeConX((jlong) matrix->matrix_addr() - ((jlong) heap_base >> rs) * (stride + 1));
4385
4386 // Compute addr part
4387 Node* adr_idx = _gvn.transform(new CastP2XNode(control(), adr));
4388 adr_idx = _gvn.transform(new URShiftXNode(adr_idx, intcon(rs)));
4389
4390 // Compute new_val part
4391 Node* val_idx = _gvn.transform(new CastP2XNode(control(), not_null_val));
4392 val_idx = _gvn.transform(new URShiftXNode(val_idx, intcon(rs)));
4393 val_idx = _gvn.transform(new MulXNode(val_idx, MakeConX(stride)));
4394
4395 // Add everything up
4396 adr_idx = _gvn.transform(new AddXNode(adr_idx, val_idx));
4397 adr_idx = _gvn.transform(new CastX2PNode(adr_idx));
4398 Node* matrix_adr = _gvn.transform(new AddPNode(top(), adr_idx, magic_con));
4399
4400 // Load current value
4401 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
4402 Node* current = _gvn.transform(LoadNode::make(_gvn, control(), memory(Compile::AliasIdxRaw),
4403 matrix_adr, adr_type, TypeInt::INT, T_BYTE, MemNode::unordered));
4404
4405 // Check if already set
4406 Node* cmp_set = _gvn.transform(new CmpINode(current, intcon(0)));
4407 Node* cmp_set_bool = _gvn.transform(new BoolNode(cmp_set, BoolTest::eq));
4408 IfNode* cmp_iff = create_and_map_if(control(), cmp_set_bool, PROB_MIN, COUNT_UNKNOWN);
4409
4410 Node* if_not_set = _gvn.transform(new IfTrueNode(cmp_iff));
4411 Node* if_set = _gvn.transform(new IfFalseNode(cmp_iff));
4412
4413 // Already set, exit
4414 set_control(if_set);
4415 region->init_req(_already_set_path, control());
4416 memphi->init_req(_already_set_path, prev_mem);
4417
4418 // Not set: do the store, and finish up
4419 set_control(if_not_set);
4420 Node* store = _gvn.transform(StoreNode::make(_gvn, control(), memory(Compile::AliasIdxRaw),
4421 matrix_adr, adr_type, intcon(1), T_BYTE, MemNode::unordered));
4422 region->init_req(_set_path, control());
4423 memphi->init_req(_set_path, store);
4424
4425 // Merge control flows and memory.
4426 set_control(_gvn.transform(region));
4427 record_for_igvn(region);
4428 set_memory(_gvn.transform(memphi), Compile::AliasIdxRaw);
4429 }
4430
4431 /*
4432 * G1 similar to any GC with a Young Generation requires a way to keep track of
4433 * references from Old Generation to Young Generation to make sure all live
4434 * objects are found. G1 also requires to keep track of object references
4435 * between different regions to enable evacuation of old regions, which is done
4436 * as part of mixed collections. References are tracked in remembered sets and
4437 * is continuously updated as reference are written to with the help of the
4438 * post-barrier.
4439 *
4440 * To reduce the number of updates to the remembered set the post-barrier
4441 * filters updates to fields in objects located in the Young Generation,
4442 * the same region as the reference, when the NULL is being written or
4443 * if the card is already marked as dirty by an earlier write.
4444 *
4445 * Under certain circumstances it is possible to avoid generating the
4446 * post-barrier completely if it is possible during compile time to prove
4447 * the object is newly allocated and that no safepoint exists between the
4448 * allocation and the store.
4449 *
4450 * In the case of slow allocation the allocation code must handle the barrier
4451 * as part of the allocation in the case the allocated object is not located
4452 * in the nursery, this would happen for humongous objects. This is similar to
4453 * how CMS is required to handle this case, see the comments for the method
4454 * CardTableBarrierSet::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier.
4455 * A deferred card mark is required for these objects and handled in the above
4456 * mentioned methods.
4457 *
4458 * Returns true if the post barrier can be removed
4459 */
4460 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4461 Node* adr) {
4462 intptr_t offset = 0;
4463 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4464 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4465
4466 if (offset == Type::OffsetBot) {
4467 return false; // cannot unalias unless there are precise offsets
4468 }
4469
4470 if (alloc == NULL) {
4471 return false; // No allocation found
4472 }
4473
4474 // Start search from Store node
4475 Node* mem = store->in(MemNode::Control);
4476 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4477
4478 InitializeNode* st_init = mem->in(0)->as_Initialize();
4479 AllocateNode* st_alloc = st_init->allocation();
4480
4481 // Make sure we are looking at the same allocation
4482 if (alloc == st_alloc) {
4483 return true;
4484 }
4485 }
4486
4487 return false;
4488 }
4489
4490 //
4491 // Update the card table and add card address to the queue
4492 //
4493 void GraphKit::g1_mark_card(IdealKit& ideal,
4494 Node* card_adr,
4495 Node* oop_store,
4496 uint oop_alias_idx,
4497 Node* index,
4498 Node* index_adr,
4499 Node* buffer,
4500 const TypeFunc* tf) {
4501
4502 Node* zero = __ ConI(0);
4503 Node* zeroX = __ ConX(0);
4504 Node* no_base = __ top();
4505 BasicType card_bt = T_BYTE;
4506 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4507 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4508
4509 // Now do the queue work
4510 __ if_then(index, BoolTest::ne, zeroX); {
4511
4512 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4513 Node* log_addr = __ AddP(no_base, buffer, next_index);
4514
4515 // Order, see storeCM.
4516 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4517 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4518
4519 } __ else_(); {
4520 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4521 } __ end_if();
4522
4523 }
4524
4525 void GraphKit::g1_write_barrier_post(Node* oop_store,
4526 Node* obj,
4527 Node* adr,
4528 uint alias_idx,
4529 Node* val,
4530 BasicType bt,
4531 bool use_precise) {
4532 // If we are writing a NULL then we need no post barrier
4533
4534 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4535 // Must be NULL
4536 const Type* t = val->bottom_type();
4537 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4538 // No post barrier if writing NULLx
4539 return;
4540 }
4541
4542 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4543 // We can skip marks on a freshly-allocated object in Eden.
4544 // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
4545 // That routine informs GC to take appropriate compensating steps,
4546 // upon a slow-path allocation, so as to make this card-mark
4547 // elision safe.
4548 return;
4549 }
4550
4551 if (use_ReduceInitialCardMarks()
4552 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4553 return;
4554 }
4555
4556 if (!use_precise) {
4557 // All card marks for a (non-array) instance are in one place:
4558 adr = obj;
4559 }
4560 // (Else it's an array (or unknown), and we want more precise card marks.)
4561 assert(adr != NULL, "");
4562
4563 IdealKit ideal(this, true);
4564
4565 Node* tls = __ thread(); // ThreadLocalStorage
4566
4567 Node* no_base = __ top();
4568 float likely = PROB_LIKELY(0.999);
4569 float unlikely = PROB_UNLIKELY(0.999);
4570 Node* young_card = __ ConI((jint)G1CardTable::g1_young_card_val());
4571 Node* dirty_card = __ ConI((jint)CardTable::dirty_card_val());
4572 Node* zeroX = __ ConX(0);
4573
4574 // Get the alias_index for raw card-mark memory
4575 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4576
4577 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4578
4579 // Offsets into the thread
4580 const int index_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
4581 const int buffer_offset = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
4582
4583 // Pointers into the thread
4584
4585 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4586 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4587
4588 // Now some values
4589 // Use ctrl to avoid hoisting these values past a safepoint, which could
4590 // potentially reset these fields in the JavaThread.
4591 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4592 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4593
4594 // Convert the store obj pointer to an int prior to doing math on it
4595 // Must use ctrl to prevent "integerized oop" existing across safepoint
4596 Node* cast = __ CastPX(__ ctrl(), adr);
4597
4598 // Divide pointer by card size
4599 Node* card_offset = __ URShiftX( cast, __ ConI(CardTable::card_shift) );
4600
4601 // Combine card table base and card offset
4602 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4603
4604 // If we know the value being stored does it cross regions?
4605
4606 if (val != NULL) {
4607 // Does the store cause us to cross regions?
4608
4609 // Should be able to do an unsigned compare of region_size instead of
4610 // and extra shift. Do we have an unsigned compare??
4611 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4612 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4613
4614 // if (xor_res == 0) same region so skip
4615 __ if_then(xor_res, BoolTest::ne, zeroX); {
4616
4617 // No barrier if we are storing a NULL
4618 __ if_then(val, BoolTest::ne, null(), unlikely); {
4619
4620 // Ok must mark the card if not already dirty
4621
4622 // load the original value of the card
4623 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4624
4625 __ if_then(card_val, BoolTest::ne, young_card); {
4626 sync_kit(ideal);
4627 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4628 insert_mem_bar(Op_MemBarVolatile, oop_store);
4629 __ sync_kit(this);
4630
4631 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4632 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4633 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4634 } __ end_if();
4635 } __ end_if();
4636 } __ end_if();
4637 } __ end_if();
4638 } else {
4639 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4640 // We don't need a barrier here if the destination is a newly allocated object
4641 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4642 // are set to 'g1_young_gen' (see G1CardTable::verify_g1_young_region()).
4643 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4644 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4645 __ if_then(card_val, BoolTest::ne, young_card); {
4646 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4647 } __ end_if();
4648 }
4649
4650 // Final sync IdealKit and GraphKit.
4651 final_sync(ideal);
4652 }
4653 #undef __
4654
4655
4656 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4657 Node* len = load_array_length(load_String_value(ctrl, str));
4658 Node* coder = load_String_coder(ctrl, str);
4659 // Divide length by 2 if coder is UTF16
4660 return _gvn.transform(new RShiftINode(len, coder));
4661 }
4662
4663 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4664 int value_offset = java_lang_String::value_offset_in_bytes();
4665 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4666 false, NULL, 0);
4667 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4668 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4669 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4670 ciTypeArrayKlass::make(T_BYTE), true, 0);
4671 int value_field_idx = C->get_alias_index(value_field_type);
4672
4673 if (!ShenandoahOptimizeInstanceFinals) {
4674 str = shenandoah_read_barrier(str);
4675 }
4676
4677 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4678 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4679 // String.value field is known to be @Stable.
4680 if (UseImplicitStableValues) {
4681 load = cast_array_to_stable(load, value_type);
4682 }
4683 return load;
4684 }
4685
4686 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4687 if (!CompactStrings) {
4688 return intcon(java_lang_String::CODER_UTF16);
4689 }
4690 int coder_offset = java_lang_String::coder_offset_in_bytes();
4691 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4692 false, NULL, 0);
4693 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4694 int coder_field_idx = C->get_alias_index(coder_field_type);
4695
4696 if (!ShenandoahOptimizeInstanceFinals) {
4697 str = shenandoah_read_barrier(str);
4698 }
4699
4700 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4701 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4702 }
4703
4704 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4705 int value_offset = java_lang_String::value_offset_in_bytes();
4706 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4707 false, NULL, 0);
4708 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4709
4710 str = shenandoah_write_barrier(str);
4711
4712 store_oop_to_object(control(), str, basic_plus_adr(str, value_offset), value_field_type,
4713 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4714 }
4715
4716 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4717 int coder_offset = java_lang_String::coder_offset_in_bytes();
4718 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4719 false, NULL, 0);
4720
4721 str = shenandoah_write_barrier(str);
4722
4723 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4724 int coder_field_idx = C->get_alias_index(coder_field_type);
4725 store_to_memory(control(), basic_plus_adr(str, coder_offset),
4726 value, T_BYTE, coder_field_idx, MemNode::unordered);
4727 }
4728
4729 // Capture src and dst memory state with a MergeMemNode
4730 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4731 if (src_type == dst_type) {
4732 // Types are equal, we don't need a MergeMemNode
4733 return memory(src_type);
4734 }
4735 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4736 record_for_igvn(merge); // fold it up later, if possible
4737 int src_idx = C->get_alias_index(src_type);
4738 int dst_idx = C->get_alias_index(dst_type);
4739 merge->set_memory_at(src_idx, memory(src_idx));
4740 merge->set_memory_at(dst_idx, memory(dst_idx));
4741 return merge;
4742 }
4743
4744 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4745 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4746 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4747 // If input and output memory types differ, capture both states to preserve
4748 // the dependency between preceding and subsequent loads/stores.
4749 // For example, the following program:
4750 // StoreB
4751 // compress_string
4752 // LoadB
4753 // has this memory graph (use->def):
4754 // LoadB -> compress_string -> CharMem
4755 // ... -> StoreB -> ByteMem
4756 // The intrinsic hides the dependency between LoadB and StoreB, causing
4757 // the load to read from memory not containing the result of the StoreB.
4758 // The correct memory graph should look like this:
4759 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4760 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4761 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4762 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4763 set_memory(res_mem, TypeAryPtr::BYTES);
4764 return str;
4765 }
4766
4767 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4768 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4769 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4770 // Capture src and dst memory (see comment in 'compress_string').
4771 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4772 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4773 set_memory(_gvn.transform(str), dst_type);
4774 }
4775
4776 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4777
4778 src = shenandoah_read_barrier(src);
4779 dst = shenandoah_write_barrier(dst);
4780
4781 /**
4782 * int i_char = start;
4783 * for (int i_byte = 0; i_byte < count; i_byte++) {
4784 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4785 * }
4786 */
4787 add_predicate();
4788 RegionNode* head = new RegionNode(3);
4789 head->init_req(1, control());
4790 gvn().set_type(head, Type::CONTROL);
4791 record_for_igvn(head);
4792
4793 Node* i_byte = new PhiNode(head, TypeInt::INT);
4794 i_byte->init_req(1, intcon(0));
4795 gvn().set_type(i_byte, TypeInt::INT);
4796 record_for_igvn(i_byte);
4797
4798 Node* i_char = new PhiNode(head, TypeInt::INT);
4799 i_char->init_req(1, start);
4800 gvn().set_type(i_char, TypeInt::INT);
4801 record_for_igvn(i_char);
4802
4803 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4804 gvn().set_type(mem, Type::MEMORY);
4805 record_for_igvn(mem);
4806 set_control(head);
4807 set_memory(mem, TypeAryPtr::BYTES);
4808 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4809 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4810 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4811 false, false, true /* mismatched */);
4812
4813 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4814 head->init_req(2, IfTrue(iff));
4815 mem->init_req(2, st);
4816 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4817 i_char->init_req(2, AddI(i_char, intcon(2)));
4818
4819 set_control(IfFalse(iff));
4820 set_memory(st, TypeAryPtr::BYTES);
4821 }
4822
4823 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4824 if (!field->is_constant()) {
4825 return NULL; // Field not marked as constant.
4826 }
4827 ciInstance* holder = NULL;
4828 if (!field->is_static()) {
4829 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4830 if (const_oop != NULL && const_oop->is_instance()) {
4831 holder = const_oop->as_instance();
4832 }
4833 }
4834 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4835 /*is_unsigned_load=*/false);
4836 if (con_type != NULL) {
4837 return makecon(con_type);
4838 }
4839 return NULL;
4840 }
4841
4842 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4843 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4844 // assumption of CCP analysis.
4845 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4846 }
4847
4848 Node* GraphKit::shenandoah_read_barrier(Node* obj) {
4849 if (UseShenandoahGC && ShenandoahReadBarrier) {
4850 return shenandoah_read_barrier_impl(obj, false, true, true);
4851 } else {
4852 return obj;
4853 }
4854 }
4855
4856 Node* GraphKit::shenandoah_storeval_barrier(Node* obj) {
4857 if (UseShenandoahGC) {
4858 if (ShenandoahStoreValEnqueueBarrier) {
4859 obj = shenandoah_write_barrier(obj);
4860 shenandoah_enqueue_barrier(obj);
4861 }
4862 if (ShenandoahStoreValReadBarrier) {
4863 obj = shenandoah_read_barrier_impl(obj, true, false, false);
4864 }
4865 }
4866 return obj;
4867 }
4868
4869 Node* GraphKit::shenandoah_read_barrier_acmp(Node* obj) {
4870 return shenandoah_read_barrier_impl(obj, true, true, false);
4871 }
4872
4873 Node* GraphKit::shenandoah_read_barrier_impl(Node* obj, bool use_ctrl, bool use_mem, bool allow_fromspace) {
4874
4875 const Type* obj_type = obj->bottom_type();
4876 if (obj_type->higher_equal(TypePtr::NULL_PTR)) {
4877 return obj;
4878 }
4879 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type);
4880 Node* mem = use_mem ? memory(adr_type) : immutable_memory();
4881
4882 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, mem, allow_fromspace)) {
4883 // We know it is null, no barrier needed.
4884 return obj;
4885 }
4886
4887
4888 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) {
4889
4890 // We don't know if it's null or not. Need null-check.
4891 enum { _not_null_path = 1, _null_path, PATH_LIMIT };
4892 RegionNode* region = new RegionNode(PATH_LIMIT);
4893 Node* phi = new PhiNode(region, obj_type);
4894 Node* null_ctrl = top();
4895 Node* not_null_obj = null_check_oop(obj, &null_ctrl);
4896
4897 region->init_req(_null_path, null_ctrl);
4898 phi ->init_req(_null_path, zerocon(T_OBJECT));
4899
4900 Node* ctrl = use_ctrl ? control() : NULL;
4901 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, not_null_obj, allow_fromspace);
4902 Node* n = _gvn.transform(rb);
4903
4904 region->init_req(_not_null_path, control());
4905 phi ->init_req(_not_null_path, n);
4906
4907 set_control(_gvn.transform(region));
4908 record_for_igvn(region);
4909 return _gvn.transform(phi);
4910
4911 } else {
4912 // We know it is not null. Simple barrier is sufficient.
4913 Node* ctrl = use_ctrl ? control() : NULL;
4914 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, obj, allow_fromspace);
4915 Node* n = _gvn.transform(rb);
4916 record_for_igvn(n);
4917 return n;
4918 }
4919 }
4920
4921 Node* GraphKit::shenandoah_write_barrier_helper(GraphKit& kit, Node* obj, const TypePtr* adr_type) {
4922 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(kit.C, kit.control(), kit.memory(adr_type), obj);
4923 Node* n = kit.gvn().transform(wb);
4924 if (n == wb) { // New barrier needs memory projection.
4925 Node* proj = kit.gvn().transform(new ShenandoahWBMemProjNode(n));
4926 kit.set_memory(proj, adr_type);
4927 }
4928
4929 return n;
4930 }
4931
4932 Node* GraphKit::shenandoah_write_barrier(Node* obj) {
4933
4934 if (UseShenandoahGC && ShenandoahWriteBarrier) {
4935 return shenandoah_write_barrier_impl(obj);
4936 } else {
4937 return obj;
4938 }
4939 }
4940
4941 Node* GraphKit::shenandoah_write_barrier_impl(Node* obj) {
4942 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, NULL, true)) {
4943 return obj;
4944 }
4945 const Type* obj_type = obj->bottom_type();
4946 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type);
4947 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) {
4948 // We don't know if it's null or not. Need null-check.
4949 enum { _not_null_path = 1, _null_path, PATH_LIMIT };
4950 RegionNode* region = new RegionNode(PATH_LIMIT);
4951 Node* phi = new PhiNode(region, obj_type);
4952 Node* memphi = PhiNode::make(region, memory(adr_type), Type::MEMORY, C->alias_type(adr_type)->adr_type());
4953
4954 Node* prev_mem = memory(adr_type);
4955 Node* null_ctrl = top();
4956 Node* not_null_obj = null_check_oop(obj, &null_ctrl);
4957
4958 region->init_req(_null_path, null_ctrl);
4959 phi ->init_req(_null_path, zerocon(T_OBJECT));
4960 memphi->init_req(_null_path, prev_mem);
4961
4962 Node* n = shenandoah_write_barrier_helper(*this, not_null_obj, adr_type);
4963
4964 region->init_req(_not_null_path, control());
4965 phi ->init_req(_not_null_path, n);
4966 memphi->init_req(_not_null_path, memory(adr_type));
4967
4968 set_control(_gvn.transform(region));
4969 record_for_igvn(region);
4970 set_memory(_gvn.transform(memphi), adr_type);
4971
4972 Node* res_val = _gvn.transform(phi);
4973 // replace_in_map(obj, res_val);
4974 return res_val;
4975 } else {
4976 // We know it is not null. Simple barrier is sufficient.
4977 Node* n = shenandoah_write_barrier_helper(*this, obj, adr_type);
4978 // replace_in_map(obj, n);
4979 record_for_igvn(n);
4980 return n;
4981 }
4982 }