1 /*
2 * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc/g1/heapRegion.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/cardTableModRefBS.hpp"
31 #include "gc/shared/collectedHeap.hpp"
32 #include "gc/shenandoah/brooksPointer.hpp"
33 #include "gc/shenandoah/shenandoahConnectionMatrix.hpp"
34 #include "gc/shenandoah/shenandoahHeap.hpp"
35 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/intrinsicnode.hpp"
43 #include "opto/locknode.hpp"
44 #include "opto/machnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/shenandoahSupport.hpp"
50 #include "runtime/deoptimization.hpp"
51 #include "runtime/sharedRuntime.hpp"
52
53 //----------------------------GraphKit-----------------------------------------
54 // Main utility constructor.
55 GraphKit::GraphKit(JVMState* jvms)
56 : Phase(Phase::Parser),
57 _env(C->env()),
58 _gvn(*C->initial_gvn())
59 {
60 _exceptions = jvms->map()->next_exception();
61 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
62 set_jvms(jvms);
63 }
64
65 // Private constructor for parser.
66 GraphKit::GraphKit()
67 : Phase(Phase::Parser),
68 _env(C->env()),
69 _gvn(*C->initial_gvn())
70 {
71 _exceptions = NULL;
72 set_map(NULL);
73 debug_only(_sp = -99);
74 debug_only(set_bci(-99));
75 }
76
77
78
79 //---------------------------clean_stack---------------------------------------
80 // Clear away rubbish from the stack area of the JVM state.
81 // This destroys any arguments that may be waiting on the stack.
82 void GraphKit::clean_stack(int from_sp) {
83 SafePointNode* map = this->map();
84 JVMState* jvms = this->jvms();
85 int stk_size = jvms->stk_size();
86 int stkoff = jvms->stkoff();
87 Node* top = this->top();
88 for (int i = from_sp; i < stk_size; i++) {
89 if (map->in(stkoff + i) != top) {
90 map->set_req(stkoff + i, top);
91 }
92 }
93 }
94
95
96 //--------------------------------sync_jvms-----------------------------------
97 // Make sure our current jvms agrees with our parse state.
98 JVMState* GraphKit::sync_jvms() const {
99 JVMState* jvms = this->jvms();
100 jvms->set_bci(bci()); // Record the new bci in the JVMState
101 jvms->set_sp(sp()); // Record the new sp in the JVMState
102 assert(jvms_in_sync(), "jvms is now in sync");
103 return jvms;
104 }
105
106 //--------------------------------sync_jvms_for_reexecute---------------------
107 // Make sure our current jvms agrees with our parse state. This version
108 // uses the reexecute_sp for reexecuting bytecodes.
109 JVMState* GraphKit::sync_jvms_for_reexecute() {
110 JVMState* jvms = this->jvms();
111 jvms->set_bci(bci()); // Record the new bci in the JVMState
112 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
113 return jvms;
114 }
115
116 #ifdef ASSERT
117 bool GraphKit::jvms_in_sync() const {
118 Parse* parse = is_Parse();
119 if (parse == NULL) {
120 if (bci() != jvms()->bci()) return false;
121 if (sp() != (int)jvms()->sp()) return false;
122 return true;
123 }
124 if (jvms()->method() != parse->method()) return false;
125 if (jvms()->bci() != parse->bci()) return false;
126 int jvms_sp = jvms()->sp();
127 if (jvms_sp != parse->sp()) return false;
128 int jvms_depth = jvms()->depth();
129 if (jvms_depth != parse->depth()) return false;
130 return true;
131 }
132
133 // Local helper checks for special internal merge points
134 // used to accumulate and merge exception states.
135 // They are marked by the region's in(0) edge being the map itself.
136 // Such merge points must never "escape" into the parser at large,
137 // until they have been handed to gvn.transform.
138 static bool is_hidden_merge(Node* reg) {
139 if (reg == NULL) return false;
140 if (reg->is_Phi()) {
141 reg = reg->in(0);
142 if (reg == NULL) return false;
143 }
144 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
145 }
146
147 void GraphKit::verify_map() const {
148 if (map() == NULL) return; // null map is OK
149 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
150 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
151 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
152 }
153
154 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
155 assert(ex_map->next_exception() == NULL, "not already part of a chain");
156 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
157 }
158 #endif
159
160 //---------------------------stop_and_kill_map---------------------------------
161 // Set _map to NULL, signalling a stop to further bytecode execution.
162 // First smash the current map's control to a constant, to mark it dead.
163 void GraphKit::stop_and_kill_map() {
164 SafePointNode* dead_map = stop();
165 if (dead_map != NULL) {
166 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
167 assert(dead_map->is_killed(), "must be so marked");
168 }
169 }
170
171
172 //--------------------------------stopped--------------------------------------
173 // Tell if _map is NULL, or control is top.
174 bool GraphKit::stopped() {
175 if (map() == NULL) return true;
176 else if (control() == top()) return true;
177 else return false;
178 }
179
180
181 //-----------------------------has_ex_handler----------------------------------
182 // Tell if this method or any caller method has exception handlers.
183 bool GraphKit::has_ex_handler() {
184 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
185 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
186 return true;
187 }
188 }
189 return false;
190 }
191
192 //------------------------------save_ex_oop------------------------------------
193 // Save an exception without blowing stack contents or other JVM state.
194 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
195 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
196 ex_map->add_req(ex_oop);
197 debug_only(verify_exception_state(ex_map));
198 }
199
200 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
201 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
202 Node* ex_oop = ex_map->in(ex_map->req()-1);
203 if (clear_it) ex_map->del_req(ex_map->req()-1);
204 return ex_oop;
205 }
206
207 //-----------------------------saved_ex_oop------------------------------------
208 // Recover a saved exception from its map.
209 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
210 return common_saved_ex_oop(ex_map, false);
211 }
212
213 //--------------------------clear_saved_ex_oop---------------------------------
214 // Erase a previously saved exception from its map.
215 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
216 return common_saved_ex_oop(ex_map, true);
217 }
218
219 #ifdef ASSERT
220 //---------------------------has_saved_ex_oop----------------------------------
221 // Erase a previously saved exception from its map.
222 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
223 return ex_map->req() == ex_map->jvms()->endoff()+1;
224 }
225 #endif
226
227 //-------------------------make_exception_state--------------------------------
228 // Turn the current JVM state into an exception state, appending the ex_oop.
229 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
230 sync_jvms();
231 SafePointNode* ex_map = stop(); // do not manipulate this map any more
232 set_saved_ex_oop(ex_map, ex_oop);
233 return ex_map;
234 }
235
236
237 //--------------------------add_exception_state--------------------------------
238 // Add an exception to my list of exceptions.
239 void GraphKit::add_exception_state(SafePointNode* ex_map) {
240 if (ex_map == NULL || ex_map->control() == top()) {
241 return;
242 }
243 #ifdef ASSERT
244 verify_exception_state(ex_map);
245 if (has_exceptions()) {
246 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
247 }
248 #endif
249
250 // If there is already an exception of exactly this type, merge with it.
251 // In particular, null-checks and other low-level exceptions common up here.
252 Node* ex_oop = saved_ex_oop(ex_map);
253 const Type* ex_type = _gvn.type(ex_oop);
254 if (ex_oop == top()) {
255 // No action needed.
256 return;
257 }
258 assert(ex_type->isa_instptr(), "exception must be an instance");
259 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
260 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
261 // We check sp also because call bytecodes can generate exceptions
262 // both before and after arguments are popped!
263 if (ex_type2 == ex_type
264 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
265 combine_exception_states(ex_map, e2);
266 return;
267 }
268 }
269
270 // No pre-existing exception of the same type. Chain it on the list.
271 push_exception_state(ex_map);
272 }
273
274 //-----------------------add_exception_states_from-----------------------------
275 void GraphKit::add_exception_states_from(JVMState* jvms) {
276 SafePointNode* ex_map = jvms->map()->next_exception();
277 if (ex_map != NULL) {
278 jvms->map()->set_next_exception(NULL);
279 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
280 next_map = ex_map->next_exception();
281 ex_map->set_next_exception(NULL);
282 add_exception_state(ex_map);
283 }
284 }
285 }
286
287 //-----------------------transfer_exceptions_into_jvms-------------------------
288 JVMState* GraphKit::transfer_exceptions_into_jvms() {
289 if (map() == NULL) {
290 // We need a JVMS to carry the exceptions, but the map has gone away.
291 // Create a scratch JVMS, cloned from any of the exception states...
292 if (has_exceptions()) {
293 _map = _exceptions;
294 _map = clone_map();
295 _map->set_next_exception(NULL);
296 clear_saved_ex_oop(_map);
297 debug_only(verify_map());
298 } else {
299 // ...or created from scratch
300 JVMState* jvms = new (C) JVMState(_method, NULL);
301 jvms->set_bci(_bci);
302 jvms->set_sp(_sp);
303 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
304 set_jvms(jvms);
305 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
306 set_all_memory(top());
307 while (map()->req() < jvms->endoff()) map()->add_req(top());
308 }
309 // (This is a kludge, in case you didn't notice.)
310 set_control(top());
311 }
312 JVMState* jvms = sync_jvms();
313 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
314 jvms->map()->set_next_exception(_exceptions);
315 _exceptions = NULL; // done with this set of exceptions
316 return jvms;
317 }
318
319 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
320 assert(is_hidden_merge(dstphi), "must be a special merge node");
321 assert(is_hidden_merge(srcphi), "must be a special merge node");
322 uint limit = srcphi->req();
323 for (uint i = PhiNode::Input; i < limit; i++) {
324 dstphi->add_req(srcphi->in(i));
325 }
326 }
327 static inline void add_one_req(Node* dstphi, Node* src) {
328 assert(is_hidden_merge(dstphi), "must be a special merge node");
329 assert(!is_hidden_merge(src), "must not be a special merge node");
330 dstphi->add_req(src);
331 }
332
333 //-----------------------combine_exception_states------------------------------
334 // This helper function combines exception states by building phis on a
335 // specially marked state-merging region. These regions and phis are
336 // untransformed, and can build up gradually. The region is marked by
337 // having a control input of its exception map, rather than NULL. Such
338 // regions do not appear except in this function, and in use_exception_state.
339 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
340 if (failing()) return; // dying anyway...
341 JVMState* ex_jvms = ex_map->_jvms;
342 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
343 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
344 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
345 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
346 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
347 assert(ex_map->req() == phi_map->req(), "matching maps");
348 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
349 Node* hidden_merge_mark = root();
350 Node* region = phi_map->control();
351 MergeMemNode* phi_mem = phi_map->merged_memory();
352 MergeMemNode* ex_mem = ex_map->merged_memory();
353 if (region->in(0) != hidden_merge_mark) {
354 // The control input is not (yet) a specially-marked region in phi_map.
355 // Make it so, and build some phis.
356 region = new RegionNode(2);
357 _gvn.set_type(region, Type::CONTROL);
358 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
359 region->init_req(1, phi_map->control());
360 phi_map->set_control(region);
361 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
362 record_for_igvn(io_phi);
363 _gvn.set_type(io_phi, Type::ABIO);
364 phi_map->set_i_o(io_phi);
365 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
366 Node* m = mms.memory();
367 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
368 record_for_igvn(m_phi);
369 _gvn.set_type(m_phi, Type::MEMORY);
370 mms.set_memory(m_phi);
371 }
372 }
373
374 // Either or both of phi_map and ex_map might already be converted into phis.
375 Node* ex_control = ex_map->control();
376 // if there is special marking on ex_map also, we add multiple edges from src
377 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
378 // how wide was the destination phi_map, originally?
379 uint orig_width = region->req();
380
381 if (add_multiple) {
382 add_n_reqs(region, ex_control);
383 add_n_reqs(phi_map->i_o(), ex_map->i_o());
384 } else {
385 // ex_map has no merges, so we just add single edges everywhere
386 add_one_req(region, ex_control);
387 add_one_req(phi_map->i_o(), ex_map->i_o());
388 }
389 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
390 if (mms.is_empty()) {
391 // get a copy of the base memory, and patch some inputs into it
392 const TypePtr* adr_type = mms.adr_type(C);
393 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
394 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
395 mms.set_memory(phi);
396 // Prepare to append interesting stuff onto the newly sliced phi:
397 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
398 }
399 // Append stuff from ex_map:
400 if (add_multiple) {
401 add_n_reqs(mms.memory(), mms.memory2());
402 } else {
403 add_one_req(mms.memory(), mms.memory2());
404 }
405 }
406 uint limit = ex_map->req();
407 for (uint i = TypeFunc::Parms; i < limit; i++) {
408 // Skip everything in the JVMS after tos. (The ex_oop follows.)
409 if (i == tos) i = ex_jvms->monoff();
410 Node* src = ex_map->in(i);
411 Node* dst = phi_map->in(i);
412 if (src != dst) {
413 PhiNode* phi;
414 if (dst->in(0) != region) {
415 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
416 record_for_igvn(phi);
417 _gvn.set_type(phi, phi->type());
418 phi_map->set_req(i, dst);
419 // Prepare to append interesting stuff onto the new phi:
420 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
421 } else {
422 assert(dst->is_Phi(), "nobody else uses a hidden region");
423 phi = dst->as_Phi();
424 }
425 if (add_multiple && src->in(0) == ex_control) {
426 // Both are phis.
427 add_n_reqs(dst, src);
428 } else {
429 while (dst->req() < region->req()) add_one_req(dst, src);
430 }
431 const Type* srctype = _gvn.type(src);
432 if (phi->type() != srctype) {
433 const Type* dsttype = phi->type()->meet_speculative(srctype);
434 if (phi->type() != dsttype) {
435 phi->set_type(dsttype);
436 _gvn.set_type(phi, dsttype);
437 }
438 }
439 }
440 }
441 phi_map->merge_replaced_nodes_with(ex_map);
442 }
443
444 //--------------------------use_exception_state--------------------------------
445 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
446 if (failing()) { stop(); return top(); }
447 Node* region = phi_map->control();
448 Node* hidden_merge_mark = root();
449 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
450 Node* ex_oop = clear_saved_ex_oop(phi_map);
451 if (region->in(0) == hidden_merge_mark) {
452 // Special marking for internal ex-states. Process the phis now.
453 region->set_req(0, region); // now it's an ordinary region
454 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
455 // Note: Setting the jvms also sets the bci and sp.
456 set_control(_gvn.transform(region));
457 uint tos = jvms()->stkoff() + sp();
458 for (uint i = 1; i < tos; i++) {
459 Node* x = phi_map->in(i);
460 if (x->in(0) == region) {
461 assert(x->is_Phi(), "expected a special phi");
462 phi_map->set_req(i, _gvn.transform(x));
463 }
464 }
465 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
466 Node* x = mms.memory();
467 if (x->in(0) == region) {
468 assert(x->is_Phi(), "nobody else uses a hidden region");
469 mms.set_memory(_gvn.transform(x));
470 }
471 }
472 if (ex_oop->in(0) == region) {
473 assert(ex_oop->is_Phi(), "expected a special phi");
474 ex_oop = _gvn.transform(ex_oop);
475 }
476 } else {
477 set_jvms(phi_map->jvms());
478 }
479
480 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
481 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
482 return ex_oop;
483 }
484
485 //---------------------------------java_bc-------------------------------------
486 Bytecodes::Code GraphKit::java_bc() const {
487 ciMethod* method = this->method();
488 int bci = this->bci();
489 if (method != NULL && bci != InvocationEntryBci)
490 return method->java_code_at_bci(bci);
491 else
492 return Bytecodes::_illegal;
493 }
494
495 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
496 bool must_throw) {
497 // if the exception capability is set, then we will generate code
498 // to check the JavaThread.should_post_on_exceptions flag to see
499 // if we actually need to report exception events (for this
500 // thread). If we don't need to report exception events, we will
501 // take the normal fast path provided by add_exception_events. If
502 // exception event reporting is enabled for this thread, we will
503 // take the uncommon_trap in the BuildCutout below.
504
505 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
506 Node* jthread = _gvn.transform(new ThreadLocalNode());
507 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
508 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
509
510 // Test the should_post_on_exceptions_flag vs. 0
511 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
512 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
513
514 // Branch to slow_path if should_post_on_exceptions_flag was true
515 { BuildCutout unless(this, tst, PROB_MAX);
516 // Do not try anything fancy if we're notifying the VM on every throw.
517 // Cf. case Bytecodes::_athrow in parse2.cpp.
518 uncommon_trap(reason, Deoptimization::Action_none,
519 (ciKlass*)NULL, (char*)NULL, must_throw);
520 }
521
522 }
523
524 //------------------------------builtin_throw----------------------------------
525 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
526 bool must_throw = true;
527
528 if (env()->jvmti_can_post_on_exceptions()) {
529 // check if we must post exception events, take uncommon trap if so
530 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
531 // here if should_post_on_exceptions is false
532 // continue on with the normal codegen
533 }
534
535 // If this particular condition has not yet happened at this
536 // bytecode, then use the uncommon trap mechanism, and allow for
537 // a future recompilation if several traps occur here.
538 // If the throw is hot, try to use a more complicated inline mechanism
539 // which keeps execution inside the compiled code.
540 bool treat_throw_as_hot = false;
541 ciMethodData* md = method()->method_data();
542
543 if (ProfileTraps) {
544 if (too_many_traps(reason)) {
545 treat_throw_as_hot = true;
546 }
547 // (If there is no MDO at all, assume it is early in
548 // execution, and that any deopts are part of the
549 // startup transient, and don't need to be remembered.)
550
551 // Also, if there is a local exception handler, treat all throws
552 // as hot if there has been at least one in this method.
553 if (C->trap_count(reason) != 0
554 && method()->method_data()->trap_count(reason) != 0
555 && has_ex_handler()) {
556 treat_throw_as_hot = true;
557 }
558 }
559
560 // If this throw happens frequently, an uncommon trap might cause
561 // a performance pothole. If there is a local exception handler,
562 // and if this particular bytecode appears to be deoptimizing often,
563 // let us handle the throw inline, with a preconstructed instance.
564 // Note: If the deopt count has blown up, the uncommon trap
565 // runtime is going to flush this nmethod, not matter what.
566 if (treat_throw_as_hot
567 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
568 // If the throw is local, we use a pre-existing instance and
569 // punt on the backtrace. This would lead to a missing backtrace
570 // (a repeat of 4292742) if the backtrace object is ever asked
571 // for its backtrace.
572 // Fixing this remaining case of 4292742 requires some flavor of
573 // escape analysis. Leave that for the future.
574 ciInstance* ex_obj = NULL;
575 switch (reason) {
576 case Deoptimization::Reason_null_check:
577 ex_obj = env()->NullPointerException_instance();
578 break;
579 case Deoptimization::Reason_div0_check:
580 ex_obj = env()->ArithmeticException_instance();
581 break;
582 case Deoptimization::Reason_range_check:
583 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
584 break;
585 case Deoptimization::Reason_class_check:
586 if (java_bc() == Bytecodes::_aastore) {
587 ex_obj = env()->ArrayStoreException_instance();
588 } else {
589 ex_obj = env()->ClassCastException_instance();
590 }
591 break;
592 default:
593 break;
594 }
595 if (failing()) { stop(); return; } // exception allocation might fail
596 if (ex_obj != NULL) {
597 // Cheat with a preallocated exception object.
598 if (C->log() != NULL)
599 C->log()->elem("hot_throw preallocated='1' reason='%s'",
600 Deoptimization::trap_reason_name(reason));
601 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
602 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
603
604 ex_node = shenandoah_write_barrier(ex_node);
605
606 // Clear the detail message of the preallocated exception object.
607 // Weblogic sometimes mutates the detail message of exceptions
608 // using reflection.
609 int offset = java_lang_Throwable::get_detailMessage_offset();
610 const TypePtr* adr_typ = ex_con->add_offset(offset);
611
612 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
613 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
614 // Conservatively release stores of object references.
615 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
616
617 add_exception_state(make_exception_state(ex_node));
618 return;
619 }
620 }
621
622 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
623 // It won't be much cheaper than bailing to the interp., since we'll
624 // have to pass up all the debug-info, and the runtime will have to
625 // create the stack trace.
626
627 // Usual case: Bail to interpreter.
628 // Reserve the right to recompile if we haven't seen anything yet.
629
630 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
631 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
632 if (treat_throw_as_hot
633 && (method()->method_data()->trap_recompiled_at(bci(), m)
634 || C->too_many_traps(reason))) {
635 // We cannot afford to take more traps here. Suffer in the interpreter.
636 if (C->log() != NULL)
637 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
638 Deoptimization::trap_reason_name(reason),
639 C->trap_count(reason));
640 action = Deoptimization::Action_none;
641 }
642
643 // "must_throw" prunes the JVM state to include only the stack, if there
644 // are no local exception handlers. This should cut down on register
645 // allocation time and code size, by drastically reducing the number
646 // of in-edges on the call to the uncommon trap.
647
648 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
649 }
650
651
652 //----------------------------PreserveJVMState---------------------------------
653 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
654 debug_only(kit->verify_map());
655 _kit = kit;
656 _map = kit->map(); // preserve the map
657 _sp = kit->sp();
658 kit->set_map(clone_map ? kit->clone_map() : NULL);
659 #ifdef ASSERT
660 _bci = kit->bci();
661 Parse* parser = kit->is_Parse();
662 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
663 _block = block;
664 #endif
665 }
666 PreserveJVMState::~PreserveJVMState() {
667 GraphKit* kit = _kit;
668 #ifdef ASSERT
669 assert(kit->bci() == _bci, "bci must not shift");
670 Parse* parser = kit->is_Parse();
671 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
672 assert(block == _block, "block must not shift");
673 #endif
674 kit->set_map(_map);
675 kit->set_sp(_sp);
676 }
677
678
679 //-----------------------------BuildCutout-------------------------------------
680 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
681 : PreserveJVMState(kit)
682 {
683 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
684 SafePointNode* outer_map = _map; // preserved map is caller's
685 SafePointNode* inner_map = kit->map();
686 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
687 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
688 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
689 }
690 BuildCutout::~BuildCutout() {
691 GraphKit* kit = _kit;
692 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
693 }
694
695 //---------------------------PreserveReexecuteState----------------------------
696 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
697 assert(!kit->stopped(), "must call stopped() before");
698 _kit = kit;
699 _sp = kit->sp();
700 _reexecute = kit->jvms()->_reexecute;
701 }
702 PreserveReexecuteState::~PreserveReexecuteState() {
703 if (_kit->stopped()) return;
704 _kit->jvms()->_reexecute = _reexecute;
705 _kit->set_sp(_sp);
706 }
707
708 //------------------------------clone_map--------------------------------------
709 // Implementation of PreserveJVMState
710 //
711 // Only clone_map(...) here. If this function is only used in the
712 // PreserveJVMState class we may want to get rid of this extra
713 // function eventually and do it all there.
714
715 SafePointNode* GraphKit::clone_map() {
716 if (map() == NULL) return NULL;
717
718 // Clone the memory edge first
719 Node* mem = MergeMemNode::make(map()->memory());
720 gvn().set_type_bottom(mem);
721
722 SafePointNode *clonemap = (SafePointNode*)map()->clone();
723 JVMState* jvms = this->jvms();
724 JVMState* clonejvms = jvms->clone_shallow(C);
725 clonemap->set_memory(mem);
726 clonemap->set_jvms(clonejvms);
727 clonejvms->set_map(clonemap);
728 record_for_igvn(clonemap);
729 gvn().set_type_bottom(clonemap);
730 return clonemap;
731 }
732
733
734 //-----------------------------set_map_clone-----------------------------------
735 void GraphKit::set_map_clone(SafePointNode* m) {
736 _map = m;
737 _map = clone_map();
738 _map->set_next_exception(NULL);
739 debug_only(verify_map());
740 }
741
742
743 //----------------------------kill_dead_locals---------------------------------
744 // Detect any locals which are known to be dead, and force them to top.
745 void GraphKit::kill_dead_locals() {
746 // Consult the liveness information for the locals. If any
747 // of them are unused, then they can be replaced by top(). This
748 // should help register allocation time and cut down on the size
749 // of the deoptimization information.
750
751 // This call is made from many of the bytecode handling
752 // subroutines called from the Big Switch in do_one_bytecode.
753 // Every bytecode which might include a slow path is responsible
754 // for killing its dead locals. The more consistent we
755 // are about killing deads, the fewer useless phis will be
756 // constructed for them at various merge points.
757
758 // bci can be -1 (InvocationEntryBci). We return the entry
759 // liveness for the method.
760
761 if (method() == NULL || method()->code_size() == 0) {
762 // We are building a graph for a call to a native method.
763 // All locals are live.
764 return;
765 }
766
767 ResourceMark rm;
768
769 // Consult the liveness information for the locals. If any
770 // of them are unused, then they can be replaced by top(). This
771 // should help register allocation time and cut down on the size
772 // of the deoptimization information.
773 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
774
775 int len = (int)live_locals.size();
776 assert(len <= jvms()->loc_size(), "too many live locals");
777 for (int local = 0; local < len; local++) {
778 if (!live_locals.at(local)) {
779 set_local(local, top());
780 }
781 }
782 }
783
784 #ifdef ASSERT
785 //-------------------------dead_locals_are_killed------------------------------
786 // Return true if all dead locals are set to top in the map.
787 // Used to assert "clean" debug info at various points.
788 bool GraphKit::dead_locals_are_killed() {
789 if (method() == NULL || method()->code_size() == 0) {
790 // No locals need to be dead, so all is as it should be.
791 return true;
792 }
793
794 // Make sure somebody called kill_dead_locals upstream.
795 ResourceMark rm;
796 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
797 if (jvms->loc_size() == 0) continue; // no locals to consult
798 SafePointNode* map = jvms->map();
799 ciMethod* method = jvms->method();
800 int bci = jvms->bci();
801 if (jvms == this->jvms()) {
802 bci = this->bci(); // it might not yet be synched
803 }
804 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
805 int len = (int)live_locals.size();
806 if (!live_locals.is_valid() || len == 0)
807 // This method is trivial, or is poisoned by a breakpoint.
808 return true;
809 assert(len == jvms->loc_size(), "live map consistent with locals map");
810 for (int local = 0; local < len; local++) {
811 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
812 if (PrintMiscellaneous && (Verbose || WizardMode)) {
813 tty->print_cr("Zombie local %d: ", local);
814 jvms->dump();
815 }
816 return false;
817 }
818 }
819 }
820 return true;
821 }
822
823 #endif //ASSERT
824
825 // Helper function for enforcing certain bytecodes to reexecute if
826 // deoptimization happens
827 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
828 ciMethod* cur_method = jvms->method();
829 int cur_bci = jvms->bci();
830 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
831 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
832 return Interpreter::bytecode_should_reexecute(code) ||
833 (is_anewarray && code == Bytecodes::_multianewarray);
834 // Reexecute _multianewarray bytecode which was replaced with
835 // sequence of [a]newarray. See Parse::do_multianewarray().
836 //
837 // Note: interpreter should not have it set since this optimization
838 // is limited by dimensions and guarded by flag so in some cases
839 // multianewarray() runtime calls will be generated and
840 // the bytecode should not be reexecutes (stack will not be reset).
841 } else
842 return false;
843 }
844
845 // Helper function for adding JVMState and debug information to node
846 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
847 // Add the safepoint edges to the call (or other safepoint).
848
849 // Make sure dead locals are set to top. This
850 // should help register allocation time and cut down on the size
851 // of the deoptimization information.
852 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
853
854 // Walk the inline list to fill in the correct set of JVMState's
855 // Also fill in the associated edges for each JVMState.
856
857 // If the bytecode needs to be reexecuted we need to put
858 // the arguments back on the stack.
859 const bool should_reexecute = jvms()->should_reexecute();
860 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
861
862 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
863 // undefined if the bci is different. This is normal for Parse but it
864 // should not happen for LibraryCallKit because only one bci is processed.
865 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
866 "in LibraryCallKit the reexecute bit should not change");
867
868 // If we are guaranteed to throw, we can prune everything but the
869 // input to the current bytecode.
870 bool can_prune_locals = false;
871 uint stack_slots_not_pruned = 0;
872 int inputs = 0, depth = 0;
873 if (must_throw) {
874 assert(method() == youngest_jvms->method(), "sanity");
875 if (compute_stack_effects(inputs, depth)) {
876 can_prune_locals = true;
877 stack_slots_not_pruned = inputs;
878 }
879 }
880
881 if (env()->should_retain_local_variables()) {
882 // At any safepoint, this method can get breakpointed, which would
883 // then require an immediate deoptimization.
884 can_prune_locals = false; // do not prune locals
885 stack_slots_not_pruned = 0;
886 }
887
888 // do not scribble on the input jvms
889 JVMState* out_jvms = youngest_jvms->clone_deep(C);
890 call->set_jvms(out_jvms); // Start jvms list for call node
891
892 // For a known set of bytecodes, the interpreter should reexecute them if
893 // deoptimization happens. We set the reexecute state for them here
894 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
895 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
896 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
897 }
898
899 // Presize the call:
900 DEBUG_ONLY(uint non_debug_edges = call->req());
901 call->add_req_batch(top(), youngest_jvms->debug_depth());
902 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
903
904 // Set up edges so that the call looks like this:
905 // Call [state:] ctl io mem fptr retadr
906 // [parms:] parm0 ... parmN
907 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
908 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
909 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
910 // Note that caller debug info precedes callee debug info.
911
912 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
913 uint debug_ptr = call->req();
914
915 // Loop over the map input edges associated with jvms, add them
916 // to the call node, & reset all offsets to match call node array.
917 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
918 uint debug_end = debug_ptr;
919 uint debug_start = debug_ptr - in_jvms->debug_size();
920 debug_ptr = debug_start; // back up the ptr
921
922 uint p = debug_start; // walks forward in [debug_start, debug_end)
923 uint j, k, l;
924 SafePointNode* in_map = in_jvms->map();
925 out_jvms->set_map(call);
926
927 if (can_prune_locals) {
928 assert(in_jvms->method() == out_jvms->method(), "sanity");
929 // If the current throw can reach an exception handler in this JVMS,
930 // then we must keep everything live that can reach that handler.
931 // As a quick and dirty approximation, we look for any handlers at all.
932 if (in_jvms->method()->has_exception_handlers()) {
933 can_prune_locals = false;
934 }
935 }
936
937 // Add the Locals
938 k = in_jvms->locoff();
939 l = in_jvms->loc_size();
940 out_jvms->set_locoff(p);
941 if (!can_prune_locals) {
942 for (j = 0; j < l; j++)
943 call->set_req(p++, in_map->in(k+j));
944 } else {
945 p += l; // already set to top above by add_req_batch
946 }
947
948 // Add the Expression Stack
949 k = in_jvms->stkoff();
950 l = in_jvms->sp();
951 out_jvms->set_stkoff(p);
952 if (!can_prune_locals) {
953 for (j = 0; j < l; j++)
954 call->set_req(p++, in_map->in(k+j));
955 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
956 // Divide stack into {S0,...,S1}, where S0 is set to top.
957 uint s1 = stack_slots_not_pruned;
958 stack_slots_not_pruned = 0; // for next iteration
959 if (s1 > l) s1 = l;
960 uint s0 = l - s1;
961 p += s0; // skip the tops preinstalled by add_req_batch
962 for (j = s0; j < l; j++)
963 call->set_req(p++, in_map->in(k+j));
964 } else {
965 p += l; // already set to top above by add_req_batch
966 }
967
968 // Add the Monitors
969 k = in_jvms->monoff();
970 l = in_jvms->mon_size();
971 out_jvms->set_monoff(p);
972 for (j = 0; j < l; j++)
973 call->set_req(p++, in_map->in(k+j));
974
975 // Copy any scalar object fields.
976 k = in_jvms->scloff();
977 l = in_jvms->scl_size();
978 out_jvms->set_scloff(p);
979 for (j = 0; j < l; j++)
980 call->set_req(p++, in_map->in(k+j));
981
982 // Finish the new jvms.
983 out_jvms->set_endoff(p);
984
985 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
986 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
987 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
988 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
989 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
990 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
991
992 // Update the two tail pointers in parallel.
993 out_jvms = out_jvms->caller();
994 in_jvms = in_jvms->caller();
995 }
996
997 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
998
999 // Test the correctness of JVMState::debug_xxx accessors:
1000 assert(call->jvms()->debug_start() == non_debug_edges, "");
1001 assert(call->jvms()->debug_end() == call->req(), "");
1002 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1003 }
1004
1005 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1006 Bytecodes::Code code = java_bc();
1007 if (code == Bytecodes::_wide) {
1008 code = method()->java_code_at_bci(bci() + 1);
1009 }
1010
1011 BasicType rtype = T_ILLEGAL;
1012 int rsize = 0;
1013
1014 if (code != Bytecodes::_illegal) {
1015 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1016 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1017 if (rtype < T_CONFLICT)
1018 rsize = type2size[rtype];
1019 }
1020
1021 switch (code) {
1022 case Bytecodes::_illegal:
1023 return false;
1024
1025 case Bytecodes::_ldc:
1026 case Bytecodes::_ldc_w:
1027 case Bytecodes::_ldc2_w:
1028 inputs = 0;
1029 break;
1030
1031 case Bytecodes::_dup: inputs = 1; break;
1032 case Bytecodes::_dup_x1: inputs = 2; break;
1033 case Bytecodes::_dup_x2: inputs = 3; break;
1034 case Bytecodes::_dup2: inputs = 2; break;
1035 case Bytecodes::_dup2_x1: inputs = 3; break;
1036 case Bytecodes::_dup2_x2: inputs = 4; break;
1037 case Bytecodes::_swap: inputs = 2; break;
1038 case Bytecodes::_arraylength: inputs = 1; break;
1039
1040 case Bytecodes::_getstatic:
1041 case Bytecodes::_putstatic:
1042 case Bytecodes::_getfield:
1043 case Bytecodes::_putfield:
1044 {
1045 bool ignored_will_link;
1046 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1047 int size = field->type()->size();
1048 bool is_get = (depth >= 0), is_static = (depth & 1);
1049 inputs = (is_static ? 0 : 1);
1050 if (is_get) {
1051 depth = size - inputs;
1052 } else {
1053 inputs += size; // putxxx pops the value from the stack
1054 depth = - inputs;
1055 }
1056 }
1057 break;
1058
1059 case Bytecodes::_invokevirtual:
1060 case Bytecodes::_invokespecial:
1061 case Bytecodes::_invokestatic:
1062 case Bytecodes::_invokedynamic:
1063 case Bytecodes::_invokeinterface:
1064 {
1065 bool ignored_will_link;
1066 ciSignature* declared_signature = NULL;
1067 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1068 assert(declared_signature != NULL, "cannot be null");
1069 inputs = declared_signature->arg_size_for_bc(code);
1070 int size = declared_signature->return_type()->size();
1071 depth = size - inputs;
1072 }
1073 break;
1074
1075 case Bytecodes::_multianewarray:
1076 {
1077 ciBytecodeStream iter(method());
1078 iter.reset_to_bci(bci());
1079 iter.next();
1080 inputs = iter.get_dimensions();
1081 assert(rsize == 1, "");
1082 depth = rsize - inputs;
1083 }
1084 break;
1085
1086 case Bytecodes::_ireturn:
1087 case Bytecodes::_lreturn:
1088 case Bytecodes::_freturn:
1089 case Bytecodes::_dreturn:
1090 case Bytecodes::_areturn:
1091 assert(rsize == -depth, "");
1092 inputs = rsize;
1093 break;
1094
1095 case Bytecodes::_jsr:
1096 case Bytecodes::_jsr_w:
1097 inputs = 0;
1098 depth = 1; // S.B. depth=1, not zero
1099 break;
1100
1101 default:
1102 // bytecode produces a typed result
1103 inputs = rsize - depth;
1104 assert(inputs >= 0, "");
1105 break;
1106 }
1107
1108 #ifdef ASSERT
1109 // spot check
1110 int outputs = depth + inputs;
1111 assert(outputs >= 0, "sanity");
1112 switch (code) {
1113 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1114 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1115 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1116 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1117 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1118 default: break;
1119 }
1120 #endif //ASSERT
1121
1122 return true;
1123 }
1124
1125
1126
1127 //------------------------------basic_plus_adr---------------------------------
1128 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1129 // short-circuit a common case
1130 if (offset == intcon(0)) return ptr;
1131 return _gvn.transform( new AddPNode(base, ptr, offset) );
1132 }
1133
1134 Node* GraphKit::ConvI2L(Node* offset) {
1135 // short-circuit a common case
1136 jint offset_con = find_int_con(offset, Type::OffsetBot);
1137 if (offset_con != Type::OffsetBot) {
1138 return longcon((jlong) offset_con);
1139 }
1140 return _gvn.transform( new ConvI2LNode(offset));
1141 }
1142
1143 Node* GraphKit::ConvI2UL(Node* offset) {
1144 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1145 if (offset_con != (juint) Type::OffsetBot) {
1146 return longcon((julong) offset_con);
1147 }
1148 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1149 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1150 return _gvn.transform( new AndLNode(conv, mask) );
1151 }
1152
1153 Node* GraphKit::ConvL2I(Node* offset) {
1154 // short-circuit a common case
1155 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1156 if (offset_con != (jlong)Type::OffsetBot) {
1157 return intcon((int) offset_con);
1158 }
1159 return _gvn.transform( new ConvL2INode(offset));
1160 }
1161
1162 //-------------------------load_object_klass-----------------------------------
1163 Node* GraphKit::load_object_klass(Node* obj) {
1164 // Special-case a fresh allocation to avoid building nodes:
1165 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1166 if (akls != NULL) return akls;
1167 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1168 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1169 }
1170
1171 //-------------------------load_array_length-----------------------------------
1172 Node* GraphKit::load_array_length(Node* array) {
1173 // Special-case a fresh allocation to avoid building nodes:
1174 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1175 Node *alen;
1176 if (alloc == NULL) {
1177 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1178 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1179 } else {
1180 alen = alloc->Ideal_length();
1181 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1182 if (ccast != alen) {
1183 alen = _gvn.transform(ccast);
1184 }
1185 }
1186 return alen;
1187 }
1188
1189 //------------------------------do_null_check----------------------------------
1190 // Helper function to do a NULL pointer check. Returned value is
1191 // the incoming address with NULL casted away. You are allowed to use the
1192 // not-null value only if you are control dependent on the test.
1193 #ifndef PRODUCT
1194 extern int explicit_null_checks_inserted,
1195 explicit_null_checks_elided;
1196 #endif
1197 Node* GraphKit::null_check_common(Node* value, BasicType type,
1198 // optional arguments for variations:
1199 bool assert_null,
1200 Node* *null_control,
1201 bool speculative) {
1202 assert(!assert_null || null_control == NULL, "not both at once");
1203 if (stopped()) return top();
1204 NOT_PRODUCT(explicit_null_checks_inserted++);
1205
1206 // Construct NULL check
1207 Node *chk = NULL;
1208 switch(type) {
1209 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1210 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1211 case T_ARRAY : // fall through
1212 type = T_OBJECT; // simplify further tests
1213 case T_OBJECT : {
1214 const Type *t = _gvn.type( value );
1215
1216 const TypeOopPtr* tp = t->isa_oopptr();
1217 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1218 // Only for do_null_check, not any of its siblings:
1219 && !assert_null && null_control == NULL) {
1220 // Usually, any field access or invocation on an unloaded oop type
1221 // will simply fail to link, since the statically linked class is
1222 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1223 // the static class is loaded but the sharper oop type is not.
1224 // Rather than checking for this obscure case in lots of places,
1225 // we simply observe that a null check on an unloaded class
1226 // will always be followed by a nonsense operation, so we
1227 // can just issue the uncommon trap here.
1228 // Our access to the unloaded class will only be correct
1229 // after it has been loaded and initialized, which requires
1230 // a trip through the interpreter.
1231 #ifndef PRODUCT
1232 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1233 #endif
1234 uncommon_trap(Deoptimization::Reason_unloaded,
1235 Deoptimization::Action_reinterpret,
1236 tp->klass(), "!loaded");
1237 return top();
1238 }
1239
1240 if (assert_null) {
1241 // See if the type is contained in NULL_PTR.
1242 // If so, then the value is already null.
1243 if (t->higher_equal(TypePtr::NULL_PTR)) {
1244 NOT_PRODUCT(explicit_null_checks_elided++);
1245 return value; // Elided null assert quickly!
1246 }
1247 } else {
1248 // See if mixing in the NULL pointer changes type.
1249 // If so, then the NULL pointer was not allowed in the original
1250 // type. In other words, "value" was not-null.
1251 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1252 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1253 NOT_PRODUCT(explicit_null_checks_elided++);
1254 return value; // Elided null check quickly!
1255 }
1256 }
1257 chk = new CmpPNode( value, null() );
1258 break;
1259 }
1260
1261 default:
1262 fatal("unexpected type: %s", type2name(type));
1263 }
1264 assert(chk != NULL, "sanity check");
1265 chk = _gvn.transform(chk);
1266
1267 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1268 BoolNode *btst = new BoolNode( chk, btest);
1269 Node *tst = _gvn.transform( btst );
1270
1271 //-----------
1272 // if peephole optimizations occurred, a prior test existed.
1273 // If a prior test existed, maybe it dominates as we can avoid this test.
1274 if (tst != btst && type == T_OBJECT) {
1275 // At this point we want to scan up the CFG to see if we can
1276 // find an identical test (and so avoid this test altogether).
1277 Node *cfg = control();
1278 int depth = 0;
1279 while( depth < 16 ) { // Limit search depth for speed
1280 if( cfg->Opcode() == Op_IfTrue &&
1281 cfg->in(0)->in(1) == tst ) {
1282 // Found prior test. Use "cast_not_null" to construct an identical
1283 // CastPP (and hence hash to) as already exists for the prior test.
1284 // Return that casted value.
1285 if (assert_null) {
1286 replace_in_map(value, null());
1287 return null(); // do not issue the redundant test
1288 }
1289 Node *oldcontrol = control();
1290 set_control(cfg);
1291 Node *res = cast_not_null(value);
1292 set_control(oldcontrol);
1293 NOT_PRODUCT(explicit_null_checks_elided++);
1294 return res;
1295 }
1296 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1297 if (cfg == NULL) break; // Quit at region nodes
1298 depth++;
1299 }
1300 }
1301
1302 //-----------
1303 // Branch to failure if null
1304 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1305 Deoptimization::DeoptReason reason;
1306 if (assert_null) {
1307 reason = Deoptimization::reason_null_assert(speculative);
1308 } else if (type == T_OBJECT) {
1309 reason = Deoptimization::reason_null_check(speculative);
1310 } else {
1311 reason = Deoptimization::Reason_div0_check;
1312 }
1313 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1314 // ciMethodData::has_trap_at will return a conservative -1 if any
1315 // must-be-null assertion has failed. This could cause performance
1316 // problems for a method after its first do_null_assert failure.
1317 // Consider using 'Reason_class_check' instead?
1318
1319 // To cause an implicit null check, we set the not-null probability
1320 // to the maximum (PROB_MAX). For an explicit check the probability
1321 // is set to a smaller value.
1322 if (null_control != NULL || too_many_traps(reason)) {
1323 // probability is less likely
1324 ok_prob = PROB_LIKELY_MAG(3);
1325 } else if (!assert_null &&
1326 (ImplicitNullCheckThreshold > 0) &&
1327 method() != NULL &&
1328 (method()->method_data()->trap_count(reason)
1329 >= (uint)ImplicitNullCheckThreshold)) {
1330 ok_prob = PROB_LIKELY_MAG(3);
1331 }
1332
1333 if (null_control != NULL) {
1334 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1335 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1336 set_control( _gvn.transform( new IfTrueNode(iff)));
1337 #ifndef PRODUCT
1338 if (null_true == top()) {
1339 explicit_null_checks_elided++;
1340 }
1341 #endif
1342 (*null_control) = null_true;
1343 } else {
1344 BuildCutout unless(this, tst, ok_prob);
1345 // Check for optimizer eliding test at parse time
1346 if (stopped()) {
1347 // Failure not possible; do not bother making uncommon trap.
1348 NOT_PRODUCT(explicit_null_checks_elided++);
1349 } else if (assert_null) {
1350 uncommon_trap(reason,
1351 Deoptimization::Action_make_not_entrant,
1352 NULL, "assert_null");
1353 } else {
1354 replace_in_map(value, zerocon(type));
1355 builtin_throw(reason);
1356 }
1357 }
1358
1359 // Must throw exception, fall-thru not possible?
1360 if (stopped()) {
1361 return top(); // No result
1362 }
1363
1364 if (assert_null) {
1365 // Cast obj to null on this path.
1366 replace_in_map(value, zerocon(type));
1367 return zerocon(type);
1368 }
1369
1370 // Cast obj to not-null on this path, if there is no null_control.
1371 // (If there is a null_control, a non-null value may come back to haunt us.)
1372 if (type == T_OBJECT) {
1373 Node* cast = cast_not_null(value, false);
1374 if (null_control == NULL || (*null_control) == top())
1375 replace_in_map(value, cast);
1376 value = cast;
1377 }
1378
1379 return value;
1380 }
1381
1382
1383 //------------------------------cast_not_null----------------------------------
1384 // Cast obj to not-null on this path
1385 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1386 const Type *t = _gvn.type(obj);
1387 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1388 // Object is already not-null?
1389 if( t == t_not_null ) return obj;
1390
1391 Node *cast = new CastPPNode(obj,t_not_null);
1392 cast->init_req(0, control());
1393 cast = _gvn.transform( cast );
1394
1395 // Scan for instances of 'obj' in the current JVM mapping.
1396 // These instances are known to be not-null after the test.
1397 if (do_replace_in_map)
1398 replace_in_map(obj, cast);
1399
1400 return cast; // Return casted value
1401 }
1402
1403 // Sometimes in intrinsics, we implicitly know an object is not null
1404 // (there's no actual null check) so we can cast it to not null. In
1405 // the course of optimizations, the input to the cast can become null.
1406 // In that case that data path will die and we need the control path
1407 // to become dead as well to keep the graph consistent. So we have to
1408 // add a check for null for which one branch can't be taken. It uses
1409 // an Opaque4 node that will cause the check to be removed after loop
1410 // opts so the test goes away and the compiled code doesn't execute a
1411 // useless check.
1412 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1413 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1414 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1415 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1416 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1417 _gvn.set_type(iff, iff->Value(&_gvn));
1418 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1419 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1420 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1421 C->root()->add_req(halt);
1422 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1423 set_control(if_t);
1424 return cast_not_null(value, do_replace_in_map);
1425 }
1426
1427
1428 //--------------------------replace_in_map-------------------------------------
1429 void GraphKit::replace_in_map(Node* old, Node* neww) {
1430 if (old == neww) {
1431 return;
1432 }
1433
1434 map()->replace_edge(old, neww);
1435
1436 // Note: This operation potentially replaces any edge
1437 // on the map. This includes locals, stack, and monitors
1438 // of the current (innermost) JVM state.
1439
1440 // don't let inconsistent types from profiling escape this
1441 // method
1442
1443 const Type* told = _gvn.type(old);
1444 const Type* tnew = _gvn.type(neww);
1445
1446 if (!tnew->higher_equal(told)) {
1447 return;
1448 }
1449
1450 map()->record_replaced_node(old, neww);
1451 }
1452
1453
1454 //=============================================================================
1455 //--------------------------------memory---------------------------------------
1456 Node* GraphKit::memory(uint alias_idx) {
1457 MergeMemNode* mem = merged_memory();
1458 Node* p = mem->memory_at(alias_idx);
1459 _gvn.set_type(p, Type::MEMORY); // must be mapped
1460 return p;
1461 }
1462
1463 //-----------------------------reset_memory------------------------------------
1464 Node* GraphKit::reset_memory() {
1465 Node* mem = map()->memory();
1466 // do not use this node for any more parsing!
1467 debug_only( map()->set_memory((Node*)NULL) );
1468 return _gvn.transform( mem );
1469 }
1470
1471 //------------------------------set_all_memory---------------------------------
1472 void GraphKit::set_all_memory(Node* newmem) {
1473 Node* mergemem = MergeMemNode::make(newmem);
1474 gvn().set_type_bottom(mergemem);
1475 map()->set_memory(mergemem);
1476 }
1477
1478 //------------------------------set_all_memory_call----------------------------
1479 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1480 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1481 set_all_memory(newmem);
1482 }
1483
1484 //=============================================================================
1485 //
1486 // parser factory methods for MemNodes
1487 //
1488 // These are layered on top of the factory methods in LoadNode and StoreNode,
1489 // and integrate with the parser's memory state and _gvn engine.
1490 //
1491
1492 // factory methods in "int adr_idx"
1493 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1494 int adr_idx,
1495 MemNode::MemOrd mo,
1496 LoadNode::ControlDependency control_dependency,
1497 bool require_atomic_access,
1498 bool unaligned,
1499 bool mismatched) {
1500 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1501 const TypePtr* adr_type = NULL; // debug-mode-only argument
1502 debug_only(adr_type = C->get_adr_type(adr_idx));
1503 Node* mem = memory(adr_idx);
1504 Node* ld;
1505 if (require_atomic_access && bt == T_LONG) {
1506 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1507 } else if (require_atomic_access && bt == T_DOUBLE) {
1508 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched);
1509 } else {
1510 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched);
1511 }
1512 ld = _gvn.transform(ld);
1513 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1514 // Improve graph before escape analysis and boxing elimination.
1515 record_for_igvn(ld);
1516 }
1517 return ld;
1518 }
1519
1520 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1521 int adr_idx,
1522 MemNode::MemOrd mo,
1523 bool require_atomic_access,
1524 bool unaligned,
1525 bool mismatched) {
1526 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1527 const TypePtr* adr_type = NULL;
1528 debug_only(adr_type = C->get_adr_type(adr_idx));
1529 Node *mem = memory(adr_idx);
1530 Node* st;
1531 if (require_atomic_access && bt == T_LONG) {
1532 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1533 } else if (require_atomic_access && bt == T_DOUBLE) {
1534 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1535 } else {
1536 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1537 }
1538 if (unaligned) {
1539 st->as_Store()->set_unaligned_access();
1540 }
1541 if (mismatched) {
1542 st->as_Store()->set_mismatched_access();
1543 }
1544 st = _gvn.transform(st);
1545 set_memory(st, adr_idx);
1546 // Back-to-back stores can only remove intermediate store with DU info
1547 // so push on worklist for optimizer.
1548 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1549 record_for_igvn(st);
1550
1551 return st;
1552 }
1553
1554
1555 void GraphKit::pre_barrier(bool do_load,
1556 Node* ctl,
1557 Node* obj,
1558 Node* adr,
1559 uint adr_idx,
1560 Node* val,
1561 const TypeOopPtr* val_type,
1562 Node* pre_val,
1563 BasicType bt) {
1564
1565 BarrierSet* bs = Universe::heap()->barrier_set();
1566 set_control(ctl);
1567 switch (bs->kind()) {
1568 case BarrierSet::G1SATBCTLogging:
1569 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1570 break;
1571 case BarrierSet::ShenandoahBarrierSet:
1572 shenandoah_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1573 break;
1574 case BarrierSet::CardTableForRS:
1575 case BarrierSet::CardTableExtension:
1576 case BarrierSet::ModRef:
1577 break;
1578
1579 default :
1580 ShouldNotReachHere();
1581
1582 }
1583 }
1584
1585 bool GraphKit::can_move_pre_barrier() const {
1586 BarrierSet* bs = Universe::heap()->barrier_set();
1587 switch (bs->kind()) {
1588 case BarrierSet::G1SATBCTLogging:
1589 case BarrierSet::ShenandoahBarrierSet:
1590 return true; // Can move it if no safepoint
1591
1592 case BarrierSet::CardTableForRS:
1593 case BarrierSet::CardTableExtension:
1594 case BarrierSet::ModRef:
1595 return true; // There is no pre-barrier
1596
1597 default :
1598 ShouldNotReachHere();
1599 }
1600 return false;
1601 }
1602
1603 void GraphKit::post_barrier(Node* ctl,
1604 Node* store,
1605 Node* obj,
1606 Node* adr,
1607 uint adr_idx,
1608 Node* val,
1609 BasicType bt,
1610 bool use_precise) {
1611 BarrierSet* bs = Universe::heap()->barrier_set();
1612 set_control(ctl);
1613 switch (bs->kind()) {
1614 case BarrierSet::G1SATBCTLogging:
1615 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1616 break;
1617
1618 case BarrierSet::CardTableForRS:
1619 case BarrierSet::CardTableExtension:
1620 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1621 break;
1622
1623 case BarrierSet::ModRef:
1624 case BarrierSet::ShenandoahBarrierSet:
1625 break;
1626
1627 default :
1628 ShouldNotReachHere();
1629
1630 }
1631 }
1632
1633 void GraphKit::keep_alive_barrier(Node* ctl, Node* obj) {
1634 BarrierSet* bs = Universe::heap()->barrier_set();
1635 switch (bs->kind()) {
1636 case BarrierSet::G1SATBCTLogging:
1637 pre_barrier(false /* do_load */,
1638 ctl,
1639 NULL /* obj */,
1640 NULL /* adr */,
1641 max_juint /* alias_idx */,
1642 NULL /* val */,
1643 NULL /* val_type */,
1644 obj /* pre_val */,
1645 T_OBJECT);
1646 break;
1647 case BarrierSet::ShenandoahBarrierSet:
1648 if (ShenandoahKeepAliveBarrier) {
1649 pre_barrier(false /* do_load */,
1650 ctl,
1651 NULL /* obj */,
1652 NULL /* adr */,
1653 max_juint /* alias_idx */,
1654 NULL /* val */,
1655 NULL /* val_type */,
1656 obj /* pre_val */,
1657 T_OBJECT);
1658 }
1659 break;
1660 case BarrierSet::CardTableForRS:
1661 case BarrierSet::CardTableExtension:
1662 case BarrierSet::ModRef:
1663 break;
1664 default :
1665 ShouldNotReachHere();
1666
1667 }
1668 }
1669
1670 Node* GraphKit::store_oop(Node* ctl,
1671 Node* obj,
1672 Node* adr,
1673 const TypePtr* adr_type,
1674 Node* val,
1675 const TypeOopPtr* val_type,
1676 BasicType bt,
1677 bool use_precise,
1678 MemNode::MemOrd mo,
1679 bool mismatched) {
1680 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1681 // could be delayed during Parse (for example, in adjust_map_after_if()).
1682 // Execute transformation here to avoid barrier generation in such case.
1683 if (_gvn.type(val) == TypePtr::NULL_PTR)
1684 val = _gvn.makecon(TypePtr::NULL_PTR);
1685
1686 set_control(ctl);
1687 if (stopped()) return top(); // Dead path ?
1688
1689 assert(bt == T_OBJECT, "sanity");
1690 assert(val != NULL, "not dead path");
1691 uint adr_idx = C->get_alias_index(adr_type);
1692 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1693
1694 val = shenandoah_storeval_barrier(val);
1695
1696 pre_barrier(true /* do_load */,
1697 control(), obj, adr, adr_idx, val, val_type,
1698 NULL /* pre_val */,
1699 bt);
1700
1701 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1702 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1703 return store;
1704 }
1705
1706 // Could be an array or object we don't know at compile time (unsafe ref.)
1707 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1708 Node* obj, // containing obj
1709 Node* adr, // actual adress to store val at
1710 const TypePtr* adr_type,
1711 Node* val,
1712 BasicType bt,
1713 MemNode::MemOrd mo,
1714 bool mismatched) {
1715 Compile::AliasType* at = C->alias_type(adr_type);
1716 const TypeOopPtr* val_type = NULL;
1717 if (adr_type->isa_instptr()) {
1718 if (at->field() != NULL) {
1719 // known field. This code is a copy of the do_put_xxx logic.
1720 ciField* field = at->field();
1721 if (!field->type()->is_loaded()) {
1722 val_type = TypeInstPtr::BOTTOM;
1723 } else {
1724 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1725 }
1726 }
1727 } else if (adr_type->isa_aryptr()) {
1728 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1729 }
1730 if (val_type == NULL) {
1731 val_type = TypeInstPtr::BOTTOM;
1732 }
1733 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1734 }
1735
1736
1737 //-------------------------array_element_address-------------------------
1738 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1739 const TypeInt* sizetype, Node* ctrl) {
1740 uint shift = exact_log2(type2aelembytes(elembt));
1741 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1742
1743 // short-circuit a common case (saves lots of confusing waste motion)
1744 jint idx_con = find_int_con(idx, -1);
1745 if (idx_con >= 0) {
1746 intptr_t offset = header + ((intptr_t)idx_con << shift);
1747 return basic_plus_adr(ary, offset);
1748 }
1749
1750 // must be correct type for alignment purposes
1751 Node* base = basic_plus_adr(ary, header);
1752 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1753 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1754 return basic_plus_adr(ary, base, scale);
1755 }
1756
1757 //-------------------------load_array_element-------------------------
1758 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1759 const Type* elemtype = arytype->elem();
1760 BasicType elembt = elemtype->array_element_basic_type();
1761 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1762 if (elembt == T_NARROWOOP) {
1763 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1764 }
1765 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1766 return ld;
1767 }
1768
1769 //-------------------------set_arguments_for_java_call-------------------------
1770 // Arguments (pre-popped from the stack) are taken from the JVMS.
1771 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1772 // Add the call arguments:
1773 uint nargs = call->method()->arg_size();
1774 for (uint i = 0; i < nargs; i++) {
1775 Node* arg = argument(i);
1776 call->init_req(i + TypeFunc::Parms, arg);
1777 }
1778 }
1779
1780 //---------------------------set_edges_for_java_call---------------------------
1781 // Connect a newly created call into the current JVMS.
1782 // A return value node (if any) is returned from set_edges_for_java_call.
1783 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1784
1785 // Add the predefined inputs:
1786 call->init_req( TypeFunc::Control, control() );
1787 call->init_req( TypeFunc::I_O , i_o() );
1788 call->init_req( TypeFunc::Memory , reset_memory() );
1789 call->init_req( TypeFunc::FramePtr, frameptr() );
1790 call->init_req( TypeFunc::ReturnAdr, top() );
1791
1792 add_safepoint_edges(call, must_throw);
1793
1794 Node* xcall = _gvn.transform(call);
1795
1796 if (xcall == top()) {
1797 set_control(top());
1798 return;
1799 }
1800 assert(xcall == call, "call identity is stable");
1801
1802 // Re-use the current map to produce the result.
1803
1804 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1805 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1806 set_all_memory_call(xcall, separate_io_proj);
1807
1808 //return xcall; // no need, caller already has it
1809 }
1810
1811 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1812 if (stopped()) return top(); // maybe the call folded up?
1813
1814 // Capture the return value, if any.
1815 Node* ret;
1816 if (call->method() == NULL ||
1817 call->method()->return_type()->basic_type() == T_VOID)
1818 ret = top();
1819 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1820
1821 // Note: Since any out-of-line call can produce an exception,
1822 // we always insert an I_O projection from the call into the result.
1823
1824 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1825
1826 if (separate_io_proj) {
1827 // The caller requested separate projections be used by the fall
1828 // through and exceptional paths, so replace the projections for
1829 // the fall through path.
1830 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1831 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1832 }
1833 return ret;
1834 }
1835
1836 //--------------------set_predefined_input_for_runtime_call--------------------
1837 // Reading and setting the memory state is way conservative here.
1838 // The real problem is that I am not doing real Type analysis on memory,
1839 // so I cannot distinguish card mark stores from other stores. Across a GC
1840 // point the Store Barrier and the card mark memory has to agree. I cannot
1841 // have a card mark store and its barrier split across the GC point from
1842 // either above or below. Here I get that to happen by reading ALL of memory.
1843 // A better answer would be to separate out card marks from other memory.
1844 // For now, return the input memory state, so that it can be reused
1845 // after the call, if this call has restricted memory effects.
1846 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call) {
1847 // Set fixed predefined input arguments
1848 Node* memory = reset_memory();
1849 call->init_req( TypeFunc::Control, control() );
1850 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1851 call->init_req( TypeFunc::Memory, memory ); // may gc ptrs
1852 call->init_req( TypeFunc::FramePtr, frameptr() );
1853 call->init_req( TypeFunc::ReturnAdr, top() );
1854 return memory;
1855 }
1856
1857 //-------------------set_predefined_output_for_runtime_call--------------------
1858 // Set control and memory (not i_o) from the call.
1859 // If keep_mem is not NULL, use it for the output state,
1860 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1861 // If hook_mem is NULL, this call produces no memory effects at all.
1862 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1863 // then only that memory slice is taken from the call.
1864 // In the last case, we must put an appropriate memory barrier before
1865 // the call, so as to create the correct anti-dependencies on loads
1866 // preceding the call.
1867 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1868 Node* keep_mem,
1869 const TypePtr* hook_mem) {
1870 // no i/o
1871 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1872 if (keep_mem) {
1873 // First clone the existing memory state
1874 set_all_memory(keep_mem);
1875 if (hook_mem != NULL) {
1876 // Make memory for the call
1877 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1878 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1879 // We also use hook_mem to extract specific effects from arraycopy stubs.
1880 set_memory(mem, hook_mem);
1881 }
1882 // ...else the call has NO memory effects.
1883
1884 // Make sure the call advertises its memory effects precisely.
1885 // This lets us build accurate anti-dependences in gcm.cpp.
1886 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1887 "call node must be constructed correctly");
1888 } else {
1889 assert(hook_mem == NULL, "");
1890 // This is not a "slow path" call; all memory comes from the call.
1891 set_all_memory_call(call);
1892 }
1893 }
1894
1895
1896 // Replace the call with the current state of the kit.
1897 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1898 JVMState* ejvms = NULL;
1899 if (has_exceptions()) {
1900 ejvms = transfer_exceptions_into_jvms();
1901 }
1902
1903 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1904 ReplacedNodes replaced_nodes_exception;
1905 Node* ex_ctl = top();
1906
1907 SafePointNode* final_state = stop();
1908
1909 // Find all the needed outputs of this call
1910 CallProjections callprojs;
1911 call->extract_projections(&callprojs, true);
1912
1913 Node* init_mem = call->in(TypeFunc::Memory);
1914 Node* final_mem = final_state->in(TypeFunc::Memory);
1915 Node* final_ctl = final_state->in(TypeFunc::Control);
1916 Node* final_io = final_state->in(TypeFunc::I_O);
1917
1918 // Replace all the old call edges with the edges from the inlining result
1919 if (callprojs.fallthrough_catchproj != NULL) {
1920 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1921 }
1922 if (callprojs.fallthrough_memproj != NULL) {
1923 if (final_mem->is_MergeMem()) {
1924 // Parser's exits MergeMem was not transformed but may be optimized
1925 final_mem = _gvn.transform(final_mem);
1926 }
1927 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1928 }
1929 if (callprojs.fallthrough_ioproj != NULL) {
1930 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1931 }
1932
1933 // Replace the result with the new result if it exists and is used
1934 if (callprojs.resproj != NULL && result != NULL) {
1935 C->gvn_replace_by(callprojs.resproj, result);
1936 }
1937
1938 if (ejvms == NULL) {
1939 // No exception edges to simply kill off those paths
1940 if (callprojs.catchall_catchproj != NULL) {
1941 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1942 }
1943 if (callprojs.catchall_memproj != NULL) {
1944 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1945 }
1946 if (callprojs.catchall_ioproj != NULL) {
1947 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1948 }
1949 // Replace the old exception object with top
1950 if (callprojs.exobj != NULL) {
1951 C->gvn_replace_by(callprojs.exobj, C->top());
1952 }
1953 } else {
1954 GraphKit ekit(ejvms);
1955
1956 // Load my combined exception state into the kit, with all phis transformed:
1957 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1958 replaced_nodes_exception = ex_map->replaced_nodes();
1959
1960 Node* ex_oop = ekit.use_exception_state(ex_map);
1961
1962 if (callprojs.catchall_catchproj != NULL) {
1963 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1964 ex_ctl = ekit.control();
1965 }
1966 if (callprojs.catchall_memproj != NULL) {
1967 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1968 }
1969 if (callprojs.catchall_ioproj != NULL) {
1970 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1971 }
1972
1973 // Replace the old exception object with the newly created one
1974 if (callprojs.exobj != NULL) {
1975 C->gvn_replace_by(callprojs.exobj, ex_oop);
1976 }
1977 }
1978
1979 // Disconnect the call from the graph
1980 call->disconnect_inputs(NULL, C);
1981 C->gvn_replace_by(call, C->top());
1982
1983 // Clean up any MergeMems that feed other MergeMems since the
1984 // optimizer doesn't like that.
1985 if (final_mem->is_MergeMem()) {
1986 Node_List wl;
1987 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1988 Node* m = i.get();
1989 if (m->is_MergeMem() && !wl.contains(m)) {
1990 wl.push(m);
1991 }
1992 }
1993 while (wl.size() > 0) {
1994 _gvn.transform(wl.pop());
1995 }
1996 }
1997
1998 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1999 replaced_nodes.apply(C, final_ctl);
2000 }
2001 if (!ex_ctl->is_top() && do_replaced_nodes) {
2002 replaced_nodes_exception.apply(C, ex_ctl);
2003 }
2004 }
2005
2006
2007 //------------------------------increment_counter------------------------------
2008 // for statistics: increment a VM counter by 1
2009
2010 void GraphKit::increment_counter(address counter_addr) {
2011 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2012 increment_counter(adr1);
2013 }
2014
2015 void GraphKit::increment_counter(Node* counter_addr) {
2016 int adr_type = Compile::AliasIdxRaw;
2017 Node* ctrl = control();
2018 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2019 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2020 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2021 }
2022
2023
2024 //------------------------------uncommon_trap----------------------------------
2025 // Bail out to the interpreter in mid-method. Implemented by calling the
2026 // uncommon_trap blob. This helper function inserts a runtime call with the
2027 // right debug info.
2028 void GraphKit::uncommon_trap(int trap_request,
2029 ciKlass* klass, const char* comment,
2030 bool must_throw,
2031 bool keep_exact_action) {
2032 if (failing()) stop();
2033 if (stopped()) return; // trap reachable?
2034
2035 // Note: If ProfileTraps is true, and if a deopt. actually
2036 // occurs here, the runtime will make sure an MDO exists. There is
2037 // no need to call method()->ensure_method_data() at this point.
2038
2039 // Set the stack pointer to the right value for reexecution:
2040 set_sp(reexecute_sp());
2041
2042 #ifdef ASSERT
2043 if (!must_throw) {
2044 // Make sure the stack has at least enough depth to execute
2045 // the current bytecode.
2046 int inputs, ignored_depth;
2047 if (compute_stack_effects(inputs, ignored_depth)) {
2048 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2049 Bytecodes::name(java_bc()), sp(), inputs);
2050 }
2051 }
2052 #endif
2053
2054 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2055 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2056
2057 switch (action) {
2058 case Deoptimization::Action_maybe_recompile:
2059 case Deoptimization::Action_reinterpret:
2060 // Temporary fix for 6529811 to allow virtual calls to be sure they
2061 // get the chance to go from mono->bi->mega
2062 if (!keep_exact_action &&
2063 Deoptimization::trap_request_index(trap_request) < 0 &&
2064 too_many_recompiles(reason)) {
2065 // This BCI is causing too many recompilations.
2066 if (C->log() != NULL) {
2067 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2068 Deoptimization::trap_reason_name(reason),
2069 Deoptimization::trap_action_name(action));
2070 }
2071 action = Deoptimization::Action_none;
2072 trap_request = Deoptimization::make_trap_request(reason, action);
2073 } else {
2074 C->set_trap_can_recompile(true);
2075 }
2076 break;
2077 case Deoptimization::Action_make_not_entrant:
2078 C->set_trap_can_recompile(true);
2079 break;
2080 case Deoptimization::Action_none:
2081 case Deoptimization::Action_make_not_compilable:
2082 break;
2083 default:
2084 #ifdef ASSERT
2085 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2086 #endif
2087 break;
2088 }
2089
2090 if (TraceOptoParse) {
2091 char buf[100];
2092 tty->print_cr("Uncommon trap %s at bci:%d",
2093 Deoptimization::format_trap_request(buf, sizeof(buf),
2094 trap_request), bci());
2095 }
2096
2097 CompileLog* log = C->log();
2098 if (log != NULL) {
2099 int kid = (klass == NULL)? -1: log->identify(klass);
2100 log->begin_elem("uncommon_trap bci='%d'", bci());
2101 char buf[100];
2102 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2103 trap_request));
2104 if (kid >= 0) log->print(" klass='%d'", kid);
2105 if (comment != NULL) log->print(" comment='%s'", comment);
2106 log->end_elem();
2107 }
2108
2109 // Make sure any guarding test views this path as very unlikely
2110 Node *i0 = control()->in(0);
2111 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2112 IfNode *iff = i0->as_If();
2113 float f = iff->_prob; // Get prob
2114 if (control()->Opcode() == Op_IfTrue) {
2115 if (f > PROB_UNLIKELY_MAG(4))
2116 iff->_prob = PROB_MIN;
2117 } else {
2118 if (f < PROB_LIKELY_MAG(4))
2119 iff->_prob = PROB_MAX;
2120 }
2121 }
2122
2123 // Clear out dead values from the debug info.
2124 kill_dead_locals();
2125
2126 // Now insert the uncommon trap subroutine call
2127 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2128 const TypePtr* no_memory_effects = NULL;
2129 // Pass the index of the class to be loaded
2130 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2131 (must_throw ? RC_MUST_THROW : 0),
2132 OptoRuntime::uncommon_trap_Type(),
2133 call_addr, "uncommon_trap", no_memory_effects,
2134 intcon(trap_request));
2135 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2136 "must extract request correctly from the graph");
2137 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2138
2139 call->set_req(TypeFunc::ReturnAdr, returnadr());
2140 // The debug info is the only real input to this call.
2141
2142 // Halt-and-catch fire here. The above call should never return!
2143 HaltNode* halt = new HaltNode(control(), frameptr());
2144 _gvn.set_type_bottom(halt);
2145 root()->add_req(halt);
2146
2147 stop_and_kill_map();
2148 }
2149
2150
2151 //--------------------------just_allocated_object------------------------------
2152 // Report the object that was just allocated.
2153 // It must be the case that there are no intervening safepoints.
2154 // We use this to determine if an object is so "fresh" that
2155 // it does not require card marks.
2156 Node* GraphKit::just_allocated_object(Node* current_control) {
2157 if (C->recent_alloc_ctl() == current_control)
2158 return C->recent_alloc_obj();
2159 return NULL;
2160 }
2161
2162
2163 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2164 // (Note: TypeFunc::make has a cache that makes this fast.)
2165 const TypeFunc* tf = TypeFunc::make(dest_method);
2166 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2167 for (int j = 0; j < nargs; j++) {
2168 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2169 if( targ->basic_type() == T_DOUBLE ) {
2170 // If any parameters are doubles, they must be rounded before
2171 // the call, dstore_rounding does gvn.transform
2172 Node *arg = argument(j);
2173 arg = dstore_rounding(arg);
2174 set_argument(j, arg);
2175 }
2176 }
2177 }
2178
2179 /**
2180 * Record profiling data exact_kls for Node n with the type system so
2181 * that it can propagate it (speculation)
2182 *
2183 * @param n node that the type applies to
2184 * @param exact_kls type from profiling
2185 * @param maybe_null did profiling see null?
2186 *
2187 * @return node with improved type
2188 */
2189 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2190 const Type* current_type = _gvn.type(n);
2191 assert(UseTypeSpeculation, "type speculation must be on");
2192
2193 const TypePtr* speculative = current_type->speculative();
2194
2195 // Should the klass from the profile be recorded in the speculative type?
2196 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2197 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2198 const TypeOopPtr* xtype = tklass->as_instance_type();
2199 assert(xtype->klass_is_exact(), "Should be exact");
2200 // Any reason to believe n is not null (from this profiling or a previous one)?
2201 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2202 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2203 // record the new speculative type's depth
2204 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2205 speculative = speculative->with_inline_depth(jvms()->depth());
2206 } else if (current_type->would_improve_ptr(ptr_kind)) {
2207 // Profiling report that null was never seen so we can change the
2208 // speculative type to non null ptr.
2209 if (ptr_kind == ProfileAlwaysNull) {
2210 speculative = TypePtr::NULL_PTR;
2211 } else {
2212 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2213 const TypePtr* ptr = TypePtr::NOTNULL;
2214 if (speculative != NULL) {
2215 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2216 } else {
2217 speculative = ptr;
2218 }
2219 }
2220 }
2221
2222 if (speculative != current_type->speculative()) {
2223 // Build a type with a speculative type (what we think we know
2224 // about the type but will need a guard when we use it)
2225 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2226 // We're changing the type, we need a new CheckCast node to carry
2227 // the new type. The new type depends on the control: what
2228 // profiling tells us is only valid from here as far as we can
2229 // tell.
2230 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2231 cast = _gvn.transform(cast);
2232 replace_in_map(n, cast);
2233 n = cast;
2234 }
2235
2236 return n;
2237 }
2238
2239 /**
2240 * Record profiling data from receiver profiling at an invoke with the
2241 * type system so that it can propagate it (speculation)
2242 *
2243 * @param n receiver node
2244 *
2245 * @return node with improved type
2246 */
2247 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2248 if (!UseTypeSpeculation) {
2249 return n;
2250 }
2251 ciKlass* exact_kls = profile_has_unique_klass();
2252 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2253 if ((java_bc() == Bytecodes::_checkcast ||
2254 java_bc() == Bytecodes::_instanceof ||
2255 java_bc() == Bytecodes::_aastore) &&
2256 method()->method_data()->is_mature()) {
2257 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2258 if (data != NULL) {
2259 if (!data->as_BitData()->null_seen()) {
2260 ptr_kind = ProfileNeverNull;
2261 } else {
2262 assert(data->is_ReceiverTypeData(), "bad profile data type");
2263 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2264 uint i = 0;
2265 for (; i < call->row_limit(); i++) {
2266 ciKlass* receiver = call->receiver(i);
2267 if (receiver != NULL) {
2268 break;
2269 }
2270 }
2271 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2272 }
2273 }
2274 }
2275 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2276 }
2277
2278 /**
2279 * Record profiling data from argument profiling at an invoke with the
2280 * type system so that it can propagate it (speculation)
2281 *
2282 * @param dest_method target method for the call
2283 * @param bc what invoke bytecode is this?
2284 */
2285 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2286 if (!UseTypeSpeculation) {
2287 return;
2288 }
2289 const TypeFunc* tf = TypeFunc::make(dest_method);
2290 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2291 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2292 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2293 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2294 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2295 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2296 ciKlass* better_type = NULL;
2297 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2298 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2299 }
2300 i++;
2301 }
2302 }
2303 }
2304
2305 /**
2306 * Record profiling data from parameter profiling at an invoke with
2307 * the type system so that it can propagate it (speculation)
2308 */
2309 void GraphKit::record_profiled_parameters_for_speculation() {
2310 if (!UseTypeSpeculation) {
2311 return;
2312 }
2313 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2314 if (_gvn.type(local(i))->isa_oopptr()) {
2315 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2316 ciKlass* better_type = NULL;
2317 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2318 record_profile_for_speculation(local(i), better_type, ptr_kind);
2319 }
2320 j++;
2321 }
2322 }
2323 }
2324
2325 /**
2326 * Record profiling data from return value profiling at an invoke with
2327 * the type system so that it can propagate it (speculation)
2328 */
2329 void GraphKit::record_profiled_return_for_speculation() {
2330 if (!UseTypeSpeculation) {
2331 return;
2332 }
2333 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2334 ciKlass* better_type = NULL;
2335 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2336 // If profiling reports a single type for the return value,
2337 // feed it to the type system so it can propagate it as a
2338 // speculative type
2339 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2340 }
2341 }
2342
2343 void GraphKit::round_double_result(ciMethod* dest_method) {
2344 // A non-strict method may return a double value which has an extended
2345 // exponent, but this must not be visible in a caller which is 'strict'
2346 // If a strict caller invokes a non-strict callee, round a double result
2347
2348 BasicType result_type = dest_method->return_type()->basic_type();
2349 assert( method() != NULL, "must have caller context");
2350 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2351 // Destination method's return value is on top of stack
2352 // dstore_rounding() does gvn.transform
2353 Node *result = pop_pair();
2354 result = dstore_rounding(result);
2355 push_pair(result);
2356 }
2357 }
2358
2359 // rounding for strict float precision conformance
2360 Node* GraphKit::precision_rounding(Node* n) {
2361 return UseStrictFP && _method->flags().is_strict()
2362 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2363 ? _gvn.transform( new RoundFloatNode(0, n) )
2364 : n;
2365 }
2366
2367 // rounding for strict double precision conformance
2368 Node* GraphKit::dprecision_rounding(Node *n) {
2369 return UseStrictFP && _method->flags().is_strict()
2370 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2371 ? _gvn.transform( new RoundDoubleNode(0, n) )
2372 : n;
2373 }
2374
2375 // rounding for non-strict double stores
2376 Node* GraphKit::dstore_rounding(Node* n) {
2377 return Matcher::strict_fp_requires_explicit_rounding
2378 && UseSSE <= 1
2379 ? _gvn.transform( new RoundDoubleNode(0, n) )
2380 : n;
2381 }
2382
2383 //=============================================================================
2384 // Generate a fast path/slow path idiom. Graph looks like:
2385 // [foo] indicates that 'foo' is a parameter
2386 //
2387 // [in] NULL
2388 // \ /
2389 // CmpP
2390 // Bool ne
2391 // If
2392 // / \
2393 // True False-<2>
2394 // / |
2395 // / cast_not_null
2396 // Load | | ^
2397 // [fast_test] | |
2398 // gvn to opt_test | |
2399 // / \ | <1>
2400 // True False |
2401 // | \\ |
2402 // [slow_call] \[fast_result]
2403 // Ctl Val \ \
2404 // | \ \
2405 // Catch <1> \ \
2406 // / \ ^ \ \
2407 // Ex No_Ex | \ \
2408 // | \ \ | \ <2> \
2409 // ... \ [slow_res] | | \ [null_result]
2410 // \ \--+--+--- | |
2411 // \ | / \ | /
2412 // --------Region Phi
2413 //
2414 //=============================================================================
2415 // Code is structured as a series of driver functions all called 'do_XXX' that
2416 // call a set of helper functions. Helper functions first, then drivers.
2417
2418 //------------------------------null_check_oop---------------------------------
2419 // Null check oop. Set null-path control into Region in slot 3.
2420 // Make a cast-not-nullness use the other not-null control. Return cast.
2421 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2422 bool never_see_null,
2423 bool safe_for_replace,
2424 bool speculative) {
2425 // Initial NULL check taken path
2426 (*null_control) = top();
2427 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2428
2429 // Generate uncommon_trap:
2430 if (never_see_null && (*null_control) != top()) {
2431 // If we see an unexpected null at a check-cast we record it and force a
2432 // recompile; the offending check-cast will be compiled to handle NULLs.
2433 // If we see more than one offending BCI, then all checkcasts in the
2434 // method will be compiled to handle NULLs.
2435 PreserveJVMState pjvms(this);
2436 set_control(*null_control);
2437 replace_in_map(value, null());
2438 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2439 uncommon_trap(reason,
2440 Deoptimization::Action_make_not_entrant);
2441 (*null_control) = top(); // NULL path is dead
2442 }
2443 if ((*null_control) == top() && safe_for_replace) {
2444 replace_in_map(value, cast);
2445 }
2446
2447 // Cast away null-ness on the result
2448 return cast;
2449 }
2450
2451 //------------------------------opt_iff----------------------------------------
2452 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2453 // Return slow-path control.
2454 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2455 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2456
2457 // Fast path taken; set region slot 2
2458 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2459 region->init_req(2,fast_taken); // Capture fast-control
2460
2461 // Fast path not-taken, i.e. slow path
2462 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2463 return slow_taken;
2464 }
2465
2466 //-----------------------------make_runtime_call-------------------------------
2467 Node* GraphKit::make_runtime_call(int flags,
2468 const TypeFunc* call_type, address call_addr,
2469 const char* call_name,
2470 const TypePtr* adr_type,
2471 // The following parms are all optional.
2472 // The first NULL ends the list.
2473 Node* parm0, Node* parm1,
2474 Node* parm2, Node* parm3,
2475 Node* parm4, Node* parm5,
2476 Node* parm6, Node* parm7) {
2477 // Slow-path call
2478 bool is_leaf = !(flags & RC_NO_LEAF);
2479 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2480 if (call_name == NULL) {
2481 assert(!is_leaf, "must supply name for leaf");
2482 call_name = OptoRuntime::stub_name(call_addr);
2483 }
2484 CallNode* call;
2485 if (!is_leaf) {
2486 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2487 bci(), adr_type);
2488 } else if (flags & RC_NO_FP) {
2489 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2490 } else {
2491 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2492 }
2493
2494 // The following is similar to set_edges_for_java_call,
2495 // except that the memory effects of the call are restricted to AliasIdxRaw.
2496
2497 // Slow path call has no side-effects, uses few values
2498 bool wide_in = !(flags & RC_NARROW_MEM);
2499 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2500
2501 Node* prev_mem = NULL;
2502 if (wide_in) {
2503 prev_mem = set_predefined_input_for_runtime_call(call);
2504 } else {
2505 assert(!wide_out, "narrow in => narrow out");
2506 Node* narrow_mem = memory(adr_type);
2507 prev_mem = reset_memory();
2508 map()->set_memory(narrow_mem);
2509 set_predefined_input_for_runtime_call(call);
2510 }
2511
2512 // Hook each parm in order. Stop looking at the first NULL.
2513 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2514 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2515 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2516 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2517 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2518 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2519 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2520 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2521 /* close each nested if ===> */ } } } } } } } }
2522 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2523
2524 if (!is_leaf) {
2525 // Non-leaves can block and take safepoints:
2526 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2527 }
2528 // Non-leaves can throw exceptions:
2529 if (has_io) {
2530 call->set_req(TypeFunc::I_O, i_o());
2531 }
2532
2533 if (flags & RC_UNCOMMON) {
2534 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2535 // (An "if" probability corresponds roughly to an unconditional count.
2536 // Sort of.)
2537 call->set_cnt(PROB_UNLIKELY_MAG(4));
2538 }
2539
2540 Node* c = _gvn.transform(call);
2541 assert(c == call, "cannot disappear");
2542
2543 if (wide_out) {
2544 // Slow path call has full side-effects.
2545 set_predefined_output_for_runtime_call(call);
2546 } else {
2547 // Slow path call has few side-effects, and/or sets few values.
2548 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2549 }
2550
2551 if (has_io) {
2552 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2553 }
2554 return call;
2555
2556 }
2557
2558 //------------------------------merge_memory-----------------------------------
2559 // Merge memory from one path into the current memory state.
2560 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2561 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2562 Node* old_slice = mms.force_memory();
2563 Node* new_slice = mms.memory2();
2564 if (old_slice != new_slice) {
2565 PhiNode* phi;
2566 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2567 if (mms.is_empty()) {
2568 // clone base memory Phi's inputs for this memory slice
2569 assert(old_slice == mms.base_memory(), "sanity");
2570 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2571 _gvn.set_type(phi, Type::MEMORY);
2572 for (uint i = 1; i < phi->req(); i++) {
2573 phi->init_req(i, old_slice->in(i));
2574 }
2575 } else {
2576 phi = old_slice->as_Phi(); // Phi was generated already
2577 }
2578 } else {
2579 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2580 _gvn.set_type(phi, Type::MEMORY);
2581 }
2582 phi->set_req(new_path, new_slice);
2583 mms.set_memory(phi);
2584 }
2585 }
2586 }
2587
2588 //------------------------------make_slow_call_ex------------------------------
2589 // Make the exception handler hookups for the slow call
2590 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2591 if (stopped()) return;
2592
2593 // Make a catch node with just two handlers: fall-through and catch-all
2594 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2595 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2596 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2597 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2598
2599 { PreserveJVMState pjvms(this);
2600 set_control(excp);
2601 set_i_o(i_o);
2602
2603 if (excp != top()) {
2604 if (deoptimize) {
2605 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2606 uncommon_trap(Deoptimization::Reason_unhandled,
2607 Deoptimization::Action_none);
2608 } else {
2609 // Create an exception state also.
2610 // Use an exact type if the caller has specified a specific exception.
2611 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2612 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2613 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2614 }
2615 }
2616 }
2617
2618 // Get the no-exception control from the CatchNode.
2619 set_control(norm);
2620 }
2621
2622 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2623 Node* cmp = NULL;
2624 switch(bt) {
2625 case T_INT: cmp = new CmpINode(in1, in2); break;
2626 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2627 default: fatal("unexpected comparison type %s", type2name(bt));
2628 }
2629 gvn->transform(cmp);
2630 Node* bol = gvn->transform(new BoolNode(cmp, test));
2631 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2632 gvn->transform(iff);
2633 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2634 return iff;
2635 }
2636
2637
2638 //-------------------------------gen_subtype_check-----------------------------
2639 // Generate a subtyping check. Takes as input the subtype and supertype.
2640 // Returns 2 values: sets the default control() to the true path and returns
2641 // the false path. Only reads invariant memory; sets no (visible) memory.
2642 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2643 // but that's not exposed to the optimizer. This call also doesn't take in an
2644 // Object; if you wish to check an Object you need to load the Object's class
2645 // prior to coming here.
2646 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2647 Compile* C = gvn->C;
2648
2649 if ((*ctrl)->is_top()) {
2650 return C->top();
2651 }
2652
2653 // Fast check for identical types, perhaps identical constants.
2654 // The types can even be identical non-constants, in cases
2655 // involving Array.newInstance, Object.clone, etc.
2656 if (subklass == superklass)
2657 return C->top(); // false path is dead; no test needed.
2658
2659 if (gvn->type(superklass)->singleton()) {
2660 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2661 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2662
2663 // In the common case of an exact superklass, try to fold up the
2664 // test before generating code. You may ask, why not just generate
2665 // the code and then let it fold up? The answer is that the generated
2666 // code will necessarily include null checks, which do not always
2667 // completely fold away. If they are also needless, then they turn
2668 // into a performance loss. Example:
2669 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2670 // Here, the type of 'fa' is often exact, so the store check
2671 // of fa[1]=x will fold up, without testing the nullness of x.
2672 switch (C->static_subtype_check(superk, subk)) {
2673 case Compile::SSC_always_false:
2674 {
2675 Node* always_fail = *ctrl;
2676 *ctrl = gvn->C->top();
2677 return always_fail;
2678 }
2679 case Compile::SSC_always_true:
2680 return C->top();
2681 case Compile::SSC_easy_test:
2682 {
2683 // Just do a direct pointer compare and be done.
2684 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2685 *ctrl = gvn->transform(new IfTrueNode(iff));
2686 return gvn->transform(new IfFalseNode(iff));
2687 }
2688 case Compile::SSC_full_test:
2689 break;
2690 default:
2691 ShouldNotReachHere();
2692 }
2693 }
2694
2695 // %%% Possible further optimization: Even if the superklass is not exact,
2696 // if the subklass is the unique subtype of the superklass, the check
2697 // will always succeed. We could leave a dependency behind to ensure this.
2698
2699 // First load the super-klass's check-offset
2700 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2701 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2702 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2703 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2704 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2705
2706 // Load from the sub-klass's super-class display list, or a 1-word cache of
2707 // the secondary superclass list, or a failing value with a sentinel offset
2708 // if the super-klass is an interface or exceptionally deep in the Java
2709 // hierarchy and we have to scan the secondary superclass list the hard way.
2710 // Worst-case type is a little odd: NULL is allowed as a result (usually
2711 // klass loads can never produce a NULL).
2712 Node *chk_off_X = chk_off;
2713 #ifdef _LP64
2714 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2715 #endif
2716 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2717 // For some types like interfaces the following loadKlass is from a 1-word
2718 // cache which is mutable so can't use immutable memory. Other
2719 // types load from the super-class display table which is immutable.
2720 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2721 Node *kmem = might_be_cache ? m : C->immutable_memory();
2722 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2723
2724 // Compile speed common case: ARE a subtype and we canNOT fail
2725 if( superklass == nkls )
2726 return C->top(); // false path is dead; no test needed.
2727
2728 // See if we get an immediate positive hit. Happens roughly 83% of the
2729 // time. Test to see if the value loaded just previously from the subklass
2730 // is exactly the superklass.
2731 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2732 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2733 *ctrl = gvn->transform(new IfFalseNode(iff1));
2734
2735 // Compile speed common case: Check for being deterministic right now. If
2736 // chk_off is a constant and not equal to cacheoff then we are NOT a
2737 // subklass. In this case we need exactly the 1 test above and we can
2738 // return those results immediately.
2739 if (!might_be_cache) {
2740 Node* not_subtype_ctrl = *ctrl;
2741 *ctrl = iftrue1; // We need exactly the 1 test above
2742 return not_subtype_ctrl;
2743 }
2744
2745 // Gather the various success & failures here
2746 RegionNode *r_ok_subtype = new RegionNode(4);
2747 gvn->record_for_igvn(r_ok_subtype);
2748 RegionNode *r_not_subtype = new RegionNode(3);
2749 gvn->record_for_igvn(r_not_subtype);
2750
2751 r_ok_subtype->init_req(1, iftrue1);
2752
2753 // Check for immediate negative hit. Happens roughly 11% of the time (which
2754 // is roughly 63% of the remaining cases). Test to see if the loaded
2755 // check-offset points into the subklass display list or the 1-element
2756 // cache. If it points to the display (and NOT the cache) and the display
2757 // missed then it's not a subtype.
2758 Node *cacheoff = gvn->intcon(cacheoff_con);
2759 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2760 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2761 *ctrl = gvn->transform(new IfFalseNode(iff2));
2762
2763 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2764 // No performance impact (too rare) but allows sharing of secondary arrays
2765 // which has some footprint reduction.
2766 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2767 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2768 *ctrl = gvn->transform(new IfFalseNode(iff3));
2769
2770 // -- Roads not taken here: --
2771 // We could also have chosen to perform the self-check at the beginning
2772 // of this code sequence, as the assembler does. This would not pay off
2773 // the same way, since the optimizer, unlike the assembler, can perform
2774 // static type analysis to fold away many successful self-checks.
2775 // Non-foldable self checks work better here in second position, because
2776 // the initial primary superclass check subsumes a self-check for most
2777 // types. An exception would be a secondary type like array-of-interface,
2778 // which does not appear in its own primary supertype display.
2779 // Finally, we could have chosen to move the self-check into the
2780 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2781 // dependent manner. But it is worthwhile to have the check here,
2782 // where it can be perhaps be optimized. The cost in code space is
2783 // small (register compare, branch).
2784
2785 // Now do a linear scan of the secondary super-klass array. Again, no real
2786 // performance impact (too rare) but it's gotta be done.
2787 // Since the code is rarely used, there is no penalty for moving it
2788 // out of line, and it can only improve I-cache density.
2789 // The decision to inline or out-of-line this final check is platform
2790 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2791 Node* psc = gvn->transform(
2792 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2793
2794 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2795 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2796 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2797
2798 // Return false path; set default control to true path.
2799 *ctrl = gvn->transform(r_ok_subtype);
2800 return gvn->transform(r_not_subtype);
2801 }
2802
2803 // Profile-driven exact type check:
2804 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2805 float prob,
2806 Node* *casted_receiver) {
2807 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2808 Node* recv_klass = load_object_klass(receiver);
2809 Node* want_klass = makecon(tklass);
2810 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass) );
2811 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2812 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2813 set_control( _gvn.transform( new IfTrueNode (iff) ));
2814 Node* fail = _gvn.transform( new IfFalseNode(iff) );
2815
2816 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2817 assert(recv_xtype->klass_is_exact(), "");
2818
2819 // Subsume downstream occurrences of receiver with a cast to
2820 // recv_xtype, since now we know what the type will be.
2821 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2822 (*casted_receiver) = _gvn.transform(cast);
2823 // (User must make the replace_in_map call.)
2824
2825 return fail;
2826 }
2827
2828
2829 //------------------------------seems_never_null-------------------------------
2830 // Use null_seen information if it is available from the profile.
2831 // If we see an unexpected null at a type check we record it and force a
2832 // recompile; the offending check will be recompiled to handle NULLs.
2833 // If we see several offending BCIs, then all checks in the
2834 // method will be recompiled.
2835 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2836 speculating = !_gvn.type(obj)->speculative_maybe_null();
2837 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2838 if (UncommonNullCast // Cutout for this technique
2839 && obj != null() // And not the -Xcomp stupid case?
2840 && !too_many_traps(reason)
2841 ) {
2842 if (speculating) {
2843 return true;
2844 }
2845 if (data == NULL)
2846 // Edge case: no mature data. Be optimistic here.
2847 return true;
2848 // If the profile has not seen a null, assume it won't happen.
2849 assert(java_bc() == Bytecodes::_checkcast ||
2850 java_bc() == Bytecodes::_instanceof ||
2851 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2852 return !data->as_BitData()->null_seen();
2853 }
2854 speculating = false;
2855 return false;
2856 }
2857
2858 //------------------------maybe_cast_profiled_receiver-------------------------
2859 // If the profile has seen exactly one type, narrow to exactly that type.
2860 // Subsequent type checks will always fold up.
2861 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2862 ciKlass* require_klass,
2863 ciKlass* spec_klass,
2864 bool safe_for_replace) {
2865 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2866
2867 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
2868
2869 // Make sure we haven't already deoptimized from this tactic.
2870 if (too_many_traps(reason) || too_many_recompiles(reason))
2871 return NULL;
2872
2873 // (No, this isn't a call, but it's enough like a virtual call
2874 // to use the same ciMethod accessor to get the profile info...)
2875 // If we have a speculative type use it instead of profiling (which
2876 // may not help us)
2877 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2878 if (exact_kls != NULL) {// no cast failures here
2879 if (require_klass == NULL ||
2880 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
2881 // If we narrow the type to match what the type profile sees or
2882 // the speculative type, we can then remove the rest of the
2883 // cast.
2884 // This is a win, even if the exact_kls is very specific,
2885 // because downstream operations, such as method calls,
2886 // will often benefit from the sharper type.
2887 Node* exact_obj = not_null_obj; // will get updated in place...
2888 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2889 &exact_obj);
2890 { PreserveJVMState pjvms(this);
2891 set_control(slow_ctl);
2892 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2893 }
2894 if (safe_for_replace) {
2895 replace_in_map(not_null_obj, exact_obj);
2896 }
2897 return exact_obj;
2898 }
2899 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
2900 }
2901
2902 return NULL;
2903 }
2904
2905 /**
2906 * Cast obj to type and emit guard unless we had too many traps here
2907 * already
2908 *
2909 * @param obj node being casted
2910 * @param type type to cast the node to
2911 * @param not_null true if we know node cannot be null
2912 */
2913 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2914 ciKlass* type,
2915 bool not_null) {
2916 if (stopped()) {
2917 return obj;
2918 }
2919
2920 // type == NULL if profiling tells us this object is always null
2921 if (type != NULL) {
2922 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2923 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
2924
2925 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2926 !too_many_traps(class_reason) &&
2927 !too_many_recompiles(class_reason)) {
2928 Node* not_null_obj = NULL;
2929 // not_null is true if we know the object is not null and
2930 // there's no need for a null check
2931 if (!not_null) {
2932 Node* null_ctl = top();
2933 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
2934 assert(null_ctl->is_top(), "no null control here");
2935 } else {
2936 not_null_obj = obj;
2937 }
2938
2939 Node* exact_obj = not_null_obj;
2940 ciKlass* exact_kls = type;
2941 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2942 &exact_obj);
2943 {
2944 PreserveJVMState pjvms(this);
2945 set_control(slow_ctl);
2946 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2947 }
2948 replace_in_map(not_null_obj, exact_obj);
2949 obj = exact_obj;
2950 }
2951 } else {
2952 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2953 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2954 Node* exact_obj = null_assert(obj);
2955 replace_in_map(obj, exact_obj);
2956 obj = exact_obj;
2957 }
2958 }
2959 return obj;
2960 }
2961
2962 //-------------------------------gen_instanceof--------------------------------
2963 // Generate an instance-of idiom. Used by both the instance-of bytecode
2964 // and the reflective instance-of call.
2965 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2966 kill_dead_locals(); // Benefit all the uncommon traps
2967 assert( !stopped(), "dead parse path should be checked in callers" );
2968 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2969 "must check for not-null not-dead klass in callers");
2970
2971 // Make the merge point
2972 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2973 RegionNode* region = new RegionNode(PATH_LIMIT);
2974 Node* phi = new PhiNode(region, TypeInt::BOOL);
2975 C->set_has_split_ifs(true); // Has chance for split-if optimization
2976
2977 ciProfileData* data = NULL;
2978 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2979 data = method()->method_data()->bci_to_data(bci());
2980 }
2981 bool speculative_not_null = false;
2982 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2983 && seems_never_null(obj, data, speculative_not_null));
2984
2985 // Null check; get casted pointer; set region slot 3
2986 Node* null_ctl = top();
2987 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
2988
2989 // If not_null_obj is dead, only null-path is taken
2990 if (stopped()) { // Doing instance-of on a NULL?
2991 set_control(null_ctl);
2992 return intcon(0);
2993 }
2994 region->init_req(_null_path, null_ctl);
2995 phi ->init_req(_null_path, intcon(0)); // Set null path value
2996 if (null_ctl == top()) {
2997 // Do this eagerly, so that pattern matches like is_diamond_phi
2998 // will work even during parsing.
2999 assert(_null_path == PATH_LIMIT-1, "delete last");
3000 region->del_req(_null_path);
3001 phi ->del_req(_null_path);
3002 }
3003
3004 // Do we know the type check always succeed?
3005 bool known_statically = false;
3006 if (_gvn.type(superklass)->singleton()) {
3007 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3008 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3009 if (subk != NULL && subk->is_loaded()) {
3010 int static_res = C->static_subtype_check(superk, subk);
3011 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3012 }
3013 }
3014
3015 if (!known_statically) {
3016 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3017 // We may not have profiling here or it may not help us. If we
3018 // have a speculative type use it to perform an exact cast.
3019 ciKlass* spec_obj_type = obj_type->speculative_type();
3020 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3021 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3022 if (stopped()) { // Profile disagrees with this path.
3023 set_control(null_ctl); // Null is the only remaining possibility.
3024 return intcon(0);
3025 }
3026 if (cast_obj != NULL) {
3027 not_null_obj = cast_obj;
3028 }
3029 }
3030 }
3031
3032 // Load the object's klass
3033 Node* obj_klass = load_object_klass(not_null_obj);
3034
3035 // Generate the subtype check
3036 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3037
3038 // Plug in the success path to the general merge in slot 1.
3039 region->init_req(_obj_path, control());
3040 phi ->init_req(_obj_path, intcon(1));
3041
3042 // Plug in the failing path to the general merge in slot 2.
3043 region->init_req(_fail_path, not_subtype_ctrl);
3044 phi ->init_req(_fail_path, intcon(0));
3045
3046 // Return final merged results
3047 set_control( _gvn.transform(region) );
3048 record_for_igvn(region);
3049
3050 // If we know the type check always succeeds then we don't use the
3051 // profiling data at this bytecode. Don't lose it, feed it to the
3052 // type system as a speculative type.
3053 if (safe_for_replace) {
3054 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3055 replace_in_map(obj, casted_obj);
3056 }
3057
3058 return _gvn.transform(phi);
3059 }
3060
3061 //-------------------------------gen_checkcast---------------------------------
3062 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3063 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3064 // uncommon-trap paths work. Adjust stack after this call.
3065 // If failure_control is supplied and not null, it is filled in with
3066 // the control edge for the cast failure. Otherwise, an appropriate
3067 // uncommon trap or exception is thrown.
3068 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3069 Node* *failure_control) {
3070 kill_dead_locals(); // Benefit all the uncommon traps
3071 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
3072 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
3073
3074 // Fast cutout: Check the case that the cast is vacuously true.
3075 // This detects the common cases where the test will short-circuit
3076 // away completely. We do this before we perform the null check,
3077 // because if the test is going to turn into zero code, we don't
3078 // want a residual null check left around. (Causes a slowdown,
3079 // for example, in some objArray manipulations, such as a[i]=a[j].)
3080 if (tk->singleton()) {
3081 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3082 if (objtp != NULL && objtp->klass() != NULL) {
3083 switch (C->static_subtype_check(tk->klass(), objtp->klass())) {
3084 case Compile::SSC_always_true:
3085 // If we know the type check always succeed then we don't use
3086 // the profiling data at this bytecode. Don't lose it, feed it
3087 // to the type system as a speculative type.
3088 return record_profiled_receiver_for_speculation(obj);
3089 case Compile::SSC_always_false:
3090 // It needs a null check because a null will *pass* the cast check.
3091 // A non-null value will always produce an exception.
3092 return null_assert(obj);
3093 }
3094 }
3095 }
3096
3097 ciProfileData* data = NULL;
3098 bool safe_for_replace = false;
3099 if (failure_control == NULL) { // use MDO in regular case only
3100 assert(java_bc() == Bytecodes::_aastore ||
3101 java_bc() == Bytecodes::_checkcast,
3102 "interpreter profiles type checks only for these BCs");
3103 data = method()->method_data()->bci_to_data(bci());
3104 safe_for_replace = true;
3105 }
3106
3107 // Make the merge point
3108 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3109 RegionNode* region = new RegionNode(PATH_LIMIT);
3110 Node* phi = new PhiNode(region, toop);
3111 C->set_has_split_ifs(true); // Has chance for split-if optimization
3112
3113 // Use null-cast information if it is available
3114 bool speculative_not_null = false;
3115 bool never_see_null = ((failure_control == NULL) // regular case only
3116 && seems_never_null(obj, data, speculative_not_null));
3117
3118 // Null check; get casted pointer; set region slot 3
3119 Node* null_ctl = top();
3120 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3121
3122 // If not_null_obj is dead, only null-path is taken
3123 if (stopped()) { // Doing instance-of on a NULL?
3124 set_control(null_ctl);
3125 return null();
3126 }
3127 region->init_req(_null_path, null_ctl);
3128 phi ->init_req(_null_path, null()); // Set null path value
3129 if (null_ctl == top()) {
3130 // Do this eagerly, so that pattern matches like is_diamond_phi
3131 // will work even during parsing.
3132 assert(_null_path == PATH_LIMIT-1, "delete last");
3133 region->del_req(_null_path);
3134 phi ->del_req(_null_path);
3135 }
3136
3137 Node* cast_obj = NULL;
3138 if (tk->klass_is_exact()) {
3139 // The following optimization tries to statically cast the speculative type of the object
3140 // (for example obtained during profiling) to the type of the superklass and then do a
3141 // dynamic check that the type of the object is what we expect. To work correctly
3142 // for checkcast and aastore the type of superklass should be exact.
3143 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3144 // We may not have profiling here or it may not help us. If we have
3145 // a speculative type use it to perform an exact cast.
3146 ciKlass* spec_obj_type = obj_type->speculative_type();
3147 if (spec_obj_type != NULL || data != NULL) {
3148 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3149 if (cast_obj != NULL) {
3150 if (failure_control != NULL) // failure is now impossible
3151 (*failure_control) = top();
3152 // adjust the type of the phi to the exact klass:
3153 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3154 }
3155 }
3156 }
3157
3158 if (cast_obj == NULL) {
3159 // Load the object's klass
3160 Node* obj_klass = load_object_klass(not_null_obj);
3161
3162 // Generate the subtype check
3163 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3164
3165 // Plug in success path into the merge
3166 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3167 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3168 if (failure_control == NULL) {
3169 if (not_subtype_ctrl != top()) { // If failure is possible
3170 PreserveJVMState pjvms(this);
3171 set_control(not_subtype_ctrl);
3172 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3173 }
3174 } else {
3175 (*failure_control) = not_subtype_ctrl;
3176 }
3177 }
3178
3179 region->init_req(_obj_path, control());
3180 phi ->init_req(_obj_path, cast_obj);
3181
3182 // A merge of NULL or Casted-NotNull obj
3183 Node* res = _gvn.transform(phi);
3184
3185 // Note I do NOT always 'replace_in_map(obj,result)' here.
3186 // if( tk->klass()->can_be_primary_super() )
3187 // This means that if I successfully store an Object into an array-of-String
3188 // I 'forget' that the Object is really now known to be a String. I have to
3189 // do this because we don't have true union types for interfaces - if I store
3190 // a Baz into an array-of-Interface and then tell the optimizer it's an
3191 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3192 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3193 // replace_in_map( obj, res );
3194
3195 // Return final merged results
3196 set_control( _gvn.transform(region) );
3197 record_for_igvn(region);
3198
3199 return record_profiled_receiver_for_speculation(res);
3200 }
3201
3202 //------------------------------next_monitor-----------------------------------
3203 // What number should be given to the next monitor?
3204 int GraphKit::next_monitor() {
3205 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3206 int next = current + C->sync_stack_slots();
3207 // Keep the toplevel high water mark current:
3208 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3209 return current;
3210 }
3211
3212 //------------------------------insert_mem_bar---------------------------------
3213 // Memory barrier to avoid floating things around
3214 // The membar serves as a pinch point between both control and all memory slices.
3215 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3216 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3217 mb->init_req(TypeFunc::Control, control());
3218 mb->init_req(TypeFunc::Memory, reset_memory());
3219 Node* membar = _gvn.transform(mb);
3220 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3221 set_all_memory_call(membar);
3222 return membar;
3223 }
3224
3225 //-------------------------insert_mem_bar_volatile----------------------------
3226 // Memory barrier to avoid floating things around
3227 // The membar serves as a pinch point between both control and memory(alias_idx).
3228 // If you want to make a pinch point on all memory slices, do not use this
3229 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3230 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3231 // When Parse::do_put_xxx updates a volatile field, it appends a series
3232 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3233 // The first membar is on the same memory slice as the field store opcode.
3234 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3235 // All the other membars (for other volatile slices, including AliasIdxBot,
3236 // which stands for all unknown volatile slices) are control-dependent
3237 // on the first membar. This prevents later volatile loads or stores
3238 // from sliding up past the just-emitted store.
3239
3240 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3241 mb->set_req(TypeFunc::Control,control());
3242 if (alias_idx == Compile::AliasIdxBot) {
3243 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3244 } else {
3245 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3246 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3247 }
3248 Node* membar = _gvn.transform(mb);
3249 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3250 if (alias_idx == Compile::AliasIdxBot) {
3251 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3252 } else {
3253 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3254 }
3255 return membar;
3256 }
3257
3258 //------------------------------shared_lock------------------------------------
3259 // Emit locking code.
3260 FastLockNode* GraphKit::shared_lock(Node* obj) {
3261 // bci is either a monitorenter bc or InvocationEntryBci
3262 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3263 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3264
3265 if( !GenerateSynchronizationCode )
3266 return NULL; // Not locking things?
3267 if (stopped()) // Dead monitor?
3268 return NULL;
3269
3270 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3271
3272 obj = shenandoah_write_barrier(obj);
3273
3274 // Box the stack location
3275 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3276 Node* mem = reset_memory();
3277
3278 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3279 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3280 // Create the counters for this fast lock.
3281 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3282 }
3283
3284 // Create the rtm counters for this fast lock if needed.
3285 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3286
3287 // Add monitor to debug info for the slow path. If we block inside the
3288 // slow path and de-opt, we need the monitor hanging around
3289 map()->push_monitor( flock );
3290
3291 const TypeFunc *tf = LockNode::lock_type();
3292 LockNode *lock = new LockNode(C, tf);
3293
3294 lock->init_req( TypeFunc::Control, control() );
3295 lock->init_req( TypeFunc::Memory , mem );
3296 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3297 lock->init_req( TypeFunc::FramePtr, frameptr() );
3298 lock->init_req( TypeFunc::ReturnAdr, top() );
3299
3300 lock->init_req(TypeFunc::Parms + 0, obj);
3301 lock->init_req(TypeFunc::Parms + 1, box);
3302 lock->init_req(TypeFunc::Parms + 2, flock);
3303 add_safepoint_edges(lock);
3304
3305 lock = _gvn.transform( lock )->as_Lock();
3306
3307 // lock has no side-effects, sets few values
3308 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3309
3310 insert_mem_bar(Op_MemBarAcquireLock);
3311
3312 // Add this to the worklist so that the lock can be eliminated
3313 record_for_igvn(lock);
3314
3315 #ifndef PRODUCT
3316 if (PrintLockStatistics) {
3317 // Update the counter for this lock. Don't bother using an atomic
3318 // operation since we don't require absolute accuracy.
3319 lock->create_lock_counter(map()->jvms());
3320 increment_counter(lock->counter()->addr());
3321 }
3322 #endif
3323
3324 return flock;
3325 }
3326
3327
3328 //------------------------------shared_unlock----------------------------------
3329 // Emit unlocking code.
3330 void GraphKit::shared_unlock(Node* box, Node* obj) {
3331 // bci is either a monitorenter bc or InvocationEntryBci
3332 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3333 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3334
3335 if( !GenerateSynchronizationCode )
3336 return;
3337 if (stopped()) { // Dead monitor?
3338 map()->pop_monitor(); // Kill monitor from debug info
3339 return;
3340 }
3341
3342 // Memory barrier to avoid floating things down past the locked region
3343 insert_mem_bar(Op_MemBarReleaseLock);
3344
3345 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3346 UnlockNode *unlock = new UnlockNode(C, tf);
3347 #ifdef ASSERT
3348 unlock->set_dbg_jvms(sync_jvms());
3349 #endif
3350 uint raw_idx = Compile::AliasIdxRaw;
3351 unlock->init_req( TypeFunc::Control, control() );
3352 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3353 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3354 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3355 unlock->init_req( TypeFunc::ReturnAdr, top() );
3356
3357 unlock->init_req(TypeFunc::Parms + 0, obj);
3358 unlock->init_req(TypeFunc::Parms + 1, box);
3359 unlock = _gvn.transform(unlock)->as_Unlock();
3360
3361 Node* mem = reset_memory();
3362
3363 // unlock has no side-effects, sets few values
3364 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3365
3366 // Kill monitor from debug info
3367 map()->pop_monitor( );
3368 }
3369
3370 //-------------------------------get_layout_helper-----------------------------
3371 // If the given klass is a constant or known to be an array,
3372 // fetch the constant layout helper value into constant_value
3373 // and return (Node*)NULL. Otherwise, load the non-constant
3374 // layout helper value, and return the node which represents it.
3375 // This two-faced routine is useful because allocation sites
3376 // almost always feature constant types.
3377 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3378 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3379 if (!StressReflectiveCode && inst_klass != NULL) {
3380 ciKlass* klass = inst_klass->klass();
3381 bool xklass = inst_klass->klass_is_exact();
3382 if (xklass || klass->is_array_klass()) {
3383 jint lhelper = klass->layout_helper();
3384 if (lhelper != Klass::_lh_neutral_value) {
3385 constant_value = lhelper;
3386 return (Node*) NULL;
3387 }
3388 }
3389 }
3390 constant_value = Klass::_lh_neutral_value; // put in a known value
3391 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3392 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3393 }
3394
3395 // We just put in an allocate/initialize with a big raw-memory effect.
3396 // Hook selected additional alias categories on the initialization.
3397 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3398 MergeMemNode* init_in_merge,
3399 Node* init_out_raw) {
3400 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3401 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3402
3403 Node* prevmem = kit.memory(alias_idx);
3404 init_in_merge->set_memory_at(alias_idx, prevmem);
3405 kit.set_memory(init_out_raw, alias_idx);
3406 }
3407
3408 //---------------------------set_output_for_allocation-------------------------
3409 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3410 const TypeOopPtr* oop_type,
3411 bool deoptimize_on_exception) {
3412 int rawidx = Compile::AliasIdxRaw;
3413 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3414 add_safepoint_edges(alloc);
3415 Node* allocx = _gvn.transform(alloc);
3416 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3417 // create memory projection for i_o
3418 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3419 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3420
3421 // create a memory projection as for the normal control path
3422 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3423 set_memory(malloc, rawidx);
3424
3425 // a normal slow-call doesn't change i_o, but an allocation does
3426 // we create a separate i_o projection for the normal control path
3427 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3428 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3429
3430 // put in an initialization barrier
3431 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3432 rawoop)->as_Initialize();
3433 assert(alloc->initialization() == init, "2-way macro link must work");
3434 assert(init ->allocation() == alloc, "2-way macro link must work");
3435 {
3436 // Extract memory strands which may participate in the new object's
3437 // initialization, and source them from the new InitializeNode.
3438 // This will allow us to observe initializations when they occur,
3439 // and link them properly (as a group) to the InitializeNode.
3440 assert(init->in(InitializeNode::Memory) == malloc, "");
3441 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3442 init->set_req(InitializeNode::Memory, minit_in);
3443 record_for_igvn(minit_in); // fold it up later, if possible
3444 Node* minit_out = memory(rawidx);
3445 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3446 if (oop_type->isa_aryptr()) {
3447 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3448 int elemidx = C->get_alias_index(telemref);
3449 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3450 } else if (oop_type->isa_instptr()) {
3451 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3452 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3453 ciField* field = ik->nonstatic_field_at(i);
3454 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3455 continue; // do not bother to track really large numbers of fields
3456 // Find (or create) the alias category for this field:
3457 int fieldidx = C->alias_type(field)->index();
3458 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3459 }
3460 }
3461 }
3462
3463 // Cast raw oop to the real thing...
3464 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3465 javaoop = _gvn.transform(javaoop);
3466 C->set_recent_alloc(control(), javaoop);
3467 assert(just_allocated_object(control()) == javaoop, "just allocated");
3468
3469 #ifdef ASSERT
3470 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3471 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3472 "Ideal_allocation works");
3473 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3474 "Ideal_allocation works");
3475 if (alloc->is_AllocateArray()) {
3476 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3477 "Ideal_allocation works");
3478 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3479 "Ideal_allocation works");
3480 } else {
3481 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3482 }
3483 }
3484 #endif //ASSERT
3485
3486 return javaoop;
3487 }
3488
3489 //---------------------------new_instance--------------------------------------
3490 // This routine takes a klass_node which may be constant (for a static type)
3491 // or may be non-constant (for reflective code). It will work equally well
3492 // for either, and the graph will fold nicely if the optimizer later reduces
3493 // the type to a constant.
3494 // The optional arguments are for specialized use by intrinsics:
3495 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3496 // - If 'return_size_val', report the the total object size to the caller.
3497 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3498 Node* GraphKit::new_instance(Node* klass_node,
3499 Node* extra_slow_test,
3500 Node* *return_size_val,
3501 bool deoptimize_on_exception) {
3502 // Compute size in doublewords
3503 // The size is always an integral number of doublewords, represented
3504 // as a positive bytewise size stored in the klass's layout_helper.
3505 // The layout_helper also encodes (in a low bit) the need for a slow path.
3506 jint layout_con = Klass::_lh_neutral_value;
3507 Node* layout_val = get_layout_helper(klass_node, layout_con);
3508 int layout_is_con = (layout_val == NULL);
3509
3510 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3511 // Generate the initial go-slow test. It's either ALWAYS (return a
3512 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3513 // case) a computed value derived from the layout_helper.
3514 Node* initial_slow_test = NULL;
3515 if (layout_is_con) {
3516 assert(!StressReflectiveCode, "stress mode does not use these paths");
3517 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3518 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3519 } else { // reflective case
3520 // This reflective path is used by Unsafe.allocateInstance.
3521 // (It may be stress-tested by specifying StressReflectiveCode.)
3522 // Basically, we want to get into the VM is there's an illegal argument.
3523 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3524 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3525 if (extra_slow_test != intcon(0)) {
3526 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3527 }
3528 // (Macro-expander will further convert this to a Bool, if necessary.)
3529 }
3530
3531 // Find the size in bytes. This is easy; it's the layout_helper.
3532 // The size value must be valid even if the slow path is taken.
3533 Node* size = NULL;
3534 if (layout_is_con) {
3535 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3536 } else { // reflective case
3537 // This reflective path is used by clone and Unsafe.allocateInstance.
3538 size = ConvI2X(layout_val);
3539
3540 // Clear the low bits to extract layout_helper_size_in_bytes:
3541 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3542 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3543 size = _gvn.transform( new AndXNode(size, mask) );
3544 }
3545 if (return_size_val != NULL) {
3546 (*return_size_val) = size;
3547 }
3548
3549 // This is a precise notnull oop of the klass.
3550 // (Actually, it need not be precise if this is a reflective allocation.)
3551 // It's what we cast the result to.
3552 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3553 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3554 const TypeOopPtr* oop_type = tklass->as_instance_type();
3555
3556 // Now generate allocation code
3557
3558 // The entire memory state is needed for slow path of the allocation
3559 // since GC and deoptimization can happened.
3560 Node *mem = reset_memory();
3561 set_all_memory(mem); // Create new memory state
3562
3563 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3564 control(), mem, i_o(),
3565 size, klass_node,
3566 initial_slow_test);
3567
3568 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3569 }
3570
3571 //-------------------------------new_array-------------------------------------
3572 // helper for both newarray and anewarray
3573 // The 'length' parameter is (obviously) the length of the array.
3574 // See comments on new_instance for the meaning of the other arguments.
3575 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3576 Node* length, // number of array elements
3577 int nargs, // number of arguments to push back for uncommon trap
3578 Node* *return_size_val,
3579 bool deoptimize_on_exception) {
3580 jint layout_con = Klass::_lh_neutral_value;
3581 Node* layout_val = get_layout_helper(klass_node, layout_con);
3582 int layout_is_con = (layout_val == NULL);
3583
3584 if (!layout_is_con && !StressReflectiveCode &&
3585 !too_many_traps(Deoptimization::Reason_class_check)) {
3586 // This is a reflective array creation site.
3587 // Optimistically assume that it is a subtype of Object[],
3588 // so that we can fold up all the address arithmetic.
3589 layout_con = Klass::array_layout_helper(T_OBJECT);
3590 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3591 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3592 { BuildCutout unless(this, bol_lh, PROB_MAX);
3593 inc_sp(nargs);
3594 uncommon_trap(Deoptimization::Reason_class_check,
3595 Deoptimization::Action_maybe_recompile);
3596 }
3597 layout_val = NULL;
3598 layout_is_con = true;
3599 }
3600
3601 // Generate the initial go-slow test. Make sure we do not overflow
3602 // if length is huge (near 2Gig) or negative! We do not need
3603 // exact double-words here, just a close approximation of needed
3604 // double-words. We can't add any offset or rounding bits, lest we
3605 // take a size -1 of bytes and make it positive. Use an unsigned
3606 // compare, so negative sizes look hugely positive.
3607 int fast_size_limit = FastAllocateSizeLimit;
3608 if (layout_is_con) {
3609 assert(!StressReflectiveCode, "stress mode does not use these paths");
3610 // Increase the size limit if we have exact knowledge of array type.
3611 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3612 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3613 }
3614
3615 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3616 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3617
3618 // --- Size Computation ---
3619 // array_size = round_to_heap(array_header + (length << elem_shift));
3620 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3621 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3622 // The rounding mask is strength-reduced, if possible.
3623 int round_mask = MinObjAlignmentInBytes - 1;
3624 Node* header_size = NULL;
3625 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3626 // (T_BYTE has the weakest alignment and size restrictions...)
3627 if (layout_is_con) {
3628 int hsize = Klass::layout_helper_header_size(layout_con);
3629 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3630 BasicType etype = Klass::layout_helper_element_type(layout_con);
3631 if ((round_mask & ~right_n_bits(eshift)) == 0)
3632 round_mask = 0; // strength-reduce it if it goes away completely
3633 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3634 assert(header_size_min <= hsize, "generic minimum is smallest");
3635 header_size_min = hsize;
3636 header_size = intcon(hsize + round_mask);
3637 } else {
3638 Node* hss = intcon(Klass::_lh_header_size_shift);
3639 Node* hsm = intcon(Klass::_lh_header_size_mask);
3640 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
3641 hsize = _gvn.transform( new AndINode(hsize, hsm) );
3642 Node* mask = intcon(round_mask);
3643 header_size = _gvn.transform( new AddINode(hsize, mask) );
3644 }
3645
3646 Node* elem_shift = NULL;
3647 if (layout_is_con) {
3648 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3649 if (eshift != 0)
3650 elem_shift = intcon(eshift);
3651 } else {
3652 // There is no need to mask or shift this value.
3653 // The semantics of LShiftINode include an implicit mask to 0x1F.
3654 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3655 elem_shift = layout_val;
3656 }
3657
3658 // Transition to native address size for all offset calculations:
3659 Node* lengthx = ConvI2X(length);
3660 Node* headerx = ConvI2X(header_size);
3661 #ifdef _LP64
3662 { const TypeInt* tilen = _gvn.find_int_type(length);
3663 if (tilen != NULL && tilen->_lo < 0) {
3664 // Add a manual constraint to a positive range. Cf. array_element_address.
3665 jint size_max = fast_size_limit;
3666 if (size_max > tilen->_hi) size_max = tilen->_hi;
3667 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3668
3669 // Only do a narrow I2L conversion if the range check passed.
3670 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3671 _gvn.transform(iff);
3672 RegionNode* region = new RegionNode(3);
3673 _gvn.set_type(region, Type::CONTROL);
3674 lengthx = new PhiNode(region, TypeLong::LONG);
3675 _gvn.set_type(lengthx, TypeLong::LONG);
3676
3677 // Range check passed. Use ConvI2L node with narrow type.
3678 Node* passed = IfFalse(iff);
3679 region->init_req(1, passed);
3680 // Make I2L conversion control dependent to prevent it from
3681 // floating above the range check during loop optimizations.
3682 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3683
3684 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3685 region->init_req(2, IfTrue(iff));
3686 lengthx->init_req(2, ConvI2X(length));
3687
3688 set_control(region);
3689 record_for_igvn(region);
3690 record_for_igvn(lengthx);
3691 }
3692 }
3693 #endif
3694
3695 // Combine header size (plus rounding) and body size. Then round down.
3696 // This computation cannot overflow, because it is used only in two
3697 // places, one where the length is sharply limited, and the other
3698 // after a successful allocation.
3699 Node* abody = lengthx;
3700 if (elem_shift != NULL)
3701 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
3702 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
3703 if (round_mask != 0) {
3704 Node* mask = MakeConX(~round_mask);
3705 size = _gvn.transform( new AndXNode(size, mask) );
3706 }
3707 // else if round_mask == 0, the size computation is self-rounding
3708
3709 if (return_size_val != NULL) {
3710 // This is the size
3711 (*return_size_val) = size;
3712 }
3713
3714 // Now generate allocation code
3715
3716 // The entire memory state is needed for slow path of the allocation
3717 // since GC and deoptimization can happened.
3718 Node *mem = reset_memory();
3719 set_all_memory(mem); // Create new memory state
3720
3721 if (initial_slow_test->is_Bool()) {
3722 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3723 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3724 }
3725
3726 // Create the AllocateArrayNode and its result projections
3727 AllocateArrayNode* alloc
3728 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3729 control(), mem, i_o(),
3730 size, klass_node,
3731 initial_slow_test,
3732 length);
3733
3734 // Cast to correct type. Note that the klass_node may be constant or not,
3735 // and in the latter case the actual array type will be inexact also.
3736 // (This happens via a non-constant argument to inline_native_newArray.)
3737 // In any case, the value of klass_node provides the desired array type.
3738 const TypeInt* length_type = _gvn.find_int_type(length);
3739 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3740 if (ary_type->isa_aryptr() && length_type != NULL) {
3741 // Try to get a better type than POS for the size
3742 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3743 }
3744
3745 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3746
3747 // Cast length on remaining path to be as narrow as possible
3748 if (map()->find_edge(length) >= 0) {
3749 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3750 if (ccast != length) {
3751 _gvn.set_type_bottom(ccast);
3752 record_for_igvn(ccast);
3753 replace_in_map(length, ccast);
3754 }
3755 }
3756
3757 return javaoop;
3758 }
3759
3760 // The following "Ideal_foo" functions are placed here because they recognize
3761 // the graph shapes created by the functions immediately above.
3762
3763 //---------------------------Ideal_allocation----------------------------------
3764 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3765 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3766 if (ptr == NULL) { // reduce dumb test in callers
3767 return NULL;
3768 }
3769
3770 // Attempt to see through Shenandoah barriers.
3771 ptr = ShenandoahBarrierNode::skip_through_barrier(ptr);
3772
3773 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3774 ptr = ptr->in(1);
3775 if (ptr == NULL) return NULL;
3776 }
3777 // Return NULL for allocations with several casts:
3778 // j.l.reflect.Array.newInstance(jobject, jint)
3779 // Object.clone()
3780 // to keep more precise type from last cast.
3781 if (ptr->is_Proj()) {
3782 Node* allo = ptr->in(0);
3783 if (allo != NULL && allo->is_Allocate()) {
3784 return allo->as_Allocate();
3785 }
3786 }
3787 // Report failure to match.
3788 return NULL;
3789 }
3790
3791 // Fancy version which also strips off an offset (and reports it to caller).
3792 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3793 intptr_t& offset) {
3794 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3795 if (base == NULL) return NULL;
3796 return Ideal_allocation(base, phase);
3797 }
3798
3799 // Trace Initialize <- Proj[Parm] <- Allocate
3800 AllocateNode* InitializeNode::allocation() {
3801 Node* rawoop = in(InitializeNode::RawAddress);
3802 if (rawoop->is_Proj()) {
3803 Node* alloc = rawoop->in(0);
3804 if (alloc->is_Allocate()) {
3805 return alloc->as_Allocate();
3806 }
3807 }
3808 return NULL;
3809 }
3810
3811 // Trace Allocate -> Proj[Parm] -> Initialize
3812 InitializeNode* AllocateNode::initialization() {
3813 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3814 if (rawoop == NULL) return NULL;
3815 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3816 Node* init = rawoop->fast_out(i);
3817 if (init->is_Initialize()) {
3818 assert(init->as_Initialize()->allocation() == this, "2-way link");
3819 return init->as_Initialize();
3820 }
3821 }
3822 return NULL;
3823 }
3824
3825 //----------------------------- loop predicates ---------------------------
3826
3827 //------------------------------add_predicate_impl----------------------------
3828 bool GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3829 // Too many traps seen?
3830 if (too_many_traps(reason)) {
3831 #ifdef ASSERT
3832 if (TraceLoopPredicate) {
3833 int tc = C->trap_count(reason);
3834 tty->print("too many traps=%s tcount=%d in ",
3835 Deoptimization::trap_reason_name(reason), tc);
3836 method()->print(); // which method has too many predicate traps
3837 tty->cr();
3838 }
3839 #endif
3840 // We cannot afford to take more traps here,
3841 // do not generate predicate.
3842 return false;
3843 }
3844
3845 Node *cont = _gvn.intcon(1);
3846 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
3847 Node *bol = _gvn.transform(new Conv2BNode(opq));
3848 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3849 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
3850 C->add_predicate_opaq(opq);
3851 {
3852 PreserveJVMState pjvms(this);
3853 set_control(iffalse);
3854 inc_sp(nargs);
3855 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3856 }
3857 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
3858 set_control(iftrue);
3859 return true;
3860 }
3861
3862 //------------------------------add_predicate---------------------------------
3863 void GraphKit::add_predicate(int nargs) {
3864 bool added = false;
3865 if (UseProfiledLoopPredicate) {
3866 added = add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
3867 }
3868 if (!added && UseLoopPredicate) {
3869 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3870 }
3871 // loop's limit check predicate should be near the loop.
3872 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3873 }
3874
3875 //----------------------------- store barriers ----------------------------
3876 #define __ ideal.
3877
3878 void GraphKit::sync_kit(IdealKit& ideal) {
3879 set_all_memory(__ merged_memory());
3880 set_i_o(__ i_o());
3881 set_control(__ ctrl());
3882 }
3883
3884 void GraphKit::final_sync(IdealKit& ideal) {
3885 // Final sync IdealKit and graphKit.
3886 sync_kit(ideal);
3887 }
3888
3889 Node* GraphKit::byte_map_base_node() {
3890 // Get base of card map
3891 CardTableModRefBS* ct =
3892 barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
3893 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust users of this code");
3894 if (ct->byte_map_base != NULL) {
3895 return makecon(TypeRawPtr::make((address)ct->byte_map_base));
3896 } else {
3897 return null();
3898 }
3899 }
3900
3901 // vanilla/CMS post barrier
3902 // Insert a write-barrier store. This is to let generational GC work; we have
3903 // to flag all oop-stores before the next GC point.
3904 void GraphKit::write_barrier_post(Node* oop_store,
3905 Node* obj,
3906 Node* adr,
3907 uint adr_idx,
3908 Node* val,
3909 bool use_precise) {
3910 // No store check needed if we're storing a NULL or an old object
3911 // (latter case is probably a string constant). The concurrent
3912 // mark sweep garbage collector, however, needs to have all nonNull
3913 // oop updates flagged via card-marks.
3914 if (val != NULL && val->is_Con()) {
3915 // must be either an oop or NULL
3916 const Type* t = val->bottom_type();
3917 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3918 // stores of null never (?) need barriers
3919 return;
3920 }
3921
3922 if (use_ReduceInitialCardMarks()
3923 && obj == just_allocated_object(control())) {
3924 // We can skip marks on a freshly-allocated object in Eden.
3925 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3926 // That routine informs GC to take appropriate compensating steps,
3927 // upon a slow-path allocation, so as to make this card-mark
3928 // elision safe.
3929 return;
3930 }
3931
3932 if (!use_precise) {
3933 // All card marks for a (non-array) instance are in one place:
3934 adr = obj;
3935 }
3936 // (Else it's an array (or unknown), and we want more precise card marks.)
3937 assert(adr != NULL, "");
3938
3939 IdealKit ideal(this, true);
3940
3941 // Convert the pointer to an int prior to doing math on it
3942 Node* cast = __ CastPX(__ ctrl(), adr);
3943
3944 // Divide by card size
3945 assert(Universe::heap()->barrier_set()->is_a(BarrierSet::CardTableModRef),
3946 "Only one we handle so far.");
3947 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3948
3949 // Combine card table base and card offset
3950 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3951
3952 // Get the alias_index for raw card-mark memory
3953 int adr_type = Compile::AliasIdxRaw;
3954 Node* zero = __ ConI(0); // Dirty card value
3955 BasicType bt = T_BYTE;
3956
3957 if (UseConcMarkSweepGC && UseCondCardMark) {
3958 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
3959 __ sync_kit(this);
3960 }
3961
3962 if (UseCondCardMark) {
3963 // The classic GC reference write barrier is typically implemented
3964 // as a store into the global card mark table. Unfortunately
3965 // unconditional stores can result in false sharing and excessive
3966 // coherence traffic as well as false transactional aborts.
3967 // UseCondCardMark enables MP "polite" conditional card mark
3968 // stores. In theory we could relax the load from ctrl() to
3969 // no_ctrl, but that doesn't buy much latitude.
3970 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3971 __ if_then(card_val, BoolTest::ne, zero);
3972 }
3973
3974 // Smash zero into card
3975 if( !UseConcMarkSweepGC ) {
3976 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3977 } else {
3978 // Specialized path for CM store barrier
3979 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3980 }
3981
3982 if (UseCondCardMark) {
3983 __ end_if();
3984 }
3985
3986 // Final sync IdealKit and GraphKit.
3987 final_sync(ideal);
3988 }
3989 /*
3990 * Determine if the G1 pre-barrier can be removed. The pre-barrier is
3991 * required by SATB to make sure all objects live at the start of the
3992 * marking are kept alive, all reference updates need to any previous
3993 * reference stored before writing.
3994 *
3995 * If the previous value is NULL there is no need to save the old value.
3996 * References that are NULL are filtered during runtime by the barrier
3997 * code to avoid unnecessary queuing.
3998 *
3999 * However in the case of newly allocated objects it might be possible to
4000 * prove that the reference about to be overwritten is NULL during compile
4001 * time and avoid adding the barrier code completely.
4002 *
4003 * The compiler needs to determine that the object in which a field is about
4004 * to be written is newly allocated, and that no prior store to the same field
4005 * has happened since the allocation.
4006 *
4007 * Returns true if the pre-barrier can be removed
4008 */
4009 bool GraphKit::g1_can_remove_pre_barrier(PhaseTransform* phase, Node* adr,
4010 BasicType bt, uint adr_idx) {
4011 intptr_t offset = 0;
4012 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4013 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4014
4015 if (offset == Type::OffsetBot) {
4016 return false; // cannot unalias unless there are precise offsets
4017 }
4018
4019 if (alloc == NULL) {
4020 return false; // No allocation found
4021 }
4022
4023 intptr_t size_in_bytes = type2aelembytes(bt);
4024
4025 Node* mem = memory(adr_idx); // start searching here...
4026
4027 for (int cnt = 0; cnt < 50; cnt++) {
4028
4029 if (mem->is_Store()) {
4030
4031 Node* st_adr = mem->in(MemNode::Address);
4032 intptr_t st_offset = 0;
4033 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
4034
4035 if (st_base == NULL) {
4036 break; // inscrutable pointer
4037 }
4038
4039 // Break we have found a store with same base and offset as ours so break
4040 if (st_base == base && st_offset == offset) {
4041 break;
4042 }
4043
4044 if (st_offset != offset && st_offset != Type::OffsetBot) {
4045 const int MAX_STORE = BytesPerLong;
4046 if (st_offset >= offset + size_in_bytes ||
4047 st_offset <= offset - MAX_STORE ||
4048 st_offset <= offset - mem->as_Store()->memory_size()) {
4049 // Success: The offsets are provably independent.
4050 // (You may ask, why not just test st_offset != offset and be done?
4051 // The answer is that stores of different sizes can co-exist
4052 // in the same sequence of RawMem effects. We sometimes initialize
4053 // a whole 'tile' of array elements with a single jint or jlong.)
4054 mem = mem->in(MemNode::Memory);
4055 continue; // advance through independent store memory
4056 }
4057 }
4058
4059 if (st_base != base
4060 && MemNode::detect_ptr_independence(base, alloc, st_base,
4061 AllocateNode::Ideal_allocation(st_base, phase),
4062 phase)) {
4063 // Success: The bases are provably independent.
4064 mem = mem->in(MemNode::Memory);
4065 continue; // advance through independent store memory
4066 }
4067 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4068
4069 InitializeNode* st_init = mem->in(0)->as_Initialize();
4070 AllocateNode* st_alloc = st_init->allocation();
4071
4072 // Make sure that we are looking at the same allocation site.
4073 // The alloc variable is guaranteed to not be null here from earlier check.
4074 if (alloc == st_alloc) {
4075 // Check that the initialization is storing NULL so that no previous store
4076 // has been moved up and directly write a reference
4077 Node* captured_store = st_init->find_captured_store(offset,
4078 type2aelembytes(T_OBJECT),
4079 phase);
4080 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
4081 return true;
4082 }
4083 }
4084 }
4085
4086 // Unless there is an explicit 'continue', we must bail out here,
4087 // because 'mem' is an inscrutable memory state (e.g., a call).
4088 break;
4089 }
4090
4091 return false;
4092 }
4093
4094 static void g1_write_barrier_pre_helper(const GraphKit& kit, Node* adr) {
4095 if (UseShenandoahGC && (ShenandoahSATBBarrier || ShenandoahConditionalSATBBarrier) && adr != NULL) {
4096 Node* c = kit.control();
4097 Node* call = c->in(1)->in(1)->in(1)->in(0);
4098 assert(call->is_g1_wb_pre_call(), "g1_wb_pre call expected");
4099 call->add_req(adr);
4100 }
4101 }
4102
4103 // G1 pre/post barriers
4104 void GraphKit::g1_write_barrier_pre(bool do_load,
4105 Node* obj,
4106 Node* adr,
4107 uint alias_idx,
4108 Node* val,
4109 const TypeOopPtr* val_type,
4110 Node* pre_val,
4111 BasicType bt) {
4112
4113 // Some sanity checks
4114 // Note: val is unused in this routine.
4115
4116 if (do_load) {
4117 // We need to generate the load of the previous value
4118 assert(obj != NULL, "must have a base");
4119 assert(adr != NULL, "where are loading from?");
4120 assert(pre_val == NULL, "loaded already?");
4121 assert(val_type != NULL, "need a type");
4122
4123 if (use_ReduceInitialCardMarks()
4124 && g1_can_remove_pre_barrier(&_gvn, adr, bt, alias_idx)) {
4125 return;
4126 }
4127
4128 } else {
4129 // In this case both val_type and alias_idx are unused.
4130 assert(pre_val != NULL, "must be loaded already");
4131 // Nothing to be done if pre_val is null.
4132 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
4133 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4134 }
4135 assert(bt == T_OBJECT, "or we shouldn't be here");
4136
4137 IdealKit ideal(this, true);
4138
4139 Node* tls = __ thread(); // ThreadLocalStorage
4140
4141 Node* no_ctrl = NULL;
4142 Node* no_base = __ top();
4143 Node* zero = __ ConI(0);
4144 Node* zeroX = __ ConX(0);
4145
4146 float likely = PROB_LIKELY(0.999);
4147 float unlikely = PROB_UNLIKELY(0.999);
4148
4149 BasicType active_type = in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
4150 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 4 || in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "flag width");
4151
4152 // Offsets into the thread
4153 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
4154 SATBMarkQueue::byte_offset_of_active());
4155 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
4156 SATBMarkQueue::byte_offset_of_index());
4157 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
4158 SATBMarkQueue::byte_offset_of_buf());
4159
4160 // Now the actual pointers into the thread
4161 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
4162 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4163 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4164
4165 // Now some of the values
4166 Node* marking;
4167 if (UseShenandoahGC) {
4168 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(JavaThread::gc_state_offset())));
4169 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
4170 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
4171 assert(ShenandoahWriteBarrierNode::is_gc_state_load(ld), "Should match the shape");
4172 } else {
4173 assert(UseG1GC, "should be");
4174 marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
4175 }
4176
4177 // if (!marking)
4178 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4179 BasicType index_bt = TypeX_X->basic_type();
4180 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4181 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4182
4183 if (do_load) {
4184 // load original value
4185 // alias_idx correct??
4186 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
4187 }
4188
4189 // if (pre_val != NULL)
4190 __ if_then(pre_val, BoolTest::ne, null()); {
4191 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4192
4193 // is the queue for this thread full?
4194 __ if_then(index, BoolTest::ne, zeroX, likely); {
4195
4196 // decrement the index
4197 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4198
4199 // Now get the buffer location we will log the previous value into and store it
4200 Node *log_addr = __ AddP(no_base, buffer, next_index);
4201 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4202 // update the index
4203 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4204
4205 } __ else_(); {
4206
4207 // logging buffer is full, call the runtime
4208 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4209 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4210 } __ end_if(); // (!index)
4211 } __ end_if(); // (pre_val != NULL)
4212 } __ end_if(); // (!marking)
4213
4214 // Final sync IdealKit and GraphKit.
4215 final_sync(ideal);
4216 g1_write_barrier_pre_helper(*this, adr);
4217 }
4218
4219 void GraphKit::shenandoah_enqueue_barrier(Node* pre_val) {
4220
4221 // Some sanity checks
4222 assert(pre_val != NULL, "must be loaded already");
4223 // Nothing to be done if pre_val is null.
4224 if (pre_val->bottom_type()->higher_equal(TypePtr::NULL_PTR)) return;
4225 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
4226
4227 IdealKit ideal(this, true);
4228
4229 Node* tls = __ thread(); // ThreadLocalStorage
4230
4231 Node* no_base = __ top();
4232 Node* zero = __ ConI(0);
4233 Node* zeroX = __ ConX(0);
4234
4235 float likely = PROB_LIKELY(0.999);
4236 float unlikely = PROB_UNLIKELY(0.999);
4237
4238 // Offsets into the thread
4239 const int gc_state_offset = in_bytes(JavaThread::gc_state_offset());
4240 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
4241 SATBMarkQueue::byte_offset_of_index());
4242 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
4243 SATBMarkQueue::byte_offset_of_buf());
4244
4245 // Now the actual pointers into the thread
4246 Node* gc_state_adr = __ AddP(no_base, tls, __ ConX(gc_state_offset));
4247 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4248 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4249
4250 const Type* obj_type = pre_val->bottom_type();
4251 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) {
4252 // dunno if it's NULL or not.
4253 // if (pre_val != NULL)
4254 __ if_then(pre_val, BoolTest::ne, null()); {
4255
4256 // Now some of the values
4257 Node* marking = __ load(__ ctrl(), gc_state_adr, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
4258 // if (!marking)
4259 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4260 BasicType index_bt = TypeX_X->basic_type();
4261 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4262 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4263
4264 // is the queue for this thread full?
4265 __ if_then(index, BoolTest::ne, zeroX, likely); {
4266
4267 // decrement the index
4268 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4269
4270 // Now get the buffer location we will log the previous value into and store it
4271 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4272 Node *log_addr = __ AddP(no_base, buffer, next_index);
4273 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4274 // update the index
4275 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4276
4277 } __ else_(); {
4278 // logging buffer is full, call the runtime
4279 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4280 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4281 } __ end_if(); // (!index)
4282 } __ end_if(); // (!marking)
4283 } __ end_if(); // (pre_val != NULL)
4284 } else {
4285 // We know it is not null.
4286 // Now some of the values
4287 Node* marking = __ load(__ ctrl(), gc_state_adr, TypeInt::BOOL, T_BOOLEAN, Compile::AliasIdxRaw);
4288
4289 // if (!marking)
4290 __ if_then(marking, BoolTest::ne, zero, unlikely); {
4291 BasicType index_bt = TypeX_X->basic_type();
4292 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
4293 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
4294
4295 // is the queue for this thread full?
4296 __ if_then(index, BoolTest::ne, zeroX, likely); {
4297
4298 // decrement the index
4299 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4300
4301 // Now get the buffer location we will log the previous value into and store it
4302 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4303 Node *log_addr = __ AddP(no_base, buffer, next_index);
4304 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
4305 // update the index
4306 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
4307
4308 } __ else_(); {
4309 // logging buffer is full, call the runtime
4310 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
4311 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
4312 } __ end_if(); // (!index)
4313 } __ end_if(); // (!marking)
4314 }
4315 // Final sync IdealKit and GraphKit.
4316 final_sync(ideal);
4317 }
4318
4319 void GraphKit::shenandoah_write_barrier_pre(bool do_load,
4320 Node* obj,
4321 Node* adr,
4322 uint alias_idx,
4323 Node* val,
4324 const TypeOopPtr* val_type,
4325 Node* pre_val,
4326 BasicType bt) {
4327
4328 IdealKit ideal(this);
4329 Node* tls = __ thread();
4330 Node* no_base = __ top();
4331 Node* no_ctrl = NULL;
4332 Node* zero = __ ConI(0);
4333 Node* offset = __ ConX(in_bytes(JavaThread::gc_state_offset()));
4334 Node* gc_state_adr = __ AddP(no_base, tls, offset);
4335 Node* gc_state = __ load(__ ctrl(), gc_state_adr, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
4336 float unlikely = PROB_UNLIKELY(0.999);
4337
4338 __ if_then(gc_state, BoolTest::ne, zero, unlikely); {
4339 sync_kit(ideal);
4340
4341 // Some sanity checks
4342 // Note: val is unused in this routine.
4343
4344 if (val != NULL) {
4345 shenandoah_update_matrix(adr, val);
4346 }
4347
4348 if (ShenandoahConditionalSATBBarrier) {
4349 enum { _set_path = 1, _not_set_path, PATH_LIMIT };
4350 RegionNode* region = new RegionNode(PATH_LIMIT);
4351 Node* prev_mem = memory(Compile::AliasIdxRaw);
4352 Node* memphi = PhiNode::make(region, prev_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
4353
4354 Node* gc_state_addr_p = _gvn.transform(new CastX2PNode(MakeConX((intptr_t) ShenandoahHeap::gc_state_addr())));
4355 Node* gc_state_addr = _gvn.transform(new AddPNode(top(), gc_state_addr_p, MakeConX(0)));
4356 Node* gc_state = _gvn.transform(LoadNode::make(_gvn, control(), memory(Compile::AliasIdxRaw), gc_state_addr, TypeRawPtr::BOTTOM, TypeInt::INT, T_BYTE, MemNode::unordered));
4357 Node* add_set = _gvn.transform(new AddINode(gc_state, intcon(ShenandoahHeap::MARKING)));
4358 Node* cmp_set = _gvn.transform(new CmpINode(add_set, intcon(0)));
4359 Node* cmp_set_bool = _gvn.transform(new BoolNode(cmp_set, BoolTest::eq));
4360 IfNode* cmp_iff = create_and_map_if(control(), cmp_set_bool, PROB_MIN, COUNT_UNKNOWN);
4361 Node* if_not_set = _gvn.transform(new IfTrueNode(cmp_iff));
4362 Node* if_set = _gvn.transform(new IfFalseNode(cmp_iff));
4363
4364 // Conc-mark not in progress. Skip SATB barrier.
4365 set_control(if_not_set);
4366 region->init_req(_not_set_path, control());
4367 memphi->init_req(_not_set_path, prev_mem);
4368
4369 // Conc-mark in progress. Do the SATB barrier.
4370 set_control(if_set);
4371 g1_write_barrier_pre(do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
4372 region->init_req(_set_path, control());
4373 memphi->init_req(_set_path, memory(Compile::AliasIdxRaw));
4374
4375 // Merge control flow and memory.
4376 set_control(_gvn.transform(region));
4377 record_for_igvn(region);
4378 set_memory(_gvn.transform(memphi), Compile::AliasIdxRaw);
4379
4380 }
4381 if (ShenandoahSATBBarrier) {
4382 g1_write_barrier_pre(do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
4383 }
4384 ideal.sync_kit(this);
4385 } __ end_if();
4386 final_sync(ideal);
4387 }
4388
4389 void GraphKit::shenandoah_update_matrix(Node* adr, Node* val) {
4390 if (!UseShenandoahMatrix) {
4391 return;
4392 }
4393
4394 assert(val != NULL, "checked before");
4395 if (adr == NULL) {
4396 return; // Nothing to do
4397 }
4398 assert(adr != NULL, "must not happen");
4399 if (val->bottom_type()->higher_equal(TypePtr::NULL_PTR)) {
4400 // Nothing to do.
4401 return;
4402 }
4403
4404 ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix();
4405
4406 enum { _set_path = 1, _already_set_path, _val_null_path, PATH_LIMIT };
4407 RegionNode* region = new RegionNode(PATH_LIMIT);
4408 Node* prev_mem = memory(Compile::AliasIdxRaw);
4409 Node* memphi = PhiNode::make(region, prev_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
4410 Node* null_ctrl = top();
4411 Node* not_null_val = null_check_oop(val, &null_ctrl);
4412
4413 // Null path: nothing to do.
4414 region->init_req(_val_null_path, null_ctrl);
4415 memphi->init_req(_val_null_path, prev_mem);
4416
4417 // Not null path. Update the matrix.
4418
4419 // This uses a fast calculation for the matrix address. For a description,
4420 // see src/share/vm/gc/shenandoah/shenandoahConnectionMatrix.inline.hpp,
4421 // ShenandoahConnectionMatrix::compute_address(const void* from, const void* to).
4422 address heap_base = ShenandoahHeap::heap()->base();
4423 jint stride = matrix->stride_jint();
4424 jint rs = ShenandoahHeapRegion::region_size_bytes_shift_jint();
4425
4426 guarantee(stride < ShenandoahHeapRegion::region_size_bytes_jint(), "sanity");
4427 guarantee(is_aligned(heap_base, ShenandoahHeapRegion::region_size_bytes()), "sanity");
4428
4429 Node* magic_con = MakeConX((jlong) matrix->matrix_addr() - ((jlong) heap_base >> rs) * (stride + 1));
4430
4431 // Compute addr part
4432 Node* adr_idx = _gvn.transform(new CastP2XNode(control(), adr));
4433 adr_idx = _gvn.transform(new URShiftXNode(adr_idx, intcon(rs)));
4434
4435 // Compute new_val part
4436 Node* val_idx = _gvn.transform(new CastP2XNode(control(), not_null_val));
4437 val_idx = _gvn.transform(new URShiftXNode(val_idx, intcon(rs)));
4438 val_idx = _gvn.transform(new MulXNode(val_idx, MakeConX(stride)));
4439
4440 // Add everything up
4441 adr_idx = _gvn.transform(new AddXNode(adr_idx, val_idx));
4442 adr_idx = _gvn.transform(new CastX2PNode(adr_idx));
4443 Node* matrix_adr = _gvn.transform(new AddPNode(top(), adr_idx, magic_con));
4444
4445 // Load current value
4446 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
4447 Node* current = _gvn.transform(LoadNode::make(_gvn, control(), memory(Compile::AliasIdxRaw),
4448 matrix_adr, adr_type, TypeInt::INT, T_BYTE, MemNode::unordered));
4449
4450 // Check if already set
4451 Node* cmp_set = _gvn.transform(new CmpINode(current, intcon(0)));
4452 Node* cmp_set_bool = _gvn.transform(new BoolNode(cmp_set, BoolTest::eq));
4453 IfNode* cmp_iff = create_and_map_if(control(), cmp_set_bool, PROB_MIN, COUNT_UNKNOWN);
4454
4455 Node* if_not_set = _gvn.transform(new IfTrueNode(cmp_iff));
4456 Node* if_set = _gvn.transform(new IfFalseNode(cmp_iff));
4457
4458 // Already set, exit
4459 set_control(if_set);
4460 region->init_req(_already_set_path, control());
4461 memphi->init_req(_already_set_path, prev_mem);
4462
4463 // Not set: do the store, and finish up
4464 set_control(if_not_set);
4465 Node* store = _gvn.transform(StoreNode::make(_gvn, control(), memory(Compile::AliasIdxRaw),
4466 matrix_adr, adr_type, intcon(1), T_BYTE, MemNode::unordered));
4467 region->init_req(_set_path, control());
4468 memphi->init_req(_set_path, store);
4469
4470 // Merge control flows and memory.
4471 set_control(_gvn.transform(region));
4472 record_for_igvn(region);
4473 set_memory(_gvn.transform(memphi), Compile::AliasIdxRaw);
4474 }
4475
4476 /*
4477 * G1 similar to any GC with a Young Generation requires a way to keep track of
4478 * references from Old Generation to Young Generation to make sure all live
4479 * objects are found. G1 also requires to keep track of object references
4480 * between different regions to enable evacuation of old regions, which is done
4481 * as part of mixed collections. References are tracked in remembered sets and
4482 * is continuously updated as reference are written to with the help of the
4483 * post-barrier.
4484 *
4485 * To reduce the number of updates to the remembered set the post-barrier
4486 * filters updates to fields in objects located in the Young Generation,
4487 * the same region as the reference, when the NULL is being written or
4488 * if the card is already marked as dirty by an earlier write.
4489 *
4490 * Under certain circumstances it is possible to avoid generating the
4491 * post-barrier completely if it is possible during compile time to prove
4492 * the object is newly allocated and that no safepoint exists between the
4493 * allocation and the store.
4494 *
4495 * In the case of slow allocation the allocation code must handle the barrier
4496 * as part of the allocation in the case the allocated object is not located
4497 * in the nursery, this would happen for humongous objects. This is similar to
4498 * how CMS is required to handle this case, see the comments for the method
4499 * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier.
4500 * A deferred card mark is required for these objects and handled in the above
4501 * mentioned methods.
4502 *
4503 * Returns true if the post barrier can be removed
4504 */
4505 bool GraphKit::g1_can_remove_post_barrier(PhaseTransform* phase, Node* store,
4506 Node* adr) {
4507 intptr_t offset = 0;
4508 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
4509 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
4510
4511 if (offset == Type::OffsetBot) {
4512 return false; // cannot unalias unless there are precise offsets
4513 }
4514
4515 if (alloc == NULL) {
4516 return false; // No allocation found
4517 }
4518
4519 // Start search from Store node
4520 Node* mem = store->in(MemNode::Control);
4521 if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
4522
4523 InitializeNode* st_init = mem->in(0)->as_Initialize();
4524 AllocateNode* st_alloc = st_init->allocation();
4525
4526 // Make sure we are looking at the same allocation
4527 if (alloc == st_alloc) {
4528 return true;
4529 }
4530 }
4531
4532 return false;
4533 }
4534
4535 //
4536 // Update the card table and add card address to the queue
4537 //
4538 void GraphKit::g1_mark_card(IdealKit& ideal,
4539 Node* card_adr,
4540 Node* oop_store,
4541 uint oop_alias_idx,
4542 Node* index,
4543 Node* index_adr,
4544 Node* buffer,
4545 const TypeFunc* tf) {
4546
4547 Node* zero = __ ConI(0);
4548 Node* zeroX = __ ConX(0);
4549 Node* no_base = __ top();
4550 BasicType card_bt = T_BYTE;
4551 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4552 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4553
4554 // Now do the queue work
4555 __ if_then(index, BoolTest::ne, zeroX); {
4556
4557 Node* next_index = _gvn.transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
4558 Node* log_addr = __ AddP(no_base, buffer, next_index);
4559
4560 // Order, see storeCM.
4561 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4562 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4563
4564 } __ else_(); {
4565 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4566 } __ end_if();
4567
4568 }
4569
4570 void GraphKit::g1_write_barrier_post(Node* oop_store,
4571 Node* obj,
4572 Node* adr,
4573 uint alias_idx,
4574 Node* val,
4575 BasicType bt,
4576 bool use_precise) {
4577 // If we are writing a NULL then we need no post barrier
4578
4579 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4580 // Must be NULL
4581 const Type* t = val->bottom_type();
4582 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4583 // No post barrier if writing NULLx
4584 return;
4585 }
4586
4587 if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
4588 // We can skip marks on a freshly-allocated object in Eden.
4589 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
4590 // That routine informs GC to take appropriate compensating steps,
4591 // upon a slow-path allocation, so as to make this card-mark
4592 // elision safe.
4593 return;
4594 }
4595
4596 if (use_ReduceInitialCardMarks()
4597 && g1_can_remove_post_barrier(&_gvn, oop_store, adr)) {
4598 return;
4599 }
4600
4601 if (!use_precise) {
4602 // All card marks for a (non-array) instance are in one place:
4603 adr = obj;
4604 }
4605 // (Else it's an array (or unknown), and we want more precise card marks.)
4606 assert(adr != NULL, "");
4607
4608 IdealKit ideal(this, true);
4609
4610 Node* tls = __ thread(); // ThreadLocalStorage
4611
4612 Node* no_base = __ top();
4613 float likely = PROB_LIKELY(0.999);
4614 float unlikely = PROB_UNLIKELY(0.999);
4615 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4616 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4617 Node* zeroX = __ ConX(0);
4618
4619 // Get the alias_index for raw card-mark memory
4620 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4621
4622 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4623
4624 // Offsets into the thread
4625 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4626 DirtyCardQueue::byte_offset_of_index());
4627 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4628 DirtyCardQueue::byte_offset_of_buf());
4629
4630 // Pointers into the thread
4631
4632 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4633 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4634
4635 // Now some values
4636 // Use ctrl to avoid hoisting these values past a safepoint, which could
4637 // potentially reset these fields in the JavaThread.
4638 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4639 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4640
4641 // Convert the store obj pointer to an int prior to doing math on it
4642 // Must use ctrl to prevent "integerized oop" existing across safepoint
4643 Node* cast = __ CastPX(__ ctrl(), adr);
4644
4645 // Divide pointer by card size
4646 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4647
4648 // Combine card table base and card offset
4649 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4650
4651 // If we know the value being stored does it cross regions?
4652
4653 if (val != NULL) {
4654 // Does the store cause us to cross regions?
4655
4656 // Should be able to do an unsigned compare of region_size instead of
4657 // and extra shift. Do we have an unsigned compare??
4658 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4659 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4660
4661 // if (xor_res == 0) same region so skip
4662 __ if_then(xor_res, BoolTest::ne, zeroX); {
4663
4664 // No barrier if we are storing a NULL
4665 __ if_then(val, BoolTest::ne, null(), unlikely); {
4666
4667 // Ok must mark the card if not already dirty
4668
4669 // load the original value of the card
4670 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4671
4672 __ if_then(card_val, BoolTest::ne, young_card); {
4673 sync_kit(ideal);
4674 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4675 insert_mem_bar(Op_MemBarVolatile, oop_store);
4676 __ sync_kit(this);
4677
4678 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4679 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4680 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4681 } __ end_if();
4682 } __ end_if();
4683 } __ end_if();
4684 } __ end_if();
4685 } else {
4686 // The Object.clone() intrinsic uses this path if !ReduceInitialCardMarks.
4687 // We don't need a barrier here if the destination is a newly allocated object
4688 // in Eden. Otherwise, GC verification breaks because we assume that cards in Eden
4689 // are set to 'g1_young_gen' (see G1SATBCardTableModRefBS::verify_g1_young_region()).
4690 assert(!use_ReduceInitialCardMarks(), "can only happen with card marking");
4691 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4692 __ if_then(card_val, BoolTest::ne, young_card); {
4693 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4694 } __ end_if();
4695 }
4696
4697 // Final sync IdealKit and GraphKit.
4698 final_sync(ideal);
4699 }
4700 #undef __
4701
4702
4703 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4704 Node* len = load_array_length(load_String_value(ctrl, str));
4705 Node* coder = load_String_coder(ctrl, str);
4706 // Divide length by 2 if coder is UTF16
4707 return _gvn.transform(new RShiftINode(len, coder));
4708 }
4709
4710 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4711 int value_offset = java_lang_String::value_offset_in_bytes();
4712 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4713 false, NULL, 0);
4714 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4715 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4716 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4717 ciTypeArrayKlass::make(T_BYTE), true, 0);
4718 int value_field_idx = C->get_alias_index(value_field_type);
4719
4720 if (!ShenandoahOptimizeInstanceFinals) {
4721 str = shenandoah_read_barrier(str);
4722 }
4723
4724 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4725 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4726 // String.value field is known to be @Stable.
4727 if (UseImplicitStableValues) {
4728 load = cast_array_to_stable(load, value_type);
4729 }
4730 return load;
4731 }
4732
4733 Node* GraphKit::load_String_coder(Node* ctrl, Node* str) {
4734 if (!CompactStrings) {
4735 return intcon(java_lang_String::CODER_UTF16);
4736 }
4737 int coder_offset = java_lang_String::coder_offset_in_bytes();
4738 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4739 false, NULL, 0);
4740 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4741 int coder_field_idx = C->get_alias_index(coder_field_type);
4742
4743 if (!ShenandoahOptimizeInstanceFinals) {
4744 str = shenandoah_read_barrier(str);
4745 }
4746
4747 return make_load(ctrl, basic_plus_adr(str, str, coder_offset),
4748 TypeInt::BYTE, T_BYTE, coder_field_idx, MemNode::unordered);
4749 }
4750
4751 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4752 int value_offset = java_lang_String::value_offset_in_bytes();
4753 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4754 false, NULL, 0);
4755 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4756
4757 str = shenandoah_write_barrier(str);
4758
4759 store_oop_to_object(control(), str, basic_plus_adr(str, value_offset), value_field_type,
4760 value, TypeAryPtr::BYTES, T_OBJECT, MemNode::unordered);
4761 }
4762
4763 void GraphKit::store_String_coder(Node* ctrl, Node* str, Node* value) {
4764 int coder_offset = java_lang_String::coder_offset_in_bytes();
4765 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4766 false, NULL, 0);
4767
4768 str = shenandoah_write_barrier(str);
4769
4770 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4771 int coder_field_idx = C->get_alias_index(coder_field_type);
4772 store_to_memory(control(), basic_plus_adr(str, coder_offset),
4773 value, T_BYTE, coder_field_idx, MemNode::unordered);
4774 }
4775
4776 // Capture src and dst memory state with a MergeMemNode
4777 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4778 if (src_type == dst_type) {
4779 // Types are equal, we don't need a MergeMemNode
4780 return memory(src_type);
4781 }
4782 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4783 record_for_igvn(merge); // fold it up later, if possible
4784 int src_idx = C->get_alias_index(src_type);
4785 int dst_idx = C->get_alias_index(dst_type);
4786 merge->set_memory_at(src_idx, memory(src_idx));
4787 merge->set_memory_at(dst_idx, memory(dst_idx));
4788 return merge;
4789 }
4790
4791 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4792 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4793 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4794 // If input and output memory types differ, capture both states to preserve
4795 // the dependency between preceding and subsequent loads/stores.
4796 // For example, the following program:
4797 // StoreB
4798 // compress_string
4799 // LoadB
4800 // has this memory graph (use->def):
4801 // LoadB -> compress_string -> CharMem
4802 // ... -> StoreB -> ByteMem
4803 // The intrinsic hides the dependency between LoadB and StoreB, causing
4804 // the load to read from memory not containing the result of the StoreB.
4805 // The correct memory graph should look like this:
4806 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4807 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4808 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4809 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4810 set_memory(res_mem, TypeAryPtr::BYTES);
4811 return str;
4812 }
4813
4814 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4815 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4816 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4817 // Capture src and dst memory (see comment in 'compress_string').
4818 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4819 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4820 set_memory(_gvn.transform(str), dst_type);
4821 }
4822
4823 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4824
4825 src = shenandoah_read_barrier(src);
4826 dst = shenandoah_write_barrier(dst);
4827
4828 /**
4829 * int i_char = start;
4830 * for (int i_byte = 0; i_byte < count; i_byte++) {
4831 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4832 * }
4833 */
4834 add_predicate();
4835 RegionNode* head = new RegionNode(3);
4836 head->init_req(1, control());
4837 gvn().set_type(head, Type::CONTROL);
4838 record_for_igvn(head);
4839
4840 Node* i_byte = new PhiNode(head, TypeInt::INT);
4841 i_byte->init_req(1, intcon(0));
4842 gvn().set_type(i_byte, TypeInt::INT);
4843 record_for_igvn(i_byte);
4844
4845 Node* i_char = new PhiNode(head, TypeInt::INT);
4846 i_char->init_req(1, start);
4847 gvn().set_type(i_char, TypeInt::INT);
4848 record_for_igvn(i_char);
4849
4850 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4851 gvn().set_type(mem, Type::MEMORY);
4852 record_for_igvn(mem);
4853 set_control(head);
4854 set_memory(mem, TypeAryPtr::BYTES);
4855 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4856 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4857 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4858 false, false, true /* mismatched */);
4859
4860 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4861 head->init_req(2, IfTrue(iff));
4862 mem->init_req(2, st);
4863 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4864 i_char->init_req(2, AddI(i_char, intcon(2)));
4865
4866 set_control(IfFalse(iff));
4867 set_memory(st, TypeAryPtr::BYTES);
4868 }
4869
4870 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4871 if (!field->is_constant()) {
4872 return NULL; // Field not marked as constant.
4873 }
4874 ciInstance* holder = NULL;
4875 if (!field->is_static()) {
4876 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4877 if (const_oop != NULL && const_oop->is_instance()) {
4878 holder = const_oop->as_instance();
4879 }
4880 }
4881 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4882 /*is_unsigned_load=*/false);
4883 if (con_type != NULL) {
4884 return makecon(con_type);
4885 }
4886 return NULL;
4887 }
4888
4889 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4890 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4891 // assumption of CCP analysis.
4892 return _gvn.transform(new CastPPNode(ary, ary_type->cast_to_stable(true)));
4893 }
4894
4895 Node* GraphKit::shenandoah_read_barrier(Node* obj) {
4896 if (UseShenandoahGC && ShenandoahReadBarrier) {
4897 return shenandoah_read_barrier_impl(obj, false, true, true);
4898 } else {
4899 return obj;
4900 }
4901 }
4902
4903 Node* GraphKit::shenandoah_storeval_barrier(Node* obj) {
4904 if (UseShenandoahGC) {
4905 if (ShenandoahStoreValWriteBarrier || ShenandoahStoreValEnqueueBarrier) {
4906 obj = shenandoah_write_barrier(obj);
4907 }
4908 if (ShenandoahStoreValEnqueueBarrier && !ShenandoahMWF) {
4909 shenandoah_enqueue_barrier(obj);
4910 }
4911 if (ShenandoahStoreValReadBarrier) {
4912 obj = shenandoah_read_barrier_impl(obj, true, false, false);
4913 }
4914 }
4915 return obj;
4916 }
4917
4918 Node* GraphKit::shenandoah_read_barrier_acmp(Node* obj) {
4919 return shenandoah_read_barrier_impl(obj, true, true, false);
4920 }
4921
4922 Node* GraphKit::shenandoah_read_barrier_impl(Node* obj, bool use_ctrl, bool use_mem, bool allow_fromspace) {
4923
4924 const Type* obj_type = obj->bottom_type();
4925 if (obj_type->higher_equal(TypePtr::NULL_PTR)) {
4926 return obj;
4927 }
4928 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type);
4929 Node* mem = use_mem ? memory(adr_type) : immutable_memory();
4930
4931 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, mem, allow_fromspace)) {
4932 // We know it is null, no barrier needed.
4933 return obj;
4934 }
4935
4936
4937 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) {
4938
4939 // We don't know if it's null or not. Need null-check.
4940 enum { _not_null_path = 1, _null_path, PATH_LIMIT };
4941 RegionNode* region = new RegionNode(PATH_LIMIT);
4942 Node* phi = new PhiNode(region, obj_type);
4943 Node* null_ctrl = top();
4944 Node* not_null_obj = null_check_oop(obj, &null_ctrl);
4945
4946 region->init_req(_null_path, null_ctrl);
4947 phi ->init_req(_null_path, zerocon(T_OBJECT));
4948
4949 Node* ctrl = use_ctrl ? control() : NULL;
4950 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, not_null_obj, allow_fromspace);
4951 Node* n = _gvn.transform(rb);
4952
4953 region->init_req(_not_null_path, control());
4954 phi ->init_req(_not_null_path, n);
4955
4956 set_control(_gvn.transform(region));
4957 record_for_igvn(region);
4958 return _gvn.transform(phi);
4959
4960 } else {
4961 // We know it is not null. Simple barrier is sufficient.
4962 Node* ctrl = use_ctrl ? control() : NULL;
4963 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, obj, allow_fromspace);
4964 Node* n = _gvn.transform(rb);
4965 record_for_igvn(n);
4966 return n;
4967 }
4968 }
4969
4970 Node* GraphKit::shenandoah_write_barrier_helper(GraphKit& kit, Node* obj, const TypePtr* adr_type) {
4971 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(kit.C, kit.control(), kit.memory(adr_type), obj);
4972 Node* n = kit.gvn().transform(wb);
4973 if (n == wb) { // New barrier needs memory projection.
4974 Node* proj = kit.gvn().transform(new ShenandoahWBMemProjNode(n));
4975 kit.set_memory(proj, adr_type);
4976 }
4977
4978 return n;
4979 }
4980
4981 Node* GraphKit::shenandoah_write_barrier(Node* obj) {
4982
4983 if (UseShenandoahGC && ShenandoahWriteBarrier) {
4984 obj = shenandoah_write_barrier_impl(obj);
4985 if (ShenandoahStoreValEnqueueBarrier && ShenandoahMWF) {
4986 shenandoah_enqueue_barrier(obj);
4987 }
4988 return obj;
4989 } else {
4990 return obj;
4991 }
4992 }
4993
4994 Node* GraphKit::shenandoah_write_barrier_impl(Node* obj) {
4995 if (! ShenandoahBarrierNode::needs_barrier(&_gvn, NULL, obj, NULL, true)) {
4996 return obj;
4997 }
4998 const Type* obj_type = obj->bottom_type();
4999 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type);
5000 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) {
5001 // We don't know if it's null or not. Need null-check.
5002 enum { _not_null_path = 1, _null_path, PATH_LIMIT };
5003 RegionNode* region = new RegionNode(PATH_LIMIT);
5004 Node* phi = new PhiNode(region, obj_type);
5005 Node* memphi = PhiNode::make(region, memory(adr_type), Type::MEMORY, C->alias_type(adr_type)->adr_type());
5006
5007 Node* prev_mem = memory(adr_type);
5008 Node* null_ctrl = top();
5009 Node* not_null_obj = null_check_oop(obj, &null_ctrl);
5010
5011 region->init_req(_null_path, null_ctrl);
5012 phi ->init_req(_null_path, zerocon(T_OBJECT));
5013 memphi->init_req(_null_path, prev_mem);
5014
5015 Node* n = shenandoah_write_barrier_helper(*this, not_null_obj, adr_type);
5016
5017 region->init_req(_not_null_path, control());
5018 phi ->init_req(_not_null_path, n);
5019 memphi->init_req(_not_null_path, memory(adr_type));
5020
5021 set_control(_gvn.transform(region));
5022 record_for_igvn(region);
5023 set_memory(_gvn.transform(memphi), adr_type);
5024
5025 Node* res_val = _gvn.transform(phi);
5026 // replace_in_map(obj, res_val);
5027 return res_val;
5028 } else {
5029 // We know it is not null. Simple barrier is sufficient.
5030 Node* n = shenandoah_write_barrier_helper(*this, obj, adr_type);
5031 // replace_in_map(obj, n);
5032 record_for_igvn(n);
5033 return n;
5034 }
5035 }