1 /*
2 * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
28 #include "gc_implementation/g1/heapRegion.hpp"
29 #include "gc_interface/collectedHeap.hpp"
30 #include "gc_implementation/shenandoah/shenandoahHeap.hpp"
31 #include "memory/barrierSet.hpp"
32 #include "memory/cardTableModRefBS.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/graphKit.hpp"
35 #include "opto/idealKit.hpp"
36 #include "opto/locknode.hpp"
37 #include "opto/machnode.hpp"
38 #include "opto/parse.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "runtime/deoptimization.hpp"
42 #include "runtime/sharedRuntime.hpp"
43
44 #if INCLUDE_ALL_GCS
45 #include "gc_implementation/shenandoah/c2/shenandoahBarrierSetC2.hpp"
46 #include "gc_implementation/shenandoah/c2/shenandoahSupport.hpp"
47 #endif
48
49 //----------------------------GraphKit-----------------------------------------
50 // Main utility constructor.
51 GraphKit::GraphKit(JVMState* jvms)
52 : Phase(Phase::Parser),
53 _env(C->env()),
54 _gvn(*C->initial_gvn())
55 {
56 _exceptions = jvms->map()->next_exception();
57 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
58 set_jvms(jvms);
59 }
60
61 // Private constructor for parser.
62 GraphKit::GraphKit()
63 : Phase(Phase::Parser),
64 _env(C->env()),
65 _gvn(*C->initial_gvn())
66 {
67 _exceptions = NULL;
68 set_map(NULL);
69 debug_only(_sp = -99);
70 debug_only(set_bci(-99));
71 }
72
73
74
75 //---------------------------clean_stack---------------------------------------
76 // Clear away rubbish from the stack area of the JVM state.
77 // This destroys any arguments that may be waiting on the stack.
78 void GraphKit::clean_stack(int from_sp) {
79 SafePointNode* map = this->map();
80 JVMState* jvms = this->jvms();
81 int stk_size = jvms->stk_size();
82 int stkoff = jvms->stkoff();
83 Node* top = this->top();
84 for (int i = from_sp; i < stk_size; i++) {
85 if (map->in(stkoff + i) != top) {
86 map->set_req(stkoff + i, top);
87 }
88 }
89 }
90
91
92 //--------------------------------sync_jvms-----------------------------------
93 // Make sure our current jvms agrees with our parse state.
94 JVMState* GraphKit::sync_jvms() const {
95 JVMState* jvms = this->jvms();
96 jvms->set_bci(bci()); // Record the new bci in the JVMState
97 jvms->set_sp(sp()); // Record the new sp in the JVMState
98 assert(jvms_in_sync(), "jvms is now in sync");
99 return jvms;
100 }
101
102 //--------------------------------sync_jvms_for_reexecute---------------------
103 // Make sure our current jvms agrees with our parse state. This version
104 // uses the reexecute_sp for reexecuting bytecodes.
105 JVMState* GraphKit::sync_jvms_for_reexecute() {
106 JVMState* jvms = this->jvms();
107 jvms->set_bci(bci()); // Record the new bci in the JVMState
108 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
109 return jvms;
110 }
111
112 #ifdef ASSERT
113 bool GraphKit::jvms_in_sync() const {
114 Parse* parse = is_Parse();
115 if (parse == NULL) {
116 if (bci() != jvms()->bci()) return false;
117 if (sp() != (int)jvms()->sp()) return false;
118 return true;
119 }
120 if (jvms()->method() != parse->method()) return false;
121 if (jvms()->bci() != parse->bci()) return false;
122 int jvms_sp = jvms()->sp();
123 if (jvms_sp != parse->sp()) return false;
124 int jvms_depth = jvms()->depth();
125 if (jvms_depth != parse->depth()) return false;
126 return true;
127 }
128
129 // Local helper checks for special internal merge points
130 // used to accumulate and merge exception states.
131 // They are marked by the region's in(0) edge being the map itself.
132 // Such merge points must never "escape" into the parser at large,
133 // until they have been handed to gvn.transform.
134 static bool is_hidden_merge(Node* reg) {
135 if (reg == NULL) return false;
136 if (reg->is_Phi()) {
137 reg = reg->in(0);
138 if (reg == NULL) return false;
139 }
140 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
141 }
142
143 void GraphKit::verify_map() const {
144 if (map() == NULL) return; // null map is OK
145 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
146 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
147 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
148 }
149
150 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
151 assert(ex_map->next_exception() == NULL, "not already part of a chain");
152 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
153 }
154 #endif
155
156 //---------------------------stop_and_kill_map---------------------------------
157 // Set _map to NULL, signalling a stop to further bytecode execution.
158 // First smash the current map's control to a constant, to mark it dead.
159 void GraphKit::stop_and_kill_map() {
160 SafePointNode* dead_map = stop();
161 if (dead_map != NULL) {
162 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
163 assert(dead_map->is_killed(), "must be so marked");
164 }
165 }
166
167
168 //--------------------------------stopped--------------------------------------
169 // Tell if _map is NULL, or control is top.
170 bool GraphKit::stopped() {
171 if (map() == NULL) return true;
172 else if (control() == top()) return true;
173 else return false;
174 }
175
176
177 //-----------------------------has_ex_handler----------------------------------
178 // Tell if this method or any caller method has exception handlers.
179 bool GraphKit::has_ex_handler() {
180 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
181 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
182 return true;
183 }
184 }
185 return false;
186 }
187
188 //------------------------------save_ex_oop------------------------------------
189 // Save an exception without blowing stack contents or other JVM state.
190 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
191 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
192 ex_map->add_req(ex_oop);
193 debug_only(verify_exception_state(ex_map));
194 }
195
196 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
197 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
198 Node* ex_oop = ex_map->in(ex_map->req()-1);
199 if (clear_it) ex_map->del_req(ex_map->req()-1);
200 return ex_oop;
201 }
202
203 //-----------------------------saved_ex_oop------------------------------------
204 // Recover a saved exception from its map.
205 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
206 return common_saved_ex_oop(ex_map, false);
207 }
208
209 //--------------------------clear_saved_ex_oop---------------------------------
210 // Erase a previously saved exception from its map.
211 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
212 return common_saved_ex_oop(ex_map, true);
213 }
214
215 #ifdef ASSERT
216 //---------------------------has_saved_ex_oop----------------------------------
217 // Erase a previously saved exception from its map.
218 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
219 return ex_map->req() == ex_map->jvms()->endoff()+1;
220 }
221 #endif
222
223 //-------------------------make_exception_state--------------------------------
224 // Turn the current JVM state into an exception state, appending the ex_oop.
225 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
226 sync_jvms();
227 SafePointNode* ex_map = stop(); // do not manipulate this map any more
228 set_saved_ex_oop(ex_map, ex_oop);
229 return ex_map;
230 }
231
232
233 //--------------------------add_exception_state--------------------------------
234 // Add an exception to my list of exceptions.
235 void GraphKit::add_exception_state(SafePointNode* ex_map) {
236 if (ex_map == NULL || ex_map->control() == top()) {
237 return;
238 }
239 #ifdef ASSERT
240 verify_exception_state(ex_map);
241 if (has_exceptions()) {
242 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
243 }
244 #endif
245
246 // If there is already an exception of exactly this type, merge with it.
247 // In particular, null-checks and other low-level exceptions common up here.
248 Node* ex_oop = saved_ex_oop(ex_map);
249 const Type* ex_type = _gvn.type(ex_oop);
250 if (ex_oop == top()) {
251 // No action needed.
252 return;
253 }
254 assert(ex_type->isa_instptr(), "exception must be an instance");
255 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
256 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
257 // We check sp also because call bytecodes can generate exceptions
258 // both before and after arguments are popped!
259 if (ex_type2 == ex_type
260 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
261 combine_exception_states(ex_map, e2);
262 return;
263 }
264 }
265
266 // No pre-existing exception of the same type. Chain it on the list.
267 push_exception_state(ex_map);
268 }
269
270 //-----------------------add_exception_states_from-----------------------------
271 void GraphKit::add_exception_states_from(JVMState* jvms) {
272 SafePointNode* ex_map = jvms->map()->next_exception();
273 if (ex_map != NULL) {
274 jvms->map()->set_next_exception(NULL);
275 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
276 next_map = ex_map->next_exception();
277 ex_map->set_next_exception(NULL);
278 add_exception_state(ex_map);
279 }
280 }
281 }
282
283 //-----------------------transfer_exceptions_into_jvms-------------------------
284 JVMState* GraphKit::transfer_exceptions_into_jvms() {
285 if (map() == NULL) {
286 // We need a JVMS to carry the exceptions, but the map has gone away.
287 // Create a scratch JVMS, cloned from any of the exception states...
288 if (has_exceptions()) {
289 _map = _exceptions;
290 _map = clone_map();
291 _map->set_next_exception(NULL);
292 clear_saved_ex_oop(_map);
293 debug_only(verify_map());
294 } else {
295 // ...or created from scratch
296 JVMState* jvms = new (C) JVMState(_method, NULL);
297 jvms->set_bci(_bci);
298 jvms->set_sp(_sp);
299 jvms->set_map(new (C) SafePointNode(TypeFunc::Parms, jvms));
300 set_jvms(jvms);
301 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
302 set_all_memory(top());
303 while (map()->req() < jvms->endoff()) map()->add_req(top());
304 }
305 // (This is a kludge, in case you didn't notice.)
306 set_control(top());
307 }
308 JVMState* jvms = sync_jvms();
309 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
310 jvms->map()->set_next_exception(_exceptions);
311 _exceptions = NULL; // done with this set of exceptions
312 return jvms;
313 }
314
315 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
316 assert(is_hidden_merge(dstphi), "must be a special merge node");
317 assert(is_hidden_merge(srcphi), "must be a special merge node");
318 uint limit = srcphi->req();
319 for (uint i = PhiNode::Input; i < limit; i++) {
320 dstphi->add_req(srcphi->in(i));
321 }
322 }
323 static inline void add_one_req(Node* dstphi, Node* src) {
324 assert(is_hidden_merge(dstphi), "must be a special merge node");
325 assert(!is_hidden_merge(src), "must not be a special merge node");
326 dstphi->add_req(src);
327 }
328
329 //-----------------------combine_exception_states------------------------------
330 // This helper function combines exception states by building phis on a
331 // specially marked state-merging region. These regions and phis are
332 // untransformed, and can build up gradually. The region is marked by
333 // having a control input of its exception map, rather than NULL. Such
334 // regions do not appear except in this function, and in use_exception_state.
335 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
336 if (failing()) return; // dying anyway...
337 JVMState* ex_jvms = ex_map->_jvms;
338 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
339 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
340 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
341 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
342 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
343 assert(ex_map->req() == phi_map->req(), "matching maps");
344 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
345 Node* hidden_merge_mark = root();
346 Node* region = phi_map->control();
347 MergeMemNode* phi_mem = phi_map->merged_memory();
348 MergeMemNode* ex_mem = ex_map->merged_memory();
349 if (region->in(0) != hidden_merge_mark) {
350 // The control input is not (yet) a specially-marked region in phi_map.
351 // Make it so, and build some phis.
352 region = new (C) RegionNode(2);
353 _gvn.set_type(region, Type::CONTROL);
354 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
355 region->init_req(1, phi_map->control());
356 phi_map->set_control(region);
357 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
358 record_for_igvn(io_phi);
359 _gvn.set_type(io_phi, Type::ABIO);
360 phi_map->set_i_o(io_phi);
361 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
362 Node* m = mms.memory();
363 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
364 record_for_igvn(m_phi);
365 _gvn.set_type(m_phi, Type::MEMORY);
366 mms.set_memory(m_phi);
367 }
368 }
369
370 // Either or both of phi_map and ex_map might already be converted into phis.
371 Node* ex_control = ex_map->control();
372 // if there is special marking on ex_map also, we add multiple edges from src
373 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
374 // how wide was the destination phi_map, originally?
375 uint orig_width = region->req();
376
377 if (add_multiple) {
378 add_n_reqs(region, ex_control);
379 add_n_reqs(phi_map->i_o(), ex_map->i_o());
380 } else {
381 // ex_map has no merges, so we just add single edges everywhere
382 add_one_req(region, ex_control);
383 add_one_req(phi_map->i_o(), ex_map->i_o());
384 }
385 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
386 if (mms.is_empty()) {
387 // get a copy of the base memory, and patch some inputs into it
388 const TypePtr* adr_type = mms.adr_type(C);
389 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
390 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
391 mms.set_memory(phi);
392 // Prepare to append interesting stuff onto the newly sliced phi:
393 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
394 }
395 // Append stuff from ex_map:
396 if (add_multiple) {
397 add_n_reqs(mms.memory(), mms.memory2());
398 } else {
399 add_one_req(mms.memory(), mms.memory2());
400 }
401 }
402 uint limit = ex_map->req();
403 for (uint i = TypeFunc::Parms; i < limit; i++) {
404 // Skip everything in the JVMS after tos. (The ex_oop follows.)
405 if (i == tos) i = ex_jvms->monoff();
406 Node* src = ex_map->in(i);
407 Node* dst = phi_map->in(i);
408 if (src != dst) {
409 PhiNode* phi;
410 if (dst->in(0) != region) {
411 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
412 record_for_igvn(phi);
413 _gvn.set_type(phi, phi->type());
414 phi_map->set_req(i, dst);
415 // Prepare to append interesting stuff onto the new phi:
416 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
417 } else {
418 assert(dst->is_Phi(), "nobody else uses a hidden region");
419 phi = dst->as_Phi();
420 }
421 if (add_multiple && src->in(0) == ex_control) {
422 // Both are phis.
423 add_n_reqs(dst, src);
424 } else {
425 while (dst->req() < region->req()) add_one_req(dst, src);
426 }
427 const Type* srctype = _gvn.type(src);
428 if (phi->type() != srctype) {
429 const Type* dsttype = phi->type()->meet_speculative(srctype);
430 if (phi->type() != dsttype) {
431 phi->set_type(dsttype);
432 _gvn.set_type(phi, dsttype);
433 }
434 }
435 }
436 }
437 phi_map->merge_replaced_nodes_with(ex_map);
438 }
439
440 //--------------------------use_exception_state--------------------------------
441 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
442 if (failing()) { stop(); return top(); }
443 Node* region = phi_map->control();
444 Node* hidden_merge_mark = root();
445 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
446 Node* ex_oop = clear_saved_ex_oop(phi_map);
447 if (region->in(0) == hidden_merge_mark) {
448 // Special marking for internal ex-states. Process the phis now.
449 region->set_req(0, region); // now it's an ordinary region
450 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
451 // Note: Setting the jvms also sets the bci and sp.
452 set_control(_gvn.transform(region));
453 uint tos = jvms()->stkoff() + sp();
454 for (uint i = 1; i < tos; i++) {
455 Node* x = phi_map->in(i);
456 if (x->in(0) == region) {
457 assert(x->is_Phi(), "expected a special phi");
458 phi_map->set_req(i, _gvn.transform(x));
459 }
460 }
461 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
462 Node* x = mms.memory();
463 if (x->in(0) == region) {
464 assert(x->is_Phi(), "nobody else uses a hidden region");
465 mms.set_memory(_gvn.transform(x));
466 }
467 }
468 if (ex_oop->in(0) == region) {
469 assert(ex_oop->is_Phi(), "expected a special phi");
470 ex_oop = _gvn.transform(ex_oop);
471 }
472 } else {
473 set_jvms(phi_map->jvms());
474 }
475
476 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
477 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
478 return ex_oop;
479 }
480
481 //---------------------------------java_bc-------------------------------------
482 Bytecodes::Code GraphKit::java_bc() const {
483 ciMethod* method = this->method();
484 int bci = this->bci();
485 if (method != NULL && bci != InvocationEntryBci)
486 return method->java_code_at_bci(bci);
487 else
488 return Bytecodes::_illegal;
489 }
490
491 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
492 bool must_throw) {
493 // if the exception capability is set, then we will generate code
494 // to check the JavaThread.should_post_on_exceptions flag to see
495 // if we actually need to report exception events (for this
496 // thread). If we don't need to report exception events, we will
497 // take the normal fast path provided by add_exception_events. If
498 // exception event reporting is enabled for this thread, we will
499 // take the uncommon_trap in the BuildCutout below.
500
501 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
502 Node* jthread = _gvn.transform(new (C) ThreadLocalNode());
503 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
504 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
505
506 // Test the should_post_on_exceptions_flag vs. 0
507 Node* chk = _gvn.transform( new (C) CmpINode(should_post_flag, intcon(0)) );
508 Node* tst = _gvn.transform( new (C) BoolNode(chk, BoolTest::eq) );
509
510 // Branch to slow_path if should_post_on_exceptions_flag was true
511 { BuildCutout unless(this, tst, PROB_MAX);
512 // Do not try anything fancy if we're notifying the VM on every throw.
513 // Cf. case Bytecodes::_athrow in parse2.cpp.
514 uncommon_trap(reason, Deoptimization::Action_none,
515 (ciKlass*)NULL, (char*)NULL, must_throw);
516 }
517
518 }
519
520 //------------------------------builtin_throw----------------------------------
521 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
522 bool must_throw = true;
523
524 if (env()->jvmti_can_post_on_exceptions()) {
525 // check if we must post exception events, take uncommon trap if so
526 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
527 // here if should_post_on_exceptions is false
528 // continue on with the normal codegen
529 }
530
531 // If this particular condition has not yet happened at this
532 // bytecode, then use the uncommon trap mechanism, and allow for
533 // a future recompilation if several traps occur here.
534 // If the throw is hot, try to use a more complicated inline mechanism
535 // which keeps execution inside the compiled code.
536 bool treat_throw_as_hot = false;
537 ciMethodData* md = method()->method_data();
538
539 if (ProfileTraps) {
540 if (too_many_traps(reason)) {
541 treat_throw_as_hot = true;
542 }
543 // (If there is no MDO at all, assume it is early in
544 // execution, and that any deopts are part of the
545 // startup transient, and don't need to be remembered.)
546
547 // Also, if there is a local exception handler, treat all throws
548 // as hot if there has been at least one in this method.
549 if (C->trap_count(reason) != 0
550 && method()->method_data()->trap_count(reason) != 0
551 && has_ex_handler()) {
552 treat_throw_as_hot = true;
553 }
554 }
555
556 // If this throw happens frequently, an uncommon trap might cause
557 // a performance pothole. If there is a local exception handler,
558 // and if this particular bytecode appears to be deoptimizing often,
559 // let us handle the throw inline, with a preconstructed instance.
560 // Note: If the deopt count has blown up, the uncommon trap
561 // runtime is going to flush this nmethod, not matter what.
562 if (treat_throw_as_hot
563 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
564 // If the throw is local, we use a pre-existing instance and
565 // punt on the backtrace. This would lead to a missing backtrace
566 // (a repeat of 4292742) if the backtrace object is ever asked
567 // for its backtrace.
568 // Fixing this remaining case of 4292742 requires some flavor of
569 // escape analysis. Leave that for the future.
570 ciInstance* ex_obj = NULL;
571 switch (reason) {
572 case Deoptimization::Reason_null_check:
573 ex_obj = env()->NullPointerException_instance();
574 break;
575 case Deoptimization::Reason_div0_check:
576 ex_obj = env()->ArithmeticException_instance();
577 break;
578 case Deoptimization::Reason_range_check:
579 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
580 break;
581 case Deoptimization::Reason_class_check:
582 if (java_bc() == Bytecodes::_aastore) {
583 ex_obj = env()->ArrayStoreException_instance();
584 } else {
585 ex_obj = env()->ClassCastException_instance();
586 }
587 break;
588 }
589 if (failing()) { stop(); return; } // exception allocation might fail
590 if (ex_obj != NULL) {
591 // Cheat with a preallocated exception object.
592 if (C->log() != NULL)
593 C->log()->elem("hot_throw preallocated='1' reason='%s'",
594 Deoptimization::trap_reason_name(reason));
595 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
596 Node* ex_node = _gvn.transform( ConNode::make(C, ex_con) );
597
598 // Clear the detail message of the preallocated exception object.
599 // Weblogic sometimes mutates the detail message of exceptions
600 // using reflection.
601 int offset = java_lang_Throwable::get_detailMessage_offset();
602 const TypePtr* adr_typ = ex_con->add_offset(offset);
603
604 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
605 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
606 // Conservatively release stores of object references.
607 Node *store = store_oop_to_object(control(), ex_node, adr, adr_typ, null(), val_type, T_OBJECT, MemNode::release);
608
609 add_exception_state(make_exception_state(ex_node));
610 return;
611 }
612 }
613
614 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
615 // It won't be much cheaper than bailing to the interp., since we'll
616 // have to pass up all the debug-info, and the runtime will have to
617 // create the stack trace.
618
619 // Usual case: Bail to interpreter.
620 // Reserve the right to recompile if we haven't seen anything yet.
621
622 assert(!Deoptimization::reason_is_speculate(reason), "unsupported");
623 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
624 if (treat_throw_as_hot
625 && (method()->method_data()->trap_recompiled_at(bci(), NULL)
626 || C->too_many_traps(reason))) {
627 // We cannot afford to take more traps here. Suffer in the interpreter.
628 if (C->log() != NULL)
629 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
630 Deoptimization::trap_reason_name(reason),
631 C->trap_count(reason));
632 action = Deoptimization::Action_none;
633 }
634
635 // "must_throw" prunes the JVM state to include only the stack, if there
636 // are no local exception handlers. This should cut down on register
637 // allocation time and code size, by drastically reducing the number
638 // of in-edges on the call to the uncommon trap.
639
640 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
641 }
642
643
644 //----------------------------PreserveJVMState---------------------------------
645 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
646 debug_only(kit->verify_map());
647 _kit = kit;
648 _map = kit->map(); // preserve the map
649 _sp = kit->sp();
650 kit->set_map(clone_map ? kit->clone_map() : NULL);
651 #ifdef ASSERT
652 _bci = kit->bci();
653 Parse* parser = kit->is_Parse();
654 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
655 _block = block;
656 #endif
657 }
658 PreserveJVMState::~PreserveJVMState() {
659 GraphKit* kit = _kit;
660 #ifdef ASSERT
661 assert(kit->bci() == _bci, "bci must not shift");
662 Parse* parser = kit->is_Parse();
663 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
664 assert(block == _block, "block must not shift");
665 #endif
666 kit->set_map(_map);
667 kit->set_sp(_sp);
668 }
669
670
671 //-----------------------------BuildCutout-------------------------------------
672 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
673 : PreserveJVMState(kit)
674 {
675 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
676 SafePointNode* outer_map = _map; // preserved map is caller's
677 SafePointNode* inner_map = kit->map();
678 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
679 outer_map->set_control(kit->gvn().transform( new (kit->C) IfTrueNode(iff) ));
680 inner_map->set_control(kit->gvn().transform( new (kit->C) IfFalseNode(iff) ));
681 }
682 BuildCutout::~BuildCutout() {
683 GraphKit* kit = _kit;
684 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
685 }
686
687 //---------------------------PreserveReexecuteState----------------------------
688 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
689 assert(!kit->stopped(), "must call stopped() before");
690 _kit = kit;
691 _sp = kit->sp();
692 _reexecute = kit->jvms()->_reexecute;
693 }
694 PreserveReexecuteState::~PreserveReexecuteState() {
695 if (_kit->stopped()) return;
696 _kit->jvms()->_reexecute = _reexecute;
697 _kit->set_sp(_sp);
698 }
699
700 //------------------------------clone_map--------------------------------------
701 // Implementation of PreserveJVMState
702 //
703 // Only clone_map(...) here. If this function is only used in the
704 // PreserveJVMState class we may want to get rid of this extra
705 // function eventually and do it all there.
706
707 SafePointNode* GraphKit::clone_map() {
708 if (map() == NULL) return NULL;
709
710 // Clone the memory edge first
711 Node* mem = MergeMemNode::make(C, map()->memory());
712 gvn().set_type_bottom(mem);
713
714 SafePointNode *clonemap = (SafePointNode*)map()->clone();
715 JVMState* jvms = this->jvms();
716 JVMState* clonejvms = jvms->clone_shallow(C);
717 clonemap->set_memory(mem);
718 clonemap->set_jvms(clonejvms);
719 clonejvms->set_map(clonemap);
720 record_for_igvn(clonemap);
721 gvn().set_type_bottom(clonemap);
722 return clonemap;
723 }
724
725
726 //-----------------------------set_map_clone-----------------------------------
727 void GraphKit::set_map_clone(SafePointNode* m) {
728 _map = m;
729 _map = clone_map();
730 _map->set_next_exception(NULL);
731 debug_only(verify_map());
732 }
733
734
735 //----------------------------kill_dead_locals---------------------------------
736 // Detect any locals which are known to be dead, and force them to top.
737 void GraphKit::kill_dead_locals() {
738 // Consult the liveness information for the locals. If any
739 // of them are unused, then they can be replaced by top(). This
740 // should help register allocation time and cut down on the size
741 // of the deoptimization information.
742
743 // This call is made from many of the bytecode handling
744 // subroutines called from the Big Switch in do_one_bytecode.
745 // Every bytecode which might include a slow path is responsible
746 // for killing its dead locals. The more consistent we
747 // are about killing deads, the fewer useless phis will be
748 // constructed for them at various merge points.
749
750 // bci can be -1 (InvocationEntryBci). We return the entry
751 // liveness for the method.
752
753 if (method() == NULL || method()->code_size() == 0) {
754 // We are building a graph for a call to a native method.
755 // All locals are live.
756 return;
757 }
758
759 ResourceMark rm;
760
761 // Consult the liveness information for the locals. If any
762 // of them are unused, then they can be replaced by top(). This
763 // should help register allocation time and cut down on the size
764 // of the deoptimization information.
765 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
766
767 int len = (int)live_locals.size();
768 assert(len <= jvms()->loc_size(), "too many live locals");
769 for (int local = 0; local < len; local++) {
770 if (!live_locals.at(local)) {
771 set_local(local, top());
772 }
773 }
774 }
775
776 #ifdef ASSERT
777 //-------------------------dead_locals_are_killed------------------------------
778 // Return true if all dead locals are set to top in the map.
779 // Used to assert "clean" debug info at various points.
780 bool GraphKit::dead_locals_are_killed() {
781 if (method() == NULL || method()->code_size() == 0) {
782 // No locals need to be dead, so all is as it should be.
783 return true;
784 }
785
786 // Make sure somebody called kill_dead_locals upstream.
787 ResourceMark rm;
788 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
789 if (jvms->loc_size() == 0) continue; // no locals to consult
790 SafePointNode* map = jvms->map();
791 ciMethod* method = jvms->method();
792 int bci = jvms->bci();
793 if (jvms == this->jvms()) {
794 bci = this->bci(); // it might not yet be synched
795 }
796 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
797 int len = (int)live_locals.size();
798 if (!live_locals.is_valid() || len == 0)
799 // This method is trivial, or is poisoned by a breakpoint.
800 return true;
801 assert(len == jvms->loc_size(), "live map consistent with locals map");
802 for (int local = 0; local < len; local++) {
803 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
804 if (PrintMiscellaneous && (Verbose || WizardMode)) {
805 tty->print_cr("Zombie local %d: ", local);
806 jvms->dump();
807 }
808 return false;
809 }
810 }
811 }
812 return true;
813 }
814
815 #endif //ASSERT
816
817 // Helper function for enforcing certain bytecodes to reexecute if
818 // deoptimization happens
819 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
820 ciMethod* cur_method = jvms->method();
821 int cur_bci = jvms->bci();
822 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
823 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
824 return Interpreter::bytecode_should_reexecute(code) ||
825 is_anewarray && code == Bytecodes::_multianewarray;
826 // Reexecute _multianewarray bytecode which was replaced with
827 // sequence of [a]newarray. See Parse::do_multianewarray().
828 //
829 // Note: interpreter should not have it set since this optimization
830 // is limited by dimensions and guarded by flag so in some cases
831 // multianewarray() runtime calls will be generated and
832 // the bytecode should not be reexecutes (stack will not be reset).
833 } else
834 return false;
835 }
836
837 // Helper function for adding JVMState and debug information to node
838 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
839 // Add the safepoint edges to the call (or other safepoint).
840
841 // Make sure dead locals are set to top. This
842 // should help register allocation time and cut down on the size
843 // of the deoptimization information.
844 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
845
846 // Walk the inline list to fill in the correct set of JVMState's
847 // Also fill in the associated edges for each JVMState.
848
849 // If the bytecode needs to be reexecuted we need to put
850 // the arguments back on the stack.
851 const bool should_reexecute = jvms()->should_reexecute();
852 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
853
854 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
855 // undefined if the bci is different. This is normal for Parse but it
856 // should not happen for LibraryCallKit because only one bci is processed.
857 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
858 "in LibraryCallKit the reexecute bit should not change");
859
860 // If we are guaranteed to throw, we can prune everything but the
861 // input to the current bytecode.
862 bool can_prune_locals = false;
863 uint stack_slots_not_pruned = 0;
864 int inputs = 0, depth = 0;
865 if (must_throw) {
866 assert(method() == youngest_jvms->method(), "sanity");
867 if (compute_stack_effects(inputs, depth)) {
868 can_prune_locals = true;
869 stack_slots_not_pruned = inputs;
870 }
871 }
872
873 if (env()->should_retain_local_variables()) {
874 // At any safepoint, this method can get breakpointed, which would
875 // then require an immediate deoptimization.
876 can_prune_locals = false; // do not prune locals
877 stack_slots_not_pruned = 0;
878 }
879
880 // do not scribble on the input jvms
881 JVMState* out_jvms = youngest_jvms->clone_deep(C);
882 call->set_jvms(out_jvms); // Start jvms list for call node
883
884 // For a known set of bytecodes, the interpreter should reexecute them if
885 // deoptimization happens. We set the reexecute state for them here
886 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
887 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
888 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
889 }
890
891 // Presize the call:
892 DEBUG_ONLY(uint non_debug_edges = call->req());
893 call->add_req_batch(top(), youngest_jvms->debug_depth());
894 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
895
896 // Set up edges so that the call looks like this:
897 // Call [state:] ctl io mem fptr retadr
898 // [parms:] parm0 ... parmN
899 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
900 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
901 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
902 // Note that caller debug info precedes callee debug info.
903
904 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
905 uint debug_ptr = call->req();
906
907 // Loop over the map input edges associated with jvms, add them
908 // to the call node, & reset all offsets to match call node array.
909 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
910 uint debug_end = debug_ptr;
911 uint debug_start = debug_ptr - in_jvms->debug_size();
912 debug_ptr = debug_start; // back up the ptr
913
914 uint p = debug_start; // walks forward in [debug_start, debug_end)
915 uint j, k, l;
916 SafePointNode* in_map = in_jvms->map();
917 out_jvms->set_map(call);
918
919 if (can_prune_locals) {
920 assert(in_jvms->method() == out_jvms->method(), "sanity");
921 // If the current throw can reach an exception handler in this JVMS,
922 // then we must keep everything live that can reach that handler.
923 // As a quick and dirty approximation, we look for any handlers at all.
924 if (in_jvms->method()->has_exception_handlers()) {
925 can_prune_locals = false;
926 }
927 }
928
929 // Add the Locals
930 k = in_jvms->locoff();
931 l = in_jvms->loc_size();
932 out_jvms->set_locoff(p);
933 if (!can_prune_locals) {
934 for (j = 0; j < l; j++)
935 call->set_req(p++, in_map->in(k+j));
936 } else {
937 p += l; // already set to top above by add_req_batch
938 }
939
940 // Add the Expression Stack
941 k = in_jvms->stkoff();
942 l = in_jvms->sp();
943 out_jvms->set_stkoff(p);
944 if (!can_prune_locals) {
945 for (j = 0; j < l; j++)
946 call->set_req(p++, in_map->in(k+j));
947 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
948 // Divide stack into {S0,...,S1}, where S0 is set to top.
949 uint s1 = stack_slots_not_pruned;
950 stack_slots_not_pruned = 0; // for next iteration
951 if (s1 > l) s1 = l;
952 uint s0 = l - s1;
953 p += s0; // skip the tops preinstalled by add_req_batch
954 for (j = s0; j < l; j++)
955 call->set_req(p++, in_map->in(k+j));
956 } else {
957 p += l; // already set to top above by add_req_batch
958 }
959
960 // Add the Monitors
961 k = in_jvms->monoff();
962 l = in_jvms->mon_size();
963 out_jvms->set_monoff(p);
964 for (j = 0; j < l; j++)
965 call->set_req(p++, in_map->in(k+j));
966
967 // Copy any scalar object fields.
968 k = in_jvms->scloff();
969 l = in_jvms->scl_size();
970 out_jvms->set_scloff(p);
971 for (j = 0; j < l; j++)
972 call->set_req(p++, in_map->in(k+j));
973
974 // Finish the new jvms.
975 out_jvms->set_endoff(p);
976
977 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
978 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
979 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
980 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
981 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
982 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
983
984 // Update the two tail pointers in parallel.
985 out_jvms = out_jvms->caller();
986 in_jvms = in_jvms->caller();
987 }
988
989 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
990
991 // Test the correctness of JVMState::debug_xxx accessors:
992 assert(call->jvms()->debug_start() == non_debug_edges, "");
993 assert(call->jvms()->debug_end() == call->req(), "");
994 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
995 }
996
997 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
998 Bytecodes::Code code = java_bc();
999 if (code == Bytecodes::_wide) {
1000 code = method()->java_code_at_bci(bci() + 1);
1001 }
1002
1003 BasicType rtype = T_ILLEGAL;
1004 int rsize = 0;
1005
1006 if (code != Bytecodes::_illegal) {
1007 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1008 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1009 if (rtype < T_CONFLICT)
1010 rsize = type2size[rtype];
1011 }
1012
1013 switch (code) {
1014 case Bytecodes::_illegal:
1015 return false;
1016
1017 case Bytecodes::_ldc:
1018 case Bytecodes::_ldc_w:
1019 case Bytecodes::_ldc2_w:
1020 inputs = 0;
1021 break;
1022
1023 case Bytecodes::_dup: inputs = 1; break;
1024 case Bytecodes::_dup_x1: inputs = 2; break;
1025 case Bytecodes::_dup_x2: inputs = 3; break;
1026 case Bytecodes::_dup2: inputs = 2; break;
1027 case Bytecodes::_dup2_x1: inputs = 3; break;
1028 case Bytecodes::_dup2_x2: inputs = 4; break;
1029 case Bytecodes::_swap: inputs = 2; break;
1030 case Bytecodes::_arraylength: inputs = 1; break;
1031
1032 case Bytecodes::_getstatic:
1033 case Bytecodes::_putstatic:
1034 case Bytecodes::_getfield:
1035 case Bytecodes::_putfield:
1036 {
1037 bool ignored_will_link;
1038 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1039 int size = field->type()->size();
1040 bool is_get = (depth >= 0), is_static = (depth & 1);
1041 inputs = (is_static ? 0 : 1);
1042 if (is_get) {
1043 depth = size - inputs;
1044 } else {
1045 inputs += size; // putxxx pops the value from the stack
1046 depth = - inputs;
1047 }
1048 }
1049 break;
1050
1051 case Bytecodes::_invokevirtual:
1052 case Bytecodes::_invokespecial:
1053 case Bytecodes::_invokestatic:
1054 case Bytecodes::_invokedynamic:
1055 case Bytecodes::_invokeinterface:
1056 {
1057 bool ignored_will_link;
1058 ciSignature* declared_signature = NULL;
1059 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1060 assert(declared_signature != NULL, "cannot be null");
1061 inputs = declared_signature->arg_size_for_bc(code);
1062 int size = declared_signature->return_type()->size();
1063 depth = size - inputs;
1064 }
1065 break;
1066
1067 case Bytecodes::_multianewarray:
1068 {
1069 ciBytecodeStream iter(method());
1070 iter.reset_to_bci(bci());
1071 iter.next();
1072 inputs = iter.get_dimensions();
1073 assert(rsize == 1, "");
1074 depth = rsize - inputs;
1075 }
1076 break;
1077
1078 case Bytecodes::_ireturn:
1079 case Bytecodes::_lreturn:
1080 case Bytecodes::_freturn:
1081 case Bytecodes::_dreturn:
1082 case Bytecodes::_areturn:
1083 assert(rsize = -depth, "");
1084 inputs = rsize;
1085 break;
1086
1087 case Bytecodes::_jsr:
1088 case Bytecodes::_jsr_w:
1089 inputs = 0;
1090 depth = 1; // S.B. depth=1, not zero
1091 break;
1092
1093 default:
1094 // bytecode produces a typed result
1095 inputs = rsize - depth;
1096 assert(inputs >= 0, "");
1097 break;
1098 }
1099
1100 #ifdef ASSERT
1101 // spot check
1102 int outputs = depth + inputs;
1103 assert(outputs >= 0, "sanity");
1104 switch (code) {
1105 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1106 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1107 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1108 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1109 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1110 }
1111 #endif //ASSERT
1112
1113 return true;
1114 }
1115
1116
1117
1118 //------------------------------basic_plus_adr---------------------------------
1119 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1120 // short-circuit a common case
1121 if (offset == intcon(0)) return ptr;
1122 return _gvn.transform( new (C) AddPNode(base, ptr, offset) );
1123 }
1124
1125 Node* GraphKit::ConvI2L(Node* offset) {
1126 // short-circuit a common case
1127 jint offset_con = find_int_con(offset, Type::OffsetBot);
1128 if (offset_con != Type::OffsetBot) {
1129 return longcon((jlong) offset_con);
1130 }
1131 return _gvn.transform( new (C) ConvI2LNode(offset));
1132 }
1133
1134 Node* GraphKit::ConvI2UL(Node* offset) {
1135 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1136 if (offset_con != (juint) Type::OffsetBot) {
1137 return longcon((julong) offset_con);
1138 }
1139 Node* conv = _gvn.transform( new (C) ConvI2LNode(offset));
1140 Node* mask = _gvn.transform( ConLNode::make(C, (julong) max_juint) );
1141 return _gvn.transform( new (C) AndLNode(conv, mask) );
1142 }
1143
1144 Node* GraphKit::ConvL2I(Node* offset) {
1145 // short-circuit a common case
1146 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1147 if (offset_con != (jlong)Type::OffsetBot) {
1148 return intcon((int) offset_con);
1149 }
1150 return _gvn.transform( new (C) ConvL2INode(offset));
1151 }
1152
1153 //-------------------------load_object_klass-----------------------------------
1154 Node* GraphKit::load_object_klass(Node* obj) {
1155 // Special-case a fresh allocation to avoid building nodes:
1156 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1157 if (akls != NULL) return akls;
1158 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1159 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1160 }
1161
1162 //-------------------------load_array_length-----------------------------------
1163 Node* GraphKit::load_array_length(Node* array) {
1164 // Special-case a fresh allocation to avoid building nodes:
1165 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1166 Node *alen;
1167 if (alloc == NULL) {
1168 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1169 alen = _gvn.transform( new (C) LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1170 } else {
1171 alen = alloc->Ideal_length();
1172 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1173 if (ccast != alen) {
1174 alen = _gvn.transform(ccast);
1175 }
1176 }
1177 return alen;
1178 }
1179
1180 //------------------------------do_null_check----------------------------------
1181 // Helper function to do a NULL pointer check. Returned value is
1182 // the incoming address with NULL casted away. You are allowed to use the
1183 // not-null value only if you are control dependent on the test.
1184 extern int explicit_null_checks_inserted,
1185 explicit_null_checks_elided;
1186 Node* GraphKit::null_check_common(Node* value, BasicType type,
1187 // optional arguments for variations:
1188 bool assert_null,
1189 Node* *null_control) {
1190 assert(!assert_null || null_control == NULL, "not both at once");
1191 if (stopped()) return top();
1192 if (!GenerateCompilerNullChecks && !assert_null && null_control == NULL) {
1193 // For some performance testing, we may wish to suppress null checking.
1194 value = cast_not_null(value); // Make it appear to be non-null (4962416).
1195 return value;
1196 }
1197 explicit_null_checks_inserted++;
1198
1199 // Construct NULL check
1200 Node *chk = NULL;
1201 switch(type) {
1202 case T_LONG : chk = new (C) CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1203 case T_INT : chk = new (C) CmpINode(value, _gvn.intcon(0)); break;
1204 case T_ARRAY : // fall through
1205 type = T_OBJECT; // simplify further tests
1206 case T_OBJECT : {
1207 const Type *t = _gvn.type( value );
1208
1209 const TypeOopPtr* tp = t->isa_oopptr();
1210 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1211 // Only for do_null_check, not any of its siblings:
1212 && !assert_null && null_control == NULL) {
1213 // Usually, any field access or invocation on an unloaded oop type
1214 // will simply fail to link, since the statically linked class is
1215 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1216 // the static class is loaded but the sharper oop type is not.
1217 // Rather than checking for this obscure case in lots of places,
1218 // we simply observe that a null check on an unloaded class
1219 // will always be followed by a nonsense operation, so we
1220 // can just issue the uncommon trap here.
1221 // Our access to the unloaded class will only be correct
1222 // after it has been loaded and initialized, which requires
1223 // a trip through the interpreter.
1224 #ifndef PRODUCT
1225 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1226 #endif
1227 uncommon_trap(Deoptimization::Reason_unloaded,
1228 Deoptimization::Action_reinterpret,
1229 tp->klass(), "!loaded");
1230 return top();
1231 }
1232
1233 if (assert_null) {
1234 // See if the type is contained in NULL_PTR.
1235 // If so, then the value is already null.
1236 if (t->higher_equal(TypePtr::NULL_PTR)) {
1237 explicit_null_checks_elided++;
1238 return value; // Elided null assert quickly!
1239 }
1240 } else {
1241 // See if mixing in the NULL pointer changes type.
1242 // If so, then the NULL pointer was not allowed in the original
1243 // type. In other words, "value" was not-null.
1244 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1245 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1246 explicit_null_checks_elided++;
1247 return value; // Elided null check quickly!
1248 }
1249 }
1250 chk = new (C) CmpPNode( value, null() );
1251 break;
1252 }
1253
1254 default:
1255 fatal(err_msg_res("unexpected type: %s", type2name(type)));
1256 }
1257 assert(chk != NULL, "sanity check");
1258 chk = _gvn.transform(chk);
1259
1260 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1261 BoolNode *btst = new (C) BoolNode( chk, btest);
1262 Node *tst = _gvn.transform( btst );
1263
1264 //-----------
1265 // if peephole optimizations occurred, a prior test existed.
1266 // If a prior test existed, maybe it dominates as we can avoid this test.
1267 if (tst != btst && type == T_OBJECT) {
1268 // At this point we want to scan up the CFG to see if we can
1269 // find an identical test (and so avoid this test altogether).
1270 Node *cfg = control();
1271 int depth = 0;
1272 while( depth < 16 ) { // Limit search depth for speed
1273 if( cfg->Opcode() == Op_IfTrue &&
1274 cfg->in(0)->in(1) == tst ) {
1275 // Found prior test. Use "cast_not_null" to construct an identical
1276 // CastPP (and hence hash to) as already exists for the prior test.
1277 // Return that casted value.
1278 if (assert_null) {
1279 replace_in_map(value, null());
1280 return null(); // do not issue the redundant test
1281 }
1282 Node *oldcontrol = control();
1283 set_control(cfg);
1284 Node *res = cast_not_null(value);
1285 set_control(oldcontrol);
1286 explicit_null_checks_elided++;
1287 return res;
1288 }
1289 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1290 if (cfg == NULL) break; // Quit at region nodes
1291 depth++;
1292 }
1293 }
1294
1295 //-----------
1296 // Branch to failure if null
1297 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1298 Deoptimization::DeoptReason reason;
1299 if (assert_null)
1300 reason = Deoptimization::Reason_null_assert;
1301 else if (type == T_OBJECT)
1302 reason = Deoptimization::Reason_null_check;
1303 else
1304 reason = Deoptimization::Reason_div0_check;
1305
1306 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1307 // ciMethodData::has_trap_at will return a conservative -1 if any
1308 // must-be-null assertion has failed. This could cause performance
1309 // problems for a method after its first do_null_assert failure.
1310 // Consider using 'Reason_class_check' instead?
1311
1312 // To cause an implicit null check, we set the not-null probability
1313 // to the maximum (PROB_MAX). For an explicit check the probability
1314 // is set to a smaller value.
1315 if (null_control != NULL || too_many_traps(reason)) {
1316 // probability is less likely
1317 ok_prob = PROB_LIKELY_MAG(3);
1318 } else if (!assert_null &&
1319 (ImplicitNullCheckThreshold > 0) &&
1320 method() != NULL &&
1321 (method()->method_data()->trap_count(reason)
1322 >= (uint)ImplicitNullCheckThreshold)) {
1323 ok_prob = PROB_LIKELY_MAG(3);
1324 }
1325
1326 if (null_control != NULL) {
1327 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1328 Node* null_true = _gvn.transform( new (C) IfFalseNode(iff));
1329 set_control( _gvn.transform( new (C) IfTrueNode(iff)));
1330 if (null_true == top())
1331 explicit_null_checks_elided++;
1332 (*null_control) = null_true;
1333 } else {
1334 BuildCutout unless(this, tst, ok_prob);
1335 // Check for optimizer eliding test at parse time
1336 if (stopped()) {
1337 // Failure not possible; do not bother making uncommon trap.
1338 explicit_null_checks_elided++;
1339 } else if (assert_null) {
1340 uncommon_trap(reason,
1341 Deoptimization::Action_make_not_entrant,
1342 NULL, "assert_null");
1343 } else {
1344 replace_in_map(value, zerocon(type));
1345 builtin_throw(reason);
1346 }
1347 }
1348
1349 // Must throw exception, fall-thru not possible?
1350 if (stopped()) {
1351 return top(); // No result
1352 }
1353
1354 if (assert_null) {
1355 // Cast obj to null on this path.
1356 replace_in_map(value, zerocon(type));
1357 return zerocon(type);
1358 }
1359
1360 // Cast obj to not-null on this path, if there is no null_control.
1361 // (If there is a null_control, a non-null value may come back to haunt us.)
1362 if (type == T_OBJECT) {
1363 Node* cast = cast_not_null(value, false);
1364 if (null_control == NULL || (*null_control) == top())
1365 replace_in_map(value, cast);
1366 value = cast;
1367 }
1368
1369 return value;
1370 }
1371
1372
1373 //------------------------------cast_not_null----------------------------------
1374 // Cast obj to not-null on this path
1375 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1376 const Type *t = _gvn.type(obj);
1377 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1378 // Object is already not-null?
1379 if( t == t_not_null ) return obj;
1380
1381 Node *cast = new (C) CastPPNode(obj,t_not_null);
1382 cast->init_req(0, control());
1383 cast = _gvn.transform( cast );
1384
1385 // Scan for instances of 'obj' in the current JVM mapping.
1386 // These instances are known to be not-null after the test.
1387 if (do_replace_in_map)
1388 replace_in_map(obj, cast);
1389
1390 return cast; // Return casted value
1391 }
1392
1393
1394 //--------------------------replace_in_map-------------------------------------
1395 void GraphKit::replace_in_map(Node* old, Node* neww) {
1396 if (old == neww) {
1397 return;
1398 }
1399
1400 map()->replace_edge(old, neww);
1401
1402 // Note: This operation potentially replaces any edge
1403 // on the map. This includes locals, stack, and monitors
1404 // of the current (innermost) JVM state.
1405
1406 // don't let inconsistent types from profiling escape this
1407 // method
1408
1409 const Type* told = _gvn.type(old);
1410 const Type* tnew = _gvn.type(neww);
1411
1412 if (!tnew->higher_equal(told)) {
1413 return;
1414 }
1415
1416 map()->record_replaced_node(old, neww);
1417 }
1418
1419
1420 //=============================================================================
1421 //--------------------------------memory---------------------------------------
1422 Node* GraphKit::memory(uint alias_idx) {
1423 MergeMemNode* mem = merged_memory();
1424 Node* p = mem->memory_at(alias_idx);
1425 _gvn.set_type(p, Type::MEMORY); // must be mapped
1426 return p;
1427 }
1428
1429 //-----------------------------reset_memory------------------------------------
1430 Node* GraphKit::reset_memory() {
1431 Node* mem = map()->memory();
1432 // do not use this node for any more parsing!
1433 debug_only( map()->set_memory((Node*)NULL) );
1434 return _gvn.transform( mem );
1435 }
1436
1437 //------------------------------set_all_memory---------------------------------
1438 void GraphKit::set_all_memory(Node* newmem) {
1439 Node* mergemem = MergeMemNode::make(C, newmem);
1440 gvn().set_type_bottom(mergemem);
1441 map()->set_memory(mergemem);
1442 }
1443
1444 //------------------------------set_all_memory_call----------------------------
1445 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1446 Node* newmem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1447 set_all_memory(newmem);
1448 }
1449
1450 //=============================================================================
1451 //
1452 // parser factory methods for MemNodes
1453 //
1454 // These are layered on top of the factory methods in LoadNode and StoreNode,
1455 // and integrate with the parser's memory state and _gvn engine.
1456 //
1457
1458 // factory methods in "int adr_idx"
1459 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1460 int adr_idx,
1461 MemNode::MemOrd mo,
1462 LoadNode::ControlDependency control_dependency,
1463 bool require_atomic_access,
1464 bool unaligned,
1465 bool mismatched) {
1466 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1467 const TypePtr* adr_type = NULL; // debug-mode-only argument
1468 debug_only(adr_type = C->get_adr_type(adr_idx));
1469 Node* mem = memory(adr_idx);
1470 Node* ld;
1471 if (require_atomic_access && bt == T_LONG) {
1472 ld = LoadLNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency);
1473 } else if (require_atomic_access && bt == T_DOUBLE) {
1474 ld = LoadDNode::make_atomic(C, ctl, mem, adr, adr_type, t, mo, control_dependency);
1475 } else {
1476 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency);
1477 }
1478 if (unaligned) {
1479 ld->as_Load()->set_unaligned_access();
1480 }
1481 if (mismatched) {
1482 ld->as_Load()->set_mismatched_access();
1483 }
1484 ld = _gvn.transform(ld);
1485 if ((bt == T_OBJECT) && C->do_escape_analysis() || C->eliminate_boxing()) {
1486 // Improve graph before escape analysis and boxing elimination.
1487 record_for_igvn(ld);
1488 }
1489 return ld;
1490 }
1491
1492 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1493 int adr_idx,
1494 MemNode::MemOrd mo,
1495 bool require_atomic_access,
1496 bool unaligned,
1497 bool mismatched) {
1498 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1499 const TypePtr* adr_type = NULL;
1500 debug_only(adr_type = C->get_adr_type(adr_idx));
1501 Node *mem = memory(adr_idx);
1502 Node* st;
1503 if (require_atomic_access && bt == T_LONG) {
1504 st = StoreLNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1505 } else if (require_atomic_access && bt == T_DOUBLE) {
1506 st = StoreDNode::make_atomic(C, ctl, mem, adr, adr_type, val, mo);
1507 } else {
1508 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1509 }
1510 if (unaligned) {
1511 st->as_Store()->set_unaligned_access();
1512 }
1513 if (mismatched) {
1514 st->as_Store()->set_mismatched_access();
1515 }
1516 st = _gvn.transform(st);
1517 set_memory(st, adr_idx);
1518 // Back-to-back stores can only remove intermediate store with DU info
1519 // so push on worklist for optimizer.
1520 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1521 record_for_igvn(st);
1522
1523 return st;
1524 }
1525
1526
1527 void GraphKit::pre_barrier(bool do_load,
1528 Node* ctl,
1529 Node* obj,
1530 Node* adr,
1531 uint adr_idx,
1532 Node* val,
1533 const TypeOopPtr* val_type,
1534 Node* pre_val,
1535 BasicType bt) {
1536
1537 BarrierSet* bs = Universe::heap()->barrier_set();
1538 set_control(ctl);
1539 switch (bs->kind()) {
1540 case BarrierSet::G1SATBCT:
1541 case BarrierSet::G1SATBCTLogging:
1542 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1543 break;
1544 case BarrierSet::ShenandoahBarrierSet:
1545 if (ShenandoahSATBBarrier) {
1546 g1_write_barrier_pre(do_load, obj, adr, adr_idx, val, val_type, pre_val, bt);
1547 }
1548 break;
1549 case BarrierSet::CardTableModRef:
1550 case BarrierSet::CardTableExtension:
1551 case BarrierSet::ModRef:
1552 break;
1553
1554 case BarrierSet::Other:
1555 default :
1556 ShouldNotReachHere();
1557
1558 }
1559 }
1560
1561 bool GraphKit::can_move_pre_barrier() const {
1562 BarrierSet* bs = Universe::heap()->barrier_set();
1563 switch (bs->kind()) {
1564 case BarrierSet::G1SATBCT:
1565 case BarrierSet::G1SATBCTLogging:
1566 case BarrierSet::ShenandoahBarrierSet:
1567 return true; // Can move it if no safepoint
1568
1569 case BarrierSet::CardTableModRef:
1570 case BarrierSet::CardTableExtension:
1571 case BarrierSet::ModRef:
1572 return true; // There is no pre-barrier
1573
1574 case BarrierSet::Other:
1575 default :
1576 ShouldNotReachHere();
1577 }
1578 return false;
1579 }
1580
1581 void GraphKit::post_barrier(Node* ctl,
1582 Node* store,
1583 Node* obj,
1584 Node* adr,
1585 uint adr_idx,
1586 Node* val,
1587 BasicType bt,
1588 bool use_precise) {
1589 BarrierSet* bs = Universe::heap()->barrier_set();
1590 set_control(ctl);
1591 switch (bs->kind()) {
1592 case BarrierSet::G1SATBCT:
1593 case BarrierSet::G1SATBCTLogging:
1594 g1_write_barrier_post(store, obj, adr, adr_idx, val, bt, use_precise);
1595 break;
1596 case BarrierSet::ShenandoahBarrierSet:
1597 if (ShenandoahStoreValEnqueueBarrier) {
1598 g1_write_barrier_pre(false, NULL, NULL, max_juint, NULL, NULL, val, bt);
1599 }
1600 break;
1601 case BarrierSet::CardTableModRef:
1602 case BarrierSet::CardTableExtension:
1603 write_barrier_post(store, obj, adr, adr_idx, val, use_precise);
1604 break;
1605
1606 case BarrierSet::ModRef:
1607 break;
1608
1609 case BarrierSet::Other:
1610 default :
1611 ShouldNotReachHere();
1612
1613 }
1614 }
1615
1616 Node* GraphKit::store_oop(Node* ctl,
1617 Node* obj,
1618 Node* adr,
1619 const TypePtr* adr_type,
1620 Node* val,
1621 const TypeOopPtr* val_type,
1622 BasicType bt,
1623 bool use_precise,
1624 MemNode::MemOrd mo,
1625 bool mismatched) {
1626 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1627 // could be delayed during Parse (for example, in adjust_map_after_if()).
1628 // Execute transformation here to avoid barrier generation in such case.
1629 if (_gvn.type(val) == TypePtr::NULL_PTR)
1630 val = _gvn.makecon(TypePtr::NULL_PTR);
1631
1632 set_control(ctl);
1633 if (stopped()) return top(); // Dead path ?
1634
1635 assert(bt == T_OBJECT, "sanity");
1636 assert(val != NULL, "not dead path");
1637 uint adr_idx = C->get_alias_index(adr_type);
1638 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1639
1640 pre_barrier(true /* do_load */,
1641 control(), obj, adr, adr_idx, val, val_type,
1642 NULL /* pre_val */,
1643 bt);
1644
1645 Node* store = store_to_memory(control(), adr, val, bt, adr_idx, mo, mismatched);
1646 post_barrier(control(), store, obj, adr, adr_idx, val, bt, use_precise);
1647 return store;
1648 }
1649
1650 // Could be an array or object we don't know at compile time (unsafe ref.)
1651 Node* GraphKit::store_oop_to_unknown(Node* ctl,
1652 Node* obj, // containing obj
1653 Node* adr, // actual adress to store val at
1654 const TypePtr* adr_type,
1655 Node* val,
1656 BasicType bt,
1657 MemNode::MemOrd mo,
1658 bool mismatched) {
1659 Compile::AliasType* at = C->alias_type(adr_type);
1660 const TypeOopPtr* val_type = NULL;
1661 if (adr_type->isa_instptr()) {
1662 if (at->field() != NULL) {
1663 // known field. This code is a copy of the do_put_xxx logic.
1664 ciField* field = at->field();
1665 if (!field->type()->is_loaded()) {
1666 val_type = TypeInstPtr::BOTTOM;
1667 } else {
1668 val_type = TypeOopPtr::make_from_klass(field->type()->as_klass());
1669 }
1670 }
1671 } else if (adr_type->isa_aryptr()) {
1672 val_type = adr_type->is_aryptr()->elem()->make_oopptr();
1673 }
1674 if (val_type == NULL) {
1675 val_type = TypeInstPtr::BOTTOM;
1676 }
1677 return store_oop(ctl, obj, adr, adr_type, val, val_type, bt, true, mo, mismatched);
1678 }
1679
1680
1681 //-------------------------array_element_address-------------------------
1682 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1683 const TypeInt* sizetype, Node* ctrl) {
1684 uint shift = exact_log2(type2aelembytes(elembt));
1685 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1686
1687 // short-circuit a common case (saves lots of confusing waste motion)
1688 jint idx_con = find_int_con(idx, -1);
1689 if (idx_con >= 0) {
1690 intptr_t offset = header + ((intptr_t)idx_con << shift);
1691 return basic_plus_adr(ary, offset);
1692 }
1693
1694 // must be correct type for alignment purposes
1695 Node* base = basic_plus_adr(ary, header);
1696 #ifdef _LP64
1697 // The scaled index operand to AddP must be a clean 64-bit value.
1698 // Java allows a 32-bit int to be incremented to a negative
1699 // value, which appears in a 64-bit register as a large
1700 // positive number. Using that large positive number as an
1701 // operand in pointer arithmetic has bad consequences.
1702 // On the other hand, 32-bit overflow is rare, and the possibility
1703 // can often be excluded, if we annotate the ConvI2L node with
1704 // a type assertion that its value is known to be a small positive
1705 // number. (The prior range check has ensured this.)
1706 // This assertion is used by ConvI2LNode::Ideal.
1707 int index_max = max_jint - 1; // array size is max_jint, index is one less
1708 if (sizetype != NULL) index_max = sizetype->_hi - 1;
1709 const TypeInt* iidxtype = TypeInt::make(0, index_max, Type::WidenMax);
1710 idx = C->constrained_convI2L(&_gvn, idx, iidxtype, ctrl);
1711 #endif
1712 Node* scale = _gvn.transform( new (C) LShiftXNode(idx, intcon(shift)) );
1713 return basic_plus_adr(ary, base, scale);
1714 }
1715
1716 //-------------------------load_array_element-------------------------
1717 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1718 const Type* elemtype = arytype->elem();
1719 BasicType elembt = elemtype->array_element_basic_type();
1720 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1721 if (elembt == T_NARROWOOP) {
1722 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1723 }
1724 assert(elembt != T_OBJECT && elembt != T_ARRAY, "sanity");
1725 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1726 return ld;
1727 }
1728
1729 //-------------------------set_arguments_for_java_call-------------------------
1730 // Arguments (pre-popped from the stack) are taken from the JVMS.
1731 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1732 // Add the call arguments:
1733 uint nargs = call->method()->arg_size();
1734 for (uint i = 0; i < nargs; i++) {
1735 Node* arg = argument(i);
1736 call->init_req(i + TypeFunc::Parms, arg);
1737 }
1738 }
1739
1740 //---------------------------set_edges_for_java_call---------------------------
1741 // Connect a newly created call into the current JVMS.
1742 // A return value node (if any) is returned from set_edges_for_java_call.
1743 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1744
1745 // Add the predefined inputs:
1746 call->init_req( TypeFunc::Control, control() );
1747 call->init_req( TypeFunc::I_O , i_o() );
1748 call->init_req( TypeFunc::Memory , reset_memory() );
1749 call->init_req( TypeFunc::FramePtr, frameptr() );
1750 call->init_req( TypeFunc::ReturnAdr, top() );
1751
1752 add_safepoint_edges(call, must_throw);
1753
1754 Node* xcall = _gvn.transform(call);
1755
1756 if (xcall == top()) {
1757 set_control(top());
1758 return;
1759 }
1760 assert(xcall == call, "call identity is stable");
1761
1762 // Re-use the current map to produce the result.
1763
1764 set_control(_gvn.transform(new (C) ProjNode(call, TypeFunc::Control)));
1765 set_i_o( _gvn.transform(new (C) ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1766 set_all_memory_call(xcall, separate_io_proj);
1767
1768 //return xcall; // no need, caller already has it
1769 }
1770
1771 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj) {
1772 if (stopped()) return top(); // maybe the call folded up?
1773
1774 // Capture the return value, if any.
1775 Node* ret;
1776 if (call->method() == NULL ||
1777 call->method()->return_type()->basic_type() == T_VOID)
1778 ret = top();
1779 else ret = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1780
1781 // Note: Since any out-of-line call can produce an exception,
1782 // we always insert an I_O projection from the call into the result.
1783
1784 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj);
1785
1786 if (separate_io_proj) {
1787 // The caller requested separate projections be used by the fall
1788 // through and exceptional paths, so replace the projections for
1789 // the fall through path.
1790 set_i_o(_gvn.transform( new (C) ProjNode(call, TypeFunc::I_O) ));
1791 set_all_memory(_gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) ));
1792 }
1793 return ret;
1794 }
1795
1796 //--------------------set_predefined_input_for_runtime_call--------------------
1797 // Reading and setting the memory state is way conservative here.
1798 // The real problem is that I am not doing real Type analysis on memory,
1799 // so I cannot distinguish card mark stores from other stores. Across a GC
1800 // point the Store Barrier and the card mark memory has to agree. I cannot
1801 // have a card mark store and its barrier split across the GC point from
1802 // either above or below. Here I get that to happen by reading ALL of memory.
1803 // A better answer would be to separate out card marks from other memory.
1804 // For now, return the input memory state, so that it can be reused
1805 // after the call, if this call has restricted memory effects.
1806 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1807 // Set fixed predefined input arguments
1808 Node* memory = reset_memory();
1809 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1810 call->init_req( TypeFunc::Control, control() );
1811 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1812 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1813 call->init_req( TypeFunc::FramePtr, frameptr() );
1814 call->init_req( TypeFunc::ReturnAdr, top() );
1815 return memory;
1816 }
1817
1818 //-------------------set_predefined_output_for_runtime_call--------------------
1819 // Set control and memory (not i_o) from the call.
1820 // If keep_mem is not NULL, use it for the output state,
1821 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1822 // If hook_mem is NULL, this call produces no memory effects at all.
1823 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1824 // then only that memory slice is taken from the call.
1825 // In the last case, we must put an appropriate memory barrier before
1826 // the call, so as to create the correct anti-dependencies on loads
1827 // preceding the call.
1828 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1829 Node* keep_mem,
1830 const TypePtr* hook_mem) {
1831 // no i/o
1832 set_control(_gvn.transform( new (C) ProjNode(call,TypeFunc::Control) ));
1833 if (keep_mem) {
1834 // First clone the existing memory state
1835 set_all_memory(keep_mem);
1836 if (hook_mem != NULL) {
1837 // Make memory for the call
1838 Node* mem = _gvn.transform( new (C) ProjNode(call, TypeFunc::Memory) );
1839 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1840 // We also use hook_mem to extract specific effects from arraycopy stubs.
1841 set_memory(mem, hook_mem);
1842 }
1843 // ...else the call has NO memory effects.
1844
1845 // Make sure the call advertises its memory effects precisely.
1846 // This lets us build accurate anti-dependences in gcm.cpp.
1847 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1848 "call node must be constructed correctly");
1849 } else {
1850 assert(hook_mem == NULL, "");
1851 // This is not a "slow path" call; all memory comes from the call.
1852 set_all_memory_call(call);
1853 }
1854 }
1855
1856
1857 // Replace the call with the current state of the kit.
1858 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1859 JVMState* ejvms = NULL;
1860 if (has_exceptions()) {
1861 ejvms = transfer_exceptions_into_jvms();
1862 }
1863
1864 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1865 ReplacedNodes replaced_nodes_exception;
1866 Node* ex_ctl = top();
1867
1868 SafePointNode* final_state = stop();
1869
1870 // Find all the needed outputs of this call
1871 CallProjections callprojs;
1872 call->extract_projections(&callprojs, true);
1873
1874 Node* init_mem = call->in(TypeFunc::Memory);
1875 Node* final_mem = final_state->in(TypeFunc::Memory);
1876 Node* final_ctl = final_state->in(TypeFunc::Control);
1877 Node* final_io = final_state->in(TypeFunc::I_O);
1878
1879 // Replace all the old call edges with the edges from the inlining result
1880 if (callprojs.fallthrough_catchproj != NULL) {
1881 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
1882 }
1883 if (callprojs.fallthrough_memproj != NULL) {
1884 if (final_mem->is_MergeMem()) {
1885 // Parser's exits MergeMem was not transformed but may be optimized
1886 final_mem = _gvn.transform(final_mem);
1887 }
1888 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
1889 }
1890 if (callprojs.fallthrough_ioproj != NULL) {
1891 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
1892 }
1893
1894 // Replace the result with the new result if it exists and is used
1895 if (callprojs.resproj != NULL && result != NULL) {
1896 C->gvn_replace_by(callprojs.resproj, result);
1897 }
1898
1899 if (ejvms == NULL) {
1900 // No exception edges to simply kill off those paths
1901 if (callprojs.catchall_catchproj != NULL) {
1902 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
1903 }
1904 if (callprojs.catchall_memproj != NULL) {
1905 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
1906 }
1907 if (callprojs.catchall_ioproj != NULL) {
1908 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
1909 }
1910 // Replace the old exception object with top
1911 if (callprojs.exobj != NULL) {
1912 C->gvn_replace_by(callprojs.exobj, C->top());
1913 }
1914 } else {
1915 GraphKit ekit(ejvms);
1916
1917 // Load my combined exception state into the kit, with all phis transformed:
1918 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
1919 replaced_nodes_exception = ex_map->replaced_nodes();
1920
1921 Node* ex_oop = ekit.use_exception_state(ex_map);
1922
1923 if (callprojs.catchall_catchproj != NULL) {
1924 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
1925 ex_ctl = ekit.control();
1926 }
1927 if (callprojs.catchall_memproj != NULL) {
1928 C->gvn_replace_by(callprojs.catchall_memproj, ekit.reset_memory());
1929 }
1930 if (callprojs.catchall_ioproj != NULL) {
1931 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
1932 }
1933
1934 // Replace the old exception object with the newly created one
1935 if (callprojs.exobj != NULL) {
1936 C->gvn_replace_by(callprojs.exobj, ex_oop);
1937 }
1938 }
1939
1940 // Disconnect the call from the graph
1941 call->disconnect_inputs(NULL, C);
1942 C->gvn_replace_by(call, C->top());
1943
1944 // Clean up any MergeMems that feed other MergeMems since the
1945 // optimizer doesn't like that.
1946 if (final_mem->is_MergeMem()) {
1947 Node_List wl;
1948 for (SimpleDUIterator i(final_mem); i.has_next(); i.next()) {
1949 Node* m = i.get();
1950 if (m->is_MergeMem() && !wl.contains(m)) {
1951 wl.push(m);
1952 }
1953 }
1954 while (wl.size() > 0) {
1955 _gvn.transform(wl.pop());
1956 }
1957 }
1958
1959 if (callprojs.fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
1960 replaced_nodes.apply(C, final_ctl);
1961 }
1962 if (!ex_ctl->is_top() && do_replaced_nodes) {
1963 replaced_nodes_exception.apply(C, ex_ctl);
1964 }
1965 }
1966
1967
1968 //------------------------------increment_counter------------------------------
1969 // for statistics: increment a VM counter by 1
1970
1971 void GraphKit::increment_counter(address counter_addr) {
1972 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
1973 increment_counter(adr1);
1974 }
1975
1976 void GraphKit::increment_counter(Node* counter_addr) {
1977 int adr_type = Compile::AliasIdxRaw;
1978 Node* ctrl = control();
1979 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
1980 Node* incr = _gvn.transform(new (C) AddINode(cnt, _gvn.intcon(1)));
1981 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
1982 }
1983
1984
1985 //------------------------------uncommon_trap----------------------------------
1986 // Bail out to the interpreter in mid-method. Implemented by calling the
1987 // uncommon_trap blob. This helper function inserts a runtime call with the
1988 // right debug info.
1989 void GraphKit::uncommon_trap(int trap_request,
1990 ciKlass* klass, const char* comment,
1991 bool must_throw,
1992 bool keep_exact_action) {
1993 if (failing()) stop();
1994 if (stopped()) return; // trap reachable?
1995
1996 // Note: If ProfileTraps is true, and if a deopt. actually
1997 // occurs here, the runtime will make sure an MDO exists. There is
1998 // no need to call method()->ensure_method_data() at this point.
1999
2000 // Set the stack pointer to the right value for reexecution:
2001 set_sp(reexecute_sp());
2002
2003 #ifdef ASSERT
2004 if (!must_throw) {
2005 // Make sure the stack has at least enough depth to execute
2006 // the current bytecode.
2007 int inputs, ignored_depth;
2008 if (compute_stack_effects(inputs, ignored_depth)) {
2009 assert(sp() >= inputs, err_msg_res("must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2010 Bytecodes::name(java_bc()), sp(), inputs));
2011 }
2012 }
2013 #endif
2014
2015 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2016 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2017
2018 switch (action) {
2019 case Deoptimization::Action_maybe_recompile:
2020 case Deoptimization::Action_reinterpret:
2021 // Temporary fix for 6529811 to allow virtual calls to be sure they
2022 // get the chance to go from mono->bi->mega
2023 if (!keep_exact_action &&
2024 Deoptimization::trap_request_index(trap_request) < 0 &&
2025 too_many_recompiles(reason)) {
2026 // This BCI is causing too many recompilations.
2027 if (C->log() != NULL) {
2028 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2029 Deoptimization::trap_reason_name(reason),
2030 Deoptimization::trap_action_name(action));
2031 }
2032 action = Deoptimization::Action_none;
2033 trap_request = Deoptimization::make_trap_request(reason, action);
2034 } else {
2035 C->set_trap_can_recompile(true);
2036 }
2037 break;
2038 case Deoptimization::Action_make_not_entrant:
2039 C->set_trap_can_recompile(true);
2040 break;
2041 #ifdef ASSERT
2042 case Deoptimization::Action_none:
2043 case Deoptimization::Action_make_not_compilable:
2044 break;
2045 default:
2046 fatal(err_msg_res("unknown action %d: %s", action, Deoptimization::trap_action_name(action)));
2047 break;
2048 #endif
2049 }
2050
2051 if (TraceOptoParse) {
2052 char buf[100];
2053 tty->print_cr("Uncommon trap %s at bci:%d",
2054 Deoptimization::format_trap_request(buf, sizeof(buf),
2055 trap_request), bci());
2056 }
2057
2058 CompileLog* log = C->log();
2059 if (log != NULL) {
2060 int kid = (klass == NULL)? -1: log->identify(klass);
2061 log->begin_elem("uncommon_trap bci='%d'", bci());
2062 char buf[100];
2063 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2064 trap_request));
2065 if (kid >= 0) log->print(" klass='%d'", kid);
2066 if (comment != NULL) log->print(" comment='%s'", comment);
2067 log->end_elem();
2068 }
2069
2070 // Make sure any guarding test views this path as very unlikely
2071 Node *i0 = control()->in(0);
2072 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2073 IfNode *iff = i0->as_If();
2074 float f = iff->_prob; // Get prob
2075 if (control()->Opcode() == Op_IfTrue) {
2076 if (f > PROB_UNLIKELY_MAG(4))
2077 iff->_prob = PROB_MIN;
2078 } else {
2079 if (f < PROB_LIKELY_MAG(4))
2080 iff->_prob = PROB_MAX;
2081 }
2082 }
2083
2084 // Clear out dead values from the debug info.
2085 kill_dead_locals();
2086
2087 // Now insert the uncommon trap subroutine call
2088 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2089 const TypePtr* no_memory_effects = NULL;
2090 // Pass the index of the class to be loaded
2091 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2092 (must_throw ? RC_MUST_THROW : 0),
2093 OptoRuntime::uncommon_trap_Type(),
2094 call_addr, "uncommon_trap", no_memory_effects,
2095 intcon(trap_request));
2096 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2097 "must extract request correctly from the graph");
2098 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2099
2100 call->set_req(TypeFunc::ReturnAdr, returnadr());
2101 // The debug info is the only real input to this call.
2102
2103 // Halt-and-catch fire here. The above call should never return!
2104 HaltNode* halt = new(C) HaltNode(control(), frameptr());
2105 _gvn.set_type_bottom(halt);
2106 root()->add_req(halt);
2107
2108 stop_and_kill_map();
2109 }
2110
2111
2112 //--------------------------just_allocated_object------------------------------
2113 // Report the object that was just allocated.
2114 // It must be the case that there are no intervening safepoints.
2115 // We use this to determine if an object is so "fresh" that
2116 // it does not require card marks.
2117 Node* GraphKit::just_allocated_object(Node* current_control) {
2118 if (C->recent_alloc_ctl() == current_control)
2119 return C->recent_alloc_obj();
2120 return NULL;
2121 }
2122
2123
2124 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2125 // (Note: TypeFunc::make has a cache that makes this fast.)
2126 const TypeFunc* tf = TypeFunc::make(dest_method);
2127 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2128 for (int j = 0; j < nargs; j++) {
2129 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2130 if( targ->basic_type() == T_DOUBLE ) {
2131 // If any parameters are doubles, they must be rounded before
2132 // the call, dstore_rounding does gvn.transform
2133 Node *arg = argument(j);
2134 arg = dstore_rounding(arg);
2135 set_argument(j, arg);
2136 }
2137 }
2138 }
2139
2140 /**
2141 * Record profiling data exact_kls for Node n with the type system so
2142 * that it can propagate it (speculation)
2143 *
2144 * @param n node that the type applies to
2145 * @param exact_kls type from profiling
2146 *
2147 * @return node with improved type
2148 */
2149 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls) {
2150 const Type* current_type = _gvn.type(n);
2151 assert(UseTypeSpeculation, "type speculation must be on");
2152
2153 const TypeOopPtr* speculative = current_type->speculative();
2154
2155 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2156 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2157 const TypeOopPtr* xtype = tklass->as_instance_type();
2158 assert(xtype->klass_is_exact(), "Should be exact");
2159 // record the new speculative type's depth
2160 speculative = xtype->with_inline_depth(jvms()->depth());
2161 }
2162
2163 if (speculative != current_type->speculative()) {
2164 // Build a type with a speculative type (what we think we know
2165 // about the type but will need a guard when we use it)
2166 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2167 // We're changing the type, we need a new CheckCast node to carry
2168 // the new type. The new type depends on the control: what
2169 // profiling tells us is only valid from here as far as we can
2170 // tell.
2171 Node* cast = new(C) CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2172 cast = _gvn.transform(cast);
2173 replace_in_map(n, cast);
2174 n = cast;
2175 }
2176
2177 return n;
2178 }
2179
2180 /**
2181 * Record profiling data from receiver profiling at an invoke with the
2182 * type system so that it can propagate it (speculation)
2183 *
2184 * @param n receiver node
2185 *
2186 * @return node with improved type
2187 */
2188 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2189 if (!UseTypeSpeculation) {
2190 return n;
2191 }
2192 ciKlass* exact_kls = profile_has_unique_klass();
2193 return record_profile_for_speculation(n, exact_kls);
2194 }
2195
2196 /**
2197 * Record profiling data from argument profiling at an invoke with the
2198 * type system so that it can propagate it (speculation)
2199 *
2200 * @param dest_method target method for the call
2201 * @param bc what invoke bytecode is this?
2202 */
2203 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2204 if (!UseTypeSpeculation) {
2205 return;
2206 }
2207 const TypeFunc* tf = TypeFunc::make(dest_method);
2208 int nargs = tf->_domain->_cnt - TypeFunc::Parms;
2209 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2210 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2211 const Type *targ = tf->_domain->field_at(j + TypeFunc::Parms);
2212 if (targ->basic_type() == T_OBJECT || targ->basic_type() == T_ARRAY) {
2213 ciKlass* better_type = method()->argument_profiled_type(bci(), i);
2214 if (better_type != NULL) {
2215 record_profile_for_speculation(argument(j), better_type);
2216 }
2217 i++;
2218 }
2219 }
2220 }
2221
2222 /**
2223 * Record profiling data from parameter profiling at an invoke with
2224 * the type system so that it can propagate it (speculation)
2225 */
2226 void GraphKit::record_profiled_parameters_for_speculation() {
2227 if (!UseTypeSpeculation) {
2228 return;
2229 }
2230 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2231 if (_gvn.type(local(i))->isa_oopptr()) {
2232 ciKlass* better_type = method()->parameter_profiled_type(j);
2233 if (better_type != NULL) {
2234 record_profile_for_speculation(local(i), better_type);
2235 }
2236 j++;
2237 }
2238 }
2239 }
2240
2241 void GraphKit::round_double_result(ciMethod* dest_method) {
2242 // A non-strict method may return a double value which has an extended
2243 // exponent, but this must not be visible in a caller which is 'strict'
2244 // If a strict caller invokes a non-strict callee, round a double result
2245
2246 BasicType result_type = dest_method->return_type()->basic_type();
2247 assert( method() != NULL, "must have caller context");
2248 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2249 // Destination method's return value is on top of stack
2250 // dstore_rounding() does gvn.transform
2251 Node *result = pop_pair();
2252 result = dstore_rounding(result);
2253 push_pair(result);
2254 }
2255 }
2256
2257 // rounding for strict float precision conformance
2258 Node* GraphKit::precision_rounding(Node* n) {
2259 return UseStrictFP && _method->flags().is_strict()
2260 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2261 ? _gvn.transform( new (C) RoundFloatNode(0, n) )
2262 : n;
2263 }
2264
2265 // rounding for strict double precision conformance
2266 Node* GraphKit::dprecision_rounding(Node *n) {
2267 return UseStrictFP && _method->flags().is_strict()
2268 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2269 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2270 : n;
2271 }
2272
2273 // rounding for non-strict double stores
2274 Node* GraphKit::dstore_rounding(Node* n) {
2275 return Matcher::strict_fp_requires_explicit_rounding
2276 && UseSSE <= 1
2277 ? _gvn.transform( new (C) RoundDoubleNode(0, n) )
2278 : n;
2279 }
2280
2281 //=============================================================================
2282 // Generate a fast path/slow path idiom. Graph looks like:
2283 // [foo] indicates that 'foo' is a parameter
2284 //
2285 // [in] NULL
2286 // \ /
2287 // CmpP
2288 // Bool ne
2289 // If
2290 // / \
2291 // True False-<2>
2292 // / |
2293 // / cast_not_null
2294 // Load | | ^
2295 // [fast_test] | |
2296 // gvn to opt_test | |
2297 // / \ | <1>
2298 // True False |
2299 // | \\ |
2300 // [slow_call] \[fast_result]
2301 // Ctl Val \ \
2302 // | \ \
2303 // Catch <1> \ \
2304 // / \ ^ \ \
2305 // Ex No_Ex | \ \
2306 // | \ \ | \ <2> \
2307 // ... \ [slow_res] | | \ [null_result]
2308 // \ \--+--+--- | |
2309 // \ | / \ | /
2310 // --------Region Phi
2311 //
2312 //=============================================================================
2313 // Code is structured as a series of driver functions all called 'do_XXX' that
2314 // call a set of helper functions. Helper functions first, then drivers.
2315
2316 //------------------------------null_check_oop---------------------------------
2317 // Null check oop. Set null-path control into Region in slot 3.
2318 // Make a cast-not-nullness use the other not-null control. Return cast.
2319 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2320 bool never_see_null, bool safe_for_replace) {
2321 // Initial NULL check taken path
2322 (*null_control) = top();
2323 Node* cast = null_check_common(value, T_OBJECT, false, null_control);
2324
2325 // Generate uncommon_trap:
2326 if (never_see_null && (*null_control) != top()) {
2327 // If we see an unexpected null at a check-cast we record it and force a
2328 // recompile; the offending check-cast will be compiled to handle NULLs.
2329 // If we see more than one offending BCI, then all checkcasts in the
2330 // method will be compiled to handle NULLs.
2331 PreserveJVMState pjvms(this);
2332 set_control(*null_control);
2333 replace_in_map(value, null());
2334 uncommon_trap(Deoptimization::Reason_null_check,
2335 Deoptimization::Action_make_not_entrant);
2336 (*null_control) = top(); // NULL path is dead
2337 }
2338 if ((*null_control) == top() && safe_for_replace) {
2339 replace_in_map(value, cast);
2340 }
2341
2342 // Cast away null-ness on the result
2343 return cast;
2344 }
2345
2346 //------------------------------opt_iff----------------------------------------
2347 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2348 // Return slow-path control.
2349 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2350 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2351
2352 // Fast path taken; set region slot 2
2353 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(opt_iff) );
2354 region->init_req(2,fast_taken); // Capture fast-control
2355
2356 // Fast path not-taken, i.e. slow path
2357 Node *slow_taken = _gvn.transform( new (C) IfTrueNode(opt_iff) );
2358 return slow_taken;
2359 }
2360
2361 //-----------------------------make_runtime_call-------------------------------
2362 Node* GraphKit::make_runtime_call(int flags,
2363 const TypeFunc* call_type, address call_addr,
2364 const char* call_name,
2365 const TypePtr* adr_type,
2366 // The following parms are all optional.
2367 // The first NULL ends the list.
2368 Node* parm0, Node* parm1,
2369 Node* parm2, Node* parm3,
2370 Node* parm4, Node* parm5,
2371 Node* parm6, Node* parm7) {
2372 // Slow-path call
2373 bool is_leaf = !(flags & RC_NO_LEAF);
2374 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2375 if (call_name == NULL) {
2376 assert(!is_leaf, "must supply name for leaf");
2377 call_name = OptoRuntime::stub_name(call_addr);
2378 }
2379 CallNode* call;
2380 if (!is_leaf) {
2381 call = new(C) CallStaticJavaNode(call_type, call_addr, call_name,
2382 bci(), adr_type);
2383 } else if (flags & RC_NO_FP) {
2384 call = new(C) CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2385 } else {
2386 call = new(C) CallLeafNode(call_type, call_addr, call_name, adr_type);
2387 }
2388
2389 // The following is similar to set_edges_for_java_call,
2390 // except that the memory effects of the call are restricted to AliasIdxRaw.
2391
2392 // Slow path call has no side-effects, uses few values
2393 bool wide_in = !(flags & RC_NARROW_MEM);
2394 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2395
2396 Node* prev_mem = NULL;
2397 if (wide_in) {
2398 prev_mem = set_predefined_input_for_runtime_call(call);
2399 } else {
2400 assert(!wide_out, "narrow in => narrow out");
2401 Node* narrow_mem = memory(adr_type);
2402 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2403 }
2404
2405 // Hook each parm in order. Stop looking at the first NULL.
2406 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2407 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2408 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2409 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2410 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2411 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2412 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2413 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2414 /* close each nested if ===> */ } } } } } } } }
2415 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2416
2417 if (!is_leaf) {
2418 // Non-leaves can block and take safepoints:
2419 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2420 }
2421 // Non-leaves can throw exceptions:
2422 if (has_io) {
2423 call->set_req(TypeFunc::I_O, i_o());
2424 }
2425
2426 if (flags & RC_UNCOMMON) {
2427 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2428 // (An "if" probability corresponds roughly to an unconditional count.
2429 // Sort of.)
2430 call->set_cnt(PROB_UNLIKELY_MAG(4));
2431 }
2432
2433 Node* c = _gvn.transform(call);
2434 assert(c == call, "cannot disappear");
2435
2436 if (wide_out) {
2437 // Slow path call has full side-effects.
2438 set_predefined_output_for_runtime_call(call);
2439 } else {
2440 // Slow path call has few side-effects, and/or sets few values.
2441 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2442 }
2443
2444 if (has_io) {
2445 set_i_o(_gvn.transform(new (C) ProjNode(call, TypeFunc::I_O)));
2446 }
2447 return call;
2448
2449 }
2450
2451 //------------------------------merge_memory-----------------------------------
2452 // Merge memory from one path into the current memory state.
2453 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2454 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2455 Node* old_slice = mms.force_memory();
2456 Node* new_slice = mms.memory2();
2457 if (old_slice != new_slice) {
2458 PhiNode* phi;
2459 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2460 if (mms.is_empty()) {
2461 // clone base memory Phi's inputs for this memory slice
2462 assert(old_slice == mms.base_memory(), "sanity");
2463 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2464 _gvn.set_type(phi, Type::MEMORY);
2465 for (uint i = 1; i < phi->req(); i++) {
2466 phi->init_req(i, old_slice->in(i));
2467 }
2468 } else {
2469 phi = old_slice->as_Phi(); // Phi was generated already
2470 }
2471 } else {
2472 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2473 _gvn.set_type(phi, Type::MEMORY);
2474 }
2475 phi->set_req(new_path, new_slice);
2476 mms.set_memory(phi);
2477 }
2478 }
2479 }
2480
2481 //------------------------------make_slow_call_ex------------------------------
2482 // Make the exception handler hookups for the slow call
2483 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2484 if (stopped()) return;
2485
2486 // Make a catch node with just two handlers: fall-through and catch-all
2487 Node* i_o = _gvn.transform( new (C) ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2488 Node* catc = _gvn.transform( new (C) CatchNode(control(), i_o, 2) );
2489 Node* norm = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2490 Node* excp = _gvn.transform( new (C) CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2491
2492 { PreserveJVMState pjvms(this);
2493 set_control(excp);
2494 set_i_o(i_o);
2495
2496 if (excp != top()) {
2497 if (deoptimize) {
2498 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2499 uncommon_trap(Deoptimization::Reason_unhandled,
2500 Deoptimization::Action_none);
2501 } else {
2502 // Create an exception state also.
2503 // Use an exact type if the caller has specified a specific exception.
2504 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2505 Node* ex_oop = new (C) CreateExNode(ex_type, control(), i_o);
2506 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2507 }
2508 }
2509 }
2510
2511 // Get the no-exception control from the CatchNode.
2512 set_control(norm);
2513 }
2514
2515
2516 //-------------------------------gen_subtype_check-----------------------------
2517 // Generate a subtyping check. Takes as input the subtype and supertype.
2518 // Returns 2 values: sets the default control() to the true path and returns
2519 // the false path. Only reads invariant memory; sets no (visible) memory.
2520 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2521 // but that's not exposed to the optimizer. This call also doesn't take in an
2522 // Object; if you wish to check an Object you need to load the Object's class
2523 // prior to coming here.
2524 Node* GraphKit::gen_subtype_check(Node* subklass, Node* superklass) {
2525 // Fast check for identical types, perhaps identical constants.
2526 // The types can even be identical non-constants, in cases
2527 // involving Array.newInstance, Object.clone, etc.
2528 if (subklass == superklass)
2529 return top(); // false path is dead; no test needed.
2530
2531 if (_gvn.type(superklass)->singleton()) {
2532 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2533 ciKlass* subk = _gvn.type(subklass)->is_klassptr()->klass();
2534
2535 // In the common case of an exact superklass, try to fold up the
2536 // test before generating code. You may ask, why not just generate
2537 // the code and then let it fold up? The answer is that the generated
2538 // code will necessarily include null checks, which do not always
2539 // completely fold away. If they are also needless, then they turn
2540 // into a performance loss. Example:
2541 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2542 // Here, the type of 'fa' is often exact, so the store check
2543 // of fa[1]=x will fold up, without testing the nullness of x.
2544 switch (static_subtype_check(superk, subk)) {
2545 case SSC_always_false:
2546 {
2547 Node* always_fail = control();
2548 set_control(top());
2549 return always_fail;
2550 }
2551 case SSC_always_true:
2552 return top();
2553 case SSC_easy_test:
2554 {
2555 // Just do a direct pointer compare and be done.
2556 Node* cmp = _gvn.transform( new(C) CmpPNode(subklass, superklass) );
2557 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2558 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2559 set_control( _gvn.transform( new(C) IfTrueNode (iff) ) );
2560 return _gvn.transform( new(C) IfFalseNode(iff) );
2561 }
2562 case SSC_full_test:
2563 break;
2564 default:
2565 ShouldNotReachHere();
2566 }
2567 }
2568
2569 // %%% Possible further optimization: Even if the superklass is not exact,
2570 // if the subklass is the unique subtype of the superklass, the check
2571 // will always succeed. We could leave a dependency behind to ensure this.
2572
2573 // First load the super-klass's check-offset
2574 Node *p1 = basic_plus_adr( superklass, superklass, in_bytes(Klass::super_check_offset_offset()) );
2575 Node *chk_off = _gvn.transform(new (C) LoadINode(NULL, memory(p1), p1, _gvn.type(p1)->is_ptr(),
2576 TypeInt::INT, MemNode::unordered));
2577 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2578 bool might_be_cache = (find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2579
2580 // Load from the sub-klass's super-class display list, or a 1-word cache of
2581 // the secondary superclass list, or a failing value with a sentinel offset
2582 // if the super-klass is an interface or exceptionally deep in the Java
2583 // hierarchy and we have to scan the secondary superclass list the hard way.
2584 // Worst-case type is a little odd: NULL is allowed as a result (usually
2585 // klass loads can never produce a NULL).
2586 Node *chk_off_X = ConvI2X(chk_off);
2587 Node *p2 = _gvn.transform( new (C) AddPNode(subklass,subklass,chk_off_X) );
2588 // For some types like interfaces the following loadKlass is from a 1-word
2589 // cache which is mutable so can't use immutable memory. Other
2590 // types load from the super-class display table which is immutable.
2591 Node *kmem = might_be_cache ? memory(p2) : immutable_memory();
2592 Node* nkls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, kmem, p2, _gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2593
2594 // Compile speed common case: ARE a subtype and we canNOT fail
2595 if( superklass == nkls )
2596 return top(); // false path is dead; no test needed.
2597
2598 // See if we get an immediate positive hit. Happens roughly 83% of the
2599 // time. Test to see if the value loaded just previously from the subklass
2600 // is exactly the superklass.
2601 Node *cmp1 = _gvn.transform( new (C) CmpPNode( superklass, nkls ) );
2602 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp1, BoolTest::eq ) );
2603 IfNode *iff1 = create_and_xform_if( control(), bol1, PROB_LIKELY(0.83f), COUNT_UNKNOWN );
2604 Node *iftrue1 = _gvn.transform( new (C) IfTrueNode ( iff1 ) );
2605 set_control( _gvn.transform( new (C) IfFalseNode( iff1 ) ) );
2606
2607 // Compile speed common case: Check for being deterministic right now. If
2608 // chk_off is a constant and not equal to cacheoff then we are NOT a
2609 // subklass. In this case we need exactly the 1 test above and we can
2610 // return those results immediately.
2611 if (!might_be_cache) {
2612 Node* not_subtype_ctrl = control();
2613 set_control(iftrue1); // We need exactly the 1 test above
2614 return not_subtype_ctrl;
2615 }
2616
2617 // Gather the various success & failures here
2618 RegionNode *r_ok_subtype = new (C) RegionNode(4);
2619 record_for_igvn(r_ok_subtype);
2620 RegionNode *r_not_subtype = new (C) RegionNode(3);
2621 record_for_igvn(r_not_subtype);
2622
2623 r_ok_subtype->init_req(1, iftrue1);
2624
2625 // Check for immediate negative hit. Happens roughly 11% of the time (which
2626 // is roughly 63% of the remaining cases). Test to see if the loaded
2627 // check-offset points into the subklass display list or the 1-element
2628 // cache. If it points to the display (and NOT the cache) and the display
2629 // missed then it's not a subtype.
2630 Node *cacheoff = _gvn.intcon(cacheoff_con);
2631 Node *cmp2 = _gvn.transform( new (C) CmpINode( chk_off, cacheoff ) );
2632 Node *bol2 = _gvn.transform( new (C) BoolNode( cmp2, BoolTest::ne ) );
2633 IfNode *iff2 = create_and_xform_if( control(), bol2, PROB_LIKELY(0.63f), COUNT_UNKNOWN );
2634 r_not_subtype->init_req(1, _gvn.transform( new (C) IfTrueNode (iff2) ) );
2635 set_control( _gvn.transform( new (C) IfFalseNode(iff2) ) );
2636
2637 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2638 // No performance impact (too rare) but allows sharing of secondary arrays
2639 // which has some footprint reduction.
2640 Node *cmp3 = _gvn.transform( new (C) CmpPNode( subklass, superklass ) );
2641 Node *bol3 = _gvn.transform( new (C) BoolNode( cmp3, BoolTest::eq ) );
2642 IfNode *iff3 = create_and_xform_if( control(), bol3, PROB_LIKELY(0.36f), COUNT_UNKNOWN );
2643 r_ok_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode ( iff3 ) ) );
2644 set_control( _gvn.transform( new (C) IfFalseNode( iff3 ) ) );
2645
2646 // -- Roads not taken here: --
2647 // We could also have chosen to perform the self-check at the beginning
2648 // of this code sequence, as the assembler does. This would not pay off
2649 // the same way, since the optimizer, unlike the assembler, can perform
2650 // static type analysis to fold away many successful self-checks.
2651 // Non-foldable self checks work better here in second position, because
2652 // the initial primary superclass check subsumes a self-check for most
2653 // types. An exception would be a secondary type like array-of-interface,
2654 // which does not appear in its own primary supertype display.
2655 // Finally, we could have chosen to move the self-check into the
2656 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2657 // dependent manner. But it is worthwhile to have the check here,
2658 // where it can be perhaps be optimized. The cost in code space is
2659 // small (register compare, branch).
2660
2661 // Now do a linear scan of the secondary super-klass array. Again, no real
2662 // performance impact (too rare) but it's gotta be done.
2663 // Since the code is rarely used, there is no penalty for moving it
2664 // out of line, and it can only improve I-cache density.
2665 // The decision to inline or out-of-line this final check is platform
2666 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2667 Node* psc = _gvn.transform(
2668 new (C) PartialSubtypeCheckNode(control(), subklass, superklass) );
2669
2670 Node *cmp4 = _gvn.transform( new (C) CmpPNode( psc, null() ) );
2671 Node *bol4 = _gvn.transform( new (C) BoolNode( cmp4, BoolTest::ne ) );
2672 IfNode *iff4 = create_and_xform_if( control(), bol4, PROB_FAIR, COUNT_UNKNOWN );
2673 r_not_subtype->init_req(2, _gvn.transform( new (C) IfTrueNode (iff4) ) );
2674 r_ok_subtype ->init_req(3, _gvn.transform( new (C) IfFalseNode(iff4) ) );
2675
2676 // Return false path; set default control to true path.
2677 set_control( _gvn.transform(r_ok_subtype) );
2678 return _gvn.transform(r_not_subtype);
2679 }
2680
2681 //----------------------------static_subtype_check-----------------------------
2682 // Shortcut important common cases when superklass is exact:
2683 // (0) superklass is java.lang.Object (can occur in reflective code)
2684 // (1) subklass is already limited to a subtype of superklass => always ok
2685 // (2) subklass does not overlap with superklass => always fail
2686 // (3) superklass has NO subtypes and we can check with a simple compare.
2687 int GraphKit::static_subtype_check(ciKlass* superk, ciKlass* subk) {
2688 if (StressReflectiveCode) {
2689 return SSC_full_test; // Let caller generate the general case.
2690 }
2691
2692 if (superk == env()->Object_klass()) {
2693 return SSC_always_true; // (0) this test cannot fail
2694 }
2695
2696 ciType* superelem = superk;
2697 if (superelem->is_array_klass())
2698 superelem = superelem->as_array_klass()->base_element_type();
2699
2700 if (!subk->is_interface()) { // cannot trust static interface types yet
2701 if (subk->is_subtype_of(superk)) {
2702 return SSC_always_true; // (1) false path dead; no dynamic test needed
2703 }
2704 if (!(superelem->is_klass() && superelem->as_klass()->is_interface()) &&
2705 !superk->is_subtype_of(subk)) {
2706 return SSC_always_false;
2707 }
2708 }
2709
2710 // If casting to an instance klass, it must have no subtypes
2711 if (superk->is_interface()) {
2712 // Cannot trust interfaces yet.
2713 // %%% S.B. superk->nof_implementors() == 1
2714 } else if (superelem->is_instance_klass()) {
2715 ciInstanceKlass* ik = superelem->as_instance_klass();
2716 if (!ik->has_subklass() && !ik->is_interface()) {
2717 if (!ik->is_final()) {
2718 // Add a dependency if there is a chance of a later subclass.
2719 C->dependencies()->assert_leaf_type(ik);
2720 }
2721 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2722 }
2723 } else {
2724 // A primitive array type has no subtypes.
2725 return SSC_easy_test; // (3) caller can do a simple ptr comparison
2726 }
2727
2728 return SSC_full_test;
2729 }
2730
2731 // Profile-driven exact type check:
2732 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2733 float prob,
2734 Node* *casted_receiver) {
2735 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2736 Node* recv_klass = load_object_klass(receiver);
2737 Node* want_klass = makecon(tklass);
2738 Node* cmp = _gvn.transform( new(C) CmpPNode(recv_klass, want_klass) );
2739 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
2740 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2741 set_control( _gvn.transform( new(C) IfTrueNode (iff) ));
2742 Node* fail = _gvn.transform( new(C) IfFalseNode(iff) );
2743
2744 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2745 assert(recv_xtype->klass_is_exact(), "");
2746
2747 // Subsume downstream occurrences of receiver with a cast to
2748 // recv_xtype, since now we know what the type will be.
2749 Node* cast = new(C) CheckCastPPNode(control(), receiver, recv_xtype);
2750 (*casted_receiver) = _gvn.transform(cast);
2751 // (User must make the replace_in_map call.)
2752
2753 return fail;
2754 }
2755
2756
2757 //------------------------------seems_never_null-------------------------------
2758 // Use null_seen information if it is available from the profile.
2759 // If we see an unexpected null at a type check we record it and force a
2760 // recompile; the offending check will be recompiled to handle NULLs.
2761 // If we see several offending BCIs, then all checks in the
2762 // method will be recompiled.
2763 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data) {
2764 if (UncommonNullCast // Cutout for this technique
2765 && obj != null() // And not the -Xcomp stupid case?
2766 && !too_many_traps(Deoptimization::Reason_null_check)
2767 ) {
2768 if (data == NULL)
2769 // Edge case: no mature data. Be optimistic here.
2770 return true;
2771 // If the profile has not seen a null, assume it won't happen.
2772 assert(java_bc() == Bytecodes::_checkcast ||
2773 java_bc() == Bytecodes::_instanceof ||
2774 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2775 return !data->as_BitData()->null_seen();
2776 }
2777 return false;
2778 }
2779
2780 //------------------------maybe_cast_profiled_receiver-------------------------
2781 // If the profile has seen exactly one type, narrow to exactly that type.
2782 // Subsequent type checks will always fold up.
2783 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
2784 ciKlass* require_klass,
2785 ciKlass* spec_klass,
2786 bool safe_for_replace) {
2787 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
2788
2789 Deoptimization::DeoptReason reason = spec_klass == NULL ? Deoptimization::Reason_class_check : Deoptimization::Reason_speculate_class_check;
2790
2791 // Make sure we haven't already deoptimized from this tactic.
2792 if (too_many_traps(reason) || too_many_recompiles(reason))
2793 return NULL;
2794
2795 // (No, this isn't a call, but it's enough like a virtual call
2796 // to use the same ciMethod accessor to get the profile info...)
2797 // If we have a speculative type use it instead of profiling (which
2798 // may not help us)
2799 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
2800 if (exact_kls != NULL) {// no cast failures here
2801 if (require_klass == NULL ||
2802 static_subtype_check(require_klass, exact_kls) == SSC_always_true) {
2803 // If we narrow the type to match what the type profile sees or
2804 // the speculative type, we can then remove the rest of the
2805 // cast.
2806 // This is a win, even if the exact_kls is very specific,
2807 // because downstream operations, such as method calls,
2808 // will often benefit from the sharper type.
2809 Node* exact_obj = not_null_obj; // will get updated in place...
2810 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2811 &exact_obj);
2812 { PreserveJVMState pjvms(this);
2813 set_control(slow_ctl);
2814 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2815 }
2816 if (safe_for_replace) {
2817 replace_in_map(not_null_obj, exact_obj);
2818 }
2819 return exact_obj;
2820 }
2821 // assert(ssc == SSC_always_true)... except maybe the profile lied to us.
2822 }
2823
2824 return NULL;
2825 }
2826
2827 /**
2828 * Cast obj to type and emit guard unless we had too many traps here
2829 * already
2830 *
2831 * @param obj node being casted
2832 * @param type type to cast the node to
2833 * @param not_null true if we know node cannot be null
2834 */
2835 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
2836 ciKlass* type,
2837 bool not_null) {
2838 // type == NULL if profiling tells us this object is always null
2839 if (type != NULL) {
2840 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
2841 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_null_check;
2842 if (!too_many_traps(null_reason) && !too_many_recompiles(null_reason) &&
2843 !too_many_traps(class_reason) && !too_many_recompiles(class_reason)) {
2844 Node* not_null_obj = NULL;
2845 // not_null is true if we know the object is not null and
2846 // there's no need for a null check
2847 if (!not_null) {
2848 Node* null_ctl = top();
2849 not_null_obj = null_check_oop(obj, &null_ctl, true, true);
2850 assert(null_ctl->is_top(), "no null control here");
2851 } else {
2852 not_null_obj = obj;
2853 }
2854
2855 Node* exact_obj = not_null_obj;
2856 ciKlass* exact_kls = type;
2857 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
2858 &exact_obj);
2859 {
2860 PreserveJVMState pjvms(this);
2861 set_control(slow_ctl);
2862 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
2863 }
2864 replace_in_map(not_null_obj, exact_obj);
2865 obj = exact_obj;
2866 }
2867 } else {
2868 if (!too_many_traps(Deoptimization::Reason_null_assert) &&
2869 !too_many_recompiles(Deoptimization::Reason_null_assert)) {
2870 Node* exact_obj = null_assert(obj);
2871 replace_in_map(obj, exact_obj);
2872 obj = exact_obj;
2873 }
2874 }
2875 return obj;
2876 }
2877
2878 //-------------------------------gen_instanceof--------------------------------
2879 // Generate an instance-of idiom. Used by both the instance-of bytecode
2880 // and the reflective instance-of call.
2881 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
2882 kill_dead_locals(); // Benefit all the uncommon traps
2883 assert( !stopped(), "dead parse path should be checked in callers" );
2884 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
2885 "must check for not-null not-dead klass in callers");
2886
2887 // Make the merge point
2888 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
2889 RegionNode* region = new(C) RegionNode(PATH_LIMIT);
2890 Node* phi = new(C) PhiNode(region, TypeInt::BOOL);
2891 C->set_has_split_ifs(true); // Has chance for split-if optimization
2892
2893 ciProfileData* data = NULL;
2894 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
2895 data = method()->method_data()->bci_to_data(bci());
2896 }
2897 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
2898 && seems_never_null(obj, data));
2899
2900 // Null check; get casted pointer; set region slot 3
2901 Node* null_ctl = top();
2902 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
2903
2904 // If not_null_obj is dead, only null-path is taken
2905 if (stopped()) { // Doing instance-of on a NULL?
2906 set_control(null_ctl);
2907 return intcon(0);
2908 }
2909 region->init_req(_null_path, null_ctl);
2910 phi ->init_req(_null_path, intcon(0)); // Set null path value
2911 if (null_ctl == top()) {
2912 // Do this eagerly, so that pattern matches like is_diamond_phi
2913 // will work even during parsing.
2914 assert(_null_path == PATH_LIMIT-1, "delete last");
2915 region->del_req(_null_path);
2916 phi ->del_req(_null_path);
2917 }
2918
2919 // Do we know the type check always succeed?
2920 bool known_statically = false;
2921 if (_gvn.type(superklass)->singleton()) {
2922 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
2923 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
2924 if (subk != NULL && subk->is_loaded()) {
2925 int static_res = static_subtype_check(superk, subk);
2926 known_statically = (static_res == SSC_always_true || static_res == SSC_always_false);
2927 }
2928 }
2929
2930 if (known_statically && UseTypeSpeculation) {
2931 // If we know the type check always succeeds then we don't use the
2932 // profiling data at this bytecode. Don't lose it, feed it to the
2933 // type system as a speculative type.
2934 not_null_obj = record_profiled_receiver_for_speculation(not_null_obj);
2935 } else {
2936 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2937 // We may not have profiling here or it may not help us. If we
2938 // have a speculative type use it to perform an exact cast.
2939 ciKlass* spec_obj_type = obj_type->speculative_type();
2940 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
2941 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
2942 if (stopped()) { // Profile disagrees with this path.
2943 set_control(null_ctl); // Null is the only remaining possibility.
2944 return intcon(0);
2945 }
2946 if (cast_obj != NULL) {
2947 not_null_obj = cast_obj;
2948 }
2949 }
2950 }
2951
2952 // Load the object's klass
2953 Node* obj_klass = load_object_klass(not_null_obj);
2954
2955 // Generate the subtype check
2956 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
2957
2958 // Plug in the success path to the general merge in slot 1.
2959 region->init_req(_obj_path, control());
2960 phi ->init_req(_obj_path, intcon(1));
2961
2962 // Plug in the failing path to the general merge in slot 2.
2963 region->init_req(_fail_path, not_subtype_ctrl);
2964 phi ->init_req(_fail_path, intcon(0));
2965
2966 // Return final merged results
2967 set_control( _gvn.transform(region) );
2968 record_for_igvn(region);
2969 return _gvn.transform(phi);
2970 }
2971
2972 //-------------------------------gen_checkcast---------------------------------
2973 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
2974 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
2975 // uncommon-trap paths work. Adjust stack after this call.
2976 // If failure_control is supplied and not null, it is filled in with
2977 // the control edge for the cast failure. Otherwise, an appropriate
2978 // uncommon trap or exception is thrown.
2979 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
2980 Node* *failure_control) {
2981 kill_dead_locals(); // Benefit all the uncommon traps
2982 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr();
2983 const Type *toop = TypeOopPtr::make_from_klass(tk->klass());
2984
2985 // Fast cutout: Check the case that the cast is vacuously true.
2986 // This detects the common cases where the test will short-circuit
2987 // away completely. We do this before we perform the null check,
2988 // because if the test is going to turn into zero code, we don't
2989 // want a residual null check left around. (Causes a slowdown,
2990 // for example, in some objArray manipulations, such as a[i]=a[j].)
2991 if (tk->singleton()) {
2992 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
2993 if (objtp != NULL && objtp->klass() != NULL) {
2994 switch (static_subtype_check(tk->klass(), objtp->klass())) {
2995 case SSC_always_true:
2996 // If we know the type check always succeed then we don't use
2997 // the profiling data at this bytecode. Don't lose it, feed it
2998 // to the type system as a speculative type.
2999 return record_profiled_receiver_for_speculation(obj);
3000 case SSC_always_false:
3001 // It needs a null check because a null will *pass* the cast check.
3002 // A non-null value will always produce an exception.
3003 return null_assert(obj);
3004 }
3005 }
3006 }
3007
3008 ciProfileData* data = NULL;
3009 bool safe_for_replace = false;
3010 if (failure_control == NULL) { // use MDO in regular case only
3011 assert(java_bc() == Bytecodes::_aastore ||
3012 java_bc() == Bytecodes::_checkcast,
3013 "interpreter profiles type checks only for these BCs");
3014 data = method()->method_data()->bci_to_data(bci());
3015 safe_for_replace = true;
3016 }
3017
3018 // Make the merge point
3019 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3020 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3021 Node* phi = new (C) PhiNode(region, toop);
3022 C->set_has_split_ifs(true); // Has chance for split-if optimization
3023
3024 // Use null-cast information if it is available
3025 bool never_see_null = ((failure_control == NULL) // regular case only
3026 && seems_never_null(obj, data));
3027
3028 // Null check; get casted pointer; set region slot 3
3029 Node* null_ctl = top();
3030 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace);
3031
3032 // If not_null_obj is dead, only null-path is taken
3033 if (stopped()) { // Doing instance-of on a NULL?
3034 set_control(null_ctl);
3035 return null();
3036 }
3037 region->init_req(_null_path, null_ctl);
3038 phi ->init_req(_null_path, null()); // Set null path value
3039 if (null_ctl == top()) {
3040 // Do this eagerly, so that pattern matches like is_diamond_phi
3041 // will work even during parsing.
3042 assert(_null_path == PATH_LIMIT-1, "delete last");
3043 region->del_req(_null_path);
3044 phi ->del_req(_null_path);
3045 }
3046
3047 Node* cast_obj = NULL;
3048 if (tk->klass_is_exact()) {
3049 // The following optimization tries to statically cast the speculative type of the object
3050 // (for example obtained during profiling) to the type of the superklass and then do a
3051 // dynamic check that the type of the object is what we expect. To work correctly
3052 // for checkcast and aastore the type of superklass should be exact.
3053 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3054 // We may not have profiling here or it may not help us. If we have
3055 // a speculative type use it to perform an exact cast.
3056 ciKlass* spec_obj_type = obj_type->speculative_type();
3057 if (spec_obj_type != NULL ||
3058 (data != NULL &&
3059 // Counter has never been decremented (due to cast failure).
3060 // ...This is a reasonable thing to expect. It is true of
3061 // all casts inserted by javac to implement generic types.
3062 data->as_CounterData()->count() >= 0)) {
3063 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3064 if (cast_obj != NULL) {
3065 if (failure_control != NULL) // failure is now impossible
3066 (*failure_control) = top();
3067 // adjust the type of the phi to the exact klass:
3068 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3069 }
3070 }
3071 }
3072
3073 if (cast_obj == NULL) {
3074 // Load the object's klass
3075 Node* obj_klass = load_object_klass(not_null_obj);
3076
3077 // Generate the subtype check
3078 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3079
3080 // Plug in success path into the merge
3081 cast_obj = _gvn.transform(new (C) CheckCastPPNode(control(),
3082 not_null_obj, toop));
3083 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3084 if (failure_control == NULL) {
3085 if (not_subtype_ctrl != top()) { // If failure is possible
3086 PreserveJVMState pjvms(this);
3087 set_control(not_subtype_ctrl);
3088 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3089 }
3090 } else {
3091 (*failure_control) = not_subtype_ctrl;
3092 }
3093 }
3094
3095 region->init_req(_obj_path, control());
3096 phi ->init_req(_obj_path, cast_obj);
3097
3098 // A merge of NULL or Casted-NotNull obj
3099 Node* res = _gvn.transform(phi);
3100
3101 // Note I do NOT always 'replace_in_map(obj,result)' here.
3102 // if( tk->klass()->can_be_primary_super() )
3103 // This means that if I successfully store an Object into an array-of-String
3104 // I 'forget' that the Object is really now known to be a String. I have to
3105 // do this because we don't have true union types for interfaces - if I store
3106 // a Baz into an array-of-Interface and then tell the optimizer it's an
3107 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3108 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3109 // replace_in_map( obj, res );
3110
3111 // Return final merged results
3112 set_control( _gvn.transform(region) );
3113 record_for_igvn(region);
3114 return res;
3115 }
3116
3117 //------------------------------next_monitor-----------------------------------
3118 // What number should be given to the next monitor?
3119 int GraphKit::next_monitor() {
3120 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3121 int next = current + C->sync_stack_slots();
3122 // Keep the toplevel high water mark current:
3123 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3124 return current;
3125 }
3126
3127 //------------------------------insert_mem_bar---------------------------------
3128 // Memory barrier to avoid floating things around
3129 // The membar serves as a pinch point between both control and all memory slices.
3130 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3131 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3132 mb->init_req(TypeFunc::Control, control());
3133 mb->init_req(TypeFunc::Memory, reset_memory());
3134 Node* membar = _gvn.transform(mb);
3135 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3136 set_all_memory_call(membar);
3137 return membar;
3138 }
3139
3140 //-------------------------insert_mem_bar_volatile----------------------------
3141 // Memory barrier to avoid floating things around
3142 // The membar serves as a pinch point between both control and memory(alias_idx).
3143 // If you want to make a pinch point on all memory slices, do not use this
3144 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3145 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3146 // When Parse::do_put_xxx updates a volatile field, it appends a series
3147 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3148 // The first membar is on the same memory slice as the field store opcode.
3149 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3150 // All the other membars (for other volatile slices, including AliasIdxBot,
3151 // which stands for all unknown volatile slices) are control-dependent
3152 // on the first membar. This prevents later volatile loads or stores
3153 // from sliding up past the just-emitted store.
3154
3155 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3156 mb->set_req(TypeFunc::Control,control());
3157 if (alias_idx == Compile::AliasIdxBot) {
3158 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3159 } else {
3160 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3161 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3162 }
3163 Node* membar = _gvn.transform(mb);
3164 set_control(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Control)));
3165 if (alias_idx == Compile::AliasIdxBot) {
3166 merged_memory()->set_base_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)));
3167 } else {
3168 set_memory(_gvn.transform(new (C) ProjNode(membar, TypeFunc::Memory)),alias_idx);
3169 }
3170 return membar;
3171 }
3172
3173 //------------------------------shared_lock------------------------------------
3174 // Emit locking code.
3175 FastLockNode* GraphKit::shared_lock(Node* obj) {
3176 // bci is either a monitorenter bc or InvocationEntryBci
3177 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3178 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3179
3180 if( !GenerateSynchronizationCode )
3181 return NULL; // Not locking things?
3182 if (stopped()) // Dead monitor?
3183 return NULL;
3184
3185 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3186
3187 // Box the stack location
3188 Node* box = _gvn.transform(new (C) BoxLockNode(next_monitor()));
3189 Node* mem = reset_memory();
3190
3191 FastLockNode * flock = _gvn.transform(new (C) FastLockNode(0, obj, box) )->as_FastLock();
3192 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3193 // Create the counters for this fast lock.
3194 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3195 }
3196
3197 // Create the rtm counters for this fast lock if needed.
3198 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3199
3200 // Add monitor to debug info for the slow path. If we block inside the
3201 // slow path and de-opt, we need the monitor hanging around
3202 map()->push_monitor( flock );
3203
3204 const TypeFunc *tf = LockNode::lock_type();
3205 LockNode *lock = new (C) LockNode(C, tf);
3206
3207 lock->init_req( TypeFunc::Control, control() );
3208 lock->init_req( TypeFunc::Memory , mem );
3209 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3210 lock->init_req( TypeFunc::FramePtr, frameptr() );
3211 lock->init_req( TypeFunc::ReturnAdr, top() );
3212
3213 lock->init_req(TypeFunc::Parms + 0, obj);
3214 lock->init_req(TypeFunc::Parms + 1, box);
3215 lock->init_req(TypeFunc::Parms + 2, flock);
3216 add_safepoint_edges(lock);
3217
3218 lock = _gvn.transform( lock )->as_Lock();
3219
3220 // lock has no side-effects, sets few values
3221 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3222
3223 insert_mem_bar(Op_MemBarAcquireLock);
3224
3225 // Add this to the worklist so that the lock can be eliminated
3226 record_for_igvn(lock);
3227
3228 #ifndef PRODUCT
3229 if (PrintLockStatistics) {
3230 // Update the counter for this lock. Don't bother using an atomic
3231 // operation since we don't require absolute accuracy.
3232 lock->create_lock_counter(map()->jvms());
3233 increment_counter(lock->counter()->addr());
3234 }
3235 #endif
3236
3237 return flock;
3238 }
3239
3240
3241 //------------------------------shared_unlock----------------------------------
3242 // Emit unlocking code.
3243 void GraphKit::shared_unlock(Node* box, Node* obj) {
3244 // bci is either a monitorenter bc or InvocationEntryBci
3245 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3246 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3247
3248 if( !GenerateSynchronizationCode )
3249 return;
3250 if (stopped()) { // Dead monitor?
3251 map()->pop_monitor(); // Kill monitor from debug info
3252 return;
3253 }
3254
3255 // Memory barrier to avoid floating things down past the locked region
3256 insert_mem_bar(Op_MemBarReleaseLock);
3257
3258 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3259 UnlockNode *unlock = new (C) UnlockNode(C, tf);
3260 #ifdef ASSERT
3261 unlock->set_dbg_jvms(sync_jvms());
3262 #endif
3263 uint raw_idx = Compile::AliasIdxRaw;
3264 unlock->init_req( TypeFunc::Control, control() );
3265 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3266 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3267 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3268 unlock->init_req( TypeFunc::ReturnAdr, top() );
3269
3270 unlock->init_req(TypeFunc::Parms + 0, obj);
3271 unlock->init_req(TypeFunc::Parms + 1, box);
3272 unlock = _gvn.transform(unlock)->as_Unlock();
3273
3274 Node* mem = reset_memory();
3275
3276 // unlock has no side-effects, sets few values
3277 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3278
3279 // Kill monitor from debug info
3280 map()->pop_monitor( );
3281 }
3282
3283 //-------------------------------get_layout_helper-----------------------------
3284 // If the given klass is a constant or known to be an array,
3285 // fetch the constant layout helper value into constant_value
3286 // and return (Node*)NULL. Otherwise, load the non-constant
3287 // layout helper value, and return the node which represents it.
3288 // This two-faced routine is useful because allocation sites
3289 // almost always feature constant types.
3290 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3291 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3292 if (!StressReflectiveCode && inst_klass != NULL) {
3293 ciKlass* klass = inst_klass->klass();
3294 bool xklass = inst_klass->klass_is_exact();
3295 if (xklass || klass->is_array_klass()) {
3296 jint lhelper = klass->layout_helper();
3297 if (lhelper != Klass::_lh_neutral_value) {
3298 constant_value = lhelper;
3299 return (Node*) NULL;
3300 }
3301 }
3302 }
3303 constant_value = Klass::_lh_neutral_value; // put in a known value
3304 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3305 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3306 }
3307
3308 // We just put in an allocate/initialize with a big raw-memory effect.
3309 // Hook selected additional alias categories on the initialization.
3310 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3311 MergeMemNode* init_in_merge,
3312 Node* init_out_raw) {
3313 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3314 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3315
3316 Node* prevmem = kit.memory(alias_idx);
3317 init_in_merge->set_memory_at(alias_idx, prevmem);
3318 kit.set_memory(init_out_raw, alias_idx);
3319 }
3320
3321 //---------------------------set_output_for_allocation-------------------------
3322 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3323 const TypeOopPtr* oop_type,
3324 bool deoptimize_on_exception) {
3325 int rawidx = Compile::AliasIdxRaw;
3326 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3327 add_safepoint_edges(alloc);
3328 Node* allocx = _gvn.transform(alloc);
3329 set_control( _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Control) ) );
3330 // create memory projection for i_o
3331 set_memory ( _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3332 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3333
3334 // create a memory projection as for the normal control path
3335 Node* malloc = _gvn.transform(new (C) ProjNode(allocx, TypeFunc::Memory));
3336 set_memory(malloc, rawidx);
3337
3338 // a normal slow-call doesn't change i_o, but an allocation does
3339 // we create a separate i_o projection for the normal control path
3340 set_i_o(_gvn.transform( new (C) ProjNode(allocx, TypeFunc::I_O, false) ) );
3341 Node* rawoop = _gvn.transform( new (C) ProjNode(allocx, TypeFunc::Parms) );
3342
3343 // put in an initialization barrier
3344 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3345 rawoop)->as_Initialize();
3346 assert(alloc->initialization() == init, "2-way macro link must work");
3347 assert(init ->allocation() == alloc, "2-way macro link must work");
3348 {
3349 // Extract memory strands which may participate in the new object's
3350 // initialization, and source them from the new InitializeNode.
3351 // This will allow us to observe initializations when they occur,
3352 // and link them properly (as a group) to the InitializeNode.
3353 assert(init->in(InitializeNode::Memory) == malloc, "");
3354 MergeMemNode* minit_in = MergeMemNode::make(C, malloc);
3355 init->set_req(InitializeNode::Memory, minit_in);
3356 record_for_igvn(minit_in); // fold it up later, if possible
3357 Node* minit_out = memory(rawidx);
3358 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3359 if (oop_type->isa_aryptr()) {
3360 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3361 int elemidx = C->get_alias_index(telemref);
3362 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3363 } else if (oop_type->isa_instptr()) {
3364 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3365 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3366 ciField* field = ik->nonstatic_field_at(i);
3367 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3368 continue; // do not bother to track really large numbers of fields
3369 // Find (or create) the alias category for this field:
3370 int fieldidx = C->alias_type(field)->index();
3371 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3372 }
3373 }
3374 }
3375
3376 // Cast raw oop to the real thing...
3377 Node* javaoop = new (C) CheckCastPPNode(control(), rawoop, oop_type);
3378 javaoop = _gvn.transform(javaoop);
3379 C->set_recent_alloc(control(), javaoop);
3380 assert(just_allocated_object(control()) == javaoop, "just allocated");
3381
3382 #ifdef ASSERT
3383 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3384 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3385 "Ideal_allocation works");
3386 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3387 "Ideal_allocation works");
3388 if (alloc->is_AllocateArray()) {
3389 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3390 "Ideal_allocation works");
3391 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3392 "Ideal_allocation works");
3393 } else {
3394 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3395 }
3396 }
3397 #endif //ASSERT
3398
3399 return javaoop;
3400 }
3401
3402 //---------------------------new_instance--------------------------------------
3403 // This routine takes a klass_node which may be constant (for a static type)
3404 // or may be non-constant (for reflective code). It will work equally well
3405 // for either, and the graph will fold nicely if the optimizer later reduces
3406 // the type to a constant.
3407 // The optional arguments are for specialized use by intrinsics:
3408 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3409 // - If 'return_size_val', report the the total object size to the caller.
3410 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3411 Node* GraphKit::new_instance(Node* klass_node,
3412 Node* extra_slow_test,
3413 Node* *return_size_val,
3414 bool deoptimize_on_exception) {
3415 // Compute size in doublewords
3416 // The size is always an integral number of doublewords, represented
3417 // as a positive bytewise size stored in the klass's layout_helper.
3418 // The layout_helper also encodes (in a low bit) the need for a slow path.
3419 jint layout_con = Klass::_lh_neutral_value;
3420 Node* layout_val = get_layout_helper(klass_node, layout_con);
3421 int layout_is_con = (layout_val == NULL);
3422
3423 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3424 // Generate the initial go-slow test. It's either ALWAYS (return a
3425 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3426 // case) a computed value derived from the layout_helper.
3427 Node* initial_slow_test = NULL;
3428 if (layout_is_con) {
3429 assert(!StressReflectiveCode, "stress mode does not use these paths");
3430 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3431 initial_slow_test = must_go_slow? intcon(1): extra_slow_test;
3432
3433 } else { // reflective case
3434 // This reflective path is used by Unsafe.allocateInstance.
3435 // (It may be stress-tested by specifying StressReflectiveCode.)
3436 // Basically, we want to get into the VM is there's an illegal argument.
3437 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3438 initial_slow_test = _gvn.transform( new (C) AndINode(layout_val, bit) );
3439 if (extra_slow_test != intcon(0)) {
3440 initial_slow_test = _gvn.transform( new (C) OrINode(initial_slow_test, extra_slow_test) );
3441 }
3442 // (Macro-expander will further convert this to a Bool, if necessary.)
3443 }
3444
3445 // Find the size in bytes. This is easy; it's the layout_helper.
3446 // The size value must be valid even if the slow path is taken.
3447 Node* size = NULL;
3448 if (layout_is_con) {
3449 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3450 } else { // reflective case
3451 // This reflective path is used by clone and Unsafe.allocateInstance.
3452 size = ConvI2X(layout_val);
3453
3454 // Clear the low bits to extract layout_helper_size_in_bytes:
3455 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3456 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3457 size = _gvn.transform( new (C) AndXNode(size, mask) );
3458 }
3459 if (return_size_val != NULL) {
3460 (*return_size_val) = size;
3461 }
3462
3463 // This is a precise notnull oop of the klass.
3464 // (Actually, it need not be precise if this is a reflective allocation.)
3465 // It's what we cast the result to.
3466 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3467 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3468 const TypeOopPtr* oop_type = tklass->as_instance_type();
3469
3470 // Now generate allocation code
3471
3472 // The entire memory state is needed for slow path of the allocation
3473 // since GC and deoptimization can happened.
3474 Node *mem = reset_memory();
3475 set_all_memory(mem); // Create new memory state
3476
3477 AllocateNode* alloc
3478 = new (C) AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3479 control(), mem, i_o(),
3480 size, klass_node,
3481 initial_slow_test);
3482
3483 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3484 }
3485
3486 //-------------------------------new_array-------------------------------------
3487 // helper for both newarray and anewarray
3488 // The 'length' parameter is (obviously) the length of the array.
3489 // See comments on new_instance for the meaning of the other arguments.
3490 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3491 Node* length, // number of array elements
3492 int nargs, // number of arguments to push back for uncommon trap
3493 Node* *return_size_val,
3494 bool deoptimize_on_exception) {
3495 jint layout_con = Klass::_lh_neutral_value;
3496 Node* layout_val = get_layout_helper(klass_node, layout_con);
3497 int layout_is_con = (layout_val == NULL);
3498
3499 if (!layout_is_con && !StressReflectiveCode &&
3500 !too_many_traps(Deoptimization::Reason_class_check)) {
3501 // This is a reflective array creation site.
3502 // Optimistically assume that it is a subtype of Object[],
3503 // so that we can fold up all the address arithmetic.
3504 layout_con = Klass::array_layout_helper(T_OBJECT);
3505 Node* cmp_lh = _gvn.transform( new(C) CmpINode(layout_val, intcon(layout_con)) );
3506 Node* bol_lh = _gvn.transform( new(C) BoolNode(cmp_lh, BoolTest::eq) );
3507 { BuildCutout unless(this, bol_lh, PROB_MAX);
3508 inc_sp(nargs);
3509 uncommon_trap(Deoptimization::Reason_class_check,
3510 Deoptimization::Action_maybe_recompile);
3511 }
3512 layout_val = NULL;
3513 layout_is_con = true;
3514 }
3515
3516 // Generate the initial go-slow test. Make sure we do not overflow
3517 // if length is huge (near 2Gig) or negative! We do not need
3518 // exact double-words here, just a close approximation of needed
3519 // double-words. We can't add any offset or rounding bits, lest we
3520 // take a size -1 of bytes and make it positive. Use an unsigned
3521 // compare, so negative sizes look hugely positive.
3522 int fast_size_limit = FastAllocateSizeLimit;
3523 if (layout_is_con) {
3524 assert(!StressReflectiveCode, "stress mode does not use these paths");
3525 // Increase the size limit if we have exact knowledge of array type.
3526 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3527 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3528 }
3529
3530 Node* initial_slow_cmp = _gvn.transform( new (C) CmpUNode( length, intcon( fast_size_limit ) ) );
3531 Node* initial_slow_test = _gvn.transform( new (C) BoolNode( initial_slow_cmp, BoolTest::gt ) );
3532
3533 // --- Size Computation ---
3534 // array_size = round_to_heap(array_header + (length << elem_shift));
3535 // where round_to_heap(x) == round_to(x, MinObjAlignmentInBytes)
3536 // and round_to(x, y) == ((x + y-1) & ~(y-1))
3537 // The rounding mask is strength-reduced, if possible.
3538 int round_mask = MinObjAlignmentInBytes - 1;
3539 Node* header_size = NULL;
3540 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3541 // (T_BYTE has the weakest alignment and size restrictions...)
3542 if (layout_is_con) {
3543 int hsize = Klass::layout_helper_header_size(layout_con);
3544 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3545 BasicType etype = Klass::layout_helper_element_type(layout_con);
3546 if ((round_mask & ~right_n_bits(eshift)) == 0)
3547 round_mask = 0; // strength-reduce it if it goes away completely
3548 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3549 assert(header_size_min <= hsize, "generic minimum is smallest");
3550 header_size_min = hsize;
3551 header_size = intcon(hsize + round_mask);
3552 } else {
3553 Node* hss = intcon(Klass::_lh_header_size_shift);
3554 Node* hsm = intcon(Klass::_lh_header_size_mask);
3555 Node* hsize = _gvn.transform( new(C) URShiftINode(layout_val, hss) );
3556 hsize = _gvn.transform( new(C) AndINode(hsize, hsm) );
3557 Node* mask = intcon(round_mask);
3558 header_size = _gvn.transform( new(C) AddINode(hsize, mask) );
3559 }
3560
3561 Node* elem_shift = NULL;
3562 if (layout_is_con) {
3563 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3564 if (eshift != 0)
3565 elem_shift = intcon(eshift);
3566 } else {
3567 // There is no need to mask or shift this value.
3568 // The semantics of LShiftINode include an implicit mask to 0x1F.
3569 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3570 elem_shift = layout_val;
3571 }
3572
3573 // Transition to native address size for all offset calculations:
3574 Node* lengthx = ConvI2X(length);
3575 Node* headerx = ConvI2X(header_size);
3576 #ifdef _LP64
3577 { const TypeInt* tilen = _gvn.find_int_type(length);
3578 if (tilen != NULL && tilen->_lo < 0) {
3579 // Add a manual constraint to a positive range. Cf. array_element_address.
3580 jlong size_max = fast_size_limit;
3581 if (size_max > tilen->_hi) size_max = tilen->_hi;
3582 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
3583
3584 // Only do a narrow I2L conversion if the range check passed.
3585 IfNode* iff = new (C) IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
3586 _gvn.transform(iff);
3587 RegionNode* region = new (C) RegionNode(3);
3588 _gvn.set_type(region, Type::CONTROL);
3589 lengthx = new (C) PhiNode(region, TypeLong::LONG);
3590 _gvn.set_type(lengthx, TypeLong::LONG);
3591
3592 // Range check passed. Use ConvI2L node with narrow type.
3593 Node* passed = IfFalse(iff);
3594 region->init_req(1, passed);
3595 // Make I2L conversion control dependent to prevent it from
3596 // floating above the range check during loop optimizations.
3597 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
3598
3599 // Range check failed. Use ConvI2L with wide type because length may be invalid.
3600 region->init_req(2, IfTrue(iff));
3601 lengthx->init_req(2, ConvI2X(length));
3602
3603 set_control(region);
3604 record_for_igvn(region);
3605 record_for_igvn(lengthx);
3606 }
3607 }
3608 #endif
3609
3610 // Combine header size (plus rounding) and body size. Then round down.
3611 // This computation cannot overflow, because it is used only in two
3612 // places, one where the length is sharply limited, and the other
3613 // after a successful allocation.
3614 Node* abody = lengthx;
3615 if (elem_shift != NULL)
3616 abody = _gvn.transform( new(C) LShiftXNode(lengthx, elem_shift) );
3617 Node* size = _gvn.transform( new(C) AddXNode(headerx, abody) );
3618 if (round_mask != 0) {
3619 Node* mask = MakeConX(~round_mask);
3620 size = _gvn.transform( new(C) AndXNode(size, mask) );
3621 }
3622 // else if round_mask == 0, the size computation is self-rounding
3623
3624 if (return_size_val != NULL) {
3625 // This is the size
3626 (*return_size_val) = size;
3627 }
3628
3629 // Now generate allocation code
3630
3631 // The entire memory state is needed for slow path of the allocation
3632 // since GC and deoptimization can happened.
3633 Node *mem = reset_memory();
3634 set_all_memory(mem); // Create new memory state
3635
3636 if (initial_slow_test->is_Bool()) {
3637 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
3638 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
3639 }
3640
3641 // Create the AllocateArrayNode and its result projections
3642 AllocateArrayNode* alloc
3643 = new (C) AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
3644 control(), mem, i_o(),
3645 size, klass_node,
3646 initial_slow_test,
3647 length);
3648
3649 // Cast to correct type. Note that the klass_node may be constant or not,
3650 // and in the latter case the actual array type will be inexact also.
3651 // (This happens via a non-constant argument to inline_native_newArray.)
3652 // In any case, the value of klass_node provides the desired array type.
3653 const TypeInt* length_type = _gvn.find_int_type(length);
3654 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
3655 if (ary_type->isa_aryptr() && length_type != NULL) {
3656 // Try to get a better type than POS for the size
3657 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
3658 }
3659
3660 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
3661
3662 // Cast length on remaining path to be as narrow as possible
3663 if (map()->find_edge(length) >= 0) {
3664 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
3665 if (ccast != length) {
3666 _gvn.set_type_bottom(ccast);
3667 record_for_igvn(ccast);
3668 replace_in_map(length, ccast);
3669 }
3670 }
3671
3672 return javaoop;
3673 }
3674
3675 // The following "Ideal_foo" functions are placed here because they recognize
3676 // the graph shapes created by the functions immediately above.
3677
3678 //---------------------------Ideal_allocation----------------------------------
3679 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
3680 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
3681 if (ptr == NULL) { // reduce dumb test in callers
3682 return NULL;
3683 }
3684
3685 #if INCLUDE_ALL_GCS
3686 if (UseShenandoahGC) {
3687 ptr = ShenandoahBarrierSetC2::bsc2()->step_over_gc_barrier(ptr);
3688 }
3689 #endif
3690 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
3691 ptr = ptr->in(1);
3692 if (ptr == NULL) return NULL;
3693 }
3694 // Return NULL for allocations with several casts:
3695 // j.l.reflect.Array.newInstance(jobject, jint)
3696 // Object.clone()
3697 // to keep more precise type from last cast.
3698 if (ptr->is_Proj()) {
3699 Node* allo = ptr->in(0);
3700 if (allo != NULL && allo->is_Allocate()) {
3701 return allo->as_Allocate();
3702 }
3703 }
3704 // Report failure to match.
3705 return NULL;
3706 }
3707
3708 // Fancy version which also strips off an offset (and reports it to caller).
3709 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
3710 intptr_t& offset) {
3711 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
3712 if (base == NULL) return NULL;
3713 return Ideal_allocation(base, phase);
3714 }
3715
3716 // Trace Initialize <- Proj[Parm] <- Allocate
3717 AllocateNode* InitializeNode::allocation() {
3718 Node* rawoop = in(InitializeNode::RawAddress);
3719 if (rawoop->is_Proj()) {
3720 Node* alloc = rawoop->in(0);
3721 if (alloc->is_Allocate()) {
3722 return alloc->as_Allocate();
3723 }
3724 }
3725 return NULL;
3726 }
3727
3728 // Trace Allocate -> Proj[Parm] -> Initialize
3729 InitializeNode* AllocateNode::initialization() {
3730 ProjNode* rawoop = proj_out(AllocateNode::RawAddress);
3731 if (rawoop == NULL) return NULL;
3732 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
3733 Node* init = rawoop->fast_out(i);
3734 if (init->is_Initialize()) {
3735 assert(init->as_Initialize()->allocation() == this, "2-way link");
3736 return init->as_Initialize();
3737 }
3738 }
3739 return NULL;
3740 }
3741
3742 //----------------------------- loop predicates ---------------------------
3743
3744 //------------------------------add_predicate_impl----------------------------
3745 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
3746 // Too many traps seen?
3747 if (too_many_traps(reason)) {
3748 #ifdef ASSERT
3749 if (TraceLoopPredicate) {
3750 int tc = C->trap_count(reason);
3751 tty->print("too many traps=%s tcount=%d in ",
3752 Deoptimization::trap_reason_name(reason), tc);
3753 method()->print(); // which method has too many predicate traps
3754 tty->cr();
3755 }
3756 #endif
3757 // We cannot afford to take more traps here,
3758 // do not generate predicate.
3759 return;
3760 }
3761
3762 Node *cont = _gvn.intcon(1);
3763 Node* opq = _gvn.transform(new (C) Opaque1Node(C, cont));
3764 Node *bol = _gvn.transform(new (C) Conv2BNode(opq));
3765 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
3766 Node* iffalse = _gvn.transform(new (C) IfFalseNode(iff));
3767 C->add_predicate_opaq(opq);
3768 {
3769 PreserveJVMState pjvms(this);
3770 set_control(iffalse);
3771 inc_sp(nargs);
3772 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
3773 }
3774 Node* iftrue = _gvn.transform(new (C) IfTrueNode(iff));
3775 set_control(iftrue);
3776 }
3777
3778 //------------------------------add_predicate---------------------------------
3779 void GraphKit::add_predicate(int nargs) {
3780 if (UseLoopPredicate) {
3781 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
3782 }
3783 // loop's limit check predicate should be near the loop.
3784 if (LoopLimitCheck) {
3785 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
3786 }
3787 }
3788
3789 //----------------------------- store barriers ----------------------------
3790 #define __ ideal.
3791
3792 void GraphKit::sync_kit(IdealKit& ideal) {
3793 set_all_memory(__ merged_memory());
3794 set_i_o(__ i_o());
3795 set_control(__ ctrl());
3796 }
3797
3798 void GraphKit::final_sync(IdealKit& ideal) {
3799 // Final sync IdealKit and graphKit.
3800 sync_kit(ideal);
3801 }
3802
3803 // vanilla/CMS post barrier
3804 // Insert a write-barrier store. This is to let generational GC work; we have
3805 // to flag all oop-stores before the next GC point.
3806 void GraphKit::write_barrier_post(Node* oop_store,
3807 Node* obj,
3808 Node* adr,
3809 uint adr_idx,
3810 Node* val,
3811 bool use_precise) {
3812 // No store check needed if we're storing a NULL or an old object
3813 // (latter case is probably a string constant). The concurrent
3814 // mark sweep garbage collector, however, needs to have all nonNull
3815 // oop updates flagged via card-marks.
3816 if (val != NULL && val->is_Con()) {
3817 // must be either an oop or NULL
3818 const Type* t = val->bottom_type();
3819 if (t == TypePtr::NULL_PTR || t == Type::TOP)
3820 // stores of null never (?) need barriers
3821 return;
3822 }
3823
3824 if (use_ReduceInitialCardMarks()
3825 && obj == just_allocated_object(control())) {
3826 // We can skip marks on a freshly-allocated object in Eden.
3827 // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
3828 // That routine informs GC to take appropriate compensating steps,
3829 // upon a slow-path allocation, so as to make this card-mark
3830 // elision safe.
3831 return;
3832 }
3833
3834 if (!use_precise) {
3835 // All card marks for a (non-array) instance are in one place:
3836 adr = obj;
3837 }
3838 // (Else it's an array (or unknown), and we want more precise card marks.)
3839 assert(adr != NULL, "");
3840
3841 IdealKit ideal(this, true);
3842
3843 // Convert the pointer to an int prior to doing math on it
3844 Node* cast = __ CastPX(__ ctrl(), adr);
3845
3846 // Divide by card size
3847 assert(Universe::heap()->barrier_set()->kind() == BarrierSet::CardTableModRef,
3848 "Only one we handle so far.");
3849 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
3850
3851 // Combine card table base and card offset
3852 Node* card_adr = __ AddP(__ top(), byte_map_base_node(), card_offset );
3853
3854 // Get the alias_index for raw card-mark memory
3855 int adr_type = Compile::AliasIdxRaw;
3856 Node* zero = __ ConI(0); // Dirty card value
3857 BasicType bt = T_BYTE;
3858
3859 if (UseCondCardMark) {
3860 // The classic GC reference write barrier is typically implemented
3861 // as a store into the global card mark table. Unfortunately
3862 // unconditional stores can result in false sharing and excessive
3863 // coherence traffic as well as false transactional aborts.
3864 // UseCondCardMark enables MP "polite" conditional card mark
3865 // stores. In theory we could relax the load from ctrl() to
3866 // no_ctrl, but that doesn't buy much latitude.
3867 Node* card_val = __ load( __ ctrl(), card_adr, TypeInt::BYTE, bt, adr_type);
3868 __ if_then(card_val, BoolTest::ne, zero);
3869 }
3870
3871 // Smash zero into card
3872 if( !UseConcMarkSweepGC ) {
3873 __ store(__ ctrl(), card_adr, zero, bt, adr_type, MemNode::unordered);
3874 } else {
3875 // Specialized path for CM store barrier
3876 __ storeCM(__ ctrl(), card_adr, zero, oop_store, adr_idx, bt, adr_type);
3877 }
3878
3879 if (UseCondCardMark) {
3880 __ end_if();
3881 }
3882
3883 // Final sync IdealKit and GraphKit.
3884 final_sync(ideal);
3885 }
3886
3887 static void g1_write_barrier_pre_helper(const GraphKit& kit, Node* adr) {
3888 if (UseShenandoahGC && adr != NULL) {
3889 Node* c = kit.control();
3890 Node* call = c->in(1)->in(1)->in(1)->in(0);
3891 assert(call->is_g1_wb_pre_call(), "g1_wb_pre call expected");
3892 call->add_req(adr);
3893 }
3894 }
3895
3896 // G1 pre/post barriers
3897 void GraphKit::g1_write_barrier_pre(bool do_load,
3898 Node* obj,
3899 Node* adr,
3900 uint alias_idx,
3901 Node* val,
3902 const TypeOopPtr* val_type,
3903 Node* pre_val,
3904 BasicType bt) {
3905
3906 // Some sanity checks
3907 // Note: val is unused in this routine.
3908
3909 if (do_load) {
3910 // We need to generate the load of the previous value
3911 assert(obj != NULL, "must have a base");
3912 assert(adr != NULL, "where are loading from?");
3913 assert(pre_val == NULL, "loaded already?");
3914 assert(val_type != NULL, "need a type");
3915 } else {
3916 // In this case both val_type and alias_idx are unused.
3917 assert(pre_val != NULL, "must be loaded already");
3918 // Nothing to be done if pre_val is null.
3919 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
3920 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
3921 }
3922 assert(bt == T_OBJECT, "or we shouldn't be here");
3923
3924 IdealKit ideal(this, true);
3925
3926 Node* tls = __ thread(); // ThreadLocalStorage
3927
3928 Node* no_ctrl = NULL;
3929 Node* no_base = __ top();
3930 Node* zero = __ ConI(0);
3931 Node* zeroX = __ ConX(0);
3932
3933 float likely = PROB_LIKELY(0.999);
3934 float unlikely = PROB_UNLIKELY(0.999);
3935
3936 BasicType active_type = in_bytes(PtrQueue::byte_width_of_active()) == 4 ? T_INT : T_BYTE;
3937 assert(in_bytes(PtrQueue::byte_width_of_active()) == 4 || in_bytes(PtrQueue::byte_width_of_active()) == 1, "flag width");
3938
3939 // Offsets into the thread
3940 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 648
3941 PtrQueue::byte_offset_of_active());
3942 const int index_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 656
3943 PtrQueue::byte_offset_of_index());
3944 const int buffer_offset = in_bytes(JavaThread::satb_mark_queue_offset() + // 652
3945 PtrQueue::byte_offset_of_buf());
3946
3947 // Now the actual pointers into the thread
3948 Node* marking_adr = __ AddP(no_base, tls, __ ConX(marking_offset));
3949 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
3950 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
3951
3952 // Now some of the values
3953 Node* marking;
3954 if (UseShenandoahGC) {
3955 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(JavaThread::gc_state_offset())));
3956 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
3957 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
3958 assert(ShenandoahBarrierC2Support::is_gc_state_load(ld), "Should match the shape");
3959 } else {
3960 assert(UseG1GC, "should be");
3961 marking = __ load(__ ctrl(), marking_adr, TypeInt::INT, active_type, Compile::AliasIdxRaw);
3962 }
3963
3964 // if (!marking)
3965 __ if_then(marking, BoolTest::ne, zero, unlikely); {
3966 BasicType index_bt = TypeX_X->basic_type();
3967 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 PtrQueue::_index with wrong size.");
3968 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
3969
3970 if (do_load) {
3971 // load original value
3972 // alias_idx correct??
3973 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
3974 }
3975
3976 // if (pre_val != NULL)
3977 __ if_then(pre_val, BoolTest::ne, null()); {
3978 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
3979
3980 // is the queue for this thread full?
3981 __ if_then(index, BoolTest::ne, zeroX, likely); {
3982
3983 // decrement the index
3984 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
3985
3986 // Now get the buffer location we will log the previous value into and store it
3987 Node *log_addr = __ AddP(no_base, buffer, next_index);
3988 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
3989 // update the index
3990 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
3991
3992 } __ else_(); {
3993
3994 // logging buffer is full, call the runtime
3995 const TypeFunc *tf = OptoRuntime::g1_wb_pre_Type();
3996 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), "g1_wb_pre", pre_val, tls);
3997 } __ end_if(); // (!index)
3998 } __ end_if(); // (pre_val != NULL)
3999 } __ end_if(); // (!marking)
4000
4001 // Final sync IdealKit and GraphKit.
4002 final_sync(ideal);
4003 g1_write_barrier_pre_helper(*this, adr);
4004 }
4005
4006 //
4007 // Update the card table and add card address to the queue
4008 //
4009 void GraphKit::g1_mark_card(IdealKit& ideal,
4010 Node* card_adr,
4011 Node* oop_store,
4012 uint oop_alias_idx,
4013 Node* index,
4014 Node* index_adr,
4015 Node* buffer,
4016 const TypeFunc* tf) {
4017
4018 Node* zero = __ ConI(0);
4019 Node* zeroX = __ ConX(0);
4020 Node* no_base = __ top();
4021 BasicType card_bt = T_BYTE;
4022 // Smash zero into card. MUST BE ORDERED WRT TO STORE
4023 __ storeCM(__ ctrl(), card_adr, zero, oop_store, oop_alias_idx, card_bt, Compile::AliasIdxRaw);
4024
4025 // Now do the queue work
4026 __ if_then(index, BoolTest::ne, zeroX); {
4027
4028 Node* next_index = _gvn.transform(new (C) SubXNode(index, __ ConX(sizeof(intptr_t))));
4029 Node* log_addr = __ AddP(no_base, buffer, next_index);
4030
4031 // Order, see storeCM.
4032 __ store(__ ctrl(), log_addr, card_adr, T_ADDRESS, Compile::AliasIdxRaw, MemNode::unordered);
4033 __ store(__ ctrl(), index_adr, next_index, TypeX_X->basic_type(), Compile::AliasIdxRaw, MemNode::unordered);
4034
4035 } __ else_(); {
4036 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), "g1_wb_post", card_adr, __ thread());
4037 } __ end_if();
4038
4039 }
4040
4041 void GraphKit::g1_write_barrier_post(Node* oop_store,
4042 Node* obj,
4043 Node* adr,
4044 uint alias_idx,
4045 Node* val,
4046 BasicType bt,
4047 bool use_precise) {
4048 // If we are writing a NULL then we need no post barrier
4049
4050 if (val != NULL && val->is_Con() && val->bottom_type() == TypePtr::NULL_PTR) {
4051 // Must be NULL
4052 const Type* t = val->bottom_type();
4053 assert(t == Type::TOP || t == TypePtr::NULL_PTR, "must be NULL");
4054 // No post barrier if writing NULLx
4055 return;
4056 }
4057
4058 if (!use_precise) {
4059 // All card marks for a (non-array) instance are in one place:
4060 adr = obj;
4061 }
4062 // (Else it's an array (or unknown), and we want more precise card marks.)
4063 assert(adr != NULL, "");
4064
4065 IdealKit ideal(this, true);
4066
4067 Node* tls = __ thread(); // ThreadLocalStorage
4068
4069 Node* no_base = __ top();
4070 float likely = PROB_LIKELY(0.999);
4071 float unlikely = PROB_UNLIKELY(0.999);
4072 Node* young_card = __ ConI((jint)G1SATBCardTableModRefBS::g1_young_card_val());
4073 Node* dirty_card = __ ConI((jint)CardTableModRefBS::dirty_card_val());
4074 Node* zeroX = __ ConX(0);
4075
4076 // Get the alias_index for raw card-mark memory
4077 const TypePtr* card_type = TypeRawPtr::BOTTOM;
4078
4079 const TypeFunc *tf = OptoRuntime::g1_wb_post_Type();
4080
4081 // Offsets into the thread
4082 const int index_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4083 PtrQueue::byte_offset_of_index());
4084 const int buffer_offset = in_bytes(JavaThread::dirty_card_queue_offset() +
4085 PtrQueue::byte_offset_of_buf());
4086
4087 // Pointers into the thread
4088
4089 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
4090 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
4091
4092 // Now some values
4093 // Use ctrl to avoid hoisting these values past a safepoint, which could
4094 // potentially reset these fields in the JavaThread.
4095 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, TypeX_X->basic_type(), Compile::AliasIdxRaw);
4096 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
4097
4098 // Convert the store obj pointer to an int prior to doing math on it
4099 // Must use ctrl to prevent "integerized oop" existing across safepoint
4100 Node* cast = __ CastPX(__ ctrl(), adr);
4101
4102 // Divide pointer by card size
4103 Node* card_offset = __ URShiftX( cast, __ ConI(CardTableModRefBS::card_shift) );
4104
4105 // Combine card table base and card offset
4106 Node* card_adr = __ AddP(no_base, byte_map_base_node(), card_offset );
4107
4108 // If we know the value being stored does it cross regions?
4109
4110 if (val != NULL) {
4111 // Does the store cause us to cross regions?
4112
4113 // Should be able to do an unsigned compare of region_size instead of
4114 // and extra shift. Do we have an unsigned compare??
4115 // Node* region_size = __ ConI(1 << HeapRegion::LogOfHRGrainBytes);
4116 Node* xor_res = __ URShiftX ( __ XorX( cast, __ CastPX(__ ctrl(), val)), __ ConI(HeapRegion::LogOfHRGrainBytes));
4117
4118 // if (xor_res == 0) same region so skip
4119 __ if_then(xor_res, BoolTest::ne, zeroX); {
4120
4121 // No barrier if we are storing a NULL
4122 __ if_then(val, BoolTest::ne, null(), unlikely); {
4123
4124 // Ok must mark the card if not already dirty
4125
4126 // load the original value of the card
4127 Node* card_val = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4128
4129 __ if_then(card_val, BoolTest::ne, young_card); {
4130 sync_kit(ideal);
4131 // Use Op_MemBarVolatile to achieve the effect of a StoreLoad barrier.
4132 insert_mem_bar(Op_MemBarVolatile, oop_store);
4133 __ sync_kit(this);
4134
4135 Node* card_val_reload = __ load(__ ctrl(), card_adr, TypeInt::INT, T_BYTE, Compile::AliasIdxRaw);
4136 __ if_then(card_val_reload, BoolTest::ne, dirty_card); {
4137 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4138 } __ end_if();
4139 } __ end_if();
4140 } __ end_if();
4141 } __ end_if();
4142 } else {
4143 // Object.clone() instrinsic uses this path.
4144 g1_mark_card(ideal, card_adr, oop_store, alias_idx, index, index_adr, buffer, tf);
4145 }
4146
4147 // Final sync IdealKit and GraphKit.
4148 final_sync(ideal);
4149 }
4150 #undef __
4151
4152
4153
4154 Node* GraphKit::load_String_offset(Node* ctrl, Node* str) {
4155 if (java_lang_String::has_offset_field()) {
4156 int offset_offset = java_lang_String::offset_offset_in_bytes();
4157 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4158 false, NULL, 0);
4159 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4160 int offset_field_idx = C->get_alias_index(offset_field_type);
4161 return make_load(ctrl,
4162 basic_plus_adr(str, str, offset_offset),
4163 TypeInt::INT, T_INT, offset_field_idx, MemNode::unordered);
4164 } else {
4165 return intcon(0);
4166 }
4167 }
4168
4169 Node* GraphKit::load_String_length(Node* ctrl, Node* str) {
4170 if (java_lang_String::has_count_field()) {
4171 int count_offset = java_lang_String::count_offset_in_bytes();
4172 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4173 false, NULL, 0);
4174 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4175 int count_field_idx = C->get_alias_index(count_field_type);
4176 return make_load(ctrl,
4177 basic_plus_adr(str, str, count_offset),
4178 TypeInt::INT, T_INT, count_field_idx, MemNode::unordered);
4179 } else {
4180 return load_array_length(load_String_value(ctrl, str));
4181 }
4182 }
4183
4184 Node* GraphKit::load_String_value(Node* ctrl, Node* str) {
4185 int value_offset = java_lang_String::value_offset_in_bytes();
4186 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4187 false, NULL, 0);
4188 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4189 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4190 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
4191 ciTypeArrayKlass::make(T_CHAR), true, 0);
4192 int value_field_idx = C->get_alias_index(value_field_type);
4193 Node* load = make_load(ctrl, basic_plus_adr(str, str, value_offset),
4194 value_type, T_OBJECT, value_field_idx, MemNode::unordered);
4195 #if INCLUDE_ALL_GCS
4196 if (UseShenandoahGC) {
4197 load = ShenandoahBarrierSetC2::bsc2()->load_reference_barrier(this, load);
4198 }
4199 #endif
4200 // String.value field is known to be @Stable.
4201 if (UseImplicitStableValues) {
4202 load = cast_array_to_stable(load, value_type);
4203 }
4204 return load;
4205 }
4206
4207 void GraphKit::store_String_offset(Node* ctrl, Node* str, Node* value) {
4208 int offset_offset = java_lang_String::offset_offset_in_bytes();
4209 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4210 false, NULL, 0);
4211 const TypePtr* offset_field_type = string_type->add_offset(offset_offset);
4212 int offset_field_idx = C->get_alias_index(offset_field_type);
4213 store_to_memory(ctrl, basic_plus_adr(str, offset_offset),
4214 value, T_INT, offset_field_idx, MemNode::unordered);
4215 }
4216
4217 void GraphKit::store_String_value(Node* ctrl, Node* str, Node* value) {
4218 int value_offset = java_lang_String::value_offset_in_bytes();
4219 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4220 false, NULL, 0);
4221 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4222
4223 store_oop_to_object(ctrl, str, basic_plus_adr(str, value_offset), value_field_type,
4224 value, TypeAryPtr::CHARS, T_OBJECT, MemNode::unordered);
4225 }
4226
4227 void GraphKit::store_String_length(Node* ctrl, Node* str, Node* value) {
4228 int count_offset = java_lang_String::count_offset_in_bytes();
4229 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4230 false, NULL, 0);
4231 const TypePtr* count_field_type = string_type->add_offset(count_offset);
4232 int count_field_idx = C->get_alias_index(count_field_type);
4233 store_to_memory(ctrl, basic_plus_adr(str, count_offset),
4234 value, T_INT, count_field_idx, MemNode::unordered);
4235 }
4236
4237 Node* GraphKit::cast_array_to_stable(Node* ary, const TypeAryPtr* ary_type) {
4238 // Reify the property as a CastPP node in Ideal graph to comply with monotonicity
4239 // assumption of CCP analysis.
4240 return _gvn.transform(new(C) CastPPNode(ary, ary_type->cast_to_stable(true)));
4241 }