1 /*
2 * Copyright (c) 2001, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciUtilities.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "ci/ciValueKlass.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/c2/barrierSetC2.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/idealKit.hpp"
38 #include "opto/intrinsicnode.hpp"
39 #include "opto/locknode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/valuetypenode.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49
50 //----------------------------GraphKit-----------------------------------------
51 // Main utility constructor.
52 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
53 : Phase(Phase::Parser),
54 _env(C->env()),
55 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
56 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
57 {
58 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
59 _exceptions = jvms->map()->next_exception();
60 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
61 set_jvms(jvms);
62 #ifdef ASSERT
63 if (_gvn.is_IterGVN() != NULL) {
64 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
65 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
66 _worklist_size = _gvn.C->for_igvn()->size();
67 }
68 #endif
69 }
70
71 // Private constructor for parser.
72 GraphKit::GraphKit()
73 : Phase(Phase::Parser),
74 _env(C->env()),
75 _gvn(*C->initial_gvn()),
76 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
77 {
78 _exceptions = NULL;
79 set_map(NULL);
80 debug_only(_sp = -99);
81 debug_only(set_bci(-99));
82 }
83
84
85
86 //---------------------------clean_stack---------------------------------------
87 // Clear away rubbish from the stack area of the JVM state.
88 // This destroys any arguments that may be waiting on the stack.
89 void GraphKit::clean_stack(int from_sp) {
90 SafePointNode* map = this->map();
91 JVMState* jvms = this->jvms();
92 int stk_size = jvms->stk_size();
93 int stkoff = jvms->stkoff();
94 Node* top = this->top();
95 for (int i = from_sp; i < stk_size; i++) {
96 if (map->in(stkoff + i) != top) {
97 map->set_req(stkoff + i, top);
98 }
99 }
100 }
101
102
103 //--------------------------------sync_jvms-----------------------------------
104 // Make sure our current jvms agrees with our parse state.
105 JVMState* GraphKit::sync_jvms() const {
106 JVMState* jvms = this->jvms();
107 jvms->set_bci(bci()); // Record the new bci in the JVMState
108 jvms->set_sp(sp()); // Record the new sp in the JVMState
109 assert(jvms_in_sync(), "jvms is now in sync");
110 return jvms;
111 }
112
113 //--------------------------------sync_jvms_for_reexecute---------------------
114 // Make sure our current jvms agrees with our parse state. This version
115 // uses the reexecute_sp for reexecuting bytecodes.
116 JVMState* GraphKit::sync_jvms_for_reexecute() {
117 JVMState* jvms = this->jvms();
118 jvms->set_bci(bci()); // Record the new bci in the JVMState
119 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
120 return jvms;
121 }
122
123 #ifdef ASSERT
124 bool GraphKit::jvms_in_sync() const {
125 Parse* parse = is_Parse();
126 if (parse == NULL) {
127 if (bci() != jvms()->bci()) return false;
128 if (sp() != (int)jvms()->sp()) return false;
129 return true;
130 }
131 if (jvms()->method() != parse->method()) return false;
132 if (jvms()->bci() != parse->bci()) return false;
133 int jvms_sp = jvms()->sp();
134 if (jvms_sp != parse->sp()) return false;
135 int jvms_depth = jvms()->depth();
136 if (jvms_depth != parse->depth()) return false;
137 return true;
138 }
139
140 // Local helper checks for special internal merge points
141 // used to accumulate and merge exception states.
142 // They are marked by the region's in(0) edge being the map itself.
143 // Such merge points must never "escape" into the parser at large,
144 // until they have been handed to gvn.transform.
145 static bool is_hidden_merge(Node* reg) {
146 if (reg == NULL) return false;
147 if (reg->is_Phi()) {
148 reg = reg->in(0);
149 if (reg == NULL) return false;
150 }
151 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
152 }
153
154 void GraphKit::verify_map() const {
155 if (map() == NULL) return; // null map is OK
156 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
157 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
158 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
159 }
160
161 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
162 assert(ex_map->next_exception() == NULL, "not already part of a chain");
163 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
164 }
165 #endif
166
167 //---------------------------stop_and_kill_map---------------------------------
168 // Set _map to NULL, signalling a stop to further bytecode execution.
169 // First smash the current map's control to a constant, to mark it dead.
170 void GraphKit::stop_and_kill_map() {
171 SafePointNode* dead_map = stop();
172 if (dead_map != NULL) {
173 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
174 assert(dead_map->is_killed(), "must be so marked");
175 }
176 }
177
178
179 //--------------------------------stopped--------------------------------------
180 // Tell if _map is NULL, or control is top.
181 bool GraphKit::stopped() {
182 if (map() == NULL) return true;
183 else if (control() == top()) return true;
184 else return false;
185 }
186
187
188 //-----------------------------has_ex_handler----------------------------------
189 // Tell if this method or any caller method has exception handlers.
190 bool GraphKit::has_ex_handler() {
191 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
192 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
193 return true;
194 }
195 }
196 return false;
197 }
198
199 //------------------------------save_ex_oop------------------------------------
200 // Save an exception without blowing stack contents or other JVM state.
201 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
202 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
203 ex_map->add_req(ex_oop);
204 debug_only(verify_exception_state(ex_map));
205 }
206
207 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
208 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
209 Node* ex_oop = ex_map->in(ex_map->req()-1);
210 if (clear_it) ex_map->del_req(ex_map->req()-1);
211 return ex_oop;
212 }
213
214 //-----------------------------saved_ex_oop------------------------------------
215 // Recover a saved exception from its map.
216 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
217 return common_saved_ex_oop(ex_map, false);
218 }
219
220 //--------------------------clear_saved_ex_oop---------------------------------
221 // Erase a previously saved exception from its map.
222 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
223 return common_saved_ex_oop(ex_map, true);
224 }
225
226 #ifdef ASSERT
227 //---------------------------has_saved_ex_oop----------------------------------
228 // Erase a previously saved exception from its map.
229 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
230 return ex_map->req() == ex_map->jvms()->endoff()+1;
231 }
232 #endif
233
234 //-------------------------make_exception_state--------------------------------
235 // Turn the current JVM state into an exception state, appending the ex_oop.
236 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
237 sync_jvms();
238 SafePointNode* ex_map = stop(); // do not manipulate this map any more
239 set_saved_ex_oop(ex_map, ex_oop);
240 return ex_map;
241 }
242
243
244 //--------------------------add_exception_state--------------------------------
245 // Add an exception to my list of exceptions.
246 void GraphKit::add_exception_state(SafePointNode* ex_map) {
247 if (ex_map == NULL || ex_map->control() == top()) {
248 return;
249 }
250 #ifdef ASSERT
251 verify_exception_state(ex_map);
252 if (has_exceptions()) {
253 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
254 }
255 #endif
256
257 // If there is already an exception of exactly this type, merge with it.
258 // In particular, null-checks and other low-level exceptions common up here.
259 Node* ex_oop = saved_ex_oop(ex_map);
260 const Type* ex_type = _gvn.type(ex_oop);
261 if (ex_oop == top()) {
262 // No action needed.
263 return;
264 }
265 assert(ex_type->isa_instptr(), "exception must be an instance");
266 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
267 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
268 // We check sp also because call bytecodes can generate exceptions
269 // both before and after arguments are popped!
270 if (ex_type2 == ex_type
271 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
272 combine_exception_states(ex_map, e2);
273 return;
274 }
275 }
276
277 // No pre-existing exception of the same type. Chain it on the list.
278 push_exception_state(ex_map);
279 }
280
281 //-----------------------add_exception_states_from-----------------------------
282 void GraphKit::add_exception_states_from(JVMState* jvms) {
283 SafePointNode* ex_map = jvms->map()->next_exception();
284 if (ex_map != NULL) {
285 jvms->map()->set_next_exception(NULL);
286 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
287 next_map = ex_map->next_exception();
288 ex_map->set_next_exception(NULL);
289 add_exception_state(ex_map);
290 }
291 }
292 }
293
294 //-----------------------transfer_exceptions_into_jvms-------------------------
295 JVMState* GraphKit::transfer_exceptions_into_jvms() {
296 if (map() == NULL) {
297 // We need a JVMS to carry the exceptions, but the map has gone away.
298 // Create a scratch JVMS, cloned from any of the exception states...
299 if (has_exceptions()) {
300 _map = _exceptions;
301 _map = clone_map();
302 _map->set_next_exception(NULL);
303 clear_saved_ex_oop(_map);
304 debug_only(verify_map());
305 } else {
306 // ...or created from scratch
307 JVMState* jvms = new (C) JVMState(_method, NULL);
308 jvms->set_bci(_bci);
309 jvms->set_sp(_sp);
310 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
311 set_jvms(jvms);
312 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
313 set_all_memory(top());
314 while (map()->req() < jvms->endoff()) map()->add_req(top());
315 }
316 // (This is a kludge, in case you didn't notice.)
317 set_control(top());
318 }
319 JVMState* jvms = sync_jvms();
320 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
321 jvms->map()->set_next_exception(_exceptions);
322 _exceptions = NULL; // done with this set of exceptions
323 return jvms;
324 }
325
326 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
327 assert(is_hidden_merge(dstphi), "must be a special merge node");
328 assert(is_hidden_merge(srcphi), "must be a special merge node");
329 uint limit = srcphi->req();
330 for (uint i = PhiNode::Input; i < limit; i++) {
331 dstphi->add_req(srcphi->in(i));
332 }
333 }
334 static inline void add_one_req(Node* dstphi, Node* src) {
335 assert(is_hidden_merge(dstphi), "must be a special merge node");
336 assert(!is_hidden_merge(src), "must not be a special merge node");
337 dstphi->add_req(src);
338 }
339
340 //-----------------------combine_exception_states------------------------------
341 // This helper function combines exception states by building phis on a
342 // specially marked state-merging region. These regions and phis are
343 // untransformed, and can build up gradually. The region is marked by
344 // having a control input of its exception map, rather than NULL. Such
345 // regions do not appear except in this function, and in use_exception_state.
346 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
347 if (failing()) return; // dying anyway...
348 JVMState* ex_jvms = ex_map->_jvms;
349 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
350 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
351 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
352 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
353 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
354 assert(ex_map->req() == phi_map->req(), "matching maps");
355 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
356 Node* hidden_merge_mark = root();
357 Node* region = phi_map->control();
358 MergeMemNode* phi_mem = phi_map->merged_memory();
359 MergeMemNode* ex_mem = ex_map->merged_memory();
360 if (region->in(0) != hidden_merge_mark) {
361 // The control input is not (yet) a specially-marked region in phi_map.
362 // Make it so, and build some phis.
363 region = new RegionNode(2);
364 _gvn.set_type(region, Type::CONTROL);
365 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
366 region->init_req(1, phi_map->control());
367 phi_map->set_control(region);
368 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
369 record_for_igvn(io_phi);
370 _gvn.set_type(io_phi, Type::ABIO);
371 phi_map->set_i_o(io_phi);
372 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
373 Node* m = mms.memory();
374 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
375 record_for_igvn(m_phi);
376 _gvn.set_type(m_phi, Type::MEMORY);
377 mms.set_memory(m_phi);
378 }
379 }
380
381 // Either or both of phi_map and ex_map might already be converted into phis.
382 Node* ex_control = ex_map->control();
383 // if there is special marking on ex_map also, we add multiple edges from src
384 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
385 // how wide was the destination phi_map, originally?
386 uint orig_width = region->req();
387
388 if (add_multiple) {
389 add_n_reqs(region, ex_control);
390 add_n_reqs(phi_map->i_o(), ex_map->i_o());
391 } else {
392 // ex_map has no merges, so we just add single edges everywhere
393 add_one_req(region, ex_control);
394 add_one_req(phi_map->i_o(), ex_map->i_o());
395 }
396 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
397 if (mms.is_empty()) {
398 // get a copy of the base memory, and patch some inputs into it
399 const TypePtr* adr_type = mms.adr_type(C);
400 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
401 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
402 mms.set_memory(phi);
403 // Prepare to append interesting stuff onto the newly sliced phi:
404 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
405 }
406 // Append stuff from ex_map:
407 if (add_multiple) {
408 add_n_reqs(mms.memory(), mms.memory2());
409 } else {
410 add_one_req(mms.memory(), mms.memory2());
411 }
412 }
413 uint limit = ex_map->req();
414 for (uint i = TypeFunc::Parms; i < limit; i++) {
415 // Skip everything in the JVMS after tos. (The ex_oop follows.)
416 if (i == tos) i = ex_jvms->monoff();
417 Node* src = ex_map->in(i);
418 Node* dst = phi_map->in(i);
419 if (src != dst) {
420 PhiNode* phi;
421 if (dst->in(0) != region) {
422 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
423 record_for_igvn(phi);
424 _gvn.set_type(phi, phi->type());
425 phi_map->set_req(i, dst);
426 // Prepare to append interesting stuff onto the new phi:
427 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
428 } else {
429 assert(dst->is_Phi(), "nobody else uses a hidden region");
430 phi = dst->as_Phi();
431 }
432 if (add_multiple && src->in(0) == ex_control) {
433 // Both are phis.
434 add_n_reqs(dst, src);
435 } else {
436 while (dst->req() < region->req()) add_one_req(dst, src);
437 }
438 const Type* srctype = _gvn.type(src);
439 if (phi->type() != srctype) {
440 const Type* dsttype = phi->type()->meet_speculative(srctype);
441 if (phi->type() != dsttype) {
442 phi->set_type(dsttype);
443 _gvn.set_type(phi, dsttype);
444 }
445 }
446 }
447 }
448 phi_map->merge_replaced_nodes_with(ex_map);
449 }
450
451 //--------------------------use_exception_state--------------------------------
452 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
453 if (failing()) { stop(); return top(); }
454 Node* region = phi_map->control();
455 Node* hidden_merge_mark = root();
456 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
457 Node* ex_oop = clear_saved_ex_oop(phi_map);
458 if (region->in(0) == hidden_merge_mark) {
459 // Special marking for internal ex-states. Process the phis now.
460 region->set_req(0, region); // now it's an ordinary region
461 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
462 // Note: Setting the jvms also sets the bci and sp.
463 set_control(_gvn.transform(region));
464 uint tos = jvms()->stkoff() + sp();
465 for (uint i = 1; i < tos; i++) {
466 Node* x = phi_map->in(i);
467 if (x->in(0) == region) {
468 assert(x->is_Phi(), "expected a special phi");
469 phi_map->set_req(i, _gvn.transform(x));
470 }
471 }
472 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
473 Node* x = mms.memory();
474 if (x->in(0) == region) {
475 assert(x->is_Phi(), "nobody else uses a hidden region");
476 mms.set_memory(_gvn.transform(x));
477 }
478 }
479 if (ex_oop->in(0) == region) {
480 assert(ex_oop->is_Phi(), "expected a special phi");
481 ex_oop = _gvn.transform(ex_oop);
482 }
483 } else {
484 set_jvms(phi_map->jvms());
485 }
486
487 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
488 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
489 return ex_oop;
490 }
491
492 //---------------------------------java_bc-------------------------------------
493 Bytecodes::Code GraphKit::java_bc() const {
494 ciMethod* method = this->method();
495 int bci = this->bci();
496 if (method != NULL && bci != InvocationEntryBci)
497 return method->java_code_at_bci(bci);
498 else
499 return Bytecodes::_illegal;
500 }
501
502 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
503 bool must_throw) {
504 // if the exception capability is set, then we will generate code
505 // to check the JavaThread.should_post_on_exceptions flag to see
506 // if we actually need to report exception events (for this
507 // thread). If we don't need to report exception events, we will
508 // take the normal fast path provided by add_exception_events. If
509 // exception event reporting is enabled for this thread, we will
510 // take the uncommon_trap in the BuildCutout below.
511
512 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
513 Node* jthread = _gvn.transform(new ThreadLocalNode());
514 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
515 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
516
517 // Test the should_post_on_exceptions_flag vs. 0
518 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
519 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
520
521 // Branch to slow_path if should_post_on_exceptions_flag was true
522 { BuildCutout unless(this, tst, PROB_MAX);
523 // Do not try anything fancy if we're notifying the VM on every throw.
524 // Cf. case Bytecodes::_athrow in parse2.cpp.
525 uncommon_trap(reason, Deoptimization::Action_none,
526 (ciKlass*)NULL, (char*)NULL, must_throw);
527 }
528
529 }
530
531 //------------------------------builtin_throw----------------------------------
532 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
533 bool must_throw = true;
534
535 if (env()->jvmti_can_post_on_exceptions()) {
536 // check if we must post exception events, take uncommon trap if so
537 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
538 // here if should_post_on_exceptions is false
539 // continue on with the normal codegen
540 }
541
542 // If this particular condition has not yet happened at this
543 // bytecode, then use the uncommon trap mechanism, and allow for
544 // a future recompilation if several traps occur here.
545 // If the throw is hot, try to use a more complicated inline mechanism
546 // which keeps execution inside the compiled code.
547 bool treat_throw_as_hot = false;
548 ciMethodData* md = method()->method_data();
549
550 if (ProfileTraps) {
551 if (too_many_traps(reason)) {
552 treat_throw_as_hot = true;
553 }
554 // (If there is no MDO at all, assume it is early in
555 // execution, and that any deopts are part of the
556 // startup transient, and don't need to be remembered.)
557
558 // Also, if there is a local exception handler, treat all throws
559 // as hot if there has been at least one in this method.
560 if (C->trap_count(reason) != 0
561 && method()->method_data()->trap_count(reason) != 0
562 && has_ex_handler()) {
563 treat_throw_as_hot = true;
564 }
565 }
566
567 // If this throw happens frequently, an uncommon trap might cause
568 // a performance pothole. If there is a local exception handler,
569 // and if this particular bytecode appears to be deoptimizing often,
570 // let us handle the throw inline, with a preconstructed instance.
571 // Note: If the deopt count has blown up, the uncommon trap
572 // runtime is going to flush this nmethod, not matter what.
573 if (treat_throw_as_hot
574 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
575 // If the throw is local, we use a pre-existing instance and
576 // punt on the backtrace. This would lead to a missing backtrace
577 // (a repeat of 4292742) if the backtrace object is ever asked
578 // for its backtrace.
579 // Fixing this remaining case of 4292742 requires some flavor of
580 // escape analysis. Leave that for the future.
581 ciInstance* ex_obj = NULL;
582 switch (reason) {
583 case Deoptimization::Reason_null_check:
584 ex_obj = env()->NullPointerException_instance();
585 break;
586 case Deoptimization::Reason_div0_check:
587 ex_obj = env()->ArithmeticException_instance();
588 break;
589 case Deoptimization::Reason_range_check:
590 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
591 break;
592 case Deoptimization::Reason_class_check:
593 if (java_bc() == Bytecodes::_aastore) {
594 ex_obj = env()->ArrayStoreException_instance();
595 } else {
596 ex_obj = env()->ClassCastException_instance();
597 }
598 break;
599 default:
600 break;
601 }
602 if (failing()) { stop(); return; } // exception allocation might fail
603 if (ex_obj != NULL) {
604 // Cheat with a preallocated exception object.
605 if (C->log() != NULL)
606 C->log()->elem("hot_throw preallocated='1' reason='%s'",
607 Deoptimization::trap_reason_name(reason));
608 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
609 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
610
611 // Clear the detail message of the preallocated exception object.
612 // Weblogic sometimes mutates the detail message of exceptions
613 // using reflection.
614 int offset = java_lang_Throwable::get_detailMessage_offset();
615 const TypePtr* adr_typ = ex_con->add_offset(offset);
616
617 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
618 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
619 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
620
621 add_exception_state(make_exception_state(ex_node));
622 return;
623 }
624 }
625
626 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
627 // It won't be much cheaper than bailing to the interp., since we'll
628 // have to pass up all the debug-info, and the runtime will have to
629 // create the stack trace.
630
631 // Usual case: Bail to interpreter.
632 // Reserve the right to recompile if we haven't seen anything yet.
633
634 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
635 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
636 if (treat_throw_as_hot
637 && (method()->method_data()->trap_recompiled_at(bci(), m)
638 || C->too_many_traps(reason))) {
639 // We cannot afford to take more traps here. Suffer in the interpreter.
640 if (C->log() != NULL)
641 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
642 Deoptimization::trap_reason_name(reason),
643 C->trap_count(reason));
644 action = Deoptimization::Action_none;
645 }
646
647 // "must_throw" prunes the JVM state to include only the stack, if there
648 // are no local exception handlers. This should cut down on register
649 // allocation time and code size, by drastically reducing the number
650 // of in-edges on the call to the uncommon trap.
651
652 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
653 }
654
655
656 //----------------------------PreserveJVMState---------------------------------
657 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
658 debug_only(kit->verify_map());
659 _kit = kit;
660 _map = kit->map(); // preserve the map
661 _sp = kit->sp();
662 kit->set_map(clone_map ? kit->clone_map() : NULL);
663 #ifdef ASSERT
664 _bci = kit->bci();
665 Parse* parser = kit->is_Parse();
666 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
667 _block = block;
668 #endif
669 }
670 PreserveJVMState::~PreserveJVMState() {
671 GraphKit* kit = _kit;
672 #ifdef ASSERT
673 assert(kit->bci() == _bci, "bci must not shift");
674 Parse* parser = kit->is_Parse();
675 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
676 assert(block == _block, "block must not shift");
677 #endif
678 kit->set_map(_map);
679 kit->set_sp(_sp);
680 }
681
682
683 //-----------------------------BuildCutout-------------------------------------
684 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
685 : PreserveJVMState(kit)
686 {
687 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
688 SafePointNode* outer_map = _map; // preserved map is caller's
689 SafePointNode* inner_map = kit->map();
690 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
691 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
692 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
693 }
694 BuildCutout::~BuildCutout() {
695 GraphKit* kit = _kit;
696 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
697 }
698
699 //---------------------------PreserveReexecuteState----------------------------
700 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
701 assert(!kit->stopped(), "must call stopped() before");
702 _kit = kit;
703 _sp = kit->sp();
704 _reexecute = kit->jvms()->_reexecute;
705 }
706 PreserveReexecuteState::~PreserveReexecuteState() {
707 if (_kit->stopped()) return;
708 _kit->jvms()->_reexecute = _reexecute;
709 _kit->set_sp(_sp);
710 }
711
712 //------------------------------clone_map--------------------------------------
713 // Implementation of PreserveJVMState
714 //
715 // Only clone_map(...) here. If this function is only used in the
716 // PreserveJVMState class we may want to get rid of this extra
717 // function eventually and do it all there.
718
719 SafePointNode* GraphKit::clone_map() {
720 if (map() == NULL) return NULL;
721
722 // Clone the memory edge first
723 Node* mem = MergeMemNode::make(map()->memory());
724 gvn().set_type_bottom(mem);
725
726 SafePointNode *clonemap = (SafePointNode*)map()->clone();
727 JVMState* jvms = this->jvms();
728 JVMState* clonejvms = jvms->clone_shallow(C);
729 clonemap->set_memory(mem);
730 clonemap->set_jvms(clonejvms);
731 clonejvms->set_map(clonemap);
732 record_for_igvn(clonemap);
733 gvn().set_type_bottom(clonemap);
734 return clonemap;
735 }
736
737
738 //-----------------------------set_map_clone-----------------------------------
739 void GraphKit::set_map_clone(SafePointNode* m) {
740 _map = m;
741 _map = clone_map();
742 _map->set_next_exception(NULL);
743 debug_only(verify_map());
744 }
745
746
747 //----------------------------kill_dead_locals---------------------------------
748 // Detect any locals which are known to be dead, and force them to top.
749 void GraphKit::kill_dead_locals() {
750 // Consult the liveness information for the locals. If any
751 // of them are unused, then they can be replaced by top(). This
752 // should help register allocation time and cut down on the size
753 // of the deoptimization information.
754
755 // This call is made from many of the bytecode handling
756 // subroutines called from the Big Switch in do_one_bytecode.
757 // Every bytecode which might include a slow path is responsible
758 // for killing its dead locals. The more consistent we
759 // are about killing deads, the fewer useless phis will be
760 // constructed for them at various merge points.
761
762 // bci can be -1 (InvocationEntryBci). We return the entry
763 // liveness for the method.
764
765 if (method() == NULL || method()->code_size() == 0) {
766 // We are building a graph for a call to a native method.
767 // All locals are live.
768 return;
769 }
770
771 ResourceMark rm;
772
773 // Consult the liveness information for the locals. If any
774 // of them are unused, then they can be replaced by top(). This
775 // should help register allocation time and cut down on the size
776 // of the deoptimization information.
777 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
778
779 int len = (int)live_locals.size();
780 assert(len <= jvms()->loc_size(), "too many live locals");
781 for (int local = 0; local < len; local++) {
782 if (!live_locals.at(local)) {
783 set_local(local, top());
784 }
785 }
786 }
787
788 #ifdef ASSERT
789 //-------------------------dead_locals_are_killed------------------------------
790 // Return true if all dead locals are set to top in the map.
791 // Used to assert "clean" debug info at various points.
792 bool GraphKit::dead_locals_are_killed() {
793 if (method() == NULL || method()->code_size() == 0) {
794 // No locals need to be dead, so all is as it should be.
795 return true;
796 }
797
798 // Make sure somebody called kill_dead_locals upstream.
799 ResourceMark rm;
800 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
801 if (jvms->loc_size() == 0) continue; // no locals to consult
802 SafePointNode* map = jvms->map();
803 ciMethod* method = jvms->method();
804 int bci = jvms->bci();
805 if (jvms == this->jvms()) {
806 bci = this->bci(); // it might not yet be synched
807 }
808 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
809 int len = (int)live_locals.size();
810 if (!live_locals.is_valid() || len == 0)
811 // This method is trivial, or is poisoned by a breakpoint.
812 return true;
813 assert(len == jvms->loc_size(), "live map consistent with locals map");
814 for (int local = 0; local < len; local++) {
815 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
816 if (PrintMiscellaneous && (Verbose || WizardMode)) {
817 tty->print_cr("Zombie local %d: ", local);
818 jvms->dump();
819 }
820 return false;
821 }
822 }
823 }
824 return true;
825 }
826
827 #endif //ASSERT
828
829 // Helper function for enforcing certain bytecodes to reexecute if
830 // deoptimization happens
831 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
832 ciMethod* cur_method = jvms->method();
833 int cur_bci = jvms->bci();
834 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
835 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
836 return Interpreter::bytecode_should_reexecute(code) ||
837 (is_anewarray && (code == Bytecodes::_multianewarray));
838 // Reexecute _multianewarray bytecode which was replaced with
839 // sequence of [a]newarray. See Parse::do_multianewarray().
840 //
841 // Note: interpreter should not have it set since this optimization
842 // is limited by dimensions and guarded by flag so in some cases
843 // multianewarray() runtime calls will be generated and
844 // the bytecode should not be reexecutes (stack will not be reset).
845 } else {
846 return false;
847 }
848 }
849
850 // Helper function for adding JVMState and debug information to node
851 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
852 // Add the safepoint edges to the call (or other safepoint).
853
854 // Make sure dead locals are set to top. This
855 // should help register allocation time and cut down on the size
856 // of the deoptimization information.
857 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
858
859 // Walk the inline list to fill in the correct set of JVMState's
860 // Also fill in the associated edges for each JVMState.
861
862 // If the bytecode needs to be reexecuted we need to put
863 // the arguments back on the stack.
864 const bool should_reexecute = jvms()->should_reexecute();
865 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
866
867 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
868 // undefined if the bci is different. This is normal for Parse but it
869 // should not happen for LibraryCallKit because only one bci is processed.
870 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
871 "in LibraryCallKit the reexecute bit should not change");
872
873 // If we are guaranteed to throw, we can prune everything but the
874 // input to the current bytecode.
875 bool can_prune_locals = false;
876 uint stack_slots_not_pruned = 0;
877 int inputs = 0, depth = 0;
878 if (must_throw) {
879 assert(method() == youngest_jvms->method(), "sanity");
880 if (compute_stack_effects(inputs, depth)) {
881 can_prune_locals = true;
882 stack_slots_not_pruned = inputs;
883 }
884 }
885
886 if (env()->should_retain_local_variables()) {
887 // At any safepoint, this method can get breakpointed, which would
888 // then require an immediate deoptimization.
889 can_prune_locals = false; // do not prune locals
890 stack_slots_not_pruned = 0;
891 }
892
893 // do not scribble on the input jvms
894 JVMState* out_jvms = youngest_jvms->clone_deep(C);
895 call->set_jvms(out_jvms); // Start jvms list for call node
896
897 // For a known set of bytecodes, the interpreter should reexecute them if
898 // deoptimization happens. We set the reexecute state for them here
899 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
900 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
901 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
902 }
903
904 // Presize the call:
905 DEBUG_ONLY(uint non_debug_edges = call->req());
906 call->add_req_batch(top(), youngest_jvms->debug_depth());
907 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
908
909 // Set up edges so that the call looks like this:
910 // Call [state:] ctl io mem fptr retadr
911 // [parms:] parm0 ... parmN
912 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
913 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
914 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
915 // Note that caller debug info precedes callee debug info.
916
917 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
918 uint debug_ptr = call->req();
919
920 // Loop over the map input edges associated with jvms, add them
921 // to the call node, & reset all offsets to match call node array.
922 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
923 uint debug_end = debug_ptr;
924 uint debug_start = debug_ptr - in_jvms->debug_size();
925 debug_ptr = debug_start; // back up the ptr
926
927 uint p = debug_start; // walks forward in [debug_start, debug_end)
928 uint j, k, l;
929 SafePointNode* in_map = in_jvms->map();
930 out_jvms->set_map(call);
931
932 if (can_prune_locals) {
933 assert(in_jvms->method() == out_jvms->method(), "sanity");
934 // If the current throw can reach an exception handler in this JVMS,
935 // then we must keep everything live that can reach that handler.
936 // As a quick and dirty approximation, we look for any handlers at all.
937 if (in_jvms->method()->has_exception_handlers()) {
938 can_prune_locals = false;
939 }
940 }
941
942 // Add the Locals
943 k = in_jvms->locoff();
944 l = in_jvms->loc_size();
945 out_jvms->set_locoff(p);
946 if (!can_prune_locals) {
947 for (j = 0; j < l; j++)
948 call->set_req(p++, in_map->in(k+j));
949 } else {
950 p += l; // already set to top above by add_req_batch
951 }
952
953 // Add the Expression Stack
954 k = in_jvms->stkoff();
955 l = in_jvms->sp();
956 out_jvms->set_stkoff(p);
957 if (!can_prune_locals) {
958 for (j = 0; j < l; j++)
959 call->set_req(p++, in_map->in(k+j));
960 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
961 // Divide stack into {S0,...,S1}, where S0 is set to top.
962 uint s1 = stack_slots_not_pruned;
963 stack_slots_not_pruned = 0; // for next iteration
964 if (s1 > l) s1 = l;
965 uint s0 = l - s1;
966 p += s0; // skip the tops preinstalled by add_req_batch
967 for (j = s0; j < l; j++)
968 call->set_req(p++, in_map->in(k+j));
969 } else {
970 p += l; // already set to top above by add_req_batch
971 }
972
973 // Add the Monitors
974 k = in_jvms->monoff();
975 l = in_jvms->mon_size();
976 out_jvms->set_monoff(p);
977 for (j = 0; j < l; j++)
978 call->set_req(p++, in_map->in(k+j));
979
980 // Copy any scalar object fields.
981 k = in_jvms->scloff();
982 l = in_jvms->scl_size();
983 out_jvms->set_scloff(p);
984 for (j = 0; j < l; j++)
985 call->set_req(p++, in_map->in(k+j));
986
987 // Finish the new jvms.
988 out_jvms->set_endoff(p);
989
990 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
991 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
992 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
993 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
994 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
995 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
996
997 // Update the two tail pointers in parallel.
998 out_jvms = out_jvms->caller();
999 in_jvms = in_jvms->caller();
1000 }
1001
1002 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1003
1004 // Test the correctness of JVMState::debug_xxx accessors:
1005 assert(call->jvms()->debug_start() == non_debug_edges, "");
1006 assert(call->jvms()->debug_end() == call->req(), "");
1007 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1008 }
1009
1010 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1011 Bytecodes::Code code = java_bc();
1012 if (code == Bytecodes::_wide) {
1013 code = method()->java_code_at_bci(bci() + 1);
1014 }
1015
1016 BasicType rtype = T_ILLEGAL;
1017 int rsize = 0;
1018
1019 if (code != Bytecodes::_illegal) {
1020 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1021 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1022 if (rtype < T_CONFLICT)
1023 rsize = type2size[rtype];
1024 }
1025
1026 switch (code) {
1027 case Bytecodes::_illegal:
1028 return false;
1029
1030 case Bytecodes::_ldc:
1031 case Bytecodes::_ldc_w:
1032 case Bytecodes::_ldc2_w:
1033 inputs = 0;
1034 break;
1035
1036 case Bytecodes::_dup: inputs = 1; break;
1037 case Bytecodes::_dup_x1: inputs = 2; break;
1038 case Bytecodes::_dup_x2: inputs = 3; break;
1039 case Bytecodes::_dup2: inputs = 2; break;
1040 case Bytecodes::_dup2_x1: inputs = 3; break;
1041 case Bytecodes::_dup2_x2: inputs = 4; break;
1042 case Bytecodes::_swap: inputs = 2; break;
1043 case Bytecodes::_arraylength: inputs = 1; break;
1044
1045 case Bytecodes::_getstatic:
1046 case Bytecodes::_putstatic:
1047 case Bytecodes::_getfield:
1048 case Bytecodes::_putfield:
1049 {
1050 bool ignored_will_link;
1051 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1052 int size = field->type()->size();
1053 bool is_get = (depth >= 0), is_static = (depth & 1);
1054 inputs = (is_static ? 0 : 1);
1055 if (is_get) {
1056 depth = size - inputs;
1057 } else {
1058 inputs += size; // putxxx pops the value from the stack
1059 depth = - inputs;
1060 }
1061 }
1062 break;
1063
1064 case Bytecodes::_invokevirtual:
1065 case Bytecodes::_invokespecial:
1066 case Bytecodes::_invokestatic:
1067 case Bytecodes::_invokedynamic:
1068 case Bytecodes::_invokeinterface:
1069 {
1070 bool ignored_will_link;
1071 ciSignature* declared_signature = NULL;
1072 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1073 assert(declared_signature != NULL, "cannot be null");
1074 inputs = declared_signature->arg_size_for_bc(code);
1075 int size = declared_signature->return_type()->size();
1076 depth = size - inputs;
1077 }
1078 break;
1079
1080 case Bytecodes::_multianewarray:
1081 {
1082 ciBytecodeStream iter(method());
1083 iter.reset_to_bci(bci());
1084 iter.next();
1085 inputs = iter.get_dimensions();
1086 assert(rsize == 1, "");
1087 depth = rsize - inputs;
1088 }
1089 break;
1090
1091 case Bytecodes::_withfield: {
1092 bool ignored_will_link;
1093 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1094 int size = field->type()->size();
1095 inputs = size+1;
1096 depth = rsize - inputs;
1097 break;
1098 }
1099
1100 case Bytecodes::_ireturn:
1101 case Bytecodes::_lreturn:
1102 case Bytecodes::_freturn:
1103 case Bytecodes::_dreturn:
1104 case Bytecodes::_areturn:
1105 assert(rsize == -depth, "");
1106 inputs = rsize;
1107 break;
1108
1109 case Bytecodes::_jsr:
1110 case Bytecodes::_jsr_w:
1111 inputs = 0;
1112 depth = 1; // S.B. depth=1, not zero
1113 break;
1114
1115 default:
1116 // bytecode produces a typed result
1117 inputs = rsize - depth;
1118 assert(inputs >= 0, "");
1119 break;
1120 }
1121
1122 #ifdef ASSERT
1123 // spot check
1124 int outputs = depth + inputs;
1125 assert(outputs >= 0, "sanity");
1126 switch (code) {
1127 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1128 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1129 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1130 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1131 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1132 default: break;
1133 }
1134 #endif //ASSERT
1135
1136 return true;
1137 }
1138
1139
1140
1141 //------------------------------basic_plus_adr---------------------------------
1142 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1143 // short-circuit a common case
1144 if (offset == intcon(0)) return ptr;
1145 return _gvn.transform( new AddPNode(base, ptr, offset) );
1146 }
1147
1148 Node* GraphKit::ConvI2L(Node* offset) {
1149 // short-circuit a common case
1150 jint offset_con = find_int_con(offset, Type::OffsetBot);
1151 if (offset_con != Type::OffsetBot) {
1152 return longcon((jlong) offset_con);
1153 }
1154 return _gvn.transform( new ConvI2LNode(offset));
1155 }
1156
1157 Node* GraphKit::ConvI2UL(Node* offset) {
1158 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1159 if (offset_con != (juint) Type::OffsetBot) {
1160 return longcon((julong) offset_con);
1161 }
1162 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1163 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1164 return _gvn.transform( new AndLNode(conv, mask) );
1165 }
1166
1167 Node* GraphKit::ConvL2I(Node* offset) {
1168 // short-circuit a common case
1169 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1170 if (offset_con != (jlong)Type::OffsetBot) {
1171 return intcon((int) offset_con);
1172 }
1173 return _gvn.transform( new ConvL2INode(offset));
1174 }
1175
1176 //-------------------------load_object_klass-----------------------------------
1177 Node* GraphKit::load_object_klass(Node* obj) {
1178 // Special-case a fresh allocation to avoid building nodes:
1179 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1180 if (akls != NULL) return akls;
1181 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1182 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1183 }
1184
1185 //-------------------------load_array_length-----------------------------------
1186 Node* GraphKit::load_array_length(Node* array) {
1187 // Special-case a fresh allocation to avoid building nodes:
1188 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1189 Node *alen;
1190 if (alloc == NULL) {
1191 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1192 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1193 } else {
1194 alen = alloc->Ideal_length();
1195 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1196 if (ccast != alen) {
1197 alen = _gvn.transform(ccast);
1198 }
1199 }
1200 return alen;
1201 }
1202
1203 //------------------------------do_null_check----------------------------------
1204 // Helper function to do a NULL pointer check. Returned value is
1205 // the incoming address with NULL casted away. You are allowed to use the
1206 // not-null value only if you are control dependent on the test.
1207 #ifndef PRODUCT
1208 extern int explicit_null_checks_inserted,
1209 explicit_null_checks_elided;
1210 #endif
1211 Node* GraphKit::null_check_common(Node* value, BasicType type,
1212 // optional arguments for variations:
1213 bool assert_null,
1214 Node* *null_control,
1215 bool speculative) {
1216 assert(!assert_null || null_control == NULL, "not both at once");
1217 if (stopped()) return top();
1218 NOT_PRODUCT(explicit_null_checks_inserted++);
1219
1220 // Construct NULL check
1221 Node *chk = NULL;
1222 switch(type) {
1223 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1224 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1225 case T_VALUETYPE : // fall through
1226 case T_ARRAY : // fall through
1227 type = T_OBJECT; // simplify further tests
1228 case T_OBJECT : {
1229 const Type *t = _gvn.type( value );
1230
1231 const TypeOopPtr* tp = t->isa_oopptr();
1232 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1233 // Only for do_null_check, not any of its siblings:
1234 && !assert_null && null_control == NULL) {
1235 // Usually, any field access or invocation on an unloaded oop type
1236 // will simply fail to link, since the statically linked class is
1237 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1238 // the static class is loaded but the sharper oop type is not.
1239 // Rather than checking for this obscure case in lots of places,
1240 // we simply observe that a null check on an unloaded class
1241 // will always be followed by a nonsense operation, so we
1242 // can just issue the uncommon trap here.
1243 // Our access to the unloaded class will only be correct
1244 // after it has been loaded and initialized, which requires
1245 // a trip through the interpreter.
1246 #ifndef PRODUCT
1247 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1248 #endif
1249 uncommon_trap(Deoptimization::Reason_unloaded,
1250 Deoptimization::Action_reinterpret,
1251 tp->klass(), "!loaded");
1252 return top();
1253 }
1254
1255 if (assert_null) {
1256 // See if the type is contained in NULL_PTR.
1257 // If so, then the value is already null.
1258 if (t->higher_equal(TypePtr::NULL_PTR)) {
1259 NOT_PRODUCT(explicit_null_checks_elided++);
1260 return value; // Elided null assert quickly!
1261 }
1262 } else {
1263 // See if mixing in the NULL pointer changes type.
1264 // If so, then the NULL pointer was not allowed in the original
1265 // type. In other words, "value" was not-null.
1266 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1267 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1268 NOT_PRODUCT(explicit_null_checks_elided++);
1269 return value; // Elided null check quickly!
1270 }
1271 }
1272 chk = new CmpPNode( value, null() );
1273 break;
1274 }
1275
1276 default:
1277 fatal("unexpected type: %s", type2name(type));
1278 }
1279 assert(chk != NULL, "sanity check");
1280 chk = _gvn.transform(chk);
1281
1282 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1283 BoolNode *btst = new BoolNode( chk, btest);
1284 Node *tst = _gvn.transform( btst );
1285
1286 //-----------
1287 // if peephole optimizations occurred, a prior test existed.
1288 // If a prior test existed, maybe it dominates as we can avoid this test.
1289 if (tst != btst && type == T_OBJECT) {
1290 // At this point we want to scan up the CFG to see if we can
1291 // find an identical test (and so avoid this test altogether).
1292 Node *cfg = control();
1293 int depth = 0;
1294 while( depth < 16 ) { // Limit search depth for speed
1295 if( cfg->Opcode() == Op_IfTrue &&
1296 cfg->in(0)->in(1) == tst ) {
1297 // Found prior test. Use "cast_not_null" to construct an identical
1298 // CastPP (and hence hash to) as already exists for the prior test.
1299 // Return that casted value.
1300 if (assert_null) {
1301 replace_in_map(value, null());
1302 return null(); // do not issue the redundant test
1303 }
1304 Node *oldcontrol = control();
1305 set_control(cfg);
1306 Node *res = cast_not_null(value);
1307 set_control(oldcontrol);
1308 NOT_PRODUCT(explicit_null_checks_elided++);
1309 return res;
1310 }
1311 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1312 if (cfg == NULL) break; // Quit at region nodes
1313 depth++;
1314 }
1315 }
1316
1317 //-----------
1318 // Branch to failure if null
1319 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1320 Deoptimization::DeoptReason reason;
1321 if (assert_null) {
1322 reason = Deoptimization::reason_null_assert(speculative);
1323 } else if (type == T_OBJECT) {
1324 reason = Deoptimization::reason_null_check(speculative);
1325 } else {
1326 reason = Deoptimization::Reason_div0_check;
1327 }
1328 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1329 // ciMethodData::has_trap_at will return a conservative -1 if any
1330 // must-be-null assertion has failed. This could cause performance
1331 // problems for a method after its first do_null_assert failure.
1332 // Consider using 'Reason_class_check' instead?
1333
1334 // To cause an implicit null check, we set the not-null probability
1335 // to the maximum (PROB_MAX). For an explicit check the probability
1336 // is set to a smaller value.
1337 if (null_control != NULL || too_many_traps(reason)) {
1338 // probability is less likely
1339 ok_prob = PROB_LIKELY_MAG(3);
1340 } else if (!assert_null &&
1341 (ImplicitNullCheckThreshold > 0) &&
1342 method() != NULL &&
1343 (method()->method_data()->trap_count(reason)
1344 >= (uint)ImplicitNullCheckThreshold)) {
1345 ok_prob = PROB_LIKELY_MAG(3);
1346 }
1347
1348 if (null_control != NULL) {
1349 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1350 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1351 set_control( _gvn.transform( new IfTrueNode(iff)));
1352 #ifndef PRODUCT
1353 if (null_true == top()) {
1354 explicit_null_checks_elided++;
1355 }
1356 #endif
1357 (*null_control) = null_true;
1358 } else {
1359 BuildCutout unless(this, tst, ok_prob);
1360 // Check for optimizer eliding test at parse time
1361 if (stopped()) {
1362 // Failure not possible; do not bother making uncommon trap.
1363 NOT_PRODUCT(explicit_null_checks_elided++);
1364 } else if (assert_null) {
1365 uncommon_trap(reason,
1366 Deoptimization::Action_make_not_entrant,
1367 NULL, "assert_null");
1368 } else {
1369 replace_in_map(value, zerocon(type));
1370 builtin_throw(reason);
1371 }
1372 }
1373
1374 // Must throw exception, fall-thru not possible?
1375 if (stopped()) {
1376 return top(); // No result
1377 }
1378
1379 if (assert_null) {
1380 // Cast obj to null on this path.
1381 replace_in_map(value, zerocon(type));
1382 return zerocon(type);
1383 }
1384
1385 // Cast obj to not-null on this path, if there is no null_control.
1386 // (If there is a null_control, a non-null value may come back to haunt us.)
1387 if (type == T_OBJECT) {
1388 Node* cast = cast_not_null(value, false);
1389 if (null_control == NULL || (*null_control) == top())
1390 replace_in_map(value, cast);
1391 value = cast;
1392 }
1393
1394 return value;
1395 }
1396
1397 Node* GraphKit::null2default(Node* value, ciValueKlass* vk) {
1398 Node* null_ctl = top();
1399 value = null_check_oop(value, &null_ctl);
1400 if (!null_ctl->is_top()) {
1401 // Return default value if oop is null
1402 Node* region = new RegionNode(3);
1403 region->init_req(1, control());
1404 region->init_req(2, null_ctl);
1405 value = PhiNode::make(region, value, TypeInstPtr::make(TypePtr::BotPTR, vk));
1406 value->set_req(2, ValueTypeNode::default_oop(gvn(), vk));
1407 set_control(gvn().transform(region));
1408 value = gvn().transform(value);
1409 }
1410 return value;
1411 }
1412
1413 //------------------------------cast_not_null----------------------------------
1414 // Cast obj to not-null on this path
1415 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1416 if (obj->is_ValueType()) {
1417 return obj;
1418 }
1419 const Type *t = _gvn.type(obj);
1420 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1421 // Object is already not-null?
1422 if( t == t_not_null ) return obj;
1423
1424 Node *cast = new CastPPNode(obj,t_not_null);
1425 cast->init_req(0, control());
1426 cast = _gvn.transform( cast );
1427
1428 // Scan for instances of 'obj' in the current JVM mapping.
1429 // These instances are known to be not-null after the test.
1430 if (do_replace_in_map)
1431 replace_in_map(obj, cast);
1432
1433 return cast; // Return casted value
1434 }
1435
1436 // Sometimes in intrinsics, we implicitly know an object is not null
1437 // (there's no actual null check) so we can cast it to not null. In
1438 // the course of optimizations, the input to the cast can become null.
1439 // In that case that data path will die and we need the control path
1440 // to become dead as well to keep the graph consistent. So we have to
1441 // add a check for null for which one branch can't be taken. It uses
1442 // an Opaque4 node that will cause the check to be removed after loop
1443 // opts so the test goes away and the compiled code doesn't execute a
1444 // useless check.
1445 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1446 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1447 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1448 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1449 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1450 _gvn.set_type(iff, iff->Value(&_gvn));
1451 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1452 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1453 Node *halt = _gvn.transform(new HaltNode(if_f, frame));
1454 C->root()->add_req(halt);
1455 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1456 set_control(if_t);
1457 return cast_not_null(value, do_replace_in_map);
1458 }
1459
1460
1461 //--------------------------replace_in_map-------------------------------------
1462 void GraphKit::replace_in_map(Node* old, Node* neww) {
1463 if (old == neww) {
1464 return;
1465 }
1466
1467 map()->replace_edge(old, neww);
1468
1469 // Note: This operation potentially replaces any edge
1470 // on the map. This includes locals, stack, and monitors
1471 // of the current (innermost) JVM state.
1472
1473 // don't let inconsistent types from profiling escape this
1474 // method
1475
1476 const Type* told = _gvn.type(old);
1477 const Type* tnew = _gvn.type(neww);
1478
1479 if (!tnew->higher_equal(told)) {
1480 return;
1481 }
1482
1483 map()->record_replaced_node(old, neww);
1484 }
1485
1486
1487 //=============================================================================
1488 //--------------------------------memory---------------------------------------
1489 Node* GraphKit::memory(uint alias_idx) {
1490 MergeMemNode* mem = merged_memory();
1491 Node* p = mem->memory_at(alias_idx);
1492 _gvn.set_type(p, Type::MEMORY); // must be mapped
1493 return p;
1494 }
1495
1496 //-----------------------------reset_memory------------------------------------
1497 Node* GraphKit::reset_memory() {
1498 Node* mem = map()->memory();
1499 // do not use this node for any more parsing!
1500 debug_only( map()->set_memory((Node*)NULL) );
1501 return _gvn.transform( mem );
1502 }
1503
1504 //------------------------------set_all_memory---------------------------------
1505 void GraphKit::set_all_memory(Node* newmem) {
1506 Node* mergemem = MergeMemNode::make(newmem);
1507 gvn().set_type_bottom(mergemem);
1508 map()->set_memory(mergemem);
1509 }
1510
1511 //------------------------------set_all_memory_call----------------------------
1512 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1513 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1514 set_all_memory(newmem);
1515 }
1516
1517 //=============================================================================
1518 //
1519 // parser factory methods for MemNodes
1520 //
1521 // These are layered on top of the factory methods in LoadNode and StoreNode,
1522 // and integrate with the parser's memory state and _gvn engine.
1523 //
1524
1525 // factory methods in "int adr_idx"
1526 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1527 int adr_idx,
1528 MemNode::MemOrd mo,
1529 LoadNode::ControlDependency control_dependency,
1530 bool require_atomic_access,
1531 bool unaligned,
1532 bool mismatched,
1533 bool unsafe) {
1534 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1535 const TypePtr* adr_type = NULL; // debug-mode-only argument
1536 debug_only(adr_type = C->get_adr_type(adr_idx));
1537 Node* mem = memory(adr_idx);
1538 Node* ld;
1539 if (require_atomic_access && bt == T_LONG) {
1540 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe);
1541 } else if (require_atomic_access && bt == T_DOUBLE) {
1542 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe);
1543 } else {
1544 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe);
1545 }
1546 ld = _gvn.transform(ld);
1547
1548 if (((bt == T_OBJECT || bt == T_VALUETYPE) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1549 // Improve graph before escape analysis and boxing elimination.
1550 record_for_igvn(ld);
1551 }
1552 return ld;
1553 }
1554
1555 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1556 int adr_idx,
1557 MemNode::MemOrd mo,
1558 bool require_atomic_access,
1559 bool unaligned,
1560 bool mismatched,
1561 bool unsafe) {
1562 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1563 const TypePtr* adr_type = NULL;
1564 debug_only(adr_type = C->get_adr_type(adr_idx));
1565 Node *mem = memory(adr_idx);
1566 Node* st;
1567 if (require_atomic_access && bt == T_LONG) {
1568 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1569 } else if (require_atomic_access && bt == T_DOUBLE) {
1570 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1571 } else {
1572 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1573 }
1574 if (unaligned) {
1575 st->as_Store()->set_unaligned_access();
1576 }
1577 if (mismatched) {
1578 st->as_Store()->set_mismatched_access();
1579 }
1580 if (unsafe) {
1581 st->as_Store()->set_unsafe_access();
1582 }
1583 st = _gvn.transform(st);
1584 set_memory(st, adr_idx);
1585 // Back-to-back stores can only remove intermediate store with DU info
1586 // so push on worklist for optimizer.
1587 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1588 record_for_igvn(st);
1589
1590 return st;
1591 }
1592
1593 Node* GraphKit::access_store_at(Node* obj,
1594 Node* adr,
1595 const TypePtr* adr_type,
1596 Node* val,
1597 const Type* val_type,
1598 BasicType bt,
1599 DecoratorSet decorators,
1600 bool deoptimize_on_exception,
1601 bool safe_for_replace) {
1602 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1603 // could be delayed during Parse (for example, in adjust_map_after_if()).
1604 // Execute transformation here to avoid barrier generation in such case.
1605 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1606 val = _gvn.makecon(TypePtr::NULL_PTR);
1607 }
1608
1609 if (stopped()) {
1610 return top(); // Dead path ?
1611 }
1612
1613 assert(val != NULL, "not dead path");
1614 if (val->is_ValueType()) {
1615 // Allocate value type and get oop
1616 val = val->as_ValueType()->allocate(this, deoptimize_on_exception, safe_for_replace)->get_oop();
1617 }
1618
1619 C2AccessValuePtr addr(adr, adr_type);
1620 C2AccessValue value(val, val_type);
1621 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1622 if (access.is_raw()) {
1623 return _barrier_set->BarrierSetC2::store_at(access, value);
1624 } else {
1625 return _barrier_set->store_at(access, value);
1626 }
1627 }
1628
1629 Node* GraphKit::access_load_at(Node* obj, // containing obj
1630 Node* adr, // actual adress to store val at
1631 const TypePtr* adr_type,
1632 const Type* val_type,
1633 BasicType bt,
1634 DecoratorSet decorators,
1635 Node* ctl) {
1636 if (stopped()) {
1637 return top(); // Dead path ?
1638 }
1639
1640 C2AccessValuePtr addr(adr, adr_type);
1641 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1642 if (access.is_raw()) {
1643 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1644 } else {
1645 return _barrier_set->load_at(access, val_type);
1646 }
1647 }
1648
1649 Node* GraphKit::access_load(Node* adr, // actual adress to load val at
1650 const Type* val_type,
1651 BasicType bt,
1652 DecoratorSet decorators) {
1653 if (stopped()) {
1654 return top(); // Dead path ?
1655 }
1656
1657 C2AccessValuePtr addr(adr, NULL);
1658 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1659 if (access.is_raw()) {
1660 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1661 } else {
1662 return _barrier_set->load_at(access, val_type);
1663 }
1664 }
1665
1666 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1667 Node* adr,
1668 const TypePtr* adr_type,
1669 int alias_idx,
1670 Node* expected_val,
1671 Node* new_val,
1672 const Type* value_type,
1673 BasicType bt,
1674 DecoratorSet decorators) {
1675 C2AccessValuePtr addr(adr, adr_type);
1676 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1677 bt, obj, addr, alias_idx);
1678 if (access.is_raw()) {
1679 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1680 } else {
1681 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1682 }
1683 }
1684
1685 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1686 Node* adr,
1687 const TypePtr* adr_type,
1688 int alias_idx,
1689 Node* expected_val,
1690 Node* new_val,
1691 const Type* value_type,
1692 BasicType bt,
1693 DecoratorSet decorators) {
1694 C2AccessValuePtr addr(adr, adr_type);
1695 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1696 bt, obj, addr, alias_idx);
1697 if (access.is_raw()) {
1698 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1699 } else {
1700 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1701 }
1702 }
1703
1704 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1705 Node* adr,
1706 const TypePtr* adr_type,
1707 int alias_idx,
1708 Node* new_val,
1709 const Type* value_type,
1710 BasicType bt,
1711 DecoratorSet decorators) {
1712 C2AccessValuePtr addr(adr, adr_type);
1713 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1714 bt, obj, addr, alias_idx);
1715 if (access.is_raw()) {
1716 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1717 } else {
1718 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1719 }
1720 }
1721
1722 Node* GraphKit::access_atomic_add_at(Node* obj,
1723 Node* adr,
1724 const TypePtr* adr_type,
1725 int alias_idx,
1726 Node* new_val,
1727 const Type* value_type,
1728 BasicType bt,
1729 DecoratorSet decorators) {
1730 C2AccessValuePtr addr(adr, adr_type);
1731 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1732 if (access.is_raw()) {
1733 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1734 } else {
1735 return _barrier_set->atomic_add_at(access, new_val, value_type);
1736 }
1737 }
1738
1739 void GraphKit::access_clone(Node* src_base, Node* dst_base, Node* countx, bool is_array) {
1740 return _barrier_set->clone(this, src_base, dst_base, countx, is_array);
1741 }
1742
1743 Node* GraphKit::access_resolve(Node* n, DecoratorSet decorators) {
1744 // Use stronger ACCESS_WRITE|ACCESS_READ by default.
1745 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
1746 decorators |= ACCESS_READ | ACCESS_WRITE;
1747 }
1748 return _barrier_set->resolve(this, n, decorators);
1749 }
1750
1751 //-------------------------array_element_address-------------------------
1752 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1753 const TypeInt* sizetype, Node* ctrl) {
1754 uint shift = exact_log2(type2aelembytes(elembt));
1755 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1756 if (arytype_klass != NULL && arytype_klass->is_value_array_klass()) {
1757 ciValueArrayKlass* vak = arytype_klass->as_value_array_klass();
1758 shift = vak->log2_element_size();
1759 }
1760 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1761
1762 // short-circuit a common case (saves lots of confusing waste motion)
1763 jint idx_con = find_int_con(idx, -1);
1764 if (idx_con >= 0) {
1765 intptr_t offset = header + ((intptr_t)idx_con << shift);
1766 return basic_plus_adr(ary, offset);
1767 }
1768
1769 // must be correct type for alignment purposes
1770 Node* base = basic_plus_adr(ary, header);
1771 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1772 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1773 return basic_plus_adr(ary, base, scale);
1774 }
1775
1776 //-------------------------load_array_element-------------------------
1777 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1778 const Type* elemtype = arytype->elem();
1779 BasicType elembt = elemtype->array_element_basic_type();
1780 assert(elembt != T_VALUETYPE, "value types are not supported by this method");
1781 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1782 if (elembt == T_NARROWOOP) {
1783 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1784 }
1785 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1786 return ld;
1787 }
1788
1789 //-------------------------set_arguments_for_java_call-------------------------
1790 // Arguments (pre-popped from the stack) are taken from the JVMS.
1791 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool incremental_inlining) {
1792 // Add the call arguments:
1793 const TypeTuple* domain = call->tf()->domain_sig();
1794 ExtendedSignature sig_cc = ExtendedSignature(call->method()->get_sig_cc(), SigEntryFilter());
1795 uint nargs = domain->cnt();
1796 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1797 Node* arg = argument(i-TypeFunc::Parms);
1798 const Type* t = domain->field_at(i);
1799 if (call->method()->has_scalarized_args() && t->is_valuetypeptr() && !t->maybe_null()) {
1800 // We don't pass value type arguments by reference but instead pass each field of the value type
1801 ValueTypeNode* vt = arg->as_ValueType();
1802 vt->pass_fields(this, call, sig_cc, idx);
1803 // If a value type argument is passed as fields, attach the Method* to the call site
1804 // to be able to access the extended signature later via attached_method_before_pc().
1805 // For example, see CompiledMethod::preserve_callee_argument_oops().
1806 call->set_override_symbolic_info(true);
1807 continue;
1808 } else if (arg->is_ValueType()) {
1809 // Pass value type argument via oop to callee
1810 if (!incremental_inlining) {
1811 arg = arg->as_ValueType()->allocate(this)->get_oop();
1812 } else {
1813 arg = ValueTypePtrNode::make_from_value_type(this, arg->as_ValueType());
1814 }
1815 }
1816 call->init_req(idx++, arg);
1817 // Skip reserved arguments
1818 BasicType bt = t->basic_type();
1819 while (SigEntry::next_is_reserved(sig_cc, bt, true)) {
1820 call->init_req(idx++, top());
1821 if (type2size[bt] == 2) {
1822 call->init_req(idx++, top());
1823 }
1824 }
1825 }
1826 }
1827
1828 //---------------------------set_edges_for_java_call---------------------------
1829 // Connect a newly created call into the current JVMS.
1830 // A return value node (if any) is returned from set_edges_for_java_call.
1831 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1832
1833 // Add the predefined inputs:
1834 call->init_req( TypeFunc::Control, control() );
1835 call->init_req( TypeFunc::I_O , i_o() );
1836 call->init_req( TypeFunc::Memory , reset_memory() );
1837 call->init_req( TypeFunc::FramePtr, frameptr() );
1838 call->init_req( TypeFunc::ReturnAdr, top() );
1839
1840 add_safepoint_edges(call, must_throw);
1841
1842 Node* xcall = _gvn.transform(call);
1843
1844 if (xcall == top()) {
1845 set_control(top());
1846 return;
1847 }
1848 assert(xcall == call, "call identity is stable");
1849
1850 // Re-use the current map to produce the result.
1851
1852 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1853 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1854 set_all_memory_call(xcall, separate_io_proj);
1855
1856 //return xcall; // no need, caller already has it
1857 }
1858
1859 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1860 if (stopped()) return top(); // maybe the call folded up?
1861
1862 // Note: Since any out-of-line call can produce an exception,
1863 // we always insert an I_O projection from the call into the result.
1864
1865 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1866
1867 if (separate_io_proj) {
1868 // The caller requested separate projections be used by the fall
1869 // through and exceptional paths, so replace the projections for
1870 // the fall through path.
1871 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1872 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1873 }
1874
1875 // Capture the return value, if any.
1876 Node* ret;
1877 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) {
1878 ret = top();
1879 } else if (call->tf()->returns_value_type_as_fields()) {
1880 // Return of multiple values (value type fields): we create a
1881 // ValueType node, each field is a projection from the call.
1882 ciValueKlass* vk = call->method()->return_type()->as_value_klass();
1883 const Array<SigEntry>* sig_array = vk->extended_sig();
1884 GrowableArray<SigEntry> sig = GrowableArray<SigEntry>(sig_array->length());
1885 sig.appendAll(sig_array);
1886 ExtendedSignature sig_cc = ExtendedSignature(&sig, SigEntryFilter());
1887 uint base_input = TypeFunc::Parms + 1;
1888 ret = ValueTypeNode::make_from_multi(this, call, sig_cc, vk, base_input, false);
1889 } else {
1890 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1891 }
1892
1893 return ret;
1894 }
1895
1896 //--------------------set_predefined_input_for_runtime_call--------------------
1897 // Reading and setting the memory state is way conservative here.
1898 // The real problem is that I am not doing real Type analysis on memory,
1899 // so I cannot distinguish card mark stores from other stores. Across a GC
1900 // point the Store Barrier and the card mark memory has to agree. I cannot
1901 // have a card mark store and its barrier split across the GC point from
1902 // either above or below. Here I get that to happen by reading ALL of memory.
1903 // A better answer would be to separate out card marks from other memory.
1904 // For now, return the input memory state, so that it can be reused
1905 // after the call, if this call has restricted memory effects.
1906 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1907 // Set fixed predefined input arguments
1908 Node* memory = reset_memory();
1909 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1910 call->init_req( TypeFunc::Control, control() );
1911 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1912 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1913 call->init_req( TypeFunc::FramePtr, frameptr() );
1914 call->init_req( TypeFunc::ReturnAdr, top() );
1915 return memory;
1916 }
1917
1918 //-------------------set_predefined_output_for_runtime_call--------------------
1919 // Set control and memory (not i_o) from the call.
1920 // If keep_mem is not NULL, use it for the output state,
1921 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1922 // If hook_mem is NULL, this call produces no memory effects at all.
1923 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1924 // then only that memory slice is taken from the call.
1925 // In the last case, we must put an appropriate memory barrier before
1926 // the call, so as to create the correct anti-dependencies on loads
1927 // preceding the call.
1928 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1929 Node* keep_mem,
1930 const TypePtr* hook_mem) {
1931 // no i/o
1932 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1933 if (keep_mem) {
1934 // First clone the existing memory state
1935 set_all_memory(keep_mem);
1936 if (hook_mem != NULL) {
1937 // Make memory for the call
1938 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1939 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1940 // We also use hook_mem to extract specific effects from arraycopy stubs.
1941 set_memory(mem, hook_mem);
1942 }
1943 // ...else the call has NO memory effects.
1944
1945 // Make sure the call advertises its memory effects precisely.
1946 // This lets us build accurate anti-dependences in gcm.cpp.
1947 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1948 "call node must be constructed correctly");
1949 } else {
1950 assert(hook_mem == NULL, "");
1951 // This is not a "slow path" call; all memory comes from the call.
1952 set_all_memory_call(call);
1953 }
1954 }
1955
1956 // Keep track of MergeMems feeding into other MergeMems
1957 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
1958 if (!mem->is_MergeMem()) {
1959 return;
1960 }
1961 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
1962 Node* use = i.get();
1963 if (use->is_MergeMem()) {
1964 wl.push(use);
1965 }
1966 }
1967 }
1968
1969 // Replace the call with the current state of the kit.
1970 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1971 JVMState* ejvms = NULL;
1972 if (has_exceptions()) {
1973 ejvms = transfer_exceptions_into_jvms();
1974 }
1975
1976 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1977 ReplacedNodes replaced_nodes_exception;
1978 Node* ex_ctl = top();
1979
1980 SafePointNode* final_state = stop();
1981
1982 // Find all the needed outputs of this call
1983 CallProjections* callprojs = call->extract_projections(true);
1984
1985 Unique_Node_List wl;
1986 Node* init_mem = call->in(TypeFunc::Memory);
1987 Node* final_mem = final_state->in(TypeFunc::Memory);
1988 Node* final_ctl = final_state->in(TypeFunc::Control);
1989 Node* final_io = final_state->in(TypeFunc::I_O);
1990
1991 // Replace all the old call edges with the edges from the inlining result
1992 if (callprojs->fallthrough_catchproj != NULL) {
1993 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
1994 }
1995 if (callprojs->fallthrough_memproj != NULL) {
1996 if (final_mem->is_MergeMem()) {
1997 // Parser's exits MergeMem was not transformed but may be optimized
1998 final_mem = _gvn.transform(final_mem);
1999 }
2000 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2001 add_mergemem_users_to_worklist(wl, final_mem);
2002 }
2003 if (callprojs->fallthrough_ioproj != NULL) {
2004 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2005 }
2006
2007 // Replace the result with the new result if it exists and is used
2008 if (callprojs->resproj[0] != NULL && result != NULL) {
2009 assert(callprojs->nb_resproj == 1, "unexpected number of results");
2010 C->gvn_replace_by(callprojs->resproj[0], result);
2011 }
2012
2013 if (ejvms == NULL) {
2014 // No exception edges to simply kill off those paths
2015 if (callprojs->catchall_catchproj != NULL) {
2016 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2017 }
2018 if (callprojs->catchall_memproj != NULL) {
2019 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2020 }
2021 if (callprojs->catchall_ioproj != NULL) {
2022 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2023 }
2024 // Replace the old exception object with top
2025 if (callprojs->exobj != NULL) {
2026 C->gvn_replace_by(callprojs->exobj, C->top());
2027 }
2028 } else {
2029 GraphKit ekit(ejvms);
2030
2031 // Load my combined exception state into the kit, with all phis transformed:
2032 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2033 replaced_nodes_exception = ex_map->replaced_nodes();
2034
2035 Node* ex_oop = ekit.use_exception_state(ex_map);
2036
2037 if (callprojs->catchall_catchproj != NULL) {
2038 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2039 ex_ctl = ekit.control();
2040 }
2041 if (callprojs->catchall_memproj != NULL) {
2042 Node* ex_mem = ekit.reset_memory();
2043 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2044 add_mergemem_users_to_worklist(wl, ex_mem);
2045 }
2046 if (callprojs->catchall_ioproj != NULL) {
2047 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2048 }
2049
2050 // Replace the old exception object with the newly created one
2051 if (callprojs->exobj != NULL) {
2052 C->gvn_replace_by(callprojs->exobj, ex_oop);
2053 }
2054 }
2055
2056 // Disconnect the call from the graph
2057 call->disconnect_inputs(NULL, C);
2058 C->gvn_replace_by(call, C->top());
2059
2060 // Clean up any MergeMems that feed other MergeMems since the
2061 // optimizer doesn't like that.
2062 while (wl.size() > 0) {
2063 _gvn.transform(wl.pop());
2064 }
2065
2066 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2067 replaced_nodes.apply(C, final_ctl);
2068 }
2069 if (!ex_ctl->is_top() && do_replaced_nodes) {
2070 replaced_nodes_exception.apply(C, ex_ctl);
2071 }
2072 }
2073
2074
2075 //------------------------------increment_counter------------------------------
2076 // for statistics: increment a VM counter by 1
2077
2078 void GraphKit::increment_counter(address counter_addr) {
2079 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2080 increment_counter(adr1);
2081 }
2082
2083 void GraphKit::increment_counter(Node* counter_addr) {
2084 int adr_type = Compile::AliasIdxRaw;
2085 Node* ctrl = control();
2086 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2087 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2088 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2089 }
2090
2091
2092 //------------------------------uncommon_trap----------------------------------
2093 // Bail out to the interpreter in mid-method. Implemented by calling the
2094 // uncommon_trap blob. This helper function inserts a runtime call with the
2095 // right debug info.
2096 void GraphKit::uncommon_trap(int trap_request,
2097 ciKlass* klass, const char* comment,
2098 bool must_throw,
2099 bool keep_exact_action) {
2100 if (failing()) stop();
2101 if (stopped()) return; // trap reachable?
2102
2103 // Note: If ProfileTraps is true, and if a deopt. actually
2104 // occurs here, the runtime will make sure an MDO exists. There is
2105 // no need to call method()->ensure_method_data() at this point.
2106
2107 // Set the stack pointer to the right value for reexecution:
2108 set_sp(reexecute_sp());
2109
2110 #ifdef ASSERT
2111 if (!must_throw) {
2112 // Make sure the stack has at least enough depth to execute
2113 // the current bytecode.
2114 int inputs, ignored_depth;
2115 if (compute_stack_effects(inputs, ignored_depth)) {
2116 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2117 Bytecodes::name(java_bc()), sp(), inputs);
2118 }
2119 }
2120 #endif
2121
2122 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2123 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2124
2125 switch (action) {
2126 case Deoptimization::Action_maybe_recompile:
2127 case Deoptimization::Action_reinterpret:
2128 // Temporary fix for 6529811 to allow virtual calls to be sure they
2129 // get the chance to go from mono->bi->mega
2130 if (!keep_exact_action &&
2131 Deoptimization::trap_request_index(trap_request) < 0 &&
2132 too_many_recompiles(reason)) {
2133 // This BCI is causing too many recompilations.
2134 if (C->log() != NULL) {
2135 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2136 Deoptimization::trap_reason_name(reason),
2137 Deoptimization::trap_action_name(action));
2138 }
2139 action = Deoptimization::Action_none;
2140 trap_request = Deoptimization::make_trap_request(reason, action);
2141 } else {
2142 C->set_trap_can_recompile(true);
2143 }
2144 break;
2145 case Deoptimization::Action_make_not_entrant:
2146 C->set_trap_can_recompile(true);
2147 break;
2148 case Deoptimization::Action_none:
2149 case Deoptimization::Action_make_not_compilable:
2150 break;
2151 default:
2152 #ifdef ASSERT
2153 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2154 #endif
2155 break;
2156 }
2157
2158 if (TraceOptoParse) {
2159 char buf[100];
2160 tty->print_cr("Uncommon trap %s at bci:%d",
2161 Deoptimization::format_trap_request(buf, sizeof(buf),
2162 trap_request), bci());
2163 }
2164
2165 CompileLog* log = C->log();
2166 if (log != NULL) {
2167 int kid = (klass == NULL)? -1: log->identify(klass);
2168 log->begin_elem("uncommon_trap bci='%d'", bci());
2169 char buf[100];
2170 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2171 trap_request));
2172 if (kid >= 0) log->print(" klass='%d'", kid);
2173 if (comment != NULL) log->print(" comment='%s'", comment);
2174 log->end_elem();
2175 }
2176
2177 // Make sure any guarding test views this path as very unlikely
2178 Node *i0 = control()->in(0);
2179 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2180 IfNode *iff = i0->as_If();
2181 float f = iff->_prob; // Get prob
2182 if (control()->Opcode() == Op_IfTrue) {
2183 if (f > PROB_UNLIKELY_MAG(4))
2184 iff->_prob = PROB_MIN;
2185 } else {
2186 if (f < PROB_LIKELY_MAG(4))
2187 iff->_prob = PROB_MAX;
2188 }
2189 }
2190
2191 // Clear out dead values from the debug info.
2192 kill_dead_locals();
2193
2194 // Now insert the uncommon trap subroutine call
2195 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2196 const TypePtr* no_memory_effects = NULL;
2197 // Pass the index of the class to be loaded
2198 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2199 (must_throw ? RC_MUST_THROW : 0),
2200 OptoRuntime::uncommon_trap_Type(),
2201 call_addr, "uncommon_trap", no_memory_effects,
2202 intcon(trap_request));
2203 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2204 "must extract request correctly from the graph");
2205 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2206
2207 call->set_req(TypeFunc::ReturnAdr, returnadr());
2208 // The debug info is the only real input to this call.
2209
2210 // Halt-and-catch fire here. The above call should never return!
2211 HaltNode* halt = new HaltNode(control(), frameptr());
2212 _gvn.set_type_bottom(halt);
2213 root()->add_req(halt);
2214
2215 stop_and_kill_map();
2216 }
2217
2218
2219 //--------------------------just_allocated_object------------------------------
2220 // Report the object that was just allocated.
2221 // It must be the case that there are no intervening safepoints.
2222 // We use this to determine if an object is so "fresh" that
2223 // it does not require card marks.
2224 Node* GraphKit::just_allocated_object(Node* current_control) {
2225 Node* ctrl = current_control;
2226 // Object::<init> is invoked after allocation, most of invoke nodes
2227 // will be reduced, but a region node is kept in parse time, we check
2228 // the pattern and skip the region node if it degraded to a copy.
2229 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2230 ctrl->as_Region()->is_copy()) {
2231 ctrl = ctrl->as_Region()->is_copy();
2232 }
2233 if (C->recent_alloc_ctl() == ctrl) {
2234 return C->recent_alloc_obj();
2235 }
2236 return NULL;
2237 }
2238
2239
2240 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2241 // (Note: TypeFunc::make has a cache that makes this fast.)
2242 const TypeFunc* tf = TypeFunc::make(dest_method);
2243 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2244 for (int j = 0; j < nargs; j++) {
2245 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2246 if( targ->basic_type() == T_DOUBLE ) {
2247 // If any parameters are doubles, they must be rounded before
2248 // the call, dstore_rounding does gvn.transform
2249 Node *arg = argument(j);
2250 arg = dstore_rounding(arg);
2251 set_argument(j, arg);
2252 }
2253 }
2254 }
2255
2256 /**
2257 * Record profiling data exact_kls for Node n with the type system so
2258 * that it can propagate it (speculation)
2259 *
2260 * @param n node that the type applies to
2261 * @param exact_kls type from profiling
2262 * @param maybe_null did profiling see null?
2263 *
2264 * @return node with improved type
2265 */
2266 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2267 const Type* current_type = _gvn.type(n);
2268 assert(UseTypeSpeculation, "type speculation must be on");
2269
2270 const TypePtr* speculative = current_type->speculative();
2271
2272 // Should the klass from the profile be recorded in the speculative type?
2273 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2274 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls);
2275 const TypeOopPtr* xtype = tklass->as_instance_type();
2276 assert(xtype->klass_is_exact(), "Should be exact");
2277 // Any reason to believe n is not null (from this profiling or a previous one)?
2278 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2279 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2280 // record the new speculative type's depth
2281 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2282 speculative = speculative->with_inline_depth(jvms()->depth());
2283 } else if (current_type->would_improve_ptr(ptr_kind)) {
2284 // Profiling report that null was never seen so we can change the
2285 // speculative type to non null ptr.
2286 if (ptr_kind == ProfileAlwaysNull) {
2287 speculative = TypePtr::NULL_PTR;
2288 } else {
2289 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2290 const TypePtr* ptr = TypePtr::NOTNULL;
2291 if (speculative != NULL) {
2292 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2293 } else {
2294 speculative = ptr;
2295 }
2296 }
2297 }
2298
2299 if (speculative != current_type->speculative()) {
2300 // Build a type with a speculative type (what we think we know
2301 // about the type but will need a guard when we use it)
2302 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2303 // We're changing the type, we need a new CheckCast node to carry
2304 // the new type. The new type depends on the control: what
2305 // profiling tells us is only valid from here as far as we can
2306 // tell.
2307 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2308 cast = _gvn.transform(cast);
2309 replace_in_map(n, cast);
2310 n = cast;
2311 }
2312
2313 return n;
2314 }
2315
2316 /**
2317 * Record profiling data from receiver profiling at an invoke with the
2318 * type system so that it can propagate it (speculation)
2319 *
2320 * @param n receiver node
2321 *
2322 * @return node with improved type
2323 */
2324 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2325 if (!UseTypeSpeculation) {
2326 return n;
2327 }
2328 ciKlass* exact_kls = profile_has_unique_klass();
2329 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2330 if ((java_bc() == Bytecodes::_checkcast ||
2331 java_bc() == Bytecodes::_instanceof ||
2332 java_bc() == Bytecodes::_aastore) &&
2333 method()->method_data()->is_mature()) {
2334 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2335 if (data != NULL) {
2336 if (!data->as_BitData()->null_seen()) {
2337 ptr_kind = ProfileNeverNull;
2338 } else {
2339 assert(data->is_ReceiverTypeData(), "bad profile data type");
2340 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2341 uint i = 0;
2342 for (; i < call->row_limit(); i++) {
2343 ciKlass* receiver = call->receiver(i);
2344 if (receiver != NULL) {
2345 break;
2346 }
2347 }
2348 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2349 }
2350 }
2351 }
2352 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2353 }
2354
2355 /**
2356 * Record profiling data from argument profiling at an invoke with the
2357 * type system so that it can propagate it (speculation)
2358 *
2359 * @param dest_method target method for the call
2360 * @param bc what invoke bytecode is this?
2361 */
2362 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2363 if (!UseTypeSpeculation) {
2364 return;
2365 }
2366 const TypeFunc* tf = TypeFunc::make(dest_method);
2367 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2368 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2369 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2370 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2371 if (targ->isa_oopptr()) {
2372 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2373 ciKlass* better_type = NULL;
2374 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2375 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2376 }
2377 i++;
2378 }
2379 }
2380 }
2381
2382 /**
2383 * Record profiling data from parameter profiling at an invoke with
2384 * the type system so that it can propagate it (speculation)
2385 */
2386 void GraphKit::record_profiled_parameters_for_speculation() {
2387 if (!UseTypeSpeculation) {
2388 return;
2389 }
2390 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2391 if (_gvn.type(local(i))->isa_oopptr()) {
2392 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2393 ciKlass* better_type = NULL;
2394 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2395 record_profile_for_speculation(local(i), better_type, ptr_kind);
2396 }
2397 j++;
2398 }
2399 }
2400 }
2401
2402 /**
2403 * Record profiling data from return value profiling at an invoke with
2404 * the type system so that it can propagate it (speculation)
2405 */
2406 void GraphKit::record_profiled_return_for_speculation() {
2407 if (!UseTypeSpeculation) {
2408 return;
2409 }
2410 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2411 ciKlass* better_type = NULL;
2412 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2413 // If profiling reports a single type for the return value,
2414 // feed it to the type system so it can propagate it as a
2415 // speculative type
2416 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2417 }
2418 }
2419
2420 void GraphKit::round_double_result(ciMethod* dest_method) {
2421 // A non-strict method may return a double value which has an extended
2422 // exponent, but this must not be visible in a caller which is 'strict'
2423 // If a strict caller invokes a non-strict callee, round a double result
2424
2425 BasicType result_type = dest_method->return_type()->basic_type();
2426 assert( method() != NULL, "must have caller context");
2427 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2428 // Destination method's return value is on top of stack
2429 // dstore_rounding() does gvn.transform
2430 Node *result = pop_pair();
2431 result = dstore_rounding(result);
2432 push_pair(result);
2433 }
2434 }
2435
2436 // rounding for strict float precision conformance
2437 Node* GraphKit::precision_rounding(Node* n) {
2438 return UseStrictFP && _method->flags().is_strict()
2439 && UseSSE == 0 && Matcher::strict_fp_requires_explicit_rounding
2440 ? _gvn.transform( new RoundFloatNode(0, n) )
2441 : n;
2442 }
2443
2444 // rounding for strict double precision conformance
2445 Node* GraphKit::dprecision_rounding(Node *n) {
2446 return UseStrictFP && _method->flags().is_strict()
2447 && UseSSE <= 1 && Matcher::strict_fp_requires_explicit_rounding
2448 ? _gvn.transform( new RoundDoubleNode(0, n) )
2449 : n;
2450 }
2451
2452 // rounding for non-strict double stores
2453 Node* GraphKit::dstore_rounding(Node* n) {
2454 return Matcher::strict_fp_requires_explicit_rounding
2455 && UseSSE <= 1
2456 ? _gvn.transform( new RoundDoubleNode(0, n) )
2457 : n;
2458 }
2459
2460 //=============================================================================
2461 // Generate a fast path/slow path idiom. Graph looks like:
2462 // [foo] indicates that 'foo' is a parameter
2463 //
2464 // [in] NULL
2465 // \ /
2466 // CmpP
2467 // Bool ne
2468 // If
2469 // / \
2470 // True False-<2>
2471 // / |
2472 // / cast_not_null
2473 // Load | | ^
2474 // [fast_test] | |
2475 // gvn to opt_test | |
2476 // / \ | <1>
2477 // True False |
2478 // | \\ |
2479 // [slow_call] \[fast_result]
2480 // Ctl Val \ \
2481 // | \ \
2482 // Catch <1> \ \
2483 // / \ ^ \ \
2484 // Ex No_Ex | \ \
2485 // | \ \ | \ <2> \
2486 // ... \ [slow_res] | | \ [null_result]
2487 // \ \--+--+--- | |
2488 // \ | / \ | /
2489 // --------Region Phi
2490 //
2491 //=============================================================================
2492 // Code is structured as a series of driver functions all called 'do_XXX' that
2493 // call a set of helper functions. Helper functions first, then drivers.
2494
2495 //------------------------------null_check_oop---------------------------------
2496 // Null check oop. Set null-path control into Region in slot 3.
2497 // Make a cast-not-nullness use the other not-null control. Return cast.
2498 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2499 bool never_see_null,
2500 bool safe_for_replace,
2501 bool speculative) {
2502 // Initial NULL check taken path
2503 (*null_control) = top();
2504 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2505
2506 // Generate uncommon_trap:
2507 if (never_see_null && (*null_control) != top()) {
2508 // If we see an unexpected null at a check-cast we record it and force a
2509 // recompile; the offending check-cast will be compiled to handle NULLs.
2510 // If we see more than one offending BCI, then all checkcasts in the
2511 // method will be compiled to handle NULLs.
2512 PreserveJVMState pjvms(this);
2513 set_control(*null_control);
2514 replace_in_map(value, null());
2515 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2516 uncommon_trap(reason,
2517 Deoptimization::Action_make_not_entrant);
2518 (*null_control) = top(); // NULL path is dead
2519 }
2520 if ((*null_control) == top() && safe_for_replace) {
2521 replace_in_map(value, cast);
2522 }
2523
2524 // Cast away null-ness on the result
2525 return cast;
2526 }
2527
2528 //------------------------------opt_iff----------------------------------------
2529 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2530 // Return slow-path control.
2531 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2532 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2533
2534 // Fast path taken; set region slot 2
2535 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2536 region->init_req(2,fast_taken); // Capture fast-control
2537
2538 // Fast path not-taken, i.e. slow path
2539 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2540 return slow_taken;
2541 }
2542
2543 //-----------------------------make_runtime_call-------------------------------
2544 Node* GraphKit::make_runtime_call(int flags,
2545 const TypeFunc* call_type, address call_addr,
2546 const char* call_name,
2547 const TypePtr* adr_type,
2548 // The following parms are all optional.
2549 // The first NULL ends the list.
2550 Node* parm0, Node* parm1,
2551 Node* parm2, Node* parm3,
2552 Node* parm4, Node* parm5,
2553 Node* parm6, Node* parm7) {
2554 assert(call_addr != NULL, "must not call NULL targets");
2555
2556 // Slow-path call
2557 bool is_leaf = !(flags & RC_NO_LEAF);
2558 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2559 if (call_name == NULL) {
2560 assert(!is_leaf, "must supply name for leaf");
2561 call_name = OptoRuntime::stub_name(call_addr);
2562 }
2563 CallNode* call;
2564 if (!is_leaf) {
2565 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2566 bci(), adr_type);
2567 } else if (flags & RC_NO_FP) {
2568 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2569 } else {
2570 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2571 }
2572
2573 // The following is similar to set_edges_for_java_call,
2574 // except that the memory effects of the call are restricted to AliasIdxRaw.
2575
2576 // Slow path call has no side-effects, uses few values
2577 bool wide_in = !(flags & RC_NARROW_MEM);
2578 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2579
2580 Node* prev_mem = NULL;
2581 if (wide_in) {
2582 prev_mem = set_predefined_input_for_runtime_call(call);
2583 } else {
2584 assert(!wide_out, "narrow in => narrow out");
2585 Node* narrow_mem = memory(adr_type);
2586 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2587 }
2588
2589 // Hook each parm in order. Stop looking at the first NULL.
2590 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2591 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2592 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2593 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2594 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2595 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2596 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2597 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2598 /* close each nested if ===> */ } } } } } } } }
2599 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2600
2601 if (!is_leaf) {
2602 // Non-leaves can block and take safepoints:
2603 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2604 }
2605 // Non-leaves can throw exceptions:
2606 if (has_io) {
2607 call->set_req(TypeFunc::I_O, i_o());
2608 }
2609
2610 if (flags & RC_UNCOMMON) {
2611 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2612 // (An "if" probability corresponds roughly to an unconditional count.
2613 // Sort of.)
2614 call->set_cnt(PROB_UNLIKELY_MAG(4));
2615 }
2616
2617 Node* c = _gvn.transform(call);
2618 assert(c == call, "cannot disappear");
2619
2620 if (wide_out) {
2621 // Slow path call has full side-effects.
2622 set_predefined_output_for_runtime_call(call);
2623 } else {
2624 // Slow path call has few side-effects, and/or sets few values.
2625 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2626 }
2627
2628 if (has_io) {
2629 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2630 }
2631 return call;
2632
2633 }
2634
2635 //------------------------------merge_memory-----------------------------------
2636 // Merge memory from one path into the current memory state.
2637 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2638 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2639 Node* old_slice = mms.force_memory();
2640 Node* new_slice = mms.memory2();
2641 if (old_slice != new_slice) {
2642 PhiNode* phi;
2643 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2644 if (mms.is_empty()) {
2645 // clone base memory Phi's inputs for this memory slice
2646 assert(old_slice == mms.base_memory(), "sanity");
2647 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2648 _gvn.set_type(phi, Type::MEMORY);
2649 for (uint i = 1; i < phi->req(); i++) {
2650 phi->init_req(i, old_slice->in(i));
2651 }
2652 } else {
2653 phi = old_slice->as_Phi(); // Phi was generated already
2654 }
2655 } else {
2656 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2657 _gvn.set_type(phi, Type::MEMORY);
2658 }
2659 phi->set_req(new_path, new_slice);
2660 mms.set_memory(phi);
2661 }
2662 }
2663 }
2664
2665 //------------------------------make_slow_call_ex------------------------------
2666 // Make the exception handler hookups for the slow call
2667 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2668 if (stopped()) return;
2669
2670 // Make a catch node with just two handlers: fall-through and catch-all
2671 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2672 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2673 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2674 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2675
2676 { PreserveJVMState pjvms(this);
2677 set_control(excp);
2678 set_i_o(i_o);
2679
2680 if (excp != top()) {
2681 if (deoptimize) {
2682 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2683 uncommon_trap(Deoptimization::Reason_unhandled,
2684 Deoptimization::Action_none);
2685 } else {
2686 // Create an exception state also.
2687 // Use an exact type if the caller has a specific exception.
2688 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2689 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2690 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2691 }
2692 }
2693 }
2694
2695 // Get the no-exception control from the CatchNode.
2696 set_control(norm);
2697 }
2698
2699 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN* gvn, BasicType bt) {
2700 Node* cmp = NULL;
2701 switch(bt) {
2702 case T_INT: cmp = new CmpINode(in1, in2); break;
2703 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2704 default: fatal("unexpected comparison type %s", type2name(bt));
2705 }
2706 gvn->transform(cmp);
2707 Node* bol = gvn->transform(new BoolNode(cmp, test));
2708 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2709 gvn->transform(iff);
2710 if (!bol->is_Con()) gvn->record_for_igvn(iff);
2711 return iff;
2712 }
2713
2714
2715 //-------------------------------gen_subtype_check-----------------------------
2716 // Generate a subtyping check. Takes as input the subtype and supertype.
2717 // Returns 2 values: sets the default control() to the true path and returns
2718 // the false path. Only reads invariant memory; sets no (visible) memory.
2719 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2720 // but that's not exposed to the optimizer. This call also doesn't take in an
2721 // Object; if you wish to check an Object you need to load the Object's class
2722 // prior to coming here.
2723 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, MergeMemNode* mem, PhaseGVN* gvn) {
2724 Compile* C = gvn->C;
2725
2726 if ((*ctrl)->is_top()) {
2727 return C->top();
2728 }
2729
2730 // Fast check for identical types, perhaps identical constants.
2731 // The types can even be identical non-constants, in cases
2732 // involving Array.newInstance, Object.clone, etc.
2733 if (subklass == superklass)
2734 return C->top(); // false path is dead; no test needed.
2735
2736 if (gvn->type(superklass)->singleton()) {
2737 ciKlass* superk = gvn->type(superklass)->is_klassptr()->klass();
2738 ciKlass* subk = gvn->type(subklass)->is_klassptr()->klass();
2739
2740 // In the common case of an exact superklass, try to fold up the
2741 // test before generating code. You may ask, why not just generate
2742 // the code and then let it fold up? The answer is that the generated
2743 // code will necessarily include null checks, which do not always
2744 // completely fold away. If they are also needless, then they turn
2745 // into a performance loss. Example:
2746 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2747 // Here, the type of 'fa' is often exact, so the store check
2748 // of fa[1]=x will fold up, without testing the nullness of x.
2749 switch (C->static_subtype_check(superk, subk)) {
2750 case Compile::SSC_always_false:
2751 {
2752 Node* always_fail = *ctrl;
2753 *ctrl = gvn->C->top();
2754 return always_fail;
2755 }
2756 case Compile::SSC_always_true:
2757 return C->top();
2758 case Compile::SSC_easy_test:
2759 {
2760 // Just do a direct pointer compare and be done.
2761 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2762 *ctrl = gvn->transform(new IfTrueNode(iff));
2763 return gvn->transform(new IfFalseNode(iff));
2764 }
2765 case Compile::SSC_full_test:
2766 break;
2767 default:
2768 ShouldNotReachHere();
2769 }
2770 }
2771
2772 // %%% Possible further optimization: Even if the superklass is not exact,
2773 // if the subklass is the unique subtype of the superklass, the check
2774 // will always succeed. We could leave a dependency behind to ensure this.
2775
2776 // First load the super-klass's check-offset
2777 Node *p1 = gvn->transform(new AddPNode(superklass, superklass, gvn->MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2778 Node* m = mem->memory_at(C->get_alias_index(gvn->type(p1)->is_ptr()));
2779 Node *chk_off = gvn->transform(new LoadINode(NULL, m, p1, gvn->type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2780 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2781 bool might_be_cache = (gvn->find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2782
2783 // Load from the sub-klass's super-class display list, or a 1-word cache of
2784 // the secondary superclass list, or a failing value with a sentinel offset
2785 // if the super-klass is an interface or exceptionally deep in the Java
2786 // hierarchy and we have to scan the secondary superclass list the hard way.
2787 // Worst-case type is a little odd: NULL is allowed as a result (usually
2788 // klass loads can never produce a NULL).
2789 Node *chk_off_X = chk_off;
2790 #ifdef _LP64
2791 chk_off_X = gvn->transform(new ConvI2LNode(chk_off_X));
2792 #endif
2793 Node *p2 = gvn->transform(new AddPNode(subklass,subklass,chk_off_X));
2794 // For some types like interfaces the following loadKlass is from a 1-word
2795 // cache which is mutable so can't use immutable memory. Other
2796 // types load from the super-class display table which is immutable.
2797 m = mem->memory_at(C->get_alias_index(gvn->type(p2)->is_ptr()));
2798 Node *kmem = might_be_cache ? m : C->immutable_memory();
2799 Node *nkls = gvn->transform(LoadKlassNode::make(*gvn, NULL, kmem, p2, gvn->type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2800
2801 // Compile speed common case: ARE a subtype and we canNOT fail
2802 if( superklass == nkls )
2803 return C->top(); // false path is dead; no test needed.
2804
2805 // See if we get an immediate positive hit. Happens roughly 83% of the
2806 // time. Test to see if the value loaded just previously from the subklass
2807 // is exactly the superklass.
2808 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2809 Node *iftrue1 = gvn->transform( new IfTrueNode (iff1));
2810 *ctrl = gvn->transform(new IfFalseNode(iff1));
2811
2812 // Compile speed common case: Check for being deterministic right now. If
2813 // chk_off is a constant and not equal to cacheoff then we are NOT a
2814 // subklass. In this case we need exactly the 1 test above and we can
2815 // return those results immediately.
2816 if (!might_be_cache) {
2817 Node* not_subtype_ctrl = *ctrl;
2818 *ctrl = iftrue1; // We need exactly the 1 test above
2819 return not_subtype_ctrl;
2820 }
2821
2822 // Gather the various success & failures here
2823 RegionNode *r_ok_subtype = new RegionNode(4);
2824 gvn->record_for_igvn(r_ok_subtype);
2825 RegionNode *r_not_subtype = new RegionNode(3);
2826 gvn->record_for_igvn(r_not_subtype);
2827
2828 r_ok_subtype->init_req(1, iftrue1);
2829
2830 // Check for immediate negative hit. Happens roughly 11% of the time (which
2831 // is roughly 63% of the remaining cases). Test to see if the loaded
2832 // check-offset points into the subklass display list or the 1-element
2833 // cache. If it points to the display (and NOT the cache) and the display
2834 // missed then it's not a subtype.
2835 Node *cacheoff = gvn->intcon(cacheoff_con);
2836 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2837 r_not_subtype->init_req(1, gvn->transform(new IfTrueNode (iff2)));
2838 *ctrl = gvn->transform(new IfFalseNode(iff2));
2839
2840 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2841 // No performance impact (too rare) but allows sharing of secondary arrays
2842 // which has some footprint reduction.
2843 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2844 r_ok_subtype->init_req(2, gvn->transform(new IfTrueNode(iff3)));
2845 *ctrl = gvn->transform(new IfFalseNode(iff3));
2846
2847 // -- Roads not taken here: --
2848 // We could also have chosen to perform the self-check at the beginning
2849 // of this code sequence, as the assembler does. This would not pay off
2850 // the same way, since the optimizer, unlike the assembler, can perform
2851 // static type analysis to fold away many successful self-checks.
2852 // Non-foldable self checks work better here in second position, because
2853 // the initial primary superclass check subsumes a self-check for most
2854 // types. An exception would be a secondary type like array-of-interface,
2855 // which does not appear in its own primary supertype display.
2856 // Finally, we could have chosen to move the self-check into the
2857 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2858 // dependent manner. But it is worthwhile to have the check here,
2859 // where it can be perhaps be optimized. The cost in code space is
2860 // small (register compare, branch).
2861
2862 // Now do a linear scan of the secondary super-klass array. Again, no real
2863 // performance impact (too rare) but it's gotta be done.
2864 // Since the code is rarely used, there is no penalty for moving it
2865 // out of line, and it can only improve I-cache density.
2866 // The decision to inline or out-of-line this final check is platform
2867 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2868 Node* psc = gvn->transform(
2869 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2870
2871 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn->zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2872 r_not_subtype->init_req(2, gvn->transform(new IfTrueNode (iff4)));
2873 r_ok_subtype ->init_req(3, gvn->transform(new IfFalseNode(iff4)));
2874
2875 // Return false path; set default control to true path.
2876 *ctrl = gvn->transform(r_ok_subtype);
2877 return gvn->transform(r_not_subtype);
2878 }
2879
2880 // Profile-driven exact type check:
2881 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2882 float prob,
2883 Node* *casted_receiver) {
2884 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2885 Node* recv_klass = load_object_klass(receiver);
2886 Node* fail = type_check(recv_klass, tklass, prob);
2887 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2888 assert(recv_xtype->klass_is_exact(), "");
2889
2890 // Subsume downstream occurrences of receiver with a cast to
2891 // recv_xtype, since now we know what the type will be.
2892 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2893 Node* res = _gvn.transform(cast);
2894 if (recv_xtype->is_valuetypeptr() && recv_xtype->value_klass()->is_scalarizable()) {
2895 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
2896 res = ValueTypeNode::make_from_oop(this, res, recv_xtype->value_klass());
2897 }
2898
2899 (*casted_receiver) = res;
2900 // (User must make the replace_in_map call.)
2901
2902 return fail;
2903 }
2904
2905 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
2906 float prob) {
2907 Node* want_klass = makecon(tklass);
2908 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass));
2909 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2910 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2911 set_control( _gvn.transform( new IfTrueNode (iff)));
2912 Node* fail = _gvn.transform( new IfFalseNode(iff));
2913 return fail;
2914 }
2915
2916 //------------------------------subtype_check_receiver-------------------------
2917 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
2918 Node** casted_receiver) {
2919 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2920 Node* recv_klass = load_object_klass(receiver);
2921 Node* want_klass = makecon(tklass);
2922
2923 Node* slow_ctl = gen_subtype_check(recv_klass, want_klass);
2924
2925 // Cast receiver after successful check
2926 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
2927 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
2928 (*casted_receiver) = _gvn.transform(cast);
2929
2930 return slow_ctl;
2931 }
2932
2933 //------------------------------seems_never_null-------------------------------
2934 // Use null_seen information if it is available from the profile.
2935 // If we see an unexpected null at a type check we record it and force a
2936 // recompile; the offending check will be recompiled to handle NULLs.
2937 // If we see several offending BCIs, then all checks in the
2938 // method will be recompiled.
2939 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
2940 speculating = !_gvn.type(obj)->speculative_maybe_null();
2941 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
2942 if (UncommonNullCast // Cutout for this technique
2943 && obj != null() // And not the -Xcomp stupid case?
2944 && !too_many_traps(reason)
2945 ) {
2946 if (speculating) {
2947 return true;
2948 }
2949 if (data == NULL)
2950 // Edge case: no mature data. Be optimistic here.
2951 return true;
2952 // If the profile has not seen a null, assume it won't happen.
2953 assert(java_bc() == Bytecodes::_checkcast ||
2954 java_bc() == Bytecodes::_instanceof ||
2955 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
2956 return !data->as_BitData()->null_seen();
2957 }
2958 speculating = false;
2959 return false;
2960 }
2961
2962 void GraphKit::guard_klass_being_initialized(Node* klass) {
2963 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
2964 Node* adr = basic_plus_adr(top(), klass, init_state_off);
2965 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
2966 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
2967 T_BYTE, MemNode::unordered);
2968 init_state = _gvn.transform(init_state);
2969
2970 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
2971
2972 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
2973 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
2974
2975 { BuildCutout unless(this, tst, PROB_MAX);
2976 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
2977 }
2978 }
2979
2980 void GraphKit::guard_init_thread(Node* klass) {
2981 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
2982 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
2983
2984 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
2985 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
2986 T_ADDRESS, MemNode::unordered);
2987 init_thread = _gvn.transform(init_thread);
2988
2989 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
2990
2991 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
2992 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
2993
2994 { BuildCutout unless(this, tst, PROB_MAX);
2995 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
2996 }
2997 }
2998
2999 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3000 if (ik->is_being_initialized()) {
3001 if (C->needs_clinit_barrier(ik, context)) {
3002 Node* klass = makecon(TypeKlassPtr::make(ik));
3003 guard_klass_being_initialized(klass);
3004 guard_init_thread(klass);
3005 insert_mem_bar(Op_MemBarCPUOrder);
3006 }
3007 } else if (ik->is_initialized()) {
3008 return; // no barrier needed
3009 } else {
3010 uncommon_trap(Deoptimization::Reason_uninitialized,
3011 Deoptimization::Action_reinterpret,
3012 NULL);
3013 }
3014 }
3015
3016 //------------------------maybe_cast_profiled_receiver-------------------------
3017 // If the profile has seen exactly one type, narrow to exactly that type.
3018 // Subsequent type checks will always fold up.
3019 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3020 ciKlass* require_klass,
3021 ciKlass* spec_klass,
3022 bool safe_for_replace) {
3023 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
3024
3025 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
3026
3027 // Make sure we haven't already deoptimized from this tactic.
3028 if (too_many_traps_or_recompiles(reason))
3029 return NULL;
3030
3031 // (No, this isn't a call, but it's enough like a virtual call
3032 // to use the same ciMethod accessor to get the profile info...)
3033 // If we have a speculative type use it instead of profiling (which
3034 // may not help us)
3035 ciKlass* exact_kls = spec_klass == NULL ? profile_has_unique_klass() : spec_klass;
3036 if (exact_kls != NULL) {// no cast failures here
3037 if (require_klass == NULL ||
3038 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
3039 // If we narrow the type to match what the type profile sees or
3040 // the speculative type, we can then remove the rest of the
3041 // cast.
3042 // This is a win, even if the exact_kls is very specific,
3043 // because downstream operations, such as method calls,
3044 // will often benefit from the sharper type.
3045 Node* exact_obj = not_null_obj; // will get updated in place...
3046 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3047 &exact_obj);
3048 { PreserveJVMState pjvms(this);
3049 set_control(slow_ctl);
3050 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3051 }
3052 if (safe_for_replace) {
3053 replace_in_map(not_null_obj, exact_obj);
3054 }
3055 return exact_obj;
3056 }
3057 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3058 }
3059
3060 return NULL;
3061 }
3062
3063 /**
3064 * Cast obj to type and emit guard unless we had too many traps here
3065 * already
3066 *
3067 * @param obj node being casted
3068 * @param type type to cast the node to
3069 * @param not_null true if we know node cannot be null
3070 */
3071 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3072 ciKlass* type,
3073 bool not_null) {
3074 if (stopped()) {
3075 return obj;
3076 }
3077
3078 // type == NULL if profiling tells us this object is always null
3079 if (type != NULL) {
3080 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3081 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3082
3083 if (!too_many_traps_or_recompiles(null_reason) &&
3084 !too_many_traps_or_recompiles(class_reason)) {
3085 Node* not_null_obj = NULL;
3086 // not_null is true if we know the object is not null and
3087 // there's no need for a null check
3088 if (!not_null) {
3089 Node* null_ctl = top();
3090 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3091 assert(null_ctl->is_top(), "no null control here");
3092 } else {
3093 not_null_obj = obj;
3094 }
3095
3096 Node* exact_obj = not_null_obj;
3097 ciKlass* exact_kls = type;
3098 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3099 &exact_obj);
3100 {
3101 PreserveJVMState pjvms(this);
3102 set_control(slow_ctl);
3103 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3104 }
3105 replace_in_map(not_null_obj, exact_obj);
3106 obj = exact_obj;
3107 }
3108 } else {
3109 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3110 Node* exact_obj = null_assert(obj);
3111 replace_in_map(obj, exact_obj);
3112 obj = exact_obj;
3113 }
3114 }
3115 return obj;
3116 }
3117
3118 //-------------------------------gen_instanceof--------------------------------
3119 // Generate an instance-of idiom. Used by both the instance-of bytecode
3120 // and the reflective instance-of call.
3121 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3122 kill_dead_locals(); // Benefit all the uncommon traps
3123 assert( !stopped(), "dead parse path should be checked in callers" );
3124 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3125 "must check for not-null not-dead klass in callers");
3126
3127 // Make the merge point
3128 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3129 RegionNode* region = new RegionNode(PATH_LIMIT);
3130 Node* phi = new PhiNode(region, TypeInt::BOOL);
3131 C->set_has_split_ifs(true); // Has chance for split-if optimization
3132
3133 ciProfileData* data = NULL;
3134 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3135 data = method()->method_data()->bci_to_data(bci());
3136 }
3137 bool speculative_not_null = false;
3138 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3139 && seems_never_null(obj, data, speculative_not_null));
3140 bool is_value = obj->is_ValueType();
3141
3142 // Null check; get casted pointer; set region slot 3
3143 Node* null_ctl = top();
3144 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3145
3146 // If not_null_obj is dead, only null-path is taken
3147 if (stopped()) { // Doing instance-of on a NULL?
3148 set_control(null_ctl);
3149 return intcon(0);
3150 }
3151 region->init_req(_null_path, null_ctl);
3152 phi ->init_req(_null_path, intcon(0)); // Set null path value
3153 if (null_ctl == top()) {
3154 // Do this eagerly, so that pattern matches like is_diamond_phi
3155 // will work even during parsing.
3156 assert(_null_path == PATH_LIMIT-1, "delete last");
3157 region->del_req(_null_path);
3158 phi ->del_req(_null_path);
3159 }
3160
3161 // Do we know the type check always succeed?
3162 if (!is_value) {
3163 bool known_statically = false;
3164 if (_gvn.type(superklass)->singleton()) {
3165 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3166 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3167 if (subk != NULL && subk->is_loaded()) {
3168 int static_res = C->static_subtype_check(superk, subk);
3169 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3170 }
3171 }
3172
3173 if (!known_statically) {
3174 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3175 // We may not have profiling here or it may not help us. If we
3176 // have a speculative type use it to perform an exact cast.
3177 ciKlass* spec_obj_type = obj_type->speculative_type();
3178 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3179 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3180 if (stopped()) { // Profile disagrees with this path.
3181 set_control(null_ctl); // Null is the only remaining possibility.
3182 return intcon(0);
3183 }
3184 if (cast_obj != NULL &&
3185 // A value that's sometimes null is not something we can optimize well
3186 !(cast_obj->is_ValueType() && null_ctl != top())) {
3187 not_null_obj = cast_obj;
3188 is_value = not_null_obj->is_ValueType();
3189 }
3190 }
3191 }
3192 }
3193
3194 // Load the object's klass
3195 Node* obj_klass = NULL;
3196 if (is_value) {
3197 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->value_klass()));
3198 } else {
3199 obj_klass = load_object_klass(not_null_obj);
3200 }
3201
3202 // Generate the subtype check
3203 Node* not_subtype_ctrl = gen_subtype_check(obj_klass, superklass);
3204
3205 // Plug in the success path to the general merge in slot 1.
3206 region->init_req(_obj_path, control());
3207 phi ->init_req(_obj_path, intcon(1));
3208
3209 // Plug in the failing path to the general merge in slot 2.
3210 region->init_req(_fail_path, not_subtype_ctrl);
3211 phi ->init_req(_fail_path, intcon(0));
3212
3213 // Return final merged results
3214 set_control( _gvn.transform(region) );
3215 record_for_igvn(region);
3216
3217 // If we know the type check always succeeds then we don't use the
3218 // profiling data at this bytecode. Don't lose it, feed it to the
3219 // type system as a speculative type.
3220 if (safe_for_replace && !is_value) {
3221 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3222 replace_in_map(obj, casted_obj);
3223 }
3224
3225 return _gvn.transform(phi);
3226 }
3227
3228 //-------------------------------gen_checkcast---------------------------------
3229 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3230 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3231 // uncommon-trap paths work. Adjust stack after this call.
3232 // If failure_control is supplied and not null, it is filled in with
3233 // the control edge for the cast failure. Otherwise, an appropriate
3234 // uncommon trap or exception is thrown.
3235 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool never_null) {
3236 kill_dead_locals(); // Benefit all the uncommon traps
3237 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3238 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass());
3239 assert(!never_null || toop->is_valuetypeptr(), "must be a value type pointer");
3240 bool is_value = obj->is_ValueType();
3241
3242 // Fast cutout: Check the case that the cast is vacuously true.
3243 // This detects the common cases where the test will short-circuit
3244 // away completely. We do this before we perform the null check,
3245 // because if the test is going to turn into zero code, we don't
3246 // want a residual null check left around. (Causes a slowdown,
3247 // for example, in some objArray manipulations, such as a[i]=a[j].)
3248 if (tk->singleton()) {
3249 ciKlass* klass = NULL;
3250 if (is_value) {
3251 klass = _gvn.type(obj)->value_klass();
3252 } else {
3253 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3254 if (objtp != NULL) {
3255 klass = objtp->klass();
3256 }
3257 }
3258 if (klass != NULL) {
3259 switch (C->static_subtype_check(tk->klass(), klass)) {
3260 case Compile::SSC_always_true:
3261 // If we know the type check always succeed then we don't use
3262 // the profiling data at this bytecode. Don't lose it, feed it
3263 // to the type system as a speculative type.
3264 if (!is_value) {
3265 obj = record_profiled_receiver_for_speculation(obj);
3266 if (never_null) {
3267 obj = null_check(obj);
3268 }
3269 if (toop->is_valuetypeptr() && toop->value_klass()->is_scalarizable() && !gvn().type(obj)->maybe_null()) {
3270 obj = ValueTypeNode::make_from_oop(this, obj, toop->value_klass());
3271 }
3272 }
3273 return obj;
3274 case Compile::SSC_always_false:
3275 if (is_value || never_null) {
3276 if (!is_value) {
3277 null_check(obj);
3278 }
3279 // Value type is never null. Always throw an exception.
3280 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(klass)));
3281 return top();
3282 } else {
3283 // It needs a null check because a null will *pass* the cast check.
3284 return null_assert(obj);
3285 }
3286 }
3287 }
3288 }
3289
3290 ciProfileData* data = NULL;
3291 bool safe_for_replace = false;
3292 if (failure_control == NULL) { // use MDO in regular case only
3293 assert(java_bc() == Bytecodes::_aastore ||
3294 java_bc() == Bytecodes::_checkcast,
3295 "interpreter profiles type checks only for these BCs");
3296 data = method()->method_data()->bci_to_data(bci());
3297 safe_for_replace = true;
3298 }
3299
3300 // Make the merge point
3301 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3302 RegionNode* region = new RegionNode(PATH_LIMIT);
3303 Node* phi = new PhiNode(region, toop);
3304 _gvn.set_type(region, Type::CONTROL);
3305 _gvn.set_type(phi, toop);
3306
3307 C->set_has_split_ifs(true); // Has chance for split-if optimization
3308
3309 // Use null-cast information if it is available
3310 bool speculative_not_null = false;
3311 bool never_see_null = ((failure_control == NULL) // regular case only
3312 && seems_never_null(obj, data, speculative_not_null));
3313
3314 // Null check; get casted pointer; set region slot 3
3315 Node* null_ctl = top();
3316 Node* not_null_obj = NULL;
3317 if (is_value) {
3318 not_null_obj = obj;
3319 } else if (never_null) {
3320 not_null_obj = null_check(obj);
3321 } else {
3322 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3323 }
3324
3325 // If not_null_obj is dead, only null-path is taken
3326 if (stopped()) { // Doing instance-of on a NULL?
3327 set_control(null_ctl);
3328 return null();
3329 }
3330 region->init_req(_null_path, null_ctl);
3331 phi ->init_req(_null_path, null()); // Set null path value
3332 if (null_ctl == top()) {
3333 // Do this eagerly, so that pattern matches like is_diamond_phi
3334 // will work even during parsing.
3335 assert(_null_path == PATH_LIMIT-1, "delete last");
3336 region->del_req(_null_path);
3337 phi ->del_req(_null_path);
3338 }
3339
3340 Node* cast_obj = NULL;
3341 if (!is_value && tk->klass_is_exact()) {
3342 // The following optimization tries to statically cast the speculative type of the object
3343 // (for example obtained during profiling) to the type of the superklass and then do a
3344 // dynamic check that the type of the object is what we expect. To work correctly
3345 // for checkcast and aastore the type of superklass should be exact.
3346 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3347 // We may not have profiling here or it may not help us. If we have
3348 // a speculative type use it to perform an exact cast.
3349 ciKlass* spec_obj_type = obj_type->speculative_type();
3350 if (spec_obj_type != NULL || data != NULL) {
3351 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3352 if (cast_obj != NULL && cast_obj->is_ValueType()) {
3353 if (null_ctl != top()) {
3354 cast_obj = NULL; // A value that's sometimes null is not something we can optimize well
3355 } else {
3356 return cast_obj;
3357 }
3358 }
3359 if (cast_obj != NULL) {
3360 if (failure_control != NULL) // failure is now impossible
3361 (*failure_control) = top();
3362 // adjust the type of the phi to the exact klass:
3363 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3364 }
3365 }
3366 }
3367
3368 if (cast_obj == NULL) {
3369 // Load the object's klass
3370 Node* obj_klass = NULL;
3371 if (is_value) {
3372 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->value_klass()));
3373 } else {
3374 obj_klass = load_object_klass(not_null_obj);
3375 }
3376
3377 // Generate the subtype check
3378 Node* not_subtype_ctrl = gen_subtype_check( obj_klass, superklass );
3379
3380 // Plug in success path into the merge
3381 cast_obj = is_value ? not_null_obj : _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3382 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3383 if (failure_control == NULL) {
3384 if (not_subtype_ctrl != top()) { // If failure is possible
3385 PreserveJVMState pjvms(this);
3386 set_control(not_subtype_ctrl);
3387 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3388 }
3389 } else {
3390 (*failure_control) = not_subtype_ctrl;
3391 }
3392 }
3393
3394 region->init_req(_obj_path, control());
3395 phi ->init_req(_obj_path, cast_obj);
3396
3397 // A merge of NULL or Casted-NotNull obj
3398 Node* res = _gvn.transform(phi);
3399
3400 // Note I do NOT always 'replace_in_map(obj,result)' here.
3401 // if( tk->klass()->can_be_primary_super() )
3402 // This means that if I successfully store an Object into an array-of-String
3403 // I 'forget' that the Object is really now known to be a String. I have to
3404 // do this because we don't have true union types for interfaces - if I store
3405 // a Baz into an array-of-Interface and then tell the optimizer it's an
3406 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3407 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3408 // replace_in_map( obj, res );
3409
3410 // Return final merged results
3411 set_control( _gvn.transform(region) );
3412 record_for_igvn(region);
3413
3414 bool not_null_free = !toop->can_be_value_type();
3415 bool not_flattenable = !ValueArrayFlatten || not_null_free || (toop->is_valuetypeptr() && !toop->value_klass()->flatten_array());
3416 if (EnableValhalla && not_flattenable) {
3417 // Check if obj has been loaded from an array
3418 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3419 Node* array = NULL;
3420 if (obj->isa_Load()) {
3421 Node* address = obj->in(MemNode::Address);
3422 if (address->isa_AddP()) {
3423 array = address->as_AddP()->in(AddPNode::Base);
3424 }
3425 } else if (obj->is_Phi()) {
3426 Node* region = obj->in(0);
3427 if (region->req() == 3 && region->in(1) != NULL) {
3428 IfNode* iff = region->in(1)->in(0)->isa_If();
3429 if (iff != NULL) {
3430 iff->is_flattened_array_check(&_gvn, array);
3431 }
3432 }
3433 }
3434 if (array != NULL) {
3435 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3436 if (ary_t != NULL) {
3437 if (!ary_t->is_not_null_free() && not_null_free) {
3438 // Casting array element to a non-inline-type, mark array as not null-free.
3439 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3440 replace_in_map(array, cast);
3441 } else if (!ary_t->is_not_flat()) {
3442 // Casting array element to a non-flattenable type, mark array as not flat.
3443 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3444 replace_in_map(array, cast);
3445 }
3446 }
3447 }
3448 }
3449
3450 if (!is_value) {
3451 res = record_profiled_receiver_for_speculation(res);
3452 if (toop->is_valuetypeptr() && toop->value_klass()->is_scalarizable() && !gvn().type(res)->maybe_null()) {
3453 res = ValueTypeNode::make_from_oop(this, res, toop->value_klass());
3454 }
3455 }
3456 return res;
3457 }
3458
3459 Node* GraphKit::is_always_locked(Node* obj) {
3460 Node* mark_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3461 Node* mark = make_load(NULL, mark_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3462 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern);
3463 return _gvn.transform(new AndXNode(mark, value_mask));
3464 }
3465
3466 Node* GraphKit::is_value_mirror(Node* mirror) {
3467 Node* p = basic_plus_adr(mirror, java_lang_Class::inline_mirror_offset_in_bytes());
3468 Node* inline_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR), T_OBJECT, IN_HEAP);
3469 Node* cmp = _gvn.transform(new CmpPNode(mirror, inline_mirror));
3470 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3471 }
3472
3473 // Deoptimize if 'obj' is a value type
3474 void GraphKit::gen_value_type_guard(Node* obj, int nargs) {
3475 assert(EnableValhalla, "should only be used if value types are enabled");
3476 Node* bol = NULL;
3477 if (obj->is_ValueTypeBase()) {
3478 bol = intcon(0);
3479 } else {
3480 Node* is_value = is_always_locked(obj);
3481 Node* value_mask = _gvn.MakeConX(markOopDesc::always_locked_pattern);
3482 Node* cmp = _gvn.transform(new CmpXNode(is_value, value_mask));
3483 bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3484 }
3485 { BuildCutout unless(this, bol, PROB_MAX);
3486 inc_sp(nargs);
3487 uncommon_trap(Deoptimization::Reason_class_check,
3488 Deoptimization::Action_none);
3489 }
3490 }
3491
3492 // Check if 'ary' is a null-free value type array
3493 Node* GraphKit::gen_null_free_array_check(Node* ary) {
3494 assert(EnableValhalla, "should only be used if value types are enabled");
3495 // Extract null free property from klass pointer
3496 Node* k_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
3497 const TypePtr* k_adr_type = k_adr->bottom_type()->isa_ptr();
3498 Node* klass = NULL;
3499 if (k_adr_type->is_ptr_to_narrowklass()) {
3500 klass = _gvn.transform(new LoadNKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT->make_narrowklass(), MemNode::unordered));
3501 } else {
3502 klass = _gvn.transform(new LoadKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT, MemNode::unordered));
3503 }
3504 Node* null_free = _gvn.transform(new GetNullFreePropertyNode(klass));
3505 Node* cmp = NULL;
3506 if (_gvn.type(klass)->isa_klassptr()) {
3507 cmp = _gvn.transform(new CmpLNode(null_free, zerocon(T_LONG)));
3508 } else {
3509 cmp = _gvn.transform(new CmpINode(null_free, zerocon(T_INT)));
3510 }
3511 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3512 }
3513
3514 Node* GraphKit::gen_flattened_array_test(Node* ary) {
3515 assert(EnableValhalla, "should only be used if value types are enabled");
3516 // Extract flattened property from klass pointer
3517 Node* k_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
3518 const TypePtr* k_adr_type = k_adr->bottom_type()->isa_ptr();
3519 Node* klass = NULL;
3520 if (k_adr_type->is_ptr_to_narrowklass()) {
3521 klass = _gvn.transform(new LoadNKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT->make_narrowklass(), MemNode::unordered));
3522 } else {
3523 klass = _gvn.transform(new LoadKlassNode(NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT, MemNode::unordered));
3524 }
3525 return _gvn.transform(new GetFlattenedPropertyNode(klass));
3526 }
3527
3528 // Deoptimize if 'ary' is a null-free value type array and 'val' is null
3529 Node* GraphKit::gen_value_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3530 const Type* val_t = _gvn.type(val);
3531 if (val->is_ValueType() || !TypePtr::NULL_PTR->higher_equal(val_t)) {
3532 return ary; // Never null
3533 }
3534 RegionNode* region = new RegionNode(3);
3535 Node* null_ctl = top();
3536 null_check_oop(val, &null_ctl);
3537 if (null_ctl != top()) {
3538 PreserveJVMState pjvms(this);
3539 set_control(null_ctl);
3540 // Deoptimize if null-free array
3541 Node* bol = gen_null_free_array_check(ary);
3542 { BuildCutout unless(this, bol, PROB_MAX);
3543 inc_sp(nargs);
3544 uncommon_trap(Deoptimization::Reason_null_check,
3545 Deoptimization::Action_none);
3546 }
3547 region->init_req(1, control());
3548 }
3549 region->init_req(2, control());
3550 set_control(_gvn.transform(region));
3551 record_for_igvn(region);
3552 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3553 if (val_t == TypePtr::NULL_PTR && !ary_t->is_not_null_free()) {
3554 // Since we were just successfully storing null, the array can't be null free.
3555 ary_t = ary_t->cast_to_not_null_free();
3556 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3557 if (safe_for_replace) {
3558 replace_in_map(ary, cast);
3559 }
3560 ary = cast;
3561 }
3562 return ary;
3563 }
3564
3565 Node* GraphKit::load_lh_array_tag(Node* kls) {
3566 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset()));
3567 Node* layout_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lhp, lhp->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3568
3569 return _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3570 }
3571
3572
3573 Node* GraphKit::gen_lh_array_test(Node* kls, unsigned int lh_value) {
3574 Node* layout_val = load_lh_array_tag(kls);
3575 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(lh_value)));
3576 return cmp;
3577 }
3578
3579
3580 //------------------------------next_monitor-----------------------------------
3581 // What number should be given to the next monitor?
3582 int GraphKit::next_monitor() {
3583 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3584 int next = current + C->sync_stack_slots();
3585 // Keep the toplevel high water mark current:
3586 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3587 return current;
3588 }
3589
3590 //------------------------------insert_mem_bar---------------------------------
3591 // Memory barrier to avoid floating things around
3592 // The membar serves as a pinch point between both control and all memory slices.
3593 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3594 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3595 mb->init_req(TypeFunc::Control, control());
3596 mb->init_req(TypeFunc::Memory, reset_memory());
3597 Node* membar = _gvn.transform(mb);
3598 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3599 set_all_memory_call(membar);
3600 return membar;
3601 }
3602
3603 //-------------------------insert_mem_bar_volatile----------------------------
3604 // Memory barrier to avoid floating things around
3605 // The membar serves as a pinch point between both control and memory(alias_idx).
3606 // If you want to make a pinch point on all memory slices, do not use this
3607 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3608 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3609 // When Parse::do_put_xxx updates a volatile field, it appends a series
3610 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3611 // The first membar is on the same memory slice as the field store opcode.
3612 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3613 // All the other membars (for other volatile slices, including AliasIdxBot,
3614 // which stands for all unknown volatile slices) are control-dependent
3615 // on the first membar. This prevents later volatile loads or stores
3616 // from sliding up past the just-emitted store.
3617
3618 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3619 mb->set_req(TypeFunc::Control,control());
3620 if (alias_idx == Compile::AliasIdxBot) {
3621 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3622 } else {
3623 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3624 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3625 }
3626 Node* membar = _gvn.transform(mb);
3627 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3628 if (alias_idx == Compile::AliasIdxBot) {
3629 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3630 } else {
3631 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3632 }
3633 return membar;
3634 }
3635
3636 void GraphKit::insert_store_load_for_barrier() {
3637 Node* mem = reset_memory();
3638 MemBarNode* mb = MemBarNode::make(C, Op_MemBarVolatile, Compile::AliasIdxRaw);
3639 mb->init_req(TypeFunc::Control, control());
3640 mb->init_req(TypeFunc::Memory, mem);
3641 Node* membar = _gvn.transform(mb);
3642 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3643 Node* newmem = _gvn.transform(new ProjNode(membar, TypeFunc::Memory));
3644 set_all_memory(mem);
3645 set_memory(newmem, Compile::AliasIdxRaw);
3646 }
3647
3648 //------------------------------shared_lock------------------------------------
3649 // Emit locking code.
3650 FastLockNode* GraphKit::shared_lock(Node* obj) {
3651 // bci is either a monitorenter bc or InvocationEntryBci
3652 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3653 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3654
3655 if( !GenerateSynchronizationCode )
3656 return NULL; // Not locking things?
3657
3658 if (stopped()) // Dead monitor?
3659 return NULL;
3660
3661 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3662
3663 obj = access_resolve(obj, ACCESS_READ | ACCESS_WRITE);
3664
3665 // Box the stack location
3666 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3667 Node* mem = reset_memory();
3668
3669 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3670 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3671 // Create the counters for this fast lock.
3672 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3673 }
3674
3675 // Create the rtm counters for this fast lock if needed.
3676 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3677
3678 // Add monitor to debug info for the slow path. If we block inside the
3679 // slow path and de-opt, we need the monitor hanging around
3680 map()->push_monitor( flock );
3681
3682 const TypeFunc *tf = LockNode::lock_type();
3683 LockNode *lock = new LockNode(C, tf);
3684
3685 lock->init_req( TypeFunc::Control, control() );
3686 lock->init_req( TypeFunc::Memory , mem );
3687 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3688 lock->init_req( TypeFunc::FramePtr, frameptr() );
3689 lock->init_req( TypeFunc::ReturnAdr, top() );
3690
3691 lock->init_req(TypeFunc::Parms + 0, obj);
3692 lock->init_req(TypeFunc::Parms + 1, box);
3693 lock->init_req(TypeFunc::Parms + 2, flock);
3694 add_safepoint_edges(lock);
3695
3696 lock = _gvn.transform( lock )->as_Lock();
3697
3698 // lock has no side-effects, sets few values
3699 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3700
3701 insert_mem_bar(Op_MemBarAcquireLock);
3702
3703 // Add this to the worklist so that the lock can be eliminated
3704 record_for_igvn(lock);
3705
3706 #ifndef PRODUCT
3707 if (PrintLockStatistics) {
3708 // Update the counter for this lock. Don't bother using an atomic
3709 // operation since we don't require absolute accuracy.
3710 lock->create_lock_counter(map()->jvms());
3711 increment_counter(lock->counter()->addr());
3712 }
3713 #endif
3714
3715 return flock;
3716 }
3717
3718
3719 //------------------------------shared_unlock----------------------------------
3720 // Emit unlocking code.
3721 void GraphKit::shared_unlock(Node* box, Node* obj) {
3722 // bci is either a monitorenter bc or InvocationEntryBci
3723 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3724 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3725
3726 if( !GenerateSynchronizationCode )
3727 return;
3728 if (stopped()) { // Dead monitor?
3729 map()->pop_monitor(); // Kill monitor from debug info
3730 return;
3731 }
3732 assert(!obj->is_ValueTypeBase(), "should not unlock on value type");
3733
3734 // Memory barrier to avoid floating things down past the locked region
3735 insert_mem_bar(Op_MemBarReleaseLock);
3736
3737 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3738 UnlockNode *unlock = new UnlockNode(C, tf);
3739 #ifdef ASSERT
3740 unlock->set_dbg_jvms(sync_jvms());
3741 #endif
3742 uint raw_idx = Compile::AliasIdxRaw;
3743 unlock->init_req( TypeFunc::Control, control() );
3744 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3745 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3746 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3747 unlock->init_req( TypeFunc::ReturnAdr, top() );
3748
3749 unlock->init_req(TypeFunc::Parms + 0, obj);
3750 unlock->init_req(TypeFunc::Parms + 1, box);
3751 unlock = _gvn.transform(unlock)->as_Unlock();
3752
3753 Node* mem = reset_memory();
3754
3755 // unlock has no side-effects, sets few values
3756 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3757
3758 // Kill monitor from debug info
3759 map()->pop_monitor( );
3760 }
3761
3762 //-------------------------------get_layout_helper-----------------------------
3763 // If the given klass is a constant or known to be an array,
3764 // fetch the constant layout helper value into constant_value
3765 // and return (Node*)NULL. Otherwise, load the non-constant
3766 // layout helper value, and return the node which represents it.
3767 // This two-faced routine is useful because allocation sites
3768 // almost always feature constant types.
3769 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3770 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3771 if (!StressReflectiveCode && inst_klass != NULL) {
3772 ciKlass* klass = inst_klass->klass();
3773 assert(klass != NULL, "klass should not be NULL");
3774 bool xklass = inst_klass->klass_is_exact();
3775 bool can_be_flattened = false;
3776 if (ValueArrayFlatten && klass->is_obj_array_klass()) {
3777 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
3778 can_be_flattened = elem->is_java_lang_Object() || elem->is_interface() || (elem->is_valuetype() && !klass->as_array_klass()->storage_properties().is_null_free());
3779 }
3780 if (xklass || (klass->is_array_klass() && !can_be_flattened)) {
3781 jint lhelper = klass->layout_helper();
3782 if (lhelper != Klass::_lh_neutral_value) {
3783 constant_value = lhelper;
3784 return (Node*) NULL;
3785 }
3786 }
3787 }
3788 constant_value = Klass::_lh_neutral_value; // put in a known value
3789 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3790 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3791 }
3792
3793 // We just put in an allocate/initialize with a big raw-memory effect.
3794 // Hook selected additional alias categories on the initialization.
3795 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3796 MergeMemNode* init_in_merge,
3797 Node* init_out_raw) {
3798 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3799 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3800
3801 Node* prevmem = kit.memory(alias_idx);
3802 init_in_merge->set_memory_at(alias_idx, prevmem);
3803 kit.set_memory(init_out_raw, alias_idx);
3804 }
3805
3806 //---------------------------set_output_for_allocation-------------------------
3807 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3808 const TypeOopPtr* oop_type,
3809 bool deoptimize_on_exception) {
3810 int rawidx = Compile::AliasIdxRaw;
3811 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3812 add_safepoint_edges(alloc);
3813 Node* allocx = _gvn.transform(alloc);
3814 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3815 // create memory projection for i_o
3816 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3817 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3818
3819 // create a memory projection as for the normal control path
3820 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3821 set_memory(malloc, rawidx);
3822
3823 // a normal slow-call doesn't change i_o, but an allocation does
3824 // we create a separate i_o projection for the normal control path
3825 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3826 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3827
3828 // put in an initialization barrier
3829 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3830 rawoop)->as_Initialize();
3831 assert(alloc->initialization() == init, "2-way macro link must work");
3832 assert(init ->allocation() == alloc, "2-way macro link must work");
3833 {
3834 // Extract memory strands which may participate in the new object's
3835 // initialization, and source them from the new InitializeNode.
3836 // This will allow us to observe initializations when they occur,
3837 // and link them properly (as a group) to the InitializeNode.
3838 assert(init->in(InitializeNode::Memory) == malloc, "");
3839 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3840 init->set_req(InitializeNode::Memory, minit_in);
3841 record_for_igvn(minit_in); // fold it up later, if possible
3842 _gvn.set_type(minit_in, Type::MEMORY);
3843 Node* minit_out = memory(rawidx);
3844 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3845 // Add an edge in the MergeMem for the header fields so an access
3846 // to one of those has correct memory state
3847 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3848 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3849 if (oop_type->isa_aryptr()) {
3850 const TypeAryPtr* arytype = oop_type->is_aryptr();
3851 if (arytype->klass()->is_value_array_klass()) {
3852 // Initially all flattened array accesses share a single slice
3853 // but that changes after parsing. Prepare the memory graph so
3854 // it can optimize flattened array accesses properly once they
3855 // don't share a single slice.
3856 assert(C->flattened_accesses_share_alias(), "should be set at parse time");
3857 C->set_flattened_accesses_share_alias(false);
3858 ciValueArrayKlass* vak = arytype->klass()->as_value_array_klass();
3859 ciValueKlass* vk = vak->element_klass()->as_value_klass();
3860 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3861 ciField* field = vk->nonstatic_field_at(i);
3862 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3863 continue; // do not bother to track really large numbers of fields
3864 int off_in_vt = field->offset() - vk->first_field_offset();
3865 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3866 int fieldidx = C->get_alias_index(adr_type, true);
3867 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3868 }
3869 C->set_flattened_accesses_share_alias(true);
3870 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::VALUES), minit_in, minit_out);
3871 } else {
3872 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3873 int elemidx = C->get_alias_index(telemref);
3874 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3875 }
3876 } else if (oop_type->isa_instptr()) {
3877 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
3878 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3879 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3880 ciField* field = ik->nonstatic_field_at(i);
3881 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3882 continue; // do not bother to track really large numbers of fields
3883 // Find (or create) the alias category for this field:
3884 int fieldidx = C->alias_type(field)->index();
3885 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3886 }
3887 }
3888 }
3889
3890 // Cast raw oop to the real thing...
3891 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3892 javaoop = _gvn.transform(javaoop);
3893 C->set_recent_alloc(control(), javaoop);
3894 assert(just_allocated_object(control()) == javaoop, "just allocated");
3895
3896 #ifdef ASSERT
3897 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3898 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3899 "Ideal_allocation works");
3900 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3901 "Ideal_allocation works");
3902 if (alloc->is_AllocateArray()) {
3903 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3904 "Ideal_allocation works");
3905 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3906 "Ideal_allocation works");
3907 } else {
3908 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3909 }
3910 }
3911 #endif //ASSERT
3912
3913 return javaoop;
3914 }
3915
3916 //---------------------------new_instance--------------------------------------
3917 // This routine takes a klass_node which may be constant (for a static type)
3918 // or may be non-constant (for reflective code). It will work equally well
3919 // for either, and the graph will fold nicely if the optimizer later reduces
3920 // the type to a constant.
3921 // The optional arguments are for specialized use by intrinsics:
3922 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3923 // - If 'return_size_val', report the the total object size to the caller.
3924 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3925 Node* GraphKit::new_instance(Node* klass_node,
3926 Node* extra_slow_test,
3927 Node* *return_size_val,
3928 bool deoptimize_on_exception,
3929 ValueTypeBaseNode* value_node) {
3930 // Compute size in doublewords
3931 // The size is always an integral number of doublewords, represented
3932 // as a positive bytewise size stored in the klass's layout_helper.
3933 // The layout_helper also encodes (in a low bit) the need for a slow path.
3934 jint layout_con = Klass::_lh_neutral_value;
3935 Node* layout_val = get_layout_helper(klass_node, layout_con);
3936 bool layout_is_con = (layout_val == NULL);
3937
3938 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3939 // Generate the initial go-slow test. It's either ALWAYS (return a
3940 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3941 // case) a computed value derived from the layout_helper.
3942 Node* initial_slow_test = NULL;
3943 if (layout_is_con) {
3944 assert(!StressReflectiveCode, "stress mode does not use these paths");
3945 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3946 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3947 } else { // reflective case
3948 // This reflective path is used by Unsafe.allocateInstance.
3949 // (It may be stress-tested by specifying StressReflectiveCode.)
3950 // Basically, we want to get into the VM is there's an illegal argument.
3951 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3952 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3953 if (extra_slow_test != intcon(0)) {
3954 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3955 }
3956 // (Macro-expander will further convert this to a Bool, if necessary.)
3957 }
3958
3959 // Find the size in bytes. This is easy; it's the layout_helper.
3960 // The size value must be valid even if the slow path is taken.
3961 Node* size = NULL;
3962 if (layout_is_con) {
3963 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
3964 } else { // reflective case
3965 // This reflective path is used by clone and Unsafe.allocateInstance.
3966 size = ConvI2X(layout_val);
3967
3968 // Clear the low bits to extract layout_helper_size_in_bytes:
3969 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3970 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3971 size = _gvn.transform( new AndXNode(size, mask) );
3972 }
3973 if (return_size_val != NULL) {
3974 (*return_size_val) = size;
3975 }
3976
3977 // This is a precise notnull oop of the klass.
3978 // (Actually, it need not be precise if this is a reflective allocation.)
3979 // It's what we cast the result to.
3980 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3981 if (!tklass) tklass = TypeKlassPtr::OBJECT;
3982 const TypeOopPtr* oop_type = tklass->as_instance_type();
3983
3984 // Now generate allocation code
3985
3986 // The entire memory state is needed for slow path of the allocation
3987 // since GC and deoptimization can happen.
3988 Node *mem = reset_memory();
3989 set_all_memory(mem); // Create new memory state
3990
3991 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3992 control(), mem, i_o(),
3993 size, klass_node,
3994 initial_slow_test, value_node);
3995
3996 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3997 }
3998
3999 // With compressed oops, the 64 bit init value for non flattened value
4000 // arrays is built from 2 32 bit compressed oops
4001 static Node* raw_default_for_coops(Node* default_value, GraphKit& kit) {
4002 Node* lower = kit.gvn().transform(new CastP2XNode(kit.control(), default_value));
4003 Node* upper = kit.gvn().transform(new LShiftLNode(lower, kit.intcon(32)));
4004 return kit.gvn().transform(new OrLNode(lower, upper));
4005 }
4006
4007 //-------------------------------new_array-------------------------------------
4008 // helper for newarray and anewarray
4009 // The 'length' parameter is (obviously) the length of the array.
4010 // See comments on new_instance for the meaning of the other arguments.
4011 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4012 Node* length, // number of array elements
4013 int nargs, // number of arguments to push back for uncommon trap
4014 Node* *return_size_val,
4015 bool deoptimize_on_exception,
4016 Node* elem_mirror) {
4017 jint layout_con = Klass::_lh_neutral_value;
4018 Node* layout_val = get_layout_helper(klass_node, layout_con);
4019 bool layout_is_con = (layout_val == NULL);
4020
4021 if (!layout_is_con && !StressReflectiveCode &&
4022 !too_many_traps(Deoptimization::Reason_class_check)) {
4023 // This is a reflective array creation site.
4024 // Optimistically assume that it is a subtype of Object[],
4025 // so that we can fold up all the address arithmetic.
4026 layout_con = Klass::array_layout_helper(T_OBJECT);
4027 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4028 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4029 { BuildCutout unless(this, bol_lh, PROB_MAX);
4030 inc_sp(nargs);
4031 uncommon_trap(Deoptimization::Reason_class_check,
4032 Deoptimization::Action_maybe_recompile);
4033 }
4034 layout_val = NULL;
4035 layout_is_con = true;
4036 }
4037
4038 // Generate the initial go-slow test. Make sure we do not overflow
4039 // if length is huge (near 2Gig) or negative! We do not need
4040 // exact double-words here, just a close approximation of needed
4041 // double-words. We can't add any offset or rounding bits, lest we
4042 // take a size -1 of bytes and make it positive. Use an unsigned
4043 // compare, so negative sizes look hugely positive.
4044 int fast_size_limit = FastAllocateSizeLimit;
4045 if (layout_is_con) {
4046 assert(!StressReflectiveCode, "stress mode does not use these paths");
4047 // Increase the size limit if we have exact knowledge of array type.
4048 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4049 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4050 }
4051
4052 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4053 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4054
4055 // --- Size Computation ---
4056 // array_size = round_to_heap(array_header + (length << elem_shift));
4057 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4058 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4059 // The rounding mask is strength-reduced, if possible.
4060 int round_mask = MinObjAlignmentInBytes - 1;
4061 Node* header_size = NULL;
4062 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4063 // (T_BYTE has the weakest alignment and size restrictions...)
4064 if (layout_is_con) {
4065 int hsize = Klass::layout_helper_header_size(layout_con);
4066 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4067 bool is_value_array = Klass::layout_helper_is_valueArray(layout_con);
4068 if ((round_mask & ~right_n_bits(eshift)) == 0)
4069 round_mask = 0; // strength-reduce it if it goes away completely
4070 assert(is_value_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4071 assert(header_size_min <= hsize, "generic minimum is smallest");
4072 header_size_min = hsize;
4073 header_size = intcon(hsize + round_mask);
4074 } else {
4075 Node* hss = intcon(Klass::_lh_header_size_shift);
4076 Node* hsm = intcon(Klass::_lh_header_size_mask);
4077 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
4078 hsize = _gvn.transform( new AndINode(hsize, hsm) );
4079 Node* mask = intcon(round_mask);
4080 header_size = _gvn.transform( new AddINode(hsize, mask) );
4081 }
4082
4083 Node* elem_shift = NULL;
4084 if (layout_is_con) {
4085 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4086 if (eshift != 0)
4087 elem_shift = intcon(eshift);
4088 } else {
4089 // There is no need to mask or shift this value.
4090 // The semantics of LShiftINode include an implicit mask to 0x1F.
4091 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4092 elem_shift = layout_val;
4093 }
4094
4095 // Transition to native address size for all offset calculations:
4096 Node* lengthx = ConvI2X(length);
4097 Node* headerx = ConvI2X(header_size);
4098 #ifdef _LP64
4099 { const TypeInt* tilen = _gvn.find_int_type(length);
4100 if (tilen != NULL && tilen->_lo < 0) {
4101 // Add a manual constraint to a positive range. Cf. array_element_address.
4102 jint size_max = fast_size_limit;
4103 if (size_max > tilen->_hi) size_max = tilen->_hi;
4104 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4105
4106 // Only do a narrow I2L conversion if the range check passed.
4107 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4108 _gvn.transform(iff);
4109 RegionNode* region = new RegionNode(3);
4110 _gvn.set_type(region, Type::CONTROL);
4111 lengthx = new PhiNode(region, TypeLong::LONG);
4112 _gvn.set_type(lengthx, TypeLong::LONG);
4113
4114 // Range check passed. Use ConvI2L node with narrow type.
4115 Node* passed = IfFalse(iff);
4116 region->init_req(1, passed);
4117 // Make I2L conversion control dependent to prevent it from
4118 // floating above the range check during loop optimizations.
4119 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4120
4121 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4122 region->init_req(2, IfTrue(iff));
4123 lengthx->init_req(2, ConvI2X(length));
4124
4125 set_control(region);
4126 record_for_igvn(region);
4127 record_for_igvn(lengthx);
4128 }
4129 }
4130 #endif
4131
4132 // Combine header size (plus rounding) and body size. Then round down.
4133 // This computation cannot overflow, because it is used only in two
4134 // places, one where the length is sharply limited, and the other
4135 // after a successful allocation.
4136 Node* abody = lengthx;
4137 if (elem_shift != NULL)
4138 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
4139 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
4140 if (round_mask != 0) {
4141 Node* mask = MakeConX(~round_mask);
4142 size = _gvn.transform( new AndXNode(size, mask) );
4143 }
4144 // else if round_mask == 0, the size computation is self-rounding
4145
4146 if (return_size_val != NULL) {
4147 // This is the size
4148 (*return_size_val) = size;
4149 }
4150
4151 // Now generate allocation code
4152
4153 // The entire memory state is needed for slow path of the allocation
4154 // since GC and deoptimization can happen.
4155 Node *mem = reset_memory();
4156 set_all_memory(mem); // Create new memory state
4157
4158 if (initial_slow_test->is_Bool()) {
4159 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4160 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4161 }
4162
4163 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
4164 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4165 const Type* elem = NULL;
4166 ciKlass* elem_klass = NULL;
4167
4168 // Compute default value and storage properties for value type arrays:
4169 // - null-ok: MyValue.box[] (ciObjArrayKlass "[LMyValue")
4170 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue")
4171 // - null-free, flattened: MyValue.val[] (ciValueArrayKlass "[QMyValue")
4172 Node* storage_properties = NULL;
4173 Node* default_value = NULL;
4174 Node* raw_default_value = NULL;
4175 int props_shift = UseCompressedClassPointers ? oopDesc::narrow_storage_props_shift : oopDesc::wide_storage_props_shift;
4176 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) {
4177 // Array type is known
4178 elem = ary_ptr->elem();
4179 ciArrayKlass* ary_klass = ary_ptr->klass()->as_array_klass();
4180 elem_klass = ary_klass->element_klass();
4181
4182 ArrayStorageProperties props = ary_klass->storage_properties();
4183 if (!props.is_empty() && elem_klass->is_valuetype()) {
4184 if (props.is_null_free() && !props.is_flattened()) {
4185 default_value = ValueTypeNode::default_oop(gvn(), elem_klass->as_value_klass());
4186 if (elem->isa_narrowoop()) {
4187 default_value = _gvn.transform(new EncodePNode(default_value, elem));
4188 raw_default_value = raw_default_for_coops(default_value, *this);
4189 } else {
4190 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4191 }
4192 }
4193 storage_properties = MakeConX(props.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift));
4194 }
4195 }
4196
4197 if (EnableValhalla && (elem == NULL || (elem_klass != NULL && (elem_klass->is_java_lang_Object() || elem_klass->is_valuetype()) &&
4198 !ary_type->klass_is_exact()))) {
4199 // Array type is not known, compute default value and storage properties for initialization.
4200 assert(default_value == NULL && raw_default_value == NULL && storage_properties == NULL, "shouldn't be set yet");
4201 assert(elem_mirror != NULL, "should not be null");
4202
4203 Node* r = new RegionNode(4);
4204 default_value = new PhiNode(r, TypeInstPtr::BOTTOM);
4205 storage_properties = new PhiNode(r, TypeX_X);
4206
4207 Node* empty = MakeConX(ArrayStorageProperties::empty.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift));
4208 Node* null_free = MakeConX(ArrayStorageProperties::null_free.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift));
4209 Node* flat = MakeConX(ArrayStorageProperties::flattened_and_null_free.encode<NOT_LP64(jint) LP64_ONLY(jlong)>(props_shift));
4210
4211 // Check if element mirror is a value mirror
4212 IfNode* iff = create_and_map_if(control(), is_value_mirror(elem_mirror), PROB_FAIR, COUNT_UNKNOWN);
4213
4214 // Not a value mirror but a box mirror or not a value type array, initialize with all zero
4215 r->init_req(1, _gvn.transform(new IfFalseNode(iff)));
4216 default_value->init_req(1, null());
4217 storage_properties->init_req(1, empty);
4218
4219 // Value mirror (= null-free), check if flattened
4220 set_control(_gvn.transform(new IfTrueNode(iff)));
4221 Node* cmp = gen_lh_array_test(klass_node, Klass::_lh_array_tag_vt_value);
4222 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4223 iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4224
4225 // Flattened, initialize with all zero
4226 r->init_req(2, _gvn.transform(new IfTrueNode(iff)));
4227 default_value->init_req(2, null());
4228 storage_properties->init_req(2, flat);
4229
4230 // Non-flattened, initialize with the default value
4231 set_control(_gvn.transform(new IfFalseNode(iff)));
4232 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset()));
4233 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS));
4234 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset()));
4235 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4236 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(ValueKlass::default_value_offset_offset()));
4237 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
4238 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset));
4239 Node* val = access_load_at(elem_mirror, default_value_addr, _gvn.type(default_value_addr)->is_ptr(), TypeInstPtr::BOTTOM, T_OBJECT, IN_HEAP);
4240 r->init_req(3, control());
4241 default_value->init_req(3, val);
4242 storage_properties->init_req(3, null_free);
4243
4244 set_control(_gvn.transform(r));
4245 default_value = _gvn.transform(default_value);
4246 storage_properties = _gvn.transform(storage_properties);
4247 if (UseCompressedOops) {
4248 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4249 raw_default_value = raw_default_for_coops(default_value, *this);
4250 } else {
4251 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4252 }
4253 }
4254
4255 // Create the AllocateArrayNode and its result projections
4256 AllocateArrayNode* alloc = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4257 control(), mem, i_o(),
4258 size, klass_node,
4259 initial_slow_test,
4260 length, default_value,
4261 raw_default_value,
4262 storage_properties);
4263
4264 // Cast to correct type. Note that the klass_node may be constant or not,
4265 // and in the latter case the actual array type will be inexact also.
4266 // (This happens via a non-constant argument to inline_native_newArray.)
4267 // In any case, the value of klass_node provides the desired array type.
4268 const TypeInt* length_type = _gvn.find_int_type(length);
4269 if (ary_type->isa_aryptr() && length_type != NULL) {
4270 // Try to get a better type than POS for the size
4271 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4272 }
4273
4274 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4275
4276 // Cast length on remaining path to be as narrow as possible
4277 if (map()->find_edge(length) >= 0) {
4278 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
4279 if (ccast != length) {
4280 _gvn.set_type_bottom(ccast);
4281 record_for_igvn(ccast);
4282 replace_in_map(length, ccast);
4283 }
4284 }
4285
4286 return javaoop;
4287 }
4288
4289 // The following "Ideal_foo" functions are placed here because they recognize
4290 // the graph shapes created by the functions immediately above.
4291
4292 //---------------------------Ideal_allocation----------------------------------
4293 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4294 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4295 if (ptr == NULL) { // reduce dumb test in callers
4296 return NULL;
4297 }
4298
4299 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4300 ptr = bs->step_over_gc_barrier(ptr);
4301
4302 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4303 ptr = ptr->in(1);
4304 if (ptr == NULL) return NULL;
4305 }
4306 // Return NULL for allocations with several casts:
4307 // j.l.reflect.Array.newInstance(jobject, jint)
4308 // Object.clone()
4309 // to keep more precise type from last cast.
4310 if (ptr->is_Proj()) {
4311 Node* allo = ptr->in(0);
4312 if (allo != NULL && allo->is_Allocate()) {
4313 return allo->as_Allocate();
4314 }
4315 }
4316 // Report failure to match.
4317 return NULL;
4318 }
4319
4320 // Fancy version which also strips off an offset (and reports it to caller).
4321 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
4322 intptr_t& offset) {
4323 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4324 if (base == NULL) return NULL;
4325 return Ideal_allocation(base, phase);
4326 }
4327
4328 // Trace Initialize <- Proj[Parm] <- Allocate
4329 AllocateNode* InitializeNode::allocation() {
4330 Node* rawoop = in(InitializeNode::RawAddress);
4331 if (rawoop->is_Proj()) {
4332 Node* alloc = rawoop->in(0);
4333 if (alloc->is_Allocate()) {
4334 return alloc->as_Allocate();
4335 }
4336 }
4337 return NULL;
4338 }
4339
4340 // Trace Allocate -> Proj[Parm] -> Initialize
4341 InitializeNode* AllocateNode::initialization() {
4342 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4343 if (rawoop == NULL) return NULL;
4344 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4345 Node* init = rawoop->fast_out(i);
4346 if (init->is_Initialize()) {
4347 assert(init->as_Initialize()->allocation() == this, "2-way link");
4348 return init->as_Initialize();
4349 }
4350 }
4351 return NULL;
4352 }
4353
4354 //----------------------------- loop predicates ---------------------------
4355
4356 //------------------------------add_predicate_impl----------------------------
4357 void GraphKit::add_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
4358 // Too many traps seen?
4359 if (too_many_traps(reason)) {
4360 #ifdef ASSERT
4361 if (TraceLoopPredicate) {
4362 int tc = C->trap_count(reason);
4363 tty->print("too many traps=%s tcount=%d in ",
4364 Deoptimization::trap_reason_name(reason), tc);
4365 method()->print(); // which method has too many predicate traps
4366 tty->cr();
4367 }
4368 #endif
4369 // We cannot afford to take more traps here,
4370 // do not generate predicate.
4371 return;
4372 }
4373
4374 Node *cont = _gvn.intcon(1);
4375 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
4376 Node *bol = _gvn.transform(new Conv2BNode(opq));
4377 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4378 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
4379 C->add_predicate_opaq(opq);
4380 {
4381 PreserveJVMState pjvms(this);
4382 set_control(iffalse);
4383 inc_sp(nargs);
4384 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4385 }
4386 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
4387 set_control(iftrue);
4388 }
4389
4390 //------------------------------add_predicate---------------------------------
4391 void GraphKit::add_predicate(int nargs) {
4392 if (UseLoopPredicate) {
4393 add_predicate_impl(Deoptimization::Reason_predicate, nargs);
4394 }
4395 if (UseProfiledLoopPredicate) {
4396 add_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
4397 }
4398 // loop's limit check predicate should be near the loop.
4399 add_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
4400 }
4401
4402 void GraphKit::sync_kit(IdealKit& ideal) {
4403 set_all_memory(ideal.merged_memory());
4404 set_i_o(ideal.i_o());
4405 set_control(ideal.ctrl());
4406 }
4407
4408 void GraphKit::final_sync(IdealKit& ideal) {
4409 // Final sync IdealKit and graphKit.
4410 sync_kit(ideal);
4411 }
4412
4413 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4414 Node* len = load_array_length(load_String_value(str, set_ctrl));
4415 Node* coder = load_String_coder(str, set_ctrl);
4416 // Divide length by 2 if coder is UTF16
4417 return _gvn.transform(new RShiftINode(len, coder));
4418 }
4419
4420 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4421 int value_offset = java_lang_String::value_offset_in_bytes();
4422 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4423 false, NULL, Type::Offset(0), false);
4424 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4425 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4426 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, true, true),
4427 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4428 Node* p = basic_plus_adr(str, str, value_offset);
4429 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4430 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4431 return load;
4432 }
4433
4434 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4435 if (!CompactStrings) {
4436 return intcon(java_lang_String::CODER_UTF16);
4437 }
4438 int coder_offset = java_lang_String::coder_offset_in_bytes();
4439 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4440 false, NULL, Type::Offset(0), false);
4441 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4442
4443 Node* p = basic_plus_adr(str, str, coder_offset);
4444 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4445 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4446 return load;
4447 }
4448
4449 void GraphKit::store_String_value(Node* str, Node* value) {
4450 int value_offset = java_lang_String::value_offset_in_bytes();
4451 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4452 false, NULL, Type::Offset(0), false);
4453 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4454
4455 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4456 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4457 }
4458
4459 void GraphKit::store_String_coder(Node* str, Node* value) {
4460 int coder_offset = java_lang_String::coder_offset_in_bytes();
4461 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4462 false, NULL, Type::Offset(0), false);
4463 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4464
4465 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4466 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4467 }
4468
4469 // Capture src and dst memory state with a MergeMemNode
4470 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4471 if (src_type == dst_type) {
4472 // Types are equal, we don't need a MergeMemNode
4473 return memory(src_type);
4474 }
4475 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4476 record_for_igvn(merge); // fold it up later, if possible
4477 int src_idx = C->get_alias_index(src_type);
4478 int dst_idx = C->get_alias_index(dst_type);
4479 merge->set_memory_at(src_idx, memory(src_idx));
4480 merge->set_memory_at(dst_idx, memory(dst_idx));
4481 return merge;
4482 }
4483
4484 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4485 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4486 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4487 // If input and output memory types differ, capture both states to preserve
4488 // the dependency between preceding and subsequent loads/stores.
4489 // For example, the following program:
4490 // StoreB
4491 // compress_string
4492 // LoadB
4493 // has this memory graph (use->def):
4494 // LoadB -> compress_string -> CharMem
4495 // ... -> StoreB -> ByteMem
4496 // The intrinsic hides the dependency between LoadB and StoreB, causing
4497 // the load to read from memory not containing the result of the StoreB.
4498 // The correct memory graph should look like this:
4499 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4500 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4501 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4502 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4503 set_memory(res_mem, TypeAryPtr::BYTES);
4504 return str;
4505 }
4506
4507 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4508 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4509 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4510 // Capture src and dst memory (see comment in 'compress_string').
4511 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4512 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4513 set_memory(_gvn.transform(str), dst_type);
4514 }
4515
4516 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4517 /**
4518 * int i_char = start;
4519 * for (int i_byte = 0; i_byte < count; i_byte++) {
4520 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4521 * }
4522 */
4523 src = access_resolve(src, ACCESS_READ);
4524 dst = access_resolve(dst, ACCESS_WRITE);
4525 add_predicate();
4526 RegionNode* head = new RegionNode(3);
4527 head->init_req(1, control());
4528 gvn().set_type(head, Type::CONTROL);
4529 record_for_igvn(head);
4530
4531 Node* i_byte = new PhiNode(head, TypeInt::INT);
4532 i_byte->init_req(1, intcon(0));
4533 gvn().set_type(i_byte, TypeInt::INT);
4534 record_for_igvn(i_byte);
4535
4536 Node* i_char = new PhiNode(head, TypeInt::INT);
4537 i_char->init_req(1, start);
4538 gvn().set_type(i_char, TypeInt::INT);
4539 record_for_igvn(i_char);
4540
4541 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4542 gvn().set_type(mem, Type::MEMORY);
4543 record_for_igvn(mem);
4544 set_control(head);
4545 set_memory(mem, TypeAryPtr::BYTES);
4546 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4547 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4548 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4549 false, false, true /* mismatched */);
4550
4551 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4552 head->init_req(2, IfTrue(iff));
4553 mem->init_req(2, st);
4554 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4555 i_char->init_req(2, AddI(i_char, intcon(2)));
4556
4557 set_control(IfFalse(iff));
4558 set_memory(st, TypeAryPtr::BYTES);
4559 }
4560
4561 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4562 if (!field->is_constant()) {
4563 return NULL; // Field not marked as constant.
4564 }
4565 ciInstance* holder = NULL;
4566 if (!field->is_static()) {
4567 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4568 if (const_oop != NULL && const_oop->is_instance()) {
4569 holder = const_oop->as_instance();
4570 }
4571 }
4572 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4573 /*is_unsigned_load=*/false);
4574 if (con_type != NULL) {
4575 Node* con = makecon(con_type);
4576 if (field->layout_type() == T_VALUETYPE && field->type()->as_value_klass()->is_scalarizable() && !con_type->maybe_null()) {
4577 // Load value type from constant oop
4578 con = ValueTypeNode::make_from_oop(this, con, field->type()->as_value_klass());
4579 }
4580 return con;
4581 }
4582 return NULL;
4583 }
4584
4585 //---------------------------load_mirror_from_klass----------------------------
4586 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4587 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4588 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4589 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4590 // mirror = ((OopHandle)mirror)->resolve();
4591 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4592 }