1 /*
2 * Copyright (c) 2001, 2020, 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/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciUtilities.hpp"
29 #include "classfile/javaClasses.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/c2/barrierSetC2.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/machnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/rootnode.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/subtypenode.hpp"
50 #include "runtime/deoptimization.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "utilities/bitMap.inline.hpp"
53 #include "utilities/powerOfTwo.hpp"
54
55 //----------------------------GraphKit-----------------------------------------
56 // Main utility constructor.
57 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
58 : Phase(Phase::Parser),
59 _env(C->env()),
60 _gvn((gvn != NULL) ? *gvn : *C->initial_gvn()),
61 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
62 {
63 assert(gvn == NULL || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
64 _exceptions = jvms->map()->next_exception();
65 if (_exceptions != NULL) jvms->map()->set_next_exception(NULL);
66 set_jvms(jvms);
67 #ifdef ASSERT
68 if (_gvn.is_IterGVN() != NULL) {
69 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
70 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
71 _worklist_size = _gvn.C->for_igvn()->size();
72 }
73 #endif
74 }
75
76 // Private constructor for parser.
77 GraphKit::GraphKit()
78 : Phase(Phase::Parser),
79 _env(C->env()),
80 _gvn(*C->initial_gvn()),
81 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
82 {
83 _exceptions = NULL;
84 set_map(NULL);
85 debug_only(_sp = -99);
86 debug_only(set_bci(-99));
87 }
88
89
90
91 //---------------------------clean_stack---------------------------------------
92 // Clear away rubbish from the stack area of the JVM state.
93 // This destroys any arguments that may be waiting on the stack.
94 void GraphKit::clean_stack(int from_sp) {
95 SafePointNode* map = this->map();
96 JVMState* jvms = this->jvms();
97 int stk_size = jvms->stk_size();
98 int stkoff = jvms->stkoff();
99 Node* top = this->top();
100 for (int i = from_sp; i < stk_size; i++) {
101 if (map->in(stkoff + i) != top) {
102 map->set_req(stkoff + i, top);
103 }
104 }
105 }
106
107
108 //--------------------------------sync_jvms-----------------------------------
109 // Make sure our current jvms agrees with our parse state.
110 JVMState* GraphKit::sync_jvms() const {
111 JVMState* jvms = this->jvms();
112 jvms->set_bci(bci()); // Record the new bci in the JVMState
113 jvms->set_sp(sp()); // Record the new sp in the JVMState
114 assert(jvms_in_sync(), "jvms is now in sync");
115 return jvms;
116 }
117
118 //--------------------------------sync_jvms_for_reexecute---------------------
119 // Make sure our current jvms agrees with our parse state. This version
120 // uses the reexecute_sp for reexecuting bytecodes.
121 JVMState* GraphKit::sync_jvms_for_reexecute() {
122 JVMState* jvms = this->jvms();
123 jvms->set_bci(bci()); // Record the new bci in the JVMState
124 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
125 return jvms;
126 }
127
128 #ifdef ASSERT
129 bool GraphKit::jvms_in_sync() const {
130 Parse* parse = is_Parse();
131 if (parse == NULL) {
132 if (bci() != jvms()->bci()) return false;
133 if (sp() != (int)jvms()->sp()) return false;
134 return true;
135 }
136 if (jvms()->method() != parse->method()) return false;
137 if (jvms()->bci() != parse->bci()) return false;
138 int jvms_sp = jvms()->sp();
139 if (jvms_sp != parse->sp()) return false;
140 int jvms_depth = jvms()->depth();
141 if (jvms_depth != parse->depth()) return false;
142 return true;
143 }
144
145 // Local helper checks for special internal merge points
146 // used to accumulate and merge exception states.
147 // They are marked by the region's in(0) edge being the map itself.
148 // Such merge points must never "escape" into the parser at large,
149 // until they have been handed to gvn.transform.
150 static bool is_hidden_merge(Node* reg) {
151 if (reg == NULL) return false;
152 if (reg->is_Phi()) {
153 reg = reg->in(0);
154 if (reg == NULL) return false;
155 }
156 return reg->is_Region() && reg->in(0) != NULL && reg->in(0)->is_Root();
157 }
158
159 void GraphKit::verify_map() const {
160 if (map() == NULL) return; // null map is OK
161 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
162 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
163 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
164 }
165
166 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
167 assert(ex_map->next_exception() == NULL, "not already part of a chain");
168 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
169 }
170 #endif
171
172 //---------------------------stop_and_kill_map---------------------------------
173 // Set _map to NULL, signalling a stop to further bytecode execution.
174 // First smash the current map's control to a constant, to mark it dead.
175 void GraphKit::stop_and_kill_map() {
176 SafePointNode* dead_map = stop();
177 if (dead_map != NULL) {
178 dead_map->disconnect_inputs(NULL, C); // Mark the map as killed.
179 assert(dead_map->is_killed(), "must be so marked");
180 }
181 }
182
183
184 //--------------------------------stopped--------------------------------------
185 // Tell if _map is NULL, or control is top.
186 bool GraphKit::stopped() {
187 if (map() == NULL) return true;
188 else if (control() == top()) return true;
189 else return false;
190 }
191
192
193 //-----------------------------has_ex_handler----------------------------------
194 // Tell if this method or any caller method has exception handlers.
195 bool GraphKit::has_ex_handler() {
196 for (JVMState* jvmsp = jvms(); jvmsp != NULL; jvmsp = jvmsp->caller()) {
197 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
198 return true;
199 }
200 }
201 return false;
202 }
203
204 //------------------------------save_ex_oop------------------------------------
205 // Save an exception without blowing stack contents or other JVM state.
206 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
207 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
208 ex_map->add_req(ex_oop);
209 debug_only(verify_exception_state(ex_map));
210 }
211
212 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
213 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
214 Node* ex_oop = ex_map->in(ex_map->req()-1);
215 if (clear_it) ex_map->del_req(ex_map->req()-1);
216 return ex_oop;
217 }
218
219 //-----------------------------saved_ex_oop------------------------------------
220 // Recover a saved exception from its map.
221 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
222 return common_saved_ex_oop(ex_map, false);
223 }
224
225 //--------------------------clear_saved_ex_oop---------------------------------
226 // Erase a previously saved exception from its map.
227 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
228 return common_saved_ex_oop(ex_map, true);
229 }
230
231 #ifdef ASSERT
232 //---------------------------has_saved_ex_oop----------------------------------
233 // Erase a previously saved exception from its map.
234 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
235 return ex_map->req() == ex_map->jvms()->endoff()+1;
236 }
237 #endif
238
239 //-------------------------make_exception_state--------------------------------
240 // Turn the current JVM state into an exception state, appending the ex_oop.
241 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
242 sync_jvms();
243 SafePointNode* ex_map = stop(); // do not manipulate this map any more
244 set_saved_ex_oop(ex_map, ex_oop);
245 return ex_map;
246 }
247
248
249 //--------------------------add_exception_state--------------------------------
250 // Add an exception to my list of exceptions.
251 void GraphKit::add_exception_state(SafePointNode* ex_map) {
252 if (ex_map == NULL || ex_map->control() == top()) {
253 return;
254 }
255 #ifdef ASSERT
256 verify_exception_state(ex_map);
257 if (has_exceptions()) {
258 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
259 }
260 #endif
261
262 // If there is already an exception of exactly this type, merge with it.
263 // In particular, null-checks and other low-level exceptions common up here.
264 Node* ex_oop = saved_ex_oop(ex_map);
265 const Type* ex_type = _gvn.type(ex_oop);
266 if (ex_oop == top()) {
267 // No action needed.
268 return;
269 }
270 assert(ex_type->isa_instptr(), "exception must be an instance");
271 for (SafePointNode* e2 = _exceptions; e2 != NULL; e2 = e2->next_exception()) {
272 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
273 // We check sp also because call bytecodes can generate exceptions
274 // both before and after arguments are popped!
275 if (ex_type2 == ex_type
276 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
277 combine_exception_states(ex_map, e2);
278 return;
279 }
280 }
281
282 // No pre-existing exception of the same type. Chain it on the list.
283 push_exception_state(ex_map);
284 }
285
286 //-----------------------add_exception_states_from-----------------------------
287 void GraphKit::add_exception_states_from(JVMState* jvms) {
288 SafePointNode* ex_map = jvms->map()->next_exception();
289 if (ex_map != NULL) {
290 jvms->map()->set_next_exception(NULL);
291 for (SafePointNode* next_map; ex_map != NULL; ex_map = next_map) {
292 next_map = ex_map->next_exception();
293 ex_map->set_next_exception(NULL);
294 add_exception_state(ex_map);
295 }
296 }
297 }
298
299 //-----------------------transfer_exceptions_into_jvms-------------------------
300 JVMState* GraphKit::transfer_exceptions_into_jvms() {
301 if (map() == NULL) {
302 // We need a JVMS to carry the exceptions, but the map has gone away.
303 // Create a scratch JVMS, cloned from any of the exception states...
304 if (has_exceptions()) {
305 _map = _exceptions;
306 _map = clone_map();
307 _map->set_next_exception(NULL);
308 clear_saved_ex_oop(_map);
309 debug_only(verify_map());
310 } else {
311 // ...or created from scratch
312 JVMState* jvms = new (C) JVMState(_method, NULL);
313 jvms->set_bci(_bci);
314 jvms->set_sp(_sp);
315 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
316 set_jvms(jvms);
317 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
318 set_all_memory(top());
319 while (map()->req() < jvms->endoff()) map()->add_req(top());
320 }
321 // (This is a kludge, in case you didn't notice.)
322 set_control(top());
323 }
324 JVMState* jvms = sync_jvms();
325 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
326 jvms->map()->set_next_exception(_exceptions);
327 _exceptions = NULL; // done with this set of exceptions
328 return jvms;
329 }
330
331 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
332 assert(is_hidden_merge(dstphi), "must be a special merge node");
333 assert(is_hidden_merge(srcphi), "must be a special merge node");
334 uint limit = srcphi->req();
335 for (uint i = PhiNode::Input; i < limit; i++) {
336 dstphi->add_req(srcphi->in(i));
337 }
338 }
339 static inline void add_one_req(Node* dstphi, Node* src) {
340 assert(is_hidden_merge(dstphi), "must be a special merge node");
341 assert(!is_hidden_merge(src), "must not be a special merge node");
342 dstphi->add_req(src);
343 }
344
345 //-----------------------combine_exception_states------------------------------
346 // This helper function combines exception states by building phis on a
347 // specially marked state-merging region. These regions and phis are
348 // untransformed, and can build up gradually. The region is marked by
349 // having a control input of its exception map, rather than NULL. Such
350 // regions do not appear except in this function, and in use_exception_state.
351 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
352 if (failing()) return; // dying anyway...
353 JVMState* ex_jvms = ex_map->_jvms;
354 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
355 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
356 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
357 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
358 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
359 assert(ex_map->req() == phi_map->req(), "matching maps");
360 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
361 Node* hidden_merge_mark = root();
362 Node* region = phi_map->control();
363 MergeMemNode* phi_mem = phi_map->merged_memory();
364 MergeMemNode* ex_mem = ex_map->merged_memory();
365 if (region->in(0) != hidden_merge_mark) {
366 // The control input is not (yet) a specially-marked region in phi_map.
367 // Make it so, and build some phis.
368 region = new RegionNode(2);
369 _gvn.set_type(region, Type::CONTROL);
370 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
371 region->init_req(1, phi_map->control());
372 phi_map->set_control(region);
373 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
374 record_for_igvn(io_phi);
375 _gvn.set_type(io_phi, Type::ABIO);
376 phi_map->set_i_o(io_phi);
377 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
378 Node* m = mms.memory();
379 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
380 record_for_igvn(m_phi);
381 _gvn.set_type(m_phi, Type::MEMORY);
382 mms.set_memory(m_phi);
383 }
384 }
385
386 // Either or both of phi_map and ex_map might already be converted into phis.
387 Node* ex_control = ex_map->control();
388 // if there is special marking on ex_map also, we add multiple edges from src
389 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
390 // how wide was the destination phi_map, originally?
391 uint orig_width = region->req();
392
393 if (add_multiple) {
394 add_n_reqs(region, ex_control);
395 add_n_reqs(phi_map->i_o(), ex_map->i_o());
396 } else {
397 // ex_map has no merges, so we just add single edges everywhere
398 add_one_req(region, ex_control);
399 add_one_req(phi_map->i_o(), ex_map->i_o());
400 }
401 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
402 if (mms.is_empty()) {
403 // get a copy of the base memory, and patch some inputs into it
404 const TypePtr* adr_type = mms.adr_type(C);
405 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
406 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
407 mms.set_memory(phi);
408 // Prepare to append interesting stuff onto the newly sliced phi:
409 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
410 }
411 // Append stuff from ex_map:
412 if (add_multiple) {
413 add_n_reqs(mms.memory(), mms.memory2());
414 } else {
415 add_one_req(mms.memory(), mms.memory2());
416 }
417 }
418 uint limit = ex_map->req();
419 for (uint i = TypeFunc::Parms; i < limit; i++) {
420 // Skip everything in the JVMS after tos. (The ex_oop follows.)
421 if (i == tos) i = ex_jvms->monoff();
422 Node* src = ex_map->in(i);
423 Node* dst = phi_map->in(i);
424 if (src != dst) {
425 PhiNode* phi;
426 if (dst->in(0) != region) {
427 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
428 record_for_igvn(phi);
429 _gvn.set_type(phi, phi->type());
430 phi_map->set_req(i, dst);
431 // Prepare to append interesting stuff onto the new phi:
432 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
433 } else {
434 assert(dst->is_Phi(), "nobody else uses a hidden region");
435 phi = dst->as_Phi();
436 }
437 if (add_multiple && src->in(0) == ex_control) {
438 // Both are phis.
439 add_n_reqs(dst, src);
440 } else {
441 while (dst->req() < region->req()) add_one_req(dst, src);
442 }
443 const Type* srctype = _gvn.type(src);
444 if (phi->type() != srctype) {
445 const Type* dsttype = phi->type()->meet_speculative(srctype);
446 if (phi->type() != dsttype) {
447 phi->set_type(dsttype);
448 _gvn.set_type(phi, dsttype);
449 }
450 }
451 }
452 }
453 phi_map->merge_replaced_nodes_with(ex_map);
454 }
455
456 //--------------------------use_exception_state--------------------------------
457 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
458 if (failing()) { stop(); return top(); }
459 Node* region = phi_map->control();
460 Node* hidden_merge_mark = root();
461 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
462 Node* ex_oop = clear_saved_ex_oop(phi_map);
463 if (region->in(0) == hidden_merge_mark) {
464 // Special marking for internal ex-states. Process the phis now.
465 region->set_req(0, region); // now it's an ordinary region
466 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
467 // Note: Setting the jvms also sets the bci and sp.
468 set_control(_gvn.transform(region));
469 uint tos = jvms()->stkoff() + sp();
470 for (uint i = 1; i < tos; i++) {
471 Node* x = phi_map->in(i);
472 if (x->in(0) == region) {
473 assert(x->is_Phi(), "expected a special phi");
474 phi_map->set_req(i, _gvn.transform(x));
475 }
476 }
477 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
478 Node* x = mms.memory();
479 if (x->in(0) == region) {
480 assert(x->is_Phi(), "nobody else uses a hidden region");
481 mms.set_memory(_gvn.transform(x));
482 }
483 }
484 if (ex_oop->in(0) == region) {
485 assert(ex_oop->is_Phi(), "expected a special phi");
486 ex_oop = _gvn.transform(ex_oop);
487 }
488 } else {
489 set_jvms(phi_map->jvms());
490 }
491
492 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
493 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
494 return ex_oop;
495 }
496
497 //---------------------------------java_bc-------------------------------------
498 Bytecodes::Code GraphKit::java_bc() const {
499 ciMethod* method = this->method();
500 int bci = this->bci();
501 if (method != NULL && bci != InvocationEntryBci)
502 return method->java_code_at_bci(bci);
503 else
504 return Bytecodes::_illegal;
505 }
506
507 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
508 bool must_throw) {
509 // if the exception capability is set, then we will generate code
510 // to check the JavaThread.should_post_on_exceptions flag to see
511 // if we actually need to report exception events (for this
512 // thread). If we don't need to report exception events, we will
513 // take the normal fast path provided by add_exception_events. If
514 // exception event reporting is enabled for this thread, we will
515 // take the uncommon_trap in the BuildCutout below.
516
517 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
518 Node* jthread = _gvn.transform(new ThreadLocalNode());
519 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
520 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
521
522 // Test the should_post_on_exceptions_flag vs. 0
523 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
524 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
525
526 // Branch to slow_path if should_post_on_exceptions_flag was true
527 { BuildCutout unless(this, tst, PROB_MAX);
528 // Do not try anything fancy if we're notifying the VM on every throw.
529 // Cf. case Bytecodes::_athrow in parse2.cpp.
530 uncommon_trap(reason, Deoptimization::Action_none,
531 (ciKlass*)NULL, (char*)NULL, must_throw);
532 }
533
534 }
535
536 //------------------------------builtin_throw----------------------------------
537 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason, Node* arg) {
538 bool must_throw = true;
539
540 if (env()->jvmti_can_post_on_exceptions()) {
541 // check if we must post exception events, take uncommon trap if so
542 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
543 // here if should_post_on_exceptions is false
544 // continue on with the normal codegen
545 }
546
547 // If this particular condition has not yet happened at this
548 // bytecode, then use the uncommon trap mechanism, and allow for
549 // a future recompilation if several traps occur here.
550 // If the throw is hot, try to use a more complicated inline mechanism
551 // which keeps execution inside the compiled code.
552 bool treat_throw_as_hot = false;
553 ciMethodData* md = method()->method_data();
554
555 if (ProfileTraps) {
556 if (too_many_traps(reason)) {
557 treat_throw_as_hot = true;
558 }
559 // (If there is no MDO at all, assume it is early in
560 // execution, and that any deopts are part of the
561 // startup transient, and don't need to be remembered.)
562
563 // Also, if there is a local exception handler, treat all throws
564 // as hot if there has been at least one in this method.
565 if (C->trap_count(reason) != 0
566 && method()->method_data()->trap_count(reason) != 0
567 && has_ex_handler()) {
568 treat_throw_as_hot = true;
569 }
570 }
571
572 // If this throw happens frequently, an uncommon trap might cause
573 // a performance pothole. If there is a local exception handler,
574 // and if this particular bytecode appears to be deoptimizing often,
575 // let us handle the throw inline, with a preconstructed instance.
576 // Note: If the deopt count has blown up, the uncommon trap
577 // runtime is going to flush this nmethod, not matter what.
578 if (treat_throw_as_hot
579 && (!StackTraceInThrowable || OmitStackTraceInFastThrow)) {
580 // If the throw is local, we use a pre-existing instance and
581 // punt on the backtrace. This would lead to a missing backtrace
582 // (a repeat of 4292742) if the backtrace object is ever asked
583 // for its backtrace.
584 // Fixing this remaining case of 4292742 requires some flavor of
585 // escape analysis. Leave that for the future.
586 ciInstance* ex_obj = NULL;
587 switch (reason) {
588 case Deoptimization::Reason_null_check:
589 ex_obj = env()->NullPointerException_instance();
590 break;
591 case Deoptimization::Reason_div0_check:
592 ex_obj = env()->ArithmeticException_instance();
593 break;
594 case Deoptimization::Reason_range_check:
595 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
596 break;
597 case Deoptimization::Reason_class_check:
598 if (java_bc() == Bytecodes::_aastore) {
599 ex_obj = env()->ArrayStoreException_instance();
600 } else {
601 ex_obj = env()->ClassCastException_instance();
602 }
603 break;
604 default:
605 break;
606 }
607 if (failing()) { stop(); return; } // exception allocation might fail
608 if (ex_obj != NULL) {
609 // Cheat with a preallocated exception object.
610 if (C->log() != NULL)
611 C->log()->elem("hot_throw preallocated='1' reason='%s'",
612 Deoptimization::trap_reason_name(reason));
613 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
614 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
615
616 // Clear the detail message of the preallocated exception object.
617 // Weblogic sometimes mutates the detail message of exceptions
618 // using reflection.
619 int offset = java_lang_Throwable::get_detailMessage_offset();
620 const TypePtr* adr_typ = ex_con->add_offset(offset);
621
622 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
623 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
624 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
625
626 add_exception_state(make_exception_state(ex_node));
627 return;
628 }
629 }
630
631 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
632 // It won't be much cheaper than bailing to the interp., since we'll
633 // have to pass up all the debug-info, and the runtime will have to
634 // create the stack trace.
635
636 // Usual case: Bail to interpreter.
637 // Reserve the right to recompile if we haven't seen anything yet.
638
639 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : NULL;
640 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
641 if (treat_throw_as_hot
642 && (method()->method_data()->trap_recompiled_at(bci(), m)
643 || C->too_many_traps(reason))) {
644 // We cannot afford to take more traps here. Suffer in the interpreter.
645 if (C->log() != NULL)
646 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
647 Deoptimization::trap_reason_name(reason),
648 C->trap_count(reason));
649 action = Deoptimization::Action_none;
650 }
651
652 // "must_throw" prunes the JVM state to include only the stack, if there
653 // are no local exception handlers. This should cut down on register
654 // allocation time and code size, by drastically reducing the number
655 // of in-edges on the call to the uncommon trap.
656
657 uncommon_trap(reason, action, (ciKlass*)NULL, (char*)NULL, must_throw);
658 }
659
660
661 //----------------------------PreserveJVMState---------------------------------
662 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
663 debug_only(kit->verify_map());
664 _kit = kit;
665 _map = kit->map(); // preserve the map
666 _sp = kit->sp();
667 kit->set_map(clone_map ? kit->clone_map() : NULL);
668 #ifdef ASSERT
669 _bci = kit->bci();
670 Parse* parser = kit->is_Parse();
671 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
672 _block = block;
673 #endif
674 }
675 PreserveJVMState::~PreserveJVMState() {
676 GraphKit* kit = _kit;
677 #ifdef ASSERT
678 assert(kit->bci() == _bci, "bci must not shift");
679 Parse* parser = kit->is_Parse();
680 int block = (parser == NULL || parser->block() == NULL) ? -1 : parser->block()->rpo();
681 assert(block == _block, "block must not shift");
682 #endif
683 kit->set_map(_map);
684 kit->set_sp(_sp);
685 }
686
687
688 //-----------------------------BuildCutout-------------------------------------
689 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
690 : PreserveJVMState(kit)
691 {
692 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
693 SafePointNode* outer_map = _map; // preserved map is caller's
694 SafePointNode* inner_map = kit->map();
695 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
696 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
697 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
698 }
699 BuildCutout::~BuildCutout() {
700 GraphKit* kit = _kit;
701 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
702 }
703
704 //---------------------------PreserveReexecuteState----------------------------
705 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
706 assert(!kit->stopped(), "must call stopped() before");
707 _kit = kit;
708 _sp = kit->sp();
709 _reexecute = kit->jvms()->_reexecute;
710 }
711 PreserveReexecuteState::~PreserveReexecuteState() {
712 if (_kit->stopped()) return;
713 _kit->jvms()->_reexecute = _reexecute;
714 _kit->set_sp(_sp);
715 }
716
717 //------------------------------clone_map--------------------------------------
718 // Implementation of PreserveJVMState
719 //
720 // Only clone_map(...) here. If this function is only used in the
721 // PreserveJVMState class we may want to get rid of this extra
722 // function eventually and do it all there.
723
724 SafePointNode* GraphKit::clone_map() {
725 if (map() == NULL) return NULL;
726
727 // Clone the memory edge first
728 Node* mem = MergeMemNode::make(map()->memory());
729 gvn().set_type_bottom(mem);
730
731 SafePointNode *clonemap = (SafePointNode*)map()->clone();
732 JVMState* jvms = this->jvms();
733 JVMState* clonejvms = jvms->clone_shallow(C);
734 clonemap->set_memory(mem);
735 clonemap->set_jvms(clonejvms);
736 clonejvms->set_map(clonemap);
737 record_for_igvn(clonemap);
738 gvn().set_type_bottom(clonemap);
739 return clonemap;
740 }
741
742
743 //-----------------------------set_map_clone-----------------------------------
744 void GraphKit::set_map_clone(SafePointNode* m) {
745 _map = m;
746 _map = clone_map();
747 _map->set_next_exception(NULL);
748 debug_only(verify_map());
749 }
750
751
752 //----------------------------kill_dead_locals---------------------------------
753 // Detect any locals which are known to be dead, and force them to top.
754 void GraphKit::kill_dead_locals() {
755 // Consult the liveness information for the locals. If any
756 // of them are unused, then they can be replaced by top(). This
757 // should help register allocation time and cut down on the size
758 // of the deoptimization information.
759
760 // This call is made from many of the bytecode handling
761 // subroutines called from the Big Switch in do_one_bytecode.
762 // Every bytecode which might include a slow path is responsible
763 // for killing its dead locals. The more consistent we
764 // are about killing deads, the fewer useless phis will be
765 // constructed for them at various merge points.
766
767 // bci can be -1 (InvocationEntryBci). We return the entry
768 // liveness for the method.
769
770 if (method() == NULL || method()->code_size() == 0) {
771 // We are building a graph for a call to a native method.
772 // All locals are live.
773 return;
774 }
775
776 ResourceMark rm;
777
778 // Consult the liveness information for the locals. If any
779 // of them are unused, then they can be replaced by top(). This
780 // should help register allocation time and cut down on the size
781 // of the deoptimization information.
782 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
783
784 int len = (int)live_locals.size();
785 assert(len <= jvms()->loc_size(), "too many live locals");
786 for (int local = 0; local < len; local++) {
787 if (!live_locals.at(local)) {
788 set_local(local, top());
789 }
790 }
791 }
792
793 #ifdef ASSERT
794 //-------------------------dead_locals_are_killed------------------------------
795 // Return true if all dead locals are set to top in the map.
796 // Used to assert "clean" debug info at various points.
797 bool GraphKit::dead_locals_are_killed() {
798 if (method() == NULL || method()->code_size() == 0) {
799 // No locals need to be dead, so all is as it should be.
800 return true;
801 }
802
803 // Make sure somebody called kill_dead_locals upstream.
804 ResourceMark rm;
805 for (JVMState* jvms = this->jvms(); jvms != NULL; jvms = jvms->caller()) {
806 if (jvms->loc_size() == 0) continue; // no locals to consult
807 SafePointNode* map = jvms->map();
808 ciMethod* method = jvms->method();
809 int bci = jvms->bci();
810 if (jvms == this->jvms()) {
811 bci = this->bci(); // it might not yet be synched
812 }
813 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
814 int len = (int)live_locals.size();
815 if (!live_locals.is_valid() || len == 0)
816 // This method is trivial, or is poisoned by a breakpoint.
817 return true;
818 assert(len == jvms->loc_size(), "live map consistent with locals map");
819 for (int local = 0; local < len; local++) {
820 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
821 if (PrintMiscellaneous && (Verbose || WizardMode)) {
822 tty->print_cr("Zombie local %d: ", local);
823 jvms->dump();
824 }
825 return false;
826 }
827 }
828 }
829 return true;
830 }
831
832 #endif //ASSERT
833
834 // Helper function for enforcing certain bytecodes to reexecute if
835 // deoptimization happens
836 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
837 ciMethod* cur_method = jvms->method();
838 int cur_bci = jvms->bci();
839 if (cur_method != NULL && cur_bci != InvocationEntryBci) {
840 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
841 return Interpreter::bytecode_should_reexecute(code) ||
842 (is_anewarray && (code == Bytecodes::_multianewarray));
843 // Reexecute _multianewarray bytecode which was replaced with
844 // sequence of [a]newarray. See Parse::do_multianewarray().
845 //
846 // Note: interpreter should not have it set since this optimization
847 // is limited by dimensions and guarded by flag so in some cases
848 // multianewarray() runtime calls will be generated and
849 // the bytecode should not be reexecutes (stack will not be reset).
850 } else {
851 return false;
852 }
853 }
854
855 // Helper function for adding JVMState and debug information to node
856 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
857 // Add the safepoint edges to the call (or other safepoint).
858
859 // Make sure dead locals are set to top. This
860 // should help register allocation time and cut down on the size
861 // of the deoptimization information.
862 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
863
864 // Walk the inline list to fill in the correct set of JVMState's
865 // Also fill in the associated edges for each JVMState.
866
867 // If the bytecode needs to be reexecuted we need to put
868 // the arguments back on the stack.
869 const bool should_reexecute = jvms()->should_reexecute();
870 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
871
872 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
873 // undefined if the bci is different. This is normal for Parse but it
874 // should not happen for LibraryCallKit because only one bci is processed.
875 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
876 "in LibraryCallKit the reexecute bit should not change");
877
878 // If we are guaranteed to throw, we can prune everything but the
879 // input to the current bytecode.
880 bool can_prune_locals = false;
881 uint stack_slots_not_pruned = 0;
882 int inputs = 0, depth = 0;
883 if (must_throw) {
884 assert(method() == youngest_jvms->method(), "sanity");
885 if (compute_stack_effects(inputs, depth)) {
886 can_prune_locals = true;
887 stack_slots_not_pruned = inputs;
888 }
889 }
890
891 if (env()->should_retain_local_variables()) {
892 // At any safepoint, this method can get breakpointed, which would
893 // then require an immediate deoptimization.
894 can_prune_locals = false; // do not prune locals
895 stack_slots_not_pruned = 0;
896 }
897
898 // do not scribble on the input jvms
899 JVMState* out_jvms = youngest_jvms->clone_deep(C);
900 call->set_jvms(out_jvms); // Start jvms list for call node
901
902 // For a known set of bytecodes, the interpreter should reexecute them if
903 // deoptimization happens. We set the reexecute state for them here
904 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
905 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
906 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
907 }
908
909 // Presize the call:
910 DEBUG_ONLY(uint non_debug_edges = call->req());
911 call->add_req_batch(top(), youngest_jvms->debug_depth());
912 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
913
914 // Set up edges so that the call looks like this:
915 // Call [state:] ctl io mem fptr retadr
916 // [parms:] parm0 ... parmN
917 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
918 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
919 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
920 // Note that caller debug info precedes callee debug info.
921
922 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
923 uint debug_ptr = call->req();
924
925 // Loop over the map input edges associated with jvms, add them
926 // to the call node, & reset all offsets to match call node array.
927 for (JVMState* in_jvms = youngest_jvms; in_jvms != NULL; ) {
928 uint debug_end = debug_ptr;
929 uint debug_start = debug_ptr - in_jvms->debug_size();
930 debug_ptr = debug_start; // back up the ptr
931
932 uint p = debug_start; // walks forward in [debug_start, debug_end)
933 uint j, k, l;
934 SafePointNode* in_map = in_jvms->map();
935 out_jvms->set_map(call);
936
937 if (can_prune_locals) {
938 assert(in_jvms->method() == out_jvms->method(), "sanity");
939 // If the current throw can reach an exception handler in this JVMS,
940 // then we must keep everything live that can reach that handler.
941 // As a quick and dirty approximation, we look for any handlers at all.
942 if (in_jvms->method()->has_exception_handlers()) {
943 can_prune_locals = false;
944 }
945 }
946
947 // Add the Locals
948 k = in_jvms->locoff();
949 l = in_jvms->loc_size();
950 out_jvms->set_locoff(p);
951 if (!can_prune_locals) {
952 for (j = 0; j < l; j++)
953 call->set_req(p++, in_map->in(k+j));
954 } else {
955 p += l; // already set to top above by add_req_batch
956 }
957
958 // Add the Expression Stack
959 k = in_jvms->stkoff();
960 l = in_jvms->sp();
961 out_jvms->set_stkoff(p);
962 if (!can_prune_locals) {
963 for (j = 0; j < l; j++)
964 call->set_req(p++, in_map->in(k+j));
965 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
966 // Divide stack into {S0,...,S1}, where S0 is set to top.
967 uint s1 = stack_slots_not_pruned;
968 stack_slots_not_pruned = 0; // for next iteration
969 if (s1 > l) s1 = l;
970 uint s0 = l - s1;
971 p += s0; // skip the tops preinstalled by add_req_batch
972 for (j = s0; j < l; j++)
973 call->set_req(p++, in_map->in(k+j));
974 } else {
975 p += l; // already set to top above by add_req_batch
976 }
977
978 // Add the Monitors
979 k = in_jvms->monoff();
980 l = in_jvms->mon_size();
981 out_jvms->set_monoff(p);
982 for (j = 0; j < l; j++)
983 call->set_req(p++, in_map->in(k+j));
984
985 // Copy any scalar object fields.
986 k = in_jvms->scloff();
987 l = in_jvms->scl_size();
988 out_jvms->set_scloff(p);
989 for (j = 0; j < l; j++)
990 call->set_req(p++, in_map->in(k+j));
991
992 // Finish the new jvms.
993 out_jvms->set_endoff(p);
994
995 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
996 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
997 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
998 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
999 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1000 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1001
1002 // Update the two tail pointers in parallel.
1003 out_jvms = out_jvms->caller();
1004 in_jvms = in_jvms->caller();
1005 }
1006
1007 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1008
1009 // Test the correctness of JVMState::debug_xxx accessors:
1010 assert(call->jvms()->debug_start() == non_debug_edges, "");
1011 assert(call->jvms()->debug_end() == call->req(), "");
1012 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1013 }
1014
1015 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1016 Bytecodes::Code code = java_bc();
1017 if (code == Bytecodes::_wide) {
1018 code = method()->java_code_at_bci(bci() + 1);
1019 }
1020
1021 BasicType rtype = T_ILLEGAL;
1022 int rsize = 0;
1023
1024 if (code != Bytecodes::_illegal) {
1025 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1026 rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1027 if (rtype < T_CONFLICT)
1028 rsize = type2size[rtype];
1029 }
1030
1031 switch (code) {
1032 case Bytecodes::_illegal:
1033 return false;
1034
1035 case Bytecodes::_ldc:
1036 case Bytecodes::_ldc_w:
1037 case Bytecodes::_ldc2_w:
1038 inputs = 0;
1039 break;
1040
1041 case Bytecodes::_dup: inputs = 1; break;
1042 case Bytecodes::_dup_x1: inputs = 2; break;
1043 case Bytecodes::_dup_x2: inputs = 3; break;
1044 case Bytecodes::_dup2: inputs = 2; break;
1045 case Bytecodes::_dup2_x1: inputs = 3; break;
1046 case Bytecodes::_dup2_x2: inputs = 4; break;
1047 case Bytecodes::_swap: inputs = 2; break;
1048 case Bytecodes::_arraylength: inputs = 1; break;
1049
1050 case Bytecodes::_getstatic:
1051 case Bytecodes::_putstatic:
1052 case Bytecodes::_getfield:
1053 case Bytecodes::_putfield:
1054 {
1055 bool ignored_will_link;
1056 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1057 int size = field->type()->size();
1058 bool is_get = (depth >= 0), is_static = (depth & 1);
1059 inputs = (is_static ? 0 : 1);
1060 if (is_get) {
1061 depth = size - inputs;
1062 } else {
1063 inputs += size; // putxxx pops the value from the stack
1064 depth = - inputs;
1065 }
1066 }
1067 break;
1068
1069 case Bytecodes::_invokevirtual:
1070 case Bytecodes::_invokespecial:
1071 case Bytecodes::_invokestatic:
1072 case Bytecodes::_invokedynamic:
1073 case Bytecodes::_invokeinterface:
1074 {
1075 bool ignored_will_link;
1076 ciSignature* declared_signature = NULL;
1077 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1078 assert(declared_signature != NULL, "cannot be null");
1079 inputs = declared_signature->arg_size_for_bc(code);
1080 int size = declared_signature->return_type()->size();
1081 depth = size - inputs;
1082 }
1083 break;
1084
1085 case Bytecodes::_multianewarray:
1086 {
1087 ciBytecodeStream iter(method());
1088 iter.reset_to_bci(bci());
1089 iter.next();
1090 inputs = iter.get_dimensions();
1091 assert(rsize == 1, "");
1092 depth = rsize - inputs;
1093 }
1094 break;
1095
1096 case Bytecodes::_withfield: {
1097 bool ignored_will_link;
1098 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1099 int size = field->type()->size();
1100 inputs = size+1;
1101 depth = rsize - inputs;
1102 break;
1103 }
1104
1105 case Bytecodes::_ireturn:
1106 case Bytecodes::_lreturn:
1107 case Bytecodes::_freturn:
1108 case Bytecodes::_dreturn:
1109 case Bytecodes::_areturn:
1110 assert(rsize == -depth, "");
1111 inputs = rsize;
1112 break;
1113
1114 case Bytecodes::_jsr:
1115 case Bytecodes::_jsr_w:
1116 inputs = 0;
1117 depth = 1; // S.B. depth=1, not zero
1118 break;
1119
1120 default:
1121 // bytecode produces a typed result
1122 inputs = rsize - depth;
1123 assert(inputs >= 0, "");
1124 break;
1125 }
1126
1127 #ifdef ASSERT
1128 // spot check
1129 int outputs = depth + inputs;
1130 assert(outputs >= 0, "sanity");
1131 switch (code) {
1132 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1133 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1134 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1135 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1136 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1137 default: break;
1138 }
1139 #endif //ASSERT
1140
1141 return true;
1142 }
1143
1144
1145
1146 //------------------------------basic_plus_adr---------------------------------
1147 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1148 // short-circuit a common case
1149 if (offset == intcon(0)) return ptr;
1150 return _gvn.transform( new AddPNode(base, ptr, offset) );
1151 }
1152
1153 Node* GraphKit::ConvI2L(Node* offset) {
1154 // short-circuit a common case
1155 jint offset_con = find_int_con(offset, Type::OffsetBot);
1156 if (offset_con != Type::OffsetBot) {
1157 return longcon((jlong) offset_con);
1158 }
1159 return _gvn.transform( new ConvI2LNode(offset));
1160 }
1161
1162 Node* GraphKit::ConvI2UL(Node* offset) {
1163 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1164 if (offset_con != (juint) Type::OffsetBot) {
1165 return longcon((julong) offset_con);
1166 }
1167 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1168 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1169 return _gvn.transform( new AndLNode(conv, mask) );
1170 }
1171
1172 Node* GraphKit::ConvL2I(Node* offset) {
1173 // short-circuit a common case
1174 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1175 if (offset_con != (jlong)Type::OffsetBot) {
1176 return intcon((int) offset_con);
1177 }
1178 return _gvn.transform( new ConvL2INode(offset));
1179 }
1180
1181 //-------------------------load_object_klass-----------------------------------
1182 Node* GraphKit::load_object_klass(Node* obj) {
1183 // Special-case a fresh allocation to avoid building nodes:
1184 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1185 if (akls != NULL) return akls;
1186 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1187 return _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeKlassPtr::OBJECT));
1188 }
1189
1190 //-------------------------load_array_length-----------------------------------
1191 Node* GraphKit::load_array_length(Node* array) {
1192 // Special-case a fresh allocation to avoid building nodes:
1193 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array, &_gvn);
1194 Node *alen;
1195 if (alloc == NULL) {
1196 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1197 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1198 } else {
1199 alen = alloc->Ideal_length();
1200 Node* ccast = alloc->make_ideal_length(_gvn.type(array)->is_oopptr(), &_gvn);
1201 if (ccast != alen) {
1202 alen = _gvn.transform(ccast);
1203 }
1204 }
1205 return alen;
1206 }
1207
1208 //------------------------------do_null_check----------------------------------
1209 // Helper function to do a NULL pointer check. Returned value is
1210 // the incoming address with NULL casted away. You are allowed to use the
1211 // not-null value only if you are control dependent on the test.
1212 #ifndef PRODUCT
1213 extern int explicit_null_checks_inserted,
1214 explicit_null_checks_elided;
1215 #endif
1216 Node* GraphKit::null_check_common(Node* value, BasicType type,
1217 // optional arguments for variations:
1218 bool assert_null,
1219 Node* *null_control,
1220 bool speculative) {
1221 assert(!assert_null || null_control == NULL, "not both at once");
1222 if (stopped()) return top();
1223 NOT_PRODUCT(explicit_null_checks_inserted++);
1224
1225 // Construct NULL check
1226 Node *chk = NULL;
1227 switch(type) {
1228 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1229 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1230 case T_INLINE_TYPE : // fall through
1231 case T_ARRAY : // fall through
1232 type = T_OBJECT; // simplify further tests
1233 case T_OBJECT : {
1234 const Type *t = _gvn.type( value );
1235
1236 const TypeOopPtr* tp = t->isa_oopptr();
1237 if (tp != NULL && tp->klass() != NULL && !tp->klass()->is_loaded()
1238 // Only for do_null_check, not any of its siblings:
1239 && !assert_null && null_control == NULL) {
1240 // Usually, any field access or invocation on an unloaded oop type
1241 // will simply fail to link, since the statically linked class is
1242 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1243 // the static class is loaded but the sharper oop type is not.
1244 // Rather than checking for this obscure case in lots of places,
1245 // we simply observe that a null check on an unloaded class
1246 // will always be followed by a nonsense operation, so we
1247 // can just issue the uncommon trap here.
1248 // Our access to the unloaded class will only be correct
1249 // after it has been loaded and initialized, which requires
1250 // a trip through the interpreter.
1251 #ifndef PRODUCT
1252 if (WizardMode) { tty->print("Null check of unloaded "); tp->klass()->print(); tty->cr(); }
1253 #endif
1254 uncommon_trap(Deoptimization::Reason_unloaded,
1255 Deoptimization::Action_reinterpret,
1256 tp->klass(), "!loaded");
1257 return top();
1258 }
1259
1260 if (assert_null) {
1261 // See if the type is contained in NULL_PTR.
1262 // If so, then the value is already null.
1263 if (t->higher_equal(TypePtr::NULL_PTR)) {
1264 NOT_PRODUCT(explicit_null_checks_elided++);
1265 return value; // Elided null assert quickly!
1266 }
1267 } else {
1268 // See if mixing in the NULL pointer changes type.
1269 // If so, then the NULL pointer was not allowed in the original
1270 // type. In other words, "value" was not-null.
1271 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1272 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1273 NOT_PRODUCT(explicit_null_checks_elided++);
1274 return value; // Elided null check quickly!
1275 }
1276 }
1277 chk = new CmpPNode( value, null() );
1278 break;
1279 }
1280
1281 default:
1282 fatal("unexpected type: %s", type2name(type));
1283 }
1284 assert(chk != NULL, "sanity check");
1285 chk = _gvn.transform(chk);
1286
1287 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1288 BoolNode *btst = new BoolNode( chk, btest);
1289 Node *tst = _gvn.transform( btst );
1290
1291 //-----------
1292 // if peephole optimizations occurred, a prior test existed.
1293 // If a prior test existed, maybe it dominates as we can avoid this test.
1294 if (tst != btst && type == T_OBJECT) {
1295 // At this point we want to scan up the CFG to see if we can
1296 // find an identical test (and so avoid this test altogether).
1297 Node *cfg = control();
1298 int depth = 0;
1299 while( depth < 16 ) { // Limit search depth for speed
1300 if( cfg->Opcode() == Op_IfTrue &&
1301 cfg->in(0)->in(1) == tst ) {
1302 // Found prior test. Use "cast_not_null" to construct an identical
1303 // CastPP (and hence hash to) as already exists for the prior test.
1304 // Return that casted value.
1305 if (assert_null) {
1306 replace_in_map(value, null());
1307 return null(); // do not issue the redundant test
1308 }
1309 Node *oldcontrol = control();
1310 set_control(cfg);
1311 Node *res = cast_not_null(value);
1312 set_control(oldcontrol);
1313 NOT_PRODUCT(explicit_null_checks_elided++);
1314 return res;
1315 }
1316 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1317 if (cfg == NULL) break; // Quit at region nodes
1318 depth++;
1319 }
1320 }
1321
1322 //-----------
1323 // Branch to failure if null
1324 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1325 Deoptimization::DeoptReason reason;
1326 if (assert_null) {
1327 reason = Deoptimization::reason_null_assert(speculative);
1328 } else if (type == T_OBJECT) {
1329 reason = Deoptimization::reason_null_check(speculative);
1330 } else {
1331 reason = Deoptimization::Reason_div0_check;
1332 }
1333 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1334 // ciMethodData::has_trap_at will return a conservative -1 if any
1335 // must-be-null assertion has failed. This could cause performance
1336 // problems for a method after its first do_null_assert failure.
1337 // Consider using 'Reason_class_check' instead?
1338
1339 // To cause an implicit null check, we set the not-null probability
1340 // to the maximum (PROB_MAX). For an explicit check the probability
1341 // is set to a smaller value.
1342 if (null_control != NULL || too_many_traps(reason)) {
1343 // probability is less likely
1344 ok_prob = PROB_LIKELY_MAG(3);
1345 } else if (!assert_null &&
1346 (ImplicitNullCheckThreshold > 0) &&
1347 method() != NULL &&
1348 (method()->method_data()->trap_count(reason)
1349 >= (uint)ImplicitNullCheckThreshold)) {
1350 ok_prob = PROB_LIKELY_MAG(3);
1351 }
1352
1353 if (null_control != NULL) {
1354 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1355 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1356 set_control( _gvn.transform( new IfTrueNode(iff)));
1357 #ifndef PRODUCT
1358 if (null_true == top()) {
1359 explicit_null_checks_elided++;
1360 }
1361 #endif
1362 (*null_control) = null_true;
1363 } else {
1364 BuildCutout unless(this, tst, ok_prob);
1365 // Check for optimizer eliding test at parse time
1366 if (stopped()) {
1367 // Failure not possible; do not bother making uncommon trap.
1368 NOT_PRODUCT(explicit_null_checks_elided++);
1369 } else if (assert_null) {
1370 uncommon_trap(reason,
1371 Deoptimization::Action_make_not_entrant,
1372 NULL, "assert_null");
1373 } else {
1374 replace_in_map(value, zerocon(type));
1375 builtin_throw(reason);
1376 }
1377 }
1378
1379 // Must throw exception, fall-thru not possible?
1380 if (stopped()) {
1381 return top(); // No result
1382 }
1383
1384 if (assert_null) {
1385 // Cast obj to null on this path.
1386 replace_in_map(value, zerocon(type));
1387 return zerocon(type);
1388 }
1389
1390 // Cast obj to not-null on this path, if there is no null_control.
1391 // (If there is a null_control, a non-null value may come back to haunt us.)
1392 if (type == T_OBJECT) {
1393 Node* cast = cast_not_null(value, false);
1394 if (null_control == NULL || (*null_control) == top())
1395 replace_in_map(value, cast);
1396 value = cast;
1397 }
1398
1399 return value;
1400 }
1401
1402 Node* GraphKit::null2default(Node* value, ciInlineKlass* vk) {
1403 assert(!vk->is_scalarizable(), "Should only be used for non scalarizable inline klasses");
1404 Node* null_ctl = top();
1405 value = null_check_oop(value, &null_ctl);
1406 if (!null_ctl->is_top()) {
1407 // Return default value if oop is null
1408 Node* region = new RegionNode(3);
1409 region->init_req(1, control());
1410 region->init_req(2, null_ctl);
1411 value = PhiNode::make(region, value, TypeInstPtr::make(TypePtr::BotPTR, vk));
1412 value->set_req(2, InlineTypeNode::default_oop(gvn(), vk));
1413 set_control(gvn().transform(region));
1414 value = gvn().transform(value);
1415 }
1416 return value;
1417 }
1418
1419 //------------------------------cast_not_null----------------------------------
1420 // Cast obj to not-null on this path
1421 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1422 if (obj->is_InlineType()) {
1423 return obj;
1424 }
1425 const Type *t = _gvn.type(obj);
1426 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1427 // Object is already not-null?
1428 if( t == t_not_null ) return obj;
1429
1430 Node *cast = new CastPPNode(obj,t_not_null);
1431 cast->init_req(0, control());
1432 cast = _gvn.transform( cast );
1433
1434 if (t->is_inlinetypeptr() && t->inline_klass()->is_scalarizable()) {
1435 // Scalarize inline type now that we know it's non-null
1436 cast = InlineTypeNode::make_from_oop(this, cast, t->inline_klass())->as_ptr(&gvn());
1437 }
1438
1439 // Scan for instances of 'obj' in the current JVM mapping.
1440 // These instances are known to be not-null after the test.
1441 if (do_replace_in_map)
1442 replace_in_map(obj, cast);
1443
1444 return cast; // Return casted value
1445 }
1446
1447 // Sometimes in intrinsics, we implicitly know an object is not null
1448 // (there's no actual null check) so we can cast it to not null. In
1449 // the course of optimizations, the input to the cast can become null.
1450 // In that case that data path will die and we need the control path
1451 // to become dead as well to keep the graph consistent. So we have to
1452 // add a check for null for which one branch can't be taken. It uses
1453 // an Opaque4 node that will cause the check to be removed after loop
1454 // opts so the test goes away and the compiled code doesn't execute a
1455 // useless check.
1456 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1457 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1458 return value;
1459 }
1460 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1461 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1462 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1463 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1464 _gvn.set_type(iff, iff->Value(&_gvn));
1465 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1466 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1467 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1468 C->root()->add_req(halt);
1469 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1470 set_control(if_t);
1471 return cast_not_null(value, do_replace_in_map);
1472 }
1473
1474
1475 //--------------------------replace_in_map-------------------------------------
1476 void GraphKit::replace_in_map(Node* old, Node* neww) {
1477 if (old == neww) {
1478 return;
1479 }
1480
1481 map()->replace_edge(old, neww);
1482
1483 // Note: This operation potentially replaces any edge
1484 // on the map. This includes locals, stack, and monitors
1485 // of the current (innermost) JVM state.
1486
1487 // don't let inconsistent types from profiling escape this
1488 // method
1489
1490 const Type* told = _gvn.type(old);
1491 const Type* tnew = _gvn.type(neww);
1492
1493 if (!tnew->higher_equal(told)) {
1494 return;
1495 }
1496
1497 map()->record_replaced_node(old, neww);
1498 }
1499
1500
1501 //=============================================================================
1502 //--------------------------------memory---------------------------------------
1503 Node* GraphKit::memory(uint alias_idx) {
1504 MergeMemNode* mem = merged_memory();
1505 Node* p = mem->memory_at(alias_idx);
1506 _gvn.set_type(p, Type::MEMORY); // must be mapped
1507 return p;
1508 }
1509
1510 //-----------------------------reset_memory------------------------------------
1511 Node* GraphKit::reset_memory() {
1512 Node* mem = map()->memory();
1513 // do not use this node for any more parsing!
1514 debug_only( map()->set_memory((Node*)NULL) );
1515 return _gvn.transform( mem );
1516 }
1517
1518 //------------------------------set_all_memory---------------------------------
1519 void GraphKit::set_all_memory(Node* newmem) {
1520 Node* mergemem = MergeMemNode::make(newmem);
1521 gvn().set_type_bottom(mergemem);
1522 map()->set_memory(mergemem);
1523 }
1524
1525 //------------------------------set_all_memory_call----------------------------
1526 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1527 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1528 set_all_memory(newmem);
1529 }
1530
1531 //=============================================================================
1532 //
1533 // parser factory methods for MemNodes
1534 //
1535 // These are layered on top of the factory methods in LoadNode and StoreNode,
1536 // and integrate with the parser's memory state and _gvn engine.
1537 //
1538
1539 // factory methods in "int adr_idx"
1540 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1541 int adr_idx,
1542 MemNode::MemOrd mo,
1543 LoadNode::ControlDependency control_dependency,
1544 bool require_atomic_access,
1545 bool unaligned,
1546 bool mismatched,
1547 bool unsafe,
1548 uint8_t barrier_data) {
1549 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1550 const TypePtr* adr_type = NULL; // debug-mode-only argument
1551 debug_only(adr_type = C->get_adr_type(adr_idx));
1552 Node* mem = memory(adr_idx);
1553 Node* ld;
1554 if (require_atomic_access && bt == T_LONG) {
1555 ld = LoadLNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1556 } else if (require_atomic_access && bt == T_DOUBLE) {
1557 ld = LoadDNode::make_atomic(ctl, mem, adr, adr_type, t, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1558 } else {
1559 ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, unaligned, mismatched, unsafe, barrier_data);
1560 }
1561 ld = _gvn.transform(ld);
1562
1563 if (((bt == T_OBJECT || bt == T_INLINE_TYPE) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1564 // Improve graph before escape analysis and boxing elimination.
1565 record_for_igvn(ld);
1566 }
1567 return ld;
1568 }
1569
1570 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1571 int adr_idx,
1572 MemNode::MemOrd mo,
1573 bool require_atomic_access,
1574 bool unaligned,
1575 bool mismatched,
1576 bool unsafe) {
1577 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1578 const TypePtr* adr_type = NULL;
1579 debug_only(adr_type = C->get_adr_type(adr_idx));
1580 Node *mem = memory(adr_idx);
1581 Node* st;
1582 if (require_atomic_access && bt == T_LONG) {
1583 st = StoreLNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1584 } else if (require_atomic_access && bt == T_DOUBLE) {
1585 st = StoreDNode::make_atomic(ctl, mem, adr, adr_type, val, mo);
1586 } else {
1587 st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo);
1588 }
1589 if (unaligned) {
1590 st->as_Store()->set_unaligned_access();
1591 }
1592 if (mismatched) {
1593 st->as_Store()->set_mismatched_access();
1594 }
1595 if (unsafe) {
1596 st->as_Store()->set_unsafe_access();
1597 }
1598 st = _gvn.transform(st);
1599 set_memory(st, adr_idx);
1600 // Back-to-back stores can only remove intermediate store with DU info
1601 // so push on worklist for optimizer.
1602 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1603 record_for_igvn(st);
1604
1605 return st;
1606 }
1607
1608 Node* GraphKit::access_store_at(Node* obj,
1609 Node* adr,
1610 const TypePtr* adr_type,
1611 Node* val,
1612 const Type* val_type,
1613 BasicType bt,
1614 DecoratorSet decorators,
1615 bool safe_for_replace) {
1616 // Transformation of a value which could be NULL pointer (CastPP #NULL)
1617 // could be delayed during Parse (for example, in adjust_map_after_if()).
1618 // Execute transformation here to avoid barrier generation in such case.
1619 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1620 val = _gvn.makecon(TypePtr::NULL_PTR);
1621 }
1622
1623 if (stopped()) {
1624 return top(); // Dead path ?
1625 }
1626
1627 assert(val != NULL, "not dead path");
1628 if (val->is_InlineType()) {
1629 // Store to non-flattened field. Buffer the inline type and make sure
1630 // the store is re-executed if the allocation triggers deoptimization.
1631 PreserveReexecuteState preexecs(this);
1632 jvms()->set_should_reexecute(true);
1633 val = val->as_InlineType()->buffer(this, safe_for_replace);
1634 }
1635
1636 C2AccessValuePtr addr(adr, adr_type);
1637 C2AccessValue value(val, val_type);
1638 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1639 if (access.is_raw()) {
1640 return _barrier_set->BarrierSetC2::store_at(access, value);
1641 } else {
1642 return _barrier_set->store_at(access, value);
1643 }
1644 }
1645
1646 Node* GraphKit::access_load_at(Node* obj, // containing obj
1647 Node* adr, // actual adress to store val at
1648 const TypePtr* adr_type,
1649 const Type* val_type,
1650 BasicType bt,
1651 DecoratorSet decorators,
1652 Node* ctl) {
1653 if (stopped()) {
1654 return top(); // Dead path ?
1655 }
1656
1657 C2AccessValuePtr addr(adr, adr_type);
1658 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
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_load(Node* adr, // actual adress to load val at
1667 const Type* val_type,
1668 BasicType bt,
1669 DecoratorSet decorators) {
1670 if (stopped()) {
1671 return top(); // Dead path ?
1672 }
1673
1674 C2AccessValuePtr addr(adr, NULL);
1675 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, NULL, addr);
1676 if (access.is_raw()) {
1677 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1678 } else {
1679 return _barrier_set->load_at(access, val_type);
1680 }
1681 }
1682
1683 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1684 Node* adr,
1685 const TypePtr* adr_type,
1686 int alias_idx,
1687 Node* expected_val,
1688 Node* new_val,
1689 const Type* value_type,
1690 BasicType bt,
1691 DecoratorSet decorators) {
1692 C2AccessValuePtr addr(adr, adr_type);
1693 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1694 bt, obj, addr, alias_idx);
1695 if (access.is_raw()) {
1696 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1697 } else {
1698 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1699 }
1700 }
1701
1702 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1703 Node* adr,
1704 const TypePtr* adr_type,
1705 int alias_idx,
1706 Node* expected_val,
1707 Node* new_val,
1708 const Type* value_type,
1709 BasicType bt,
1710 DecoratorSet decorators) {
1711 C2AccessValuePtr addr(adr, adr_type);
1712 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1713 bt, obj, addr, alias_idx);
1714 if (access.is_raw()) {
1715 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1716 } else {
1717 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1718 }
1719 }
1720
1721 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1722 Node* adr,
1723 const TypePtr* adr_type,
1724 int alias_idx,
1725 Node* new_val,
1726 const Type* value_type,
1727 BasicType bt,
1728 DecoratorSet decorators) {
1729 C2AccessValuePtr addr(adr, adr_type);
1730 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1731 bt, obj, addr, alias_idx);
1732 if (access.is_raw()) {
1733 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1734 } else {
1735 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1736 }
1737 }
1738
1739 Node* GraphKit::access_atomic_add_at(Node* obj,
1740 Node* adr,
1741 const TypePtr* adr_type,
1742 int alias_idx,
1743 Node* new_val,
1744 const Type* value_type,
1745 BasicType bt,
1746 DecoratorSet decorators) {
1747 C2AccessValuePtr addr(adr, adr_type);
1748 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1749 if (access.is_raw()) {
1750 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1751 } else {
1752 return _barrier_set->atomic_add_at(access, new_val, value_type);
1753 }
1754 }
1755
1756 void GraphKit::access_clone(Node* src_base, Node* dst_base, Node* countx, bool is_array) {
1757 return _barrier_set->clone(this, src_base, dst_base, countx, is_array);
1758 }
1759
1760 //-------------------------array_element_address-------------------------
1761 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1762 const TypeInt* sizetype, Node* ctrl) {
1763 uint shift = exact_log2(type2aelembytes(elembt));
1764 ciKlass* arytype_klass = _gvn.type(ary)->is_aryptr()->klass();
1765 if (arytype_klass != NULL && arytype_klass->is_flat_array_klass()) {
1766 ciFlatArrayKlass* vak = arytype_klass->as_flat_array_klass();
1767 shift = vak->log2_element_size();
1768 }
1769 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1770
1771 // short-circuit a common case (saves lots of confusing waste motion)
1772 jint idx_con = find_int_con(idx, -1);
1773 if (idx_con >= 0) {
1774 intptr_t offset = header + ((intptr_t)idx_con << shift);
1775 return basic_plus_adr(ary, offset);
1776 }
1777
1778 // must be correct type for alignment purposes
1779 Node* base = basic_plus_adr(ary, header);
1780 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1781 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1782 return basic_plus_adr(ary, base, scale);
1783 }
1784
1785 //-------------------------load_array_element-------------------------
1786 Node* GraphKit::load_array_element(Node* ctl, Node* ary, Node* idx, const TypeAryPtr* arytype) {
1787 const Type* elemtype = arytype->elem();
1788 BasicType elembt = elemtype->array_element_basic_type();
1789 assert(elembt != T_INLINE_TYPE, "inline types are not supported by this method");
1790 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1791 if (elembt == T_NARROWOOP) {
1792 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1793 }
1794 Node* ld = make_load(ctl, adr, elemtype, elembt, arytype, MemNode::unordered);
1795 return ld;
1796 }
1797
1798 //-------------------------set_arguments_for_java_call-------------------------
1799 // Arguments (pre-popped from the stack) are taken from the JVMS.
1800 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1801 PreserveReexecuteState preexecs(this);
1802 if (EnableValhalla) {
1803 // Make sure the call is re-executed, if buffering of inline type arguments triggers deoptimization
1804 jvms()->set_should_reexecute(true);
1805 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1806 inc_sp(arg_size);
1807 }
1808 // Add the call arguments
1809 const TypeTuple* domain = call->tf()->domain_sig();
1810 ExtendedSignature sig_cc = ExtendedSignature(call->method()->get_sig_cc(), SigEntryFilter());
1811 uint nargs = domain->cnt();
1812 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1813 Node* arg = argument(i-TypeFunc::Parms);
1814 const Type* t = domain->field_at(i);
1815 if (call->method()->has_scalarized_args() && t->is_inlinetypeptr() && !t->maybe_null()) {
1816 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1817 InlineTypeNode* vt = arg->as_InlineType();
1818 vt->pass_fields(this, call, sig_cc, idx);
1819 // If an inline type argument is passed as fields, attach the Method* to the call site
1820 // to be able to access the extended signature later via attached_method_before_pc().
1821 // For example, see CompiledMethod::preserve_callee_argument_oops().
1822 call->set_override_symbolic_info(true);
1823 continue;
1824 } else if (arg->is_InlineType()) {
1825 // Pass inline type argument via oop to callee
1826 arg = arg->as_InlineType()->buffer(this);
1827 if (!is_late_inline) {
1828 arg = arg->as_InlineTypePtr()->get_oop();
1829 }
1830 }
1831 call->init_req(idx++, arg);
1832 // Skip reserved arguments
1833 BasicType bt = t->basic_type();
1834 while (SigEntry::next_is_reserved(sig_cc, bt, true)) {
1835 call->init_req(idx++, top());
1836 if (type2size[bt] == 2) {
1837 call->init_req(idx++, top());
1838 }
1839 }
1840 }
1841 }
1842
1843 //---------------------------set_edges_for_java_call---------------------------
1844 // Connect a newly created call into the current JVMS.
1845 // A return value node (if any) is returned from set_edges_for_java_call.
1846 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1847
1848 // Add the predefined inputs:
1849 call->init_req( TypeFunc::Control, control() );
1850 call->init_req( TypeFunc::I_O , i_o() );
1851 call->init_req( TypeFunc::Memory , reset_memory() );
1852 call->init_req( TypeFunc::FramePtr, frameptr() );
1853 call->init_req( TypeFunc::ReturnAdr, top() );
1854
1855 add_safepoint_edges(call, must_throw);
1856
1857 Node* xcall = _gvn.transform(call);
1858
1859 if (xcall == top()) {
1860 set_control(top());
1861 return;
1862 }
1863 assert(xcall == call, "call identity is stable");
1864
1865 // Re-use the current map to produce the result.
1866
1867 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1868 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1869 set_all_memory_call(xcall, separate_io_proj);
1870
1871 //return xcall; // no need, caller already has it
1872 }
1873
1874 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1875 if (stopped()) return top(); // maybe the call folded up?
1876
1877 // Note: Since any out-of-line call can produce an exception,
1878 // we always insert an I_O projection from the call into the result.
1879
1880 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1881
1882 if (separate_io_proj) {
1883 // The caller requested separate projections be used by the fall
1884 // through and exceptional paths, so replace the projections for
1885 // the fall through path.
1886 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1887 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1888 }
1889
1890 // Capture the return value, if any.
1891 Node* ret;
1892 if (call->method() == NULL || call->method()->return_type()->basic_type() == T_VOID) {
1893 ret = top();
1894 } else if (call->tf()->returns_inline_type_as_fields()) {
1895 // Return of multiple values (inline type fields): we create a
1896 // InlineType node, each field is a projection from the call.
1897 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
1898 const Array<SigEntry>* sig_array = vk->extended_sig();
1899 GrowableArray<SigEntry> sig = GrowableArray<SigEntry>(sig_array->length());
1900 sig.appendAll(sig_array);
1901 ExtendedSignature sig_cc = ExtendedSignature(&sig, SigEntryFilter());
1902 uint base_input = TypeFunc::Parms + 1;
1903 ret = InlineTypeNode::make_from_multi(this, call, sig_cc, vk, base_input, false);
1904 } else {
1905 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1906 }
1907
1908 return ret;
1909 }
1910
1911 //--------------------set_predefined_input_for_runtime_call--------------------
1912 // Reading and setting the memory state is way conservative here.
1913 // The real problem is that I am not doing real Type analysis on memory,
1914 // so I cannot distinguish card mark stores from other stores. Across a GC
1915 // point the Store Barrier and the card mark memory has to agree. I cannot
1916 // have a card mark store and its barrier split across the GC point from
1917 // either above or below. Here I get that to happen by reading ALL of memory.
1918 // A better answer would be to separate out card marks from other memory.
1919 // For now, return the input memory state, so that it can be reused
1920 // after the call, if this call has restricted memory effects.
1921 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1922 // Set fixed predefined input arguments
1923 Node* memory = reset_memory();
1924 Node* m = narrow_mem == NULL ? memory : narrow_mem;
1925 call->init_req( TypeFunc::Control, control() );
1926 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1927 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1928 call->init_req( TypeFunc::FramePtr, frameptr() );
1929 call->init_req( TypeFunc::ReturnAdr, top() );
1930 return memory;
1931 }
1932
1933 //-------------------set_predefined_output_for_runtime_call--------------------
1934 // Set control and memory (not i_o) from the call.
1935 // If keep_mem is not NULL, use it for the output state,
1936 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1937 // If hook_mem is NULL, this call produces no memory effects at all.
1938 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1939 // then only that memory slice is taken from the call.
1940 // In the last case, we must put an appropriate memory barrier before
1941 // the call, so as to create the correct anti-dependencies on loads
1942 // preceding the call.
1943 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1944 Node* keep_mem,
1945 const TypePtr* hook_mem) {
1946 // no i/o
1947 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
1948 if (keep_mem) {
1949 // First clone the existing memory state
1950 set_all_memory(keep_mem);
1951 if (hook_mem != NULL) {
1952 // Make memory for the call
1953 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
1954 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
1955 // We also use hook_mem to extract specific effects from arraycopy stubs.
1956 set_memory(mem, hook_mem);
1957 }
1958 // ...else the call has NO memory effects.
1959
1960 // Make sure the call advertises its memory effects precisely.
1961 // This lets us build accurate anti-dependences in gcm.cpp.
1962 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
1963 "call node must be constructed correctly");
1964 } else {
1965 assert(hook_mem == NULL, "");
1966 // This is not a "slow path" call; all memory comes from the call.
1967 set_all_memory_call(call);
1968 }
1969 }
1970
1971 // Keep track of MergeMems feeding into other MergeMems
1972 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
1973 if (!mem->is_MergeMem()) {
1974 return;
1975 }
1976 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
1977 Node* use = i.get();
1978 if (use->is_MergeMem()) {
1979 wl.push(use);
1980 }
1981 }
1982 }
1983
1984 // Replace the call with the current state of the kit.
1985 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
1986 JVMState* ejvms = NULL;
1987 if (has_exceptions()) {
1988 ejvms = transfer_exceptions_into_jvms();
1989 }
1990
1991 ReplacedNodes replaced_nodes = map()->replaced_nodes();
1992 ReplacedNodes replaced_nodes_exception;
1993 Node* ex_ctl = top();
1994
1995 SafePointNode* final_state = stop();
1996
1997 // Find all the needed outputs of this call
1998 CallProjections* callprojs = call->extract_projections(true);
1999
2000 Unique_Node_List wl;
2001 Node* init_mem = call->in(TypeFunc::Memory);
2002 Node* final_mem = final_state->in(TypeFunc::Memory);
2003 Node* final_ctl = final_state->in(TypeFunc::Control);
2004 Node* final_io = final_state->in(TypeFunc::I_O);
2005
2006 // Replace all the old call edges with the edges from the inlining result
2007 if (callprojs->fallthrough_catchproj != NULL) {
2008 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2009 }
2010 if (callprojs->fallthrough_memproj != NULL) {
2011 if (final_mem->is_MergeMem()) {
2012 // Parser's exits MergeMem was not transformed but may be optimized
2013 final_mem = _gvn.transform(final_mem);
2014 }
2015 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2016 add_mergemem_users_to_worklist(wl, final_mem);
2017 }
2018 if (callprojs->fallthrough_ioproj != NULL) {
2019 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2020 }
2021
2022 // Replace the result with the new result if it exists and is used
2023 if (callprojs->resproj[0] != NULL && result != NULL) {
2024 assert(callprojs->nb_resproj == 1, "unexpected number of results");
2025 C->gvn_replace_by(callprojs->resproj[0], result);
2026 }
2027
2028 if (ejvms == NULL) {
2029 // No exception edges to simply kill off those paths
2030 if (callprojs->catchall_catchproj != NULL) {
2031 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2032 }
2033 if (callprojs->catchall_memproj != NULL) {
2034 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2035 }
2036 if (callprojs->catchall_ioproj != NULL) {
2037 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2038 }
2039 // Replace the old exception object with top
2040 if (callprojs->exobj != NULL) {
2041 C->gvn_replace_by(callprojs->exobj, C->top());
2042 }
2043 } else {
2044 GraphKit ekit(ejvms);
2045
2046 // Load my combined exception state into the kit, with all phis transformed:
2047 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2048 replaced_nodes_exception = ex_map->replaced_nodes();
2049
2050 Node* ex_oop = ekit.use_exception_state(ex_map);
2051
2052 if (callprojs->catchall_catchproj != NULL) {
2053 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2054 ex_ctl = ekit.control();
2055 }
2056 if (callprojs->catchall_memproj != NULL) {
2057 Node* ex_mem = ekit.reset_memory();
2058 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2059 add_mergemem_users_to_worklist(wl, ex_mem);
2060 }
2061 if (callprojs->catchall_ioproj != NULL) {
2062 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2063 }
2064
2065 // Replace the old exception object with the newly created one
2066 if (callprojs->exobj != NULL) {
2067 C->gvn_replace_by(callprojs->exobj, ex_oop);
2068 }
2069 }
2070
2071 // Disconnect the call from the graph
2072 call->disconnect_inputs(NULL, C);
2073 C->gvn_replace_by(call, C->top());
2074
2075 // Clean up any MergeMems that feed other MergeMems since the
2076 // optimizer doesn't like that.
2077 while (wl.size() > 0) {
2078 _gvn.transform(wl.pop());
2079 }
2080
2081 if (callprojs->fallthrough_catchproj != NULL && !final_ctl->is_top() && do_replaced_nodes) {
2082 replaced_nodes.apply(C, final_ctl);
2083 }
2084 if (!ex_ctl->is_top() && do_replaced_nodes) {
2085 replaced_nodes_exception.apply(C, ex_ctl);
2086 }
2087 }
2088
2089
2090 //------------------------------increment_counter------------------------------
2091 // for statistics: increment a VM counter by 1
2092
2093 void GraphKit::increment_counter(address counter_addr) {
2094 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2095 increment_counter(adr1);
2096 }
2097
2098 void GraphKit::increment_counter(Node* counter_addr) {
2099 int adr_type = Compile::AliasIdxRaw;
2100 Node* ctrl = control();
2101 Node* cnt = make_load(ctrl, counter_addr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
2102 Node* incr = _gvn.transform(new AddINode(cnt, _gvn.intcon(1)));
2103 store_to_memory(ctrl, counter_addr, incr, T_INT, adr_type, MemNode::unordered);
2104 }
2105
2106
2107 //------------------------------uncommon_trap----------------------------------
2108 // Bail out to the interpreter in mid-method. Implemented by calling the
2109 // uncommon_trap blob. This helper function inserts a runtime call with the
2110 // right debug info.
2111 void GraphKit::uncommon_trap(int trap_request,
2112 ciKlass* klass, const char* comment,
2113 bool must_throw,
2114 bool keep_exact_action) {
2115 if (failing()) stop();
2116 if (stopped()) return; // trap reachable?
2117
2118 // Note: If ProfileTraps is true, and if a deopt. actually
2119 // occurs here, the runtime will make sure an MDO exists. There is
2120 // no need to call method()->ensure_method_data() at this point.
2121
2122 // Set the stack pointer to the right value for reexecution:
2123 set_sp(reexecute_sp());
2124
2125 #ifdef ASSERT
2126 if (!must_throw) {
2127 // Make sure the stack has at least enough depth to execute
2128 // the current bytecode.
2129 int inputs, ignored_depth;
2130 if (compute_stack_effects(inputs, ignored_depth)) {
2131 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2132 Bytecodes::name(java_bc()), sp(), inputs);
2133 }
2134 }
2135 #endif
2136
2137 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2138 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2139
2140 switch (action) {
2141 case Deoptimization::Action_maybe_recompile:
2142 case Deoptimization::Action_reinterpret:
2143 // Temporary fix for 6529811 to allow virtual calls to be sure they
2144 // get the chance to go from mono->bi->mega
2145 if (!keep_exact_action &&
2146 Deoptimization::trap_request_index(trap_request) < 0 &&
2147 too_many_recompiles(reason)) {
2148 // This BCI is causing too many recompilations.
2149 if (C->log() != NULL) {
2150 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2151 Deoptimization::trap_reason_name(reason),
2152 Deoptimization::trap_action_name(action));
2153 }
2154 action = Deoptimization::Action_none;
2155 trap_request = Deoptimization::make_trap_request(reason, action);
2156 } else {
2157 C->set_trap_can_recompile(true);
2158 }
2159 break;
2160 case Deoptimization::Action_make_not_entrant:
2161 C->set_trap_can_recompile(true);
2162 break;
2163 case Deoptimization::Action_none:
2164 case Deoptimization::Action_make_not_compilable:
2165 break;
2166 default:
2167 #ifdef ASSERT
2168 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2169 #endif
2170 break;
2171 }
2172
2173 if (TraceOptoParse) {
2174 char buf[100];
2175 tty->print_cr("Uncommon trap %s at bci:%d",
2176 Deoptimization::format_trap_request(buf, sizeof(buf),
2177 trap_request), bci());
2178 }
2179
2180 CompileLog* log = C->log();
2181 if (log != NULL) {
2182 int kid = (klass == NULL)? -1: log->identify(klass);
2183 log->begin_elem("uncommon_trap bci='%d'", bci());
2184 char buf[100];
2185 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2186 trap_request));
2187 if (kid >= 0) log->print(" klass='%d'", kid);
2188 if (comment != NULL) log->print(" comment='%s'", comment);
2189 log->end_elem();
2190 }
2191
2192 // Make sure any guarding test views this path as very unlikely
2193 Node *i0 = control()->in(0);
2194 if (i0 != NULL && i0->is_If()) { // Found a guarding if test?
2195 IfNode *iff = i0->as_If();
2196 float f = iff->_prob; // Get prob
2197 if (control()->Opcode() == Op_IfTrue) {
2198 if (f > PROB_UNLIKELY_MAG(4))
2199 iff->_prob = PROB_MIN;
2200 } else {
2201 if (f < PROB_LIKELY_MAG(4))
2202 iff->_prob = PROB_MAX;
2203 }
2204 }
2205
2206 // Clear out dead values from the debug info.
2207 kill_dead_locals();
2208
2209 // Now insert the uncommon trap subroutine call
2210 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2211 const TypePtr* no_memory_effects = NULL;
2212 // Pass the index of the class to be loaded
2213 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2214 (must_throw ? RC_MUST_THROW : 0),
2215 OptoRuntime::uncommon_trap_Type(),
2216 call_addr, "uncommon_trap", no_memory_effects,
2217 intcon(trap_request));
2218 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2219 "must extract request correctly from the graph");
2220 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2221
2222 call->set_req(TypeFunc::ReturnAdr, returnadr());
2223 // The debug info is the only real input to this call.
2224
2225 // Halt-and-catch fire here. The above call should never return!
2226 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2227 PRODUCT_ONLY(COMMA /*reachable*/false));
2228 _gvn.set_type_bottom(halt);
2229 root()->add_req(halt);
2230
2231 stop_and_kill_map();
2232 }
2233
2234
2235 //--------------------------just_allocated_object------------------------------
2236 // Report the object that was just allocated.
2237 // It must be the case that there are no intervening safepoints.
2238 // We use this to determine if an object is so "fresh" that
2239 // it does not require card marks.
2240 Node* GraphKit::just_allocated_object(Node* current_control) {
2241 Node* ctrl = current_control;
2242 // Object::<init> is invoked after allocation, most of invoke nodes
2243 // will be reduced, but a region node is kept in parse time, we check
2244 // the pattern and skip the region node if it degraded to a copy.
2245 if (ctrl != NULL && ctrl->is_Region() && ctrl->req() == 2 &&
2246 ctrl->as_Region()->is_copy()) {
2247 ctrl = ctrl->as_Region()->is_copy();
2248 }
2249 if (C->recent_alloc_ctl() == ctrl) {
2250 return C->recent_alloc_obj();
2251 }
2252 return NULL;
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 (java_bc() == Bytecodes::_aastore) {
2337 ciKlass* array_type = NULL;
2338 ciKlass* element_type = NULL;
2339 ProfilePtrKind element_ptr = ProfileMaybeNull;
2340 bool flat_array = true;
2341 bool null_free_array = true;
2342 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2343 exact_kls = element_type;
2344 ptr_kind = element_ptr;
2345 } else {
2346 if (!data->as_BitData()->null_seen()) {
2347 ptr_kind = ProfileNeverNull;
2348 } else {
2349 assert(data->is_ReceiverTypeData(), "bad profile data type");
2350 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2351 uint i = 0;
2352 for (; i < call->row_limit(); i++) {
2353 ciKlass* receiver = call->receiver(i);
2354 if (receiver != NULL) {
2355 break;
2356 }
2357 }
2358 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2359 }
2360 }
2361 }
2362 }
2363 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2364 }
2365
2366 /**
2367 * Record profiling data from argument profiling at an invoke with the
2368 * type system so that it can propagate it (speculation)
2369 *
2370 * @param dest_method target method for the call
2371 * @param bc what invoke bytecode is this?
2372 */
2373 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2374 if (!UseTypeSpeculation) {
2375 return;
2376 }
2377 const TypeFunc* tf = TypeFunc::make(dest_method);
2378 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2379 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2380 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2381 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2382 if (is_reference_type(targ->basic_type())) {
2383 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2384 ciKlass* better_type = NULL;
2385 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2386 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2387 }
2388 i++;
2389 }
2390 }
2391 }
2392
2393 /**
2394 * Record profiling data from parameter profiling at an invoke with
2395 * the type system so that it can propagate it (speculation)
2396 */
2397 void GraphKit::record_profiled_parameters_for_speculation() {
2398 if (!UseTypeSpeculation) {
2399 return;
2400 }
2401 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2402 if (_gvn.type(local(i))->isa_oopptr()) {
2403 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2404 ciKlass* better_type = NULL;
2405 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2406 record_profile_for_speculation(local(i), better_type, ptr_kind);
2407 }
2408 j++;
2409 }
2410 }
2411 }
2412
2413 /**
2414 * Record profiling data from return value profiling at an invoke with
2415 * the type system so that it can propagate it (speculation)
2416 */
2417 void GraphKit::record_profiled_return_for_speculation() {
2418 if (!UseTypeSpeculation) {
2419 return;
2420 }
2421 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2422 ciKlass* better_type = NULL;
2423 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2424 // If profiling reports a single type for the return value,
2425 // feed it to the type system so it can propagate it as a
2426 // speculative type
2427 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2428 }
2429 }
2430
2431 void GraphKit::round_double_result(ciMethod* dest_method) {
2432 if (Matcher::strict_fp_requires_explicit_rounding) {
2433 // If a strict caller invokes a non-strict callee, round a double result.
2434 // A non-strict method may return a double value which has an extended exponent,
2435 // but this must not be visible in a caller which is strict.
2436 BasicType result_type = dest_method->return_type()->basic_type();
2437 assert(method() != NULL, "must have caller context");
2438 if( result_type == T_DOUBLE && method()->is_strict() && !dest_method->is_strict() ) {
2439 // Destination method's return value is on top of stack
2440 // dstore_rounding() does gvn.transform
2441 Node *result = pop_pair();
2442 result = dstore_rounding(result);
2443 push_pair(result);
2444 }
2445 }
2446 }
2447
2448 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2449 if (Matcher::strict_fp_requires_explicit_rounding) {
2450 // (Note: TypeFunc::make has a cache that makes this fast.)
2451 const TypeFunc* tf = TypeFunc::make(dest_method);
2452 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2453 for (int j = 0; j < nargs; j++) {
2454 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2455 if (targ->basic_type() == T_DOUBLE) {
2456 // If any parameters are doubles, they must be rounded before
2457 // the call, dstore_rounding does gvn.transform
2458 Node *arg = argument(j);
2459 arg = dstore_rounding(arg);
2460 set_argument(j, arg);
2461 }
2462 }
2463 }
2464 }
2465
2466 // rounding for strict float precision conformance
2467 Node* GraphKit::precision_rounding(Node* n) {
2468 if (Matcher::strict_fp_requires_explicit_rounding) {
2469 #ifdef IA32
2470 if (_method->flags().is_strict() && UseSSE == 0) {
2471 return _gvn.transform(new RoundFloatNode(0, n));
2472 }
2473 #else
2474 Unimplemented();
2475 #endif // IA32
2476 }
2477 return n;
2478 }
2479
2480 // rounding for strict double precision conformance
2481 Node* GraphKit::dprecision_rounding(Node *n) {
2482 if (Matcher::strict_fp_requires_explicit_rounding) {
2483 #ifdef IA32
2484 if (_method->flags().is_strict() && UseSSE < 2) {
2485 return _gvn.transform(new RoundDoubleNode(0, n));
2486 }
2487 #else
2488 Unimplemented();
2489 #endif // IA32
2490 }
2491 return n;
2492 }
2493
2494 // rounding for non-strict double stores
2495 Node* GraphKit::dstore_rounding(Node* n) {
2496 if (Matcher::strict_fp_requires_explicit_rounding) {
2497 #ifdef IA32
2498 if (UseSSE < 2) {
2499 return _gvn.transform(new RoundDoubleNode(0, n));
2500 }
2501 #else
2502 Unimplemented();
2503 #endif // IA32
2504 }
2505 return n;
2506 }
2507
2508 //=============================================================================
2509 // Generate a fast path/slow path idiom. Graph looks like:
2510 // [foo] indicates that 'foo' is a parameter
2511 //
2512 // [in] NULL
2513 // \ /
2514 // CmpP
2515 // Bool ne
2516 // If
2517 // / \
2518 // True False-<2>
2519 // / |
2520 // / cast_not_null
2521 // Load | | ^
2522 // [fast_test] | |
2523 // gvn to opt_test | |
2524 // / \ | <1>
2525 // True False |
2526 // | \\ |
2527 // [slow_call] \[fast_result]
2528 // Ctl Val \ \
2529 // | \ \
2530 // Catch <1> \ \
2531 // / \ ^ \ \
2532 // Ex No_Ex | \ \
2533 // | \ \ | \ <2> \
2534 // ... \ [slow_res] | | \ [null_result]
2535 // \ \--+--+--- | |
2536 // \ | / \ | /
2537 // --------Region Phi
2538 //
2539 //=============================================================================
2540 // Code is structured as a series of driver functions all called 'do_XXX' that
2541 // call a set of helper functions. Helper functions first, then drivers.
2542
2543 //------------------------------null_check_oop---------------------------------
2544 // Null check oop. Set null-path control into Region in slot 3.
2545 // Make a cast-not-nullness use the other not-null control. Return cast.
2546 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2547 bool never_see_null,
2548 bool safe_for_replace,
2549 bool speculative) {
2550 // Initial NULL check taken path
2551 (*null_control) = top();
2552 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2553
2554 // Generate uncommon_trap:
2555 if (never_see_null && (*null_control) != top()) {
2556 // If we see an unexpected null at a check-cast we record it and force a
2557 // recompile; the offending check-cast will be compiled to handle NULLs.
2558 // If we see more than one offending BCI, then all checkcasts in the
2559 // method will be compiled to handle NULLs.
2560 PreserveJVMState pjvms(this);
2561 set_control(*null_control);
2562 replace_in_map(value, null());
2563 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2564 uncommon_trap(reason,
2565 Deoptimization::Action_make_not_entrant);
2566 (*null_control) = top(); // NULL path is dead
2567 }
2568 if ((*null_control) == top() && safe_for_replace) {
2569 replace_in_map(value, cast);
2570 }
2571
2572 // Cast away null-ness on the result
2573 return cast;
2574 }
2575
2576 //------------------------------opt_iff----------------------------------------
2577 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2578 // Return slow-path control.
2579 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2580 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2581
2582 // Fast path taken; set region slot 2
2583 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2584 region->init_req(2,fast_taken); // Capture fast-control
2585
2586 // Fast path not-taken, i.e. slow path
2587 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2588 return slow_taken;
2589 }
2590
2591 //-----------------------------make_runtime_call-------------------------------
2592 Node* GraphKit::make_runtime_call(int flags,
2593 const TypeFunc* call_type, address call_addr,
2594 const char* call_name,
2595 const TypePtr* adr_type,
2596 // The following parms are all optional.
2597 // The first NULL ends the list.
2598 Node* parm0, Node* parm1,
2599 Node* parm2, Node* parm3,
2600 Node* parm4, Node* parm5,
2601 Node* parm6, Node* parm7) {
2602 assert(call_addr != NULL, "must not call NULL targets");
2603
2604 // Slow-path call
2605 bool is_leaf = !(flags & RC_NO_LEAF);
2606 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2607 if (call_name == NULL) {
2608 assert(!is_leaf, "must supply name for leaf");
2609 call_name = OptoRuntime::stub_name(call_addr);
2610 }
2611 CallNode* call;
2612 if (!is_leaf) {
2613 call = new CallStaticJavaNode(call_type, call_addr, call_name,
2614 bci(), adr_type);
2615 } else if (flags & RC_NO_FP) {
2616 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2617 } else {
2618 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2619 }
2620
2621 // The following is similar to set_edges_for_java_call,
2622 // except that the memory effects of the call are restricted to AliasIdxRaw.
2623
2624 // Slow path call has no side-effects, uses few values
2625 bool wide_in = !(flags & RC_NARROW_MEM);
2626 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2627
2628 Node* prev_mem = NULL;
2629 if (wide_in) {
2630 prev_mem = set_predefined_input_for_runtime_call(call);
2631 } else {
2632 assert(!wide_out, "narrow in => narrow out");
2633 Node* narrow_mem = memory(adr_type);
2634 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2635 }
2636
2637 // Hook each parm in order. Stop looking at the first NULL.
2638 if (parm0 != NULL) { call->init_req(TypeFunc::Parms+0, parm0);
2639 if (parm1 != NULL) { call->init_req(TypeFunc::Parms+1, parm1);
2640 if (parm2 != NULL) { call->init_req(TypeFunc::Parms+2, parm2);
2641 if (parm3 != NULL) { call->init_req(TypeFunc::Parms+3, parm3);
2642 if (parm4 != NULL) { call->init_req(TypeFunc::Parms+4, parm4);
2643 if (parm5 != NULL) { call->init_req(TypeFunc::Parms+5, parm5);
2644 if (parm6 != NULL) { call->init_req(TypeFunc::Parms+6, parm6);
2645 if (parm7 != NULL) { call->init_req(TypeFunc::Parms+7, parm7);
2646 /* close each nested if ===> */ } } } } } } } }
2647 assert(call->in(call->req()-1) != NULL, "must initialize all parms");
2648
2649 if (!is_leaf) {
2650 // Non-leaves can block and take safepoints:
2651 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2652 }
2653 // Non-leaves can throw exceptions:
2654 if (has_io) {
2655 call->set_req(TypeFunc::I_O, i_o());
2656 }
2657
2658 if (flags & RC_UNCOMMON) {
2659 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2660 // (An "if" probability corresponds roughly to an unconditional count.
2661 // Sort of.)
2662 call->set_cnt(PROB_UNLIKELY_MAG(4));
2663 }
2664
2665 Node* c = _gvn.transform(call);
2666 assert(c == call, "cannot disappear");
2667
2668 if (wide_out) {
2669 // Slow path call has full side-effects.
2670 set_predefined_output_for_runtime_call(call);
2671 } else {
2672 // Slow path call has few side-effects, and/or sets few values.
2673 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2674 }
2675
2676 if (has_io) {
2677 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2678 }
2679 return call;
2680
2681 }
2682
2683 //------------------------------merge_memory-----------------------------------
2684 // Merge memory from one path into the current memory state.
2685 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2686 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2687 Node* old_slice = mms.force_memory();
2688 Node* new_slice = mms.memory2();
2689 if (old_slice != new_slice) {
2690 PhiNode* phi;
2691 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2692 if (mms.is_empty()) {
2693 // clone base memory Phi's inputs for this memory slice
2694 assert(old_slice == mms.base_memory(), "sanity");
2695 phi = PhiNode::make(region, NULL, Type::MEMORY, mms.adr_type(C));
2696 _gvn.set_type(phi, Type::MEMORY);
2697 for (uint i = 1; i < phi->req(); i++) {
2698 phi->init_req(i, old_slice->in(i));
2699 }
2700 } else {
2701 phi = old_slice->as_Phi(); // Phi was generated already
2702 }
2703 } else {
2704 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2705 _gvn.set_type(phi, Type::MEMORY);
2706 }
2707 phi->set_req(new_path, new_slice);
2708 mms.set_memory(phi);
2709 }
2710 }
2711 }
2712
2713 //------------------------------make_slow_call_ex------------------------------
2714 // Make the exception handler hookups for the slow call
2715 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2716 if (stopped()) return;
2717
2718 // Make a catch node with just two handlers: fall-through and catch-all
2719 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2720 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2721 Node* norm = _gvn.transform( new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci) );
2722 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2723
2724 { PreserveJVMState pjvms(this);
2725 set_control(excp);
2726 set_i_o(i_o);
2727
2728 if (excp != top()) {
2729 if (deoptimize) {
2730 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2731 uncommon_trap(Deoptimization::Reason_unhandled,
2732 Deoptimization::Action_none);
2733 } else {
2734 // Create an exception state also.
2735 // Use an exact type if the caller has a specific exception.
2736 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2737 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2738 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2739 }
2740 }
2741 }
2742
2743 // Get the no-exception control from the CatchNode.
2744 set_control(norm);
2745 }
2746
2747 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2748 Node* cmp = NULL;
2749 switch(bt) {
2750 case T_INT: cmp = new CmpINode(in1, in2); break;
2751 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2752 default: fatal("unexpected comparison type %s", type2name(bt));
2753 }
2754 gvn.transform(cmp);
2755 Node* bol = gvn.transform(new BoolNode(cmp, test));
2756 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2757 gvn.transform(iff);
2758 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2759 return iff;
2760 }
2761
2762 //-------------------------------gen_subtype_check-----------------------------
2763 // Generate a subtyping check. Takes as input the subtype and supertype.
2764 // Returns 2 values: sets the default control() to the true path and returns
2765 // the false path. Only reads invariant memory; sets no (visible) memory.
2766 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2767 // but that's not exposed to the optimizer. This call also doesn't take in an
2768 // Object; if you wish to check an Object you need to load the Object's class
2769 // prior to coming here.
2770 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn) {
2771 Compile* C = gvn.C;
2772 if ((*ctrl)->is_top()) {
2773 return C->top();
2774 }
2775
2776 // Fast check for identical types, perhaps identical constants.
2777 // The types can even be identical non-constants, in cases
2778 // involving Array.newInstance, Object.clone, etc.
2779 if (subklass == superklass)
2780 return C->top(); // false path is dead; no test needed.
2781
2782 if (gvn.type(superklass)->singleton()) {
2783 ciKlass* superk = gvn.type(superklass)->is_klassptr()->klass();
2784 ciKlass* subk = gvn.type(subklass)->is_klassptr()->klass();
2785
2786 // In the common case of an exact superklass, try to fold up the
2787 // test before generating code. You may ask, why not just generate
2788 // the code and then let it fold up? The answer is that the generated
2789 // code will necessarily include null checks, which do not always
2790 // completely fold away. If they are also needless, then they turn
2791 // into a performance loss. Example:
2792 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2793 // Here, the type of 'fa' is often exact, so the store check
2794 // of fa[1]=x will fold up, without testing the nullness of x.
2795 switch (C->static_subtype_check(superk, subk)) {
2796 case Compile::SSC_always_false:
2797 {
2798 Node* always_fail = *ctrl;
2799 *ctrl = gvn.C->top();
2800 return always_fail;
2801 }
2802 case Compile::SSC_always_true:
2803 return C->top();
2804 case Compile::SSC_easy_test:
2805 {
2806 // Just do a direct pointer compare and be done.
2807 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2808 *ctrl = gvn.transform(new IfTrueNode(iff));
2809 return gvn.transform(new IfFalseNode(iff));
2810 }
2811 case Compile::SSC_full_test:
2812 break;
2813 default:
2814 ShouldNotReachHere();
2815 }
2816 }
2817
2818 // %%% Possible further optimization: Even if the superklass is not exact,
2819 // if the subklass is the unique subtype of the superklass, the check
2820 // will always succeed. We could leave a dependency behind to ensure this.
2821
2822 // First load the super-klass's check-offset
2823 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2824 Node* m = C->immutable_memory();
2825 Node *chk_off = gvn.transform(new LoadINode(NULL, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2826 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2827 bool might_be_cache = (gvn.find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2828
2829 // Load from the sub-klass's super-class display list, or a 1-word cache of
2830 // the secondary superclass list, or a failing value with a sentinel offset
2831 // if the super-klass is an interface or exceptionally deep in the Java
2832 // hierarchy and we have to scan the secondary superclass list the hard way.
2833 // Worst-case type is a little odd: NULL is allowed as a result (usually
2834 // klass loads can never produce a NULL).
2835 Node *chk_off_X = chk_off;
2836 #ifdef _LP64
2837 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2838 #endif
2839 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2840 // For some types like interfaces the following loadKlass is from a 1-word
2841 // cache which is mutable so can't use immutable memory. Other
2842 // types load from the super-class display table which is immutable.
2843 Node *kmem = C->immutable_memory();
2844 // secondary_super_cache is not immutable but can be treated as such because:
2845 // - no ideal node writes to it in a way that could cause an
2846 // incorrect/missed optimization of the following Load.
2847 // - it's a cache so, worse case, not reading the latest value
2848 // wouldn't cause incorrect execution
2849 if (might_be_cache && mem != NULL) {
2850 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2851 }
2852 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, NULL, kmem, p2, gvn.type(p2)->is_ptr(), TypeKlassPtr::OBJECT_OR_NULL));
2853
2854 // Compile speed common case: ARE a subtype and we canNOT fail
2855 if( superklass == nkls )
2856 return C->top(); // false path is dead; no test needed.
2857
2858 // See if we get an immediate positive hit. Happens roughly 83% of the
2859 // time. Test to see if the value loaded just previously from the subklass
2860 // is exactly the superklass.
2861 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2862 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
2863 *ctrl = gvn.transform(new IfFalseNode(iff1));
2864
2865 // Compile speed common case: Check for being deterministic right now. If
2866 // chk_off is a constant and not equal to cacheoff then we are NOT a
2867 // subklass. In this case we need exactly the 1 test above and we can
2868 // return those results immediately.
2869 if (!might_be_cache) {
2870 Node* not_subtype_ctrl = *ctrl;
2871 *ctrl = iftrue1; // We need exactly the 1 test above
2872 return not_subtype_ctrl;
2873 }
2874
2875 // Gather the various success & failures here
2876 RegionNode *r_ok_subtype = new RegionNode(4);
2877 gvn.record_for_igvn(r_ok_subtype);
2878 RegionNode *r_not_subtype = new RegionNode(3);
2879 gvn.record_for_igvn(r_not_subtype);
2880
2881 r_ok_subtype->init_req(1, iftrue1);
2882
2883 // Check for immediate negative hit. Happens roughly 11% of the time (which
2884 // is roughly 63% of the remaining cases). Test to see if the loaded
2885 // check-offset points into the subklass display list or the 1-element
2886 // cache. If it points to the display (and NOT the cache) and the display
2887 // missed then it's not a subtype.
2888 Node *cacheoff = gvn.intcon(cacheoff_con);
2889 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2890 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2891 *ctrl = gvn.transform(new IfFalseNode(iff2));
2892
2893 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2894 // No performance impact (too rare) but allows sharing of secondary arrays
2895 // which has some footprint reduction.
2896 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2897 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2898 *ctrl = gvn.transform(new IfFalseNode(iff3));
2899
2900 // -- Roads not taken here: --
2901 // We could also have chosen to perform the self-check at the beginning
2902 // of this code sequence, as the assembler does. This would not pay off
2903 // the same way, since the optimizer, unlike the assembler, can perform
2904 // static type analysis to fold away many successful self-checks.
2905 // Non-foldable self checks work better here in second position, because
2906 // the initial primary superclass check subsumes a self-check for most
2907 // types. An exception would be a secondary type like array-of-interface,
2908 // which does not appear in its own primary supertype display.
2909 // Finally, we could have chosen to move the self-check into the
2910 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2911 // dependent manner. But it is worthwhile to have the check here,
2912 // where it can be perhaps be optimized. The cost in code space is
2913 // small (register compare, branch).
2914
2915 // Now do a linear scan of the secondary super-klass array. Again, no real
2916 // performance impact (too rare) but it's gotta be done.
2917 // Since the code is rarely used, there is no penalty for moving it
2918 // out of line, and it can only improve I-cache density.
2919 // The decision to inline or out-of-line this final check is platform
2920 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2921 Node* psc = gvn.transform(
2922 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2923
2924 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2925 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2926 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2927
2928 // Return false path; set default control to true path.
2929 *ctrl = gvn.transform(r_ok_subtype);
2930 return gvn.transform(r_not_subtype);
2931 }
2932
2933 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2934 const Type* sub_t = _gvn.type(obj_or_subklass);
2935 if (sub_t->isa_inlinetype()) {
2936 obj_or_subklass = makecon(TypeKlassPtr::make(sub_t->inline_klass()));
2937 }
2938 if (ExpandSubTypeCheckAtParseTime) {
2939 MergeMemNode* mem = merged_memory();
2940 Node* ctrl = control();
2941 Node* subklass = obj_or_subklass;
2942 if (!sub_t->isa_klassptr()) {
2943 subklass = load_object_klass(obj_or_subklass);
2944 }
2945 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
2946 set_control(ctrl);
2947 return n;
2948 }
2949
2950 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
2951 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2952 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2953 set_control(_gvn.transform(new IfTrueNode(iff)));
2954 return _gvn.transform(new IfFalseNode(iff));
2955 }
2956
2957 // Profile-driven exact type check:
2958 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2959 float prob, Node* *casted_receiver) {
2960 Node* fail = top();
2961 const Type* rec_t = _gvn.type(receiver);
2962 if (false && rec_t->isa_inlinetype()) {
2963 if (klass->equals(rec_t->inline_klass())) {
2964 (*casted_receiver) = receiver; // Always passes
2965 } else {
2966 (*casted_receiver) = top(); // Always fails
2967 fail = control();
2968 set_control(top());
2969 }
2970 return fail;
2971 }
2972 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
2973 Node* recv_klass = load_object_klass(receiver);
2974 fail = type_check(recv_klass, tklass, prob);
2975 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
2976 assert(recv_xtype->klass_is_exact(), "");
2977
2978 // Subsume downstream occurrences of receiver with a cast to
2979 // recv_xtype, since now we know what the type will be.
2980 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
2981 Node* res = _gvn.transform(cast);
2982 if (recv_xtype->is_inlinetypeptr() && recv_xtype->inline_klass()->is_scalarizable()) {
2983 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
2984 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
2985 }
2986
2987 (*casted_receiver) = res;
2988 // (User must make the replace_in_map call.)
2989
2990 return fail;
2991 }
2992
2993 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
2994 float prob) {
2995 Node* want_klass = makecon(tklass);
2996 Node* cmp = _gvn.transform( new CmpPNode(recv_klass, want_klass));
2997 Node* bol = _gvn.transform( new BoolNode(cmp, BoolTest::eq) );
2998 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
2999 set_control( _gvn.transform( new IfTrueNode (iff)));
3000 Node* fail = _gvn.transform( new IfFalseNode(iff));
3001 return fail;
3002 }
3003
3004 //------------------------------subtype_check_receiver-------------------------
3005 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3006 Node** casted_receiver) {
3007 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass);
3008 Node* want_klass = makecon(tklass);
3009
3010 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3011
3012 // Cast receiver after successful check
3013 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3014 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
3015 (*casted_receiver) = _gvn.transform(cast);
3016
3017 return slow_ctl;
3018 }
3019
3020 //------------------------------seems_never_null-------------------------------
3021 // Use null_seen information if it is available from the profile.
3022 // If we see an unexpected null at a type check we record it and force a
3023 // recompile; the offending check will be recompiled to handle NULLs.
3024 // If we see several offending BCIs, then all checks in the
3025 // method will be recompiled.
3026 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3027 speculating = !_gvn.type(obj)->speculative_maybe_null();
3028 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3029 if (UncommonNullCast // Cutout for this technique
3030 && obj != null() // And not the -Xcomp stupid case?
3031 && !too_many_traps(reason)
3032 ) {
3033 if (speculating) {
3034 return true;
3035 }
3036 if (data == NULL)
3037 // Edge case: no mature data. Be optimistic here.
3038 return true;
3039 // If the profile has not seen a null, assume it won't happen.
3040 assert(java_bc() == Bytecodes::_checkcast ||
3041 java_bc() == Bytecodes::_instanceof ||
3042 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3043 if (java_bc() == Bytecodes::_aastore) {
3044 return ((ciArrayLoadStoreData*)data->as_ArrayLoadStoreData())->element()->ptr_kind() == ProfileNeverNull;
3045 }
3046 return !data->as_BitData()->null_seen();
3047 }
3048 speculating = false;
3049 return false;
3050 }
3051
3052 void GraphKit::guard_klass_being_initialized(Node* klass) {
3053 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3054 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3055 Node* init_state = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3056 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3057 T_BYTE, MemNode::unordered);
3058 init_state = _gvn.transform(init_state);
3059
3060 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3061
3062 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3063 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3064
3065 { BuildCutout unless(this, tst, PROB_MAX);
3066 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3067 }
3068 }
3069
3070 void GraphKit::guard_init_thread(Node* klass) {
3071 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3072 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3073
3074 Node* init_thread = LoadNode::make(_gvn, NULL, immutable_memory(), adr,
3075 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3076 T_ADDRESS, MemNode::unordered);
3077 init_thread = _gvn.transform(init_thread);
3078
3079 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3080
3081 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3082 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3083
3084 { BuildCutout unless(this, tst, PROB_MAX);
3085 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3086 }
3087 }
3088
3089 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3090 if (ik->is_being_initialized()) {
3091 if (C->needs_clinit_barrier(ik, context)) {
3092 Node* klass = makecon(TypeKlassPtr::make(ik));
3093 guard_klass_being_initialized(klass);
3094 guard_init_thread(klass);
3095 insert_mem_bar(Op_MemBarCPUOrder);
3096 }
3097 } else if (ik->is_initialized()) {
3098 return; // no barrier needed
3099 } else {
3100 uncommon_trap(Deoptimization::Reason_uninitialized,
3101 Deoptimization::Action_reinterpret,
3102 NULL);
3103 }
3104 }
3105
3106 //------------------------maybe_cast_profiled_receiver-------------------------
3107 // If the profile has seen exactly one type, narrow to exactly that type.
3108 // Subsequent type checks will always fold up.
3109 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3110 ciKlass* require_klass,
3111 ciKlass* spec_klass,
3112 bool safe_for_replace) {
3113 if (!UseTypeProfile || !TypeProfileCasts) return NULL;
3114
3115 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != NULL);
3116
3117 // Make sure we haven't already deoptimized from this tactic.
3118 if (too_many_traps_or_recompiles(reason))
3119 return NULL;
3120
3121 // (No, this isn't a call, but it's enough like a virtual call
3122 // to use the same ciMethod accessor to get the profile info...)
3123 // If we have a speculative type use it instead of profiling (which
3124 // may not help us)
3125 ciKlass* exact_kls = spec_klass;
3126 if (exact_kls == NULL) {
3127 if (java_bc() == Bytecodes::_aastore) {
3128 ciKlass* array_type = NULL;
3129 ciKlass* element_type = NULL;
3130 ProfilePtrKind element_ptr = ProfileMaybeNull;
3131 bool flat_array = true;
3132 bool null_free_array = true;
3133 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3134 exact_kls = element_type;
3135 } else {
3136 exact_kls = profile_has_unique_klass();
3137 }
3138 }
3139 if (exact_kls != NULL) {// no cast failures here
3140 if (require_klass == NULL ||
3141 C->static_subtype_check(require_klass, exact_kls) == Compile::SSC_always_true) {
3142 // If we narrow the type to match what the type profile sees or
3143 // the speculative type, we can then remove the rest of the
3144 // cast.
3145 // This is a win, even if the exact_kls is very specific,
3146 // because downstream operations, such as method calls,
3147 // will often benefit from the sharper type.
3148 Node* exact_obj = not_null_obj; // will get updated in place...
3149 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3150 &exact_obj);
3151 { PreserveJVMState pjvms(this);
3152 set_control(slow_ctl);
3153 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3154 }
3155 if (safe_for_replace) {
3156 replace_in_map(not_null_obj, exact_obj);
3157 }
3158 return exact_obj;
3159 }
3160 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3161 }
3162
3163 return NULL;
3164 }
3165
3166 /**
3167 * Cast obj to type and emit guard unless we had too many traps here
3168 * already
3169 *
3170 * @param obj node being casted
3171 * @param type type to cast the node to
3172 * @param not_null true if we know node cannot be null
3173 */
3174 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3175 ciKlass* type,
3176 bool not_null) {
3177 if (stopped()) {
3178 return obj;
3179 }
3180
3181 // type == NULL if profiling tells us this object is always null
3182 if (type != NULL) {
3183 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3184 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3185
3186 if (!too_many_traps_or_recompiles(null_reason) &&
3187 !too_many_traps_or_recompiles(class_reason)) {
3188 Node* not_null_obj = NULL;
3189 // not_null is true if we know the object is not null and
3190 // there's no need for a null check
3191 if (!not_null) {
3192 Node* null_ctl = top();
3193 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3194 assert(null_ctl->is_top(), "no null control here");
3195 } else {
3196 not_null_obj = obj;
3197 }
3198
3199 Node* exact_obj = not_null_obj;
3200 ciKlass* exact_kls = type;
3201 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3202 &exact_obj);
3203 {
3204 PreserveJVMState pjvms(this);
3205 set_control(slow_ctl);
3206 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3207 }
3208 replace_in_map(not_null_obj, exact_obj);
3209 obj = exact_obj;
3210 }
3211 } else {
3212 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3213 Node* exact_obj = null_assert(obj);
3214 replace_in_map(obj, exact_obj);
3215 obj = exact_obj;
3216 }
3217 }
3218 return obj;
3219 }
3220
3221 //-------------------------------gen_instanceof--------------------------------
3222 // Generate an instance-of idiom. Used by both the instance-of bytecode
3223 // and the reflective instance-of call.
3224 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3225 kill_dead_locals(); // Benefit all the uncommon traps
3226 assert( !stopped(), "dead parse path should be checked in callers" );
3227 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3228 "must check for not-null not-dead klass in callers");
3229
3230 // Make the merge point
3231 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3232 RegionNode* region = new RegionNode(PATH_LIMIT);
3233 Node* phi = new PhiNode(region, TypeInt::BOOL);
3234 C->set_has_split_ifs(true); // Has chance for split-if optimization
3235
3236 ciProfileData* data = NULL;
3237 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3238 data = method()->method_data()->bci_to_data(bci());
3239 }
3240 bool speculative_not_null = false;
3241 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3242 && seems_never_null(obj, data, speculative_not_null));
3243 bool is_value = obj->is_InlineType();
3244
3245 // Null check; get casted pointer; set region slot 3
3246 Node* null_ctl = top();
3247 Node* not_null_obj = is_value ? obj : null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3248
3249 // If not_null_obj is dead, only null-path is taken
3250 if (stopped()) { // Doing instance-of on a NULL?
3251 set_control(null_ctl);
3252 return intcon(0);
3253 }
3254 region->init_req(_null_path, null_ctl);
3255 phi ->init_req(_null_path, intcon(0)); // Set null path value
3256 if (null_ctl == top()) {
3257 // Do this eagerly, so that pattern matches like is_diamond_phi
3258 // will work even during parsing.
3259 assert(_null_path == PATH_LIMIT-1, "delete last");
3260 region->del_req(_null_path);
3261 phi ->del_req(_null_path);
3262 }
3263
3264 // Do we know the type check always succeed?
3265 if (!is_value) {
3266 bool known_statically = false;
3267 if (_gvn.type(superklass)->singleton()) {
3268 ciKlass* superk = _gvn.type(superklass)->is_klassptr()->klass();
3269 ciKlass* subk = _gvn.type(obj)->is_oopptr()->klass();
3270 if (subk != NULL && subk->is_loaded()) {
3271 int static_res = C->static_subtype_check(superk, subk);
3272 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3273 }
3274 }
3275
3276 if (!known_statically) {
3277 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3278 // We may not have profiling here or it may not help us. If we
3279 // have a speculative type use it to perform an exact cast.
3280 ciKlass* spec_obj_type = obj_type->speculative_type();
3281 if (spec_obj_type != NULL || (ProfileDynamicTypes && data != NULL)) {
3282 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, NULL, spec_obj_type, safe_for_replace);
3283 if (stopped()) { // Profile disagrees with this path.
3284 set_control(null_ctl); // Null is the only remaining possibility.
3285 return intcon(0);
3286 }
3287 if (cast_obj != NULL &&
3288 // A value that's sometimes null is not something we can optimize well
3289 !(cast_obj->is_InlineType() && null_ctl != top())) {
3290 not_null_obj = cast_obj;
3291 is_value = not_null_obj->is_InlineType();
3292 }
3293 }
3294 }
3295 }
3296
3297 // Generate the subtype check
3298 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3299
3300 // Plug in the success path to the general merge in slot 1.
3301 region->init_req(_obj_path, control());
3302 phi ->init_req(_obj_path, intcon(1));
3303
3304 // Plug in the failing path to the general merge in slot 2.
3305 region->init_req(_fail_path, not_subtype_ctrl);
3306 phi ->init_req(_fail_path, intcon(0));
3307
3308 // Return final merged results
3309 set_control( _gvn.transform(region) );
3310 record_for_igvn(region);
3311
3312 // If we know the type check always succeeds then we don't use the
3313 // profiling data at this bytecode. Don't lose it, feed it to the
3314 // type system as a speculative type.
3315 if (safe_for_replace && !is_value) {
3316 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3317 replace_in_map(obj, casted_obj);
3318 }
3319
3320 return _gvn.transform(phi);
3321 }
3322
3323 //-------------------------------gen_checkcast---------------------------------
3324 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3325 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3326 // uncommon-trap paths work. Adjust stack after this call.
3327 // If failure_control is supplied and not null, it is filled in with
3328 // the control edge for the cast failure. Otherwise, an appropriate
3329 // uncommon trap or exception is thrown.
3330 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control) {
3331 kill_dead_locals(); // Benefit all the uncommon traps
3332 const TypeKlassPtr* tk = _gvn.type(superklass)->is_klassptr();
3333 const TypeOopPtr* toop = TypeOopPtr::make_from_klass(tk->klass());
3334
3335 // Check if inline types are involved
3336 bool from_inline = obj->is_InlineType();
3337 bool to_inline = tk->klass()->is_inlinetype();
3338
3339 // Fast cutout: Check the case that the cast is vacuously true.
3340 // This detects the common cases where the test will short-circuit
3341 // away completely. We do this before we perform the null check,
3342 // because if the test is going to turn into zero code, we don't
3343 // want a residual null check left around. (Causes a slowdown,
3344 // for example, in some objArray manipulations, such as a[i]=a[j].)
3345 if (tk->singleton()) {
3346 ciKlass* klass = NULL;
3347 if (from_inline) {
3348 klass = _gvn.type(obj)->inline_klass();
3349 } else {
3350 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3351 if (objtp != NULL) {
3352 klass = objtp->klass();
3353 }
3354 }
3355 if (klass != NULL) {
3356 switch (C->static_subtype_check(tk->klass(), klass)) {
3357 case Compile::SSC_always_true:
3358 // If we know the type check always succeed then we don't use
3359 // the profiling data at this bytecode. Don't lose it, feed it
3360 // to the type system as a speculative type.
3361 if (!from_inline) {
3362 obj = record_profiled_receiver_for_speculation(obj);
3363 if (to_inline) {
3364 obj = null_check(obj);
3365 if (toop->inline_klass()->is_scalarizable()) {
3366 obj = InlineTypeNode::make_from_oop(this, obj, toop->inline_klass());
3367 }
3368 }
3369 }
3370 return obj;
3371 case Compile::SSC_always_false:
3372 if (from_inline || to_inline) {
3373 if (!from_inline) {
3374 null_check(obj);
3375 }
3376 // Inline type is never null. Always throw an exception.
3377 builtin_throw(Deoptimization::Reason_class_check, makecon(TypeKlassPtr::make(klass)));
3378 return top();
3379 } else {
3380 // It needs a null check because a null will *pass* the cast check.
3381 return null_assert(obj);
3382 }
3383 }
3384 }
3385 }
3386
3387 ciProfileData* data = NULL;
3388 bool safe_for_replace = false;
3389 if (failure_control == NULL) { // use MDO in regular case only
3390 assert(java_bc() == Bytecodes::_aastore ||
3391 java_bc() == Bytecodes::_checkcast,
3392 "interpreter profiles type checks only for these BCs");
3393 if (method()->method_data()->is_mature()) {
3394 data = method()->method_data()->bci_to_data(bci());
3395 }
3396 safe_for_replace = true;
3397 }
3398
3399 // Make the merge point
3400 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3401 RegionNode* region = new RegionNode(PATH_LIMIT);
3402 Node* phi = new PhiNode(region, toop);
3403 _gvn.set_type(region, Type::CONTROL);
3404 _gvn.set_type(phi, toop);
3405
3406 C->set_has_split_ifs(true); // Has chance for split-if optimization
3407
3408 // Use null-cast information if it is available
3409 bool speculative_not_null = false;
3410 bool never_see_null = ((failure_control == NULL) // regular case only
3411 && seems_never_null(obj, data, speculative_not_null));
3412
3413 // Null check; get casted pointer; set region slot 3
3414 Node* null_ctl = top();
3415 Node* not_null_obj = NULL;
3416 if (from_inline) {
3417 not_null_obj = obj;
3418 } else if (to_inline) {
3419 not_null_obj = null_check(obj);
3420 } else {
3421 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3422 }
3423
3424 // If not_null_obj is dead, only null-path is taken
3425 if (stopped()) { // Doing instance-of on a NULL?
3426 set_control(null_ctl);
3427 return null();
3428 }
3429 region->init_req(_null_path, null_ctl);
3430 phi ->init_req(_null_path, null()); // Set null path value
3431 if (null_ctl == top()) {
3432 // Do this eagerly, so that pattern matches like is_diamond_phi
3433 // will work even during parsing.
3434 assert(_null_path == PATH_LIMIT-1, "delete last");
3435 region->del_req(_null_path);
3436 phi ->del_req(_null_path);
3437 }
3438
3439 Node* cast_obj = NULL;
3440 if (!from_inline && tk->klass_is_exact()) {
3441 // The following optimization tries to statically cast the speculative type of the object
3442 // (for example obtained during profiling) to the type of the superklass and then do a
3443 // dynamic check that the type of the object is what we expect. To work correctly
3444 // for checkcast and aastore the type of superklass should be exact.
3445 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3446 // We may not have profiling here or it may not help us. If we have
3447 // a speculative type use it to perform an exact cast.
3448 ciKlass* spec_obj_type = obj_type->speculative_type();
3449 if (spec_obj_type != NULL || data != NULL) {
3450 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk->klass(), spec_obj_type, safe_for_replace);
3451 if (cast_obj != NULL && cast_obj->is_InlineType()) {
3452 if (null_ctl != top()) {
3453 cast_obj = NULL; // A value that's sometimes null is not something we can optimize well
3454 } else {
3455 return cast_obj;
3456 }
3457 }
3458 if (cast_obj != NULL) {
3459 if (failure_control != NULL) // failure is now impossible
3460 (*failure_control) = top();
3461 // adjust the type of the phi to the exact klass:
3462 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3463 }
3464 }
3465 }
3466
3467 if (cast_obj == NULL) {
3468 // Generate the subtype check
3469 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3470
3471 // Plug in success path into the merge
3472 cast_obj = from_inline ? not_null_obj : _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3473 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3474 if (failure_control == NULL) {
3475 if (not_subtype_ctrl != top()) { // If failure is possible
3476 PreserveJVMState pjvms(this);
3477 set_control(not_subtype_ctrl);
3478 Node* obj_klass = NULL;
3479 if (from_inline) {
3480 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3481 } else {
3482 obj_klass = load_object_klass(not_null_obj);
3483 }
3484 builtin_throw(Deoptimization::Reason_class_check, obj_klass);
3485 }
3486 } else {
3487 (*failure_control) = not_subtype_ctrl;
3488 }
3489 }
3490
3491 region->init_req(_obj_path, control());
3492 phi ->init_req(_obj_path, cast_obj);
3493
3494 // A merge of NULL or Casted-NotNull obj
3495 Node* res = _gvn.transform(phi);
3496
3497 // Note I do NOT always 'replace_in_map(obj,result)' here.
3498 // if( tk->klass()->can_be_primary_super() )
3499 // This means that if I successfully store an Object into an array-of-String
3500 // I 'forget' that the Object is really now known to be a String. I have to
3501 // do this because we don't have true union types for interfaces - if I store
3502 // a Baz into an array-of-Interface and then tell the optimizer it's an
3503 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3504 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3505 // replace_in_map( obj, res );
3506
3507 // Return final merged results
3508 set_control( _gvn.transform(region) );
3509 record_for_igvn(region);
3510
3511 bool not_inline = !toop->can_be_inline_type();
3512 bool not_flattened = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flatten_array());
3513 if (EnableValhalla && not_flattened) {
3514 // Check if obj has been loaded from an array
3515 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3516 Node* array = NULL;
3517 if (obj->isa_Load()) {
3518 Node* address = obj->in(MemNode::Address);
3519 if (address->isa_AddP()) {
3520 array = address->as_AddP()->in(AddPNode::Base);
3521 }
3522 } else if (obj->is_Phi()) {
3523 Node* region = obj->in(0);
3524 // TODO make this more robust (see JDK-8231346)
3525 if (region->req() == 3 && region->in(2) != NULL && region->in(2)->in(0) != NULL) {
3526 IfNode* iff = region->in(2)->in(0)->isa_If();
3527 if (iff != NULL) {
3528 iff->is_non_flattened_array_check(&_gvn, &array);
3529 }
3530 }
3531 }
3532 if (array != NULL) {
3533 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3534 if (ary_t != NULL) {
3535 if (!ary_t->is_not_null_free() && not_inline) {
3536 // Casting array element to a non-inline-type, mark array as not null-free.
3537 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3538 replace_in_map(array, cast);
3539 } else if (!ary_t->is_not_flat()) {
3540 // Casting array element to a non-flattened type, mark array as not flat.
3541 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3542 replace_in_map(array, cast);
3543 }
3544 }
3545 }
3546 }
3547
3548 if (!from_inline) {
3549 res = record_profiled_receiver_for_speculation(res);
3550 if (to_inline && toop->inline_klass()->is_scalarizable()) {
3551 assert(!gvn().type(res)->maybe_null(), "Inline types are null-free");
3552 res = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3553 }
3554 }
3555 return res;
3556 }
3557
3558 // Check if 'obj' is an inline type by checking if it has the always_locked markWord pattern set.
3559 Node* GraphKit::is_inline_type(Node* obj) {
3560 Node* mark_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3561 Node* mark = make_load(NULL, mark_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3562 Node* mask = _gvn.MakeConX(markWord::always_locked_pattern);
3563 Node* andx = _gvn.transform(new AndXNode(mark, mask));
3564 Node* cmp = _gvn.transform(new CmpXNode(andx, mask));
3565 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3566 }
3567
3568 // Check if 'ary' is a non-flattened array
3569 Node* GraphKit::is_non_flattened_array(Node* ary) {
3570 Node* kls = load_object_klass(ary);
3571 Node* tag = load_lh_array_tag(kls);
3572 Node* cmp = gen_lh_array_test(kls, Klass::_lh_array_tag_vt_value);
3573 return _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3574 }
3575
3576 // Check if 'ary' is a nullable array
3577 Node* GraphKit::is_nullable_array(Node* ary) {
3578 Node* kls = load_object_klass(ary);
3579 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset()));
3580 Node* layout_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lhp, lhp->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3581 Node* null_free = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_null_free_shift)));
3582 null_free = _gvn.transform(new AndINode(null_free, intcon(Klass::_lh_null_free_mask)));
3583 Node* cmp = _gvn.transform(new CmpINode(null_free, intcon(0)));
3584 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3585 }
3586
3587 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3588 Node* GraphKit::gen_inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3589 const Type* val_t = _gvn.type(val);
3590 if (val->is_InlineType() || !TypePtr::NULL_PTR->higher_equal(val_t)) {
3591 return ary; // Never null
3592 }
3593 RegionNode* region = new RegionNode(3);
3594 Node* null_ctl = top();
3595 null_check_oop(val, &null_ctl);
3596 if (null_ctl != top()) {
3597 PreserveJVMState pjvms(this);
3598 set_control(null_ctl);
3599 {
3600 // Deoptimize if null-free array
3601 BuildCutout unless(this, is_nullable_array(ary), PROB_MAX);
3602 inc_sp(nargs);
3603 uncommon_trap(Deoptimization::Reason_null_check,
3604 Deoptimization::Action_none);
3605 }
3606 region->init_req(1, control());
3607 }
3608 region->init_req(2, control());
3609 set_control(_gvn.transform(region));
3610 record_for_igvn(region);
3611 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3612 if (val_t == TypePtr::NULL_PTR && !ary_t->is_not_null_free()) {
3613 // Since we were just successfully storing null, the array can't be null free.
3614 ary_t = ary_t->cast_to_not_null_free();
3615 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3616 if (safe_for_replace) {
3617 replace_in_map(ary, cast);
3618 }
3619 ary = cast;
3620 }
3621 return ary;
3622 }
3623
3624 Node* GraphKit::load_lh_array_tag(Node* kls) {
3625 Node* lhp = basic_plus_adr(kls, in_bytes(Klass::layout_helper_offset()));
3626 Node* layout_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lhp, lhp->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3627 return _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3628 }
3629
3630 Node* GraphKit::gen_lh_array_test(Node* kls, unsigned int lh_value) {
3631 Node* layout_val = load_lh_array_tag(kls);
3632 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(lh_value)));
3633 return cmp;
3634 }
3635
3636 //------------------------------next_monitor-----------------------------------
3637 // What number should be given to the next monitor?
3638 int GraphKit::next_monitor() {
3639 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3640 int next = current + C->sync_stack_slots();
3641 // Keep the toplevel high water mark current:
3642 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3643 return current;
3644 }
3645
3646 //------------------------------insert_mem_bar---------------------------------
3647 // Memory barrier to avoid floating things around
3648 // The membar serves as a pinch point between both control and all memory slices.
3649 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3650 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3651 mb->init_req(TypeFunc::Control, control());
3652 mb->init_req(TypeFunc::Memory, reset_memory());
3653 Node* membar = _gvn.transform(mb);
3654 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3655 set_all_memory_call(membar);
3656 return membar;
3657 }
3658
3659 //-------------------------insert_mem_bar_volatile----------------------------
3660 // Memory barrier to avoid floating things around
3661 // The membar serves as a pinch point between both control and memory(alias_idx).
3662 // If you want to make a pinch point on all memory slices, do not use this
3663 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3664 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3665 // When Parse::do_put_xxx updates a volatile field, it appends a series
3666 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3667 // The first membar is on the same memory slice as the field store opcode.
3668 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3669 // All the other membars (for other volatile slices, including AliasIdxBot,
3670 // which stands for all unknown volatile slices) are control-dependent
3671 // on the first membar. This prevents later volatile loads or stores
3672 // from sliding up past the just-emitted store.
3673
3674 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3675 mb->set_req(TypeFunc::Control,control());
3676 if (alias_idx == Compile::AliasIdxBot) {
3677 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3678 } else {
3679 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3680 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3681 }
3682 Node* membar = _gvn.transform(mb);
3683 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3684 if (alias_idx == Compile::AliasIdxBot) {
3685 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3686 } else {
3687 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3688 }
3689 return membar;
3690 }
3691
3692 //------------------------------shared_lock------------------------------------
3693 // Emit locking code.
3694 FastLockNode* GraphKit::shared_lock(Node* obj) {
3695 // bci is either a monitorenter bc or InvocationEntryBci
3696 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3697 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3698
3699 if( !GenerateSynchronizationCode )
3700 return NULL; // Not locking things?
3701
3702 if (stopped()) // Dead monitor?
3703 return NULL;
3704
3705 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3706
3707 // Box the stack location
3708 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3709 Node* mem = reset_memory();
3710
3711 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3712 if (UseBiasedLocking && PrintPreciseBiasedLockingStatistics) {
3713 // Create the counters for this fast lock.
3714 flock->create_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3715 }
3716
3717 // Create the rtm counters for this fast lock if needed.
3718 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3719
3720 // Add monitor to debug info for the slow path. If we block inside the
3721 // slow path and de-opt, we need the monitor hanging around
3722 map()->push_monitor( flock );
3723
3724 const TypeFunc *tf = LockNode::lock_type();
3725 LockNode *lock = new LockNode(C, tf);
3726
3727 lock->init_req( TypeFunc::Control, control() );
3728 lock->init_req( TypeFunc::Memory , mem );
3729 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3730 lock->init_req( TypeFunc::FramePtr, frameptr() );
3731 lock->init_req( TypeFunc::ReturnAdr, top() );
3732
3733 lock->init_req(TypeFunc::Parms + 0, obj);
3734 lock->init_req(TypeFunc::Parms + 1, box);
3735 lock->init_req(TypeFunc::Parms + 2, flock);
3736 add_safepoint_edges(lock);
3737
3738 lock = _gvn.transform( lock )->as_Lock();
3739
3740 // lock has no side-effects, sets few values
3741 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3742
3743 insert_mem_bar(Op_MemBarAcquireLock);
3744
3745 // Add this to the worklist so that the lock can be eliminated
3746 record_for_igvn(lock);
3747
3748 #ifndef PRODUCT
3749 if (PrintLockStatistics) {
3750 // Update the counter for this lock. Don't bother using an atomic
3751 // operation since we don't require absolute accuracy.
3752 lock->create_lock_counter(map()->jvms());
3753 increment_counter(lock->counter()->addr());
3754 }
3755 #endif
3756
3757 return flock;
3758 }
3759
3760
3761 //------------------------------shared_unlock----------------------------------
3762 // Emit unlocking code.
3763 void GraphKit::shared_unlock(Node* box, Node* obj) {
3764 // bci is either a monitorenter bc or InvocationEntryBci
3765 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3766 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3767
3768 if( !GenerateSynchronizationCode )
3769 return;
3770 if (stopped()) { // Dead monitor?
3771 map()->pop_monitor(); // Kill monitor from debug info
3772 return;
3773 }
3774 assert(!obj->is_InlineTypeBase(), "should not unlock on inline type");
3775
3776 // Memory barrier to avoid floating things down past the locked region
3777 insert_mem_bar(Op_MemBarReleaseLock);
3778
3779 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3780 UnlockNode *unlock = new UnlockNode(C, tf);
3781 #ifdef ASSERT
3782 unlock->set_dbg_jvms(sync_jvms());
3783 #endif
3784 uint raw_idx = Compile::AliasIdxRaw;
3785 unlock->init_req( TypeFunc::Control, control() );
3786 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3787 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3788 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3789 unlock->init_req( TypeFunc::ReturnAdr, top() );
3790
3791 unlock->init_req(TypeFunc::Parms + 0, obj);
3792 unlock->init_req(TypeFunc::Parms + 1, box);
3793 unlock = _gvn.transform(unlock)->as_Unlock();
3794
3795 Node* mem = reset_memory();
3796
3797 // unlock has no side-effects, sets few values
3798 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3799
3800 // Kill monitor from debug info
3801 map()->pop_monitor( );
3802 }
3803
3804 //-------------------------------get_layout_helper-----------------------------
3805 // If the given klass is a constant or known to be an array,
3806 // fetch the constant layout helper value into constant_value
3807 // and return (Node*)NULL. Otherwise, load the non-constant
3808 // layout helper value, and return the node which represents it.
3809 // This two-faced routine is useful because allocation sites
3810 // almost always feature constant types.
3811 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3812 const TypeKlassPtr* inst_klass = _gvn.type(klass_node)->isa_klassptr();
3813 if (!StressReflectiveCode && inst_klass != NULL) {
3814 ciKlass* klass = inst_klass->klass();
3815 assert(klass != NULL, "klass should not be NULL");
3816 bool xklass = inst_klass->klass_is_exact();
3817 bool can_be_flattened = false;
3818 if (UseFlatArray && klass->is_obj_array_klass()) {
3819 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
3820 can_be_flattened = elem->can_be_inline_klass() && (!elem->is_inlinetype() || elem->flatten_array());
3821 }
3822 if (xklass || (klass->is_array_klass() && !can_be_flattened)) {
3823 jint lhelper = klass->layout_helper();
3824 if (lhelper != Klass::_lh_neutral_value) {
3825 constant_value = lhelper;
3826 return (Node*) NULL;
3827 }
3828 }
3829 }
3830 constant_value = Klass::_lh_neutral_value; // put in a known value
3831 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3832 return make_load(NULL, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3833 }
3834
3835 // We just put in an allocate/initialize with a big raw-memory effect.
3836 // Hook selected additional alias categories on the initialization.
3837 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3838 MergeMemNode* init_in_merge,
3839 Node* init_out_raw) {
3840 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3841 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3842
3843 Node* prevmem = kit.memory(alias_idx);
3844 init_in_merge->set_memory_at(alias_idx, prevmem);
3845 kit.set_memory(init_out_raw, alias_idx);
3846 }
3847
3848 //---------------------------set_output_for_allocation-------------------------
3849 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3850 const TypeOopPtr* oop_type,
3851 bool deoptimize_on_exception) {
3852 int rawidx = Compile::AliasIdxRaw;
3853 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3854 add_safepoint_edges(alloc);
3855 Node* allocx = _gvn.transform(alloc);
3856 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3857 // create memory projection for i_o
3858 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3859 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3860
3861 // create a memory projection as for the normal control path
3862 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3863 set_memory(malloc, rawidx);
3864
3865 // a normal slow-call doesn't change i_o, but an allocation does
3866 // we create a separate i_o projection for the normal control path
3867 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3868 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3869
3870 // put in an initialization barrier
3871 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3872 rawoop)->as_Initialize();
3873 assert(alloc->initialization() == init, "2-way macro link must work");
3874 assert(init ->allocation() == alloc, "2-way macro link must work");
3875 {
3876 // Extract memory strands which may participate in the new object's
3877 // initialization, and source them from the new InitializeNode.
3878 // This will allow us to observe initializations when they occur,
3879 // and link them properly (as a group) to the InitializeNode.
3880 assert(init->in(InitializeNode::Memory) == malloc, "");
3881 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3882 init->set_req(InitializeNode::Memory, minit_in);
3883 record_for_igvn(minit_in); // fold it up later, if possible
3884 _gvn.set_type(minit_in, Type::MEMORY);
3885 Node* minit_out = memory(rawidx);
3886 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3887 // Add an edge in the MergeMem for the header fields so an access
3888 // to one of those has correct memory state
3889 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3890 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3891 if (oop_type->isa_aryptr()) {
3892 const TypeAryPtr* arytype = oop_type->is_aryptr();
3893 if (arytype->klass()->is_flat_array_klass()) {
3894 // Initially all flattened array accesses share a single slice
3895 // but that changes after parsing. Prepare the memory graph so
3896 // it can optimize flattened array accesses properly once they
3897 // don't share a single slice.
3898 assert(C->flattened_accesses_share_alias(), "should be set at parse time");
3899 C->set_flattened_accesses_share_alias(false);
3900 ciFlatArrayKlass* vak = arytype->klass()->as_flat_array_klass();
3901 ciInlineKlass* vk = vak->element_klass()->as_inline_klass();
3902 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3903 ciField* field = vk->nonstatic_field_at(i);
3904 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3905 continue; // do not bother to track really large numbers of fields
3906 int off_in_vt = field->offset() - vk->first_field_offset();
3907 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3908 int fieldidx = C->get_alias_index(adr_type, true);
3909 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3910 }
3911 C->set_flattened_accesses_share_alias(true);
3912 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
3913 } else {
3914 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3915 int elemidx = C->get_alias_index(telemref);
3916 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3917 }
3918 } else if (oop_type->isa_instptr()) {
3919 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
3920 ciInstanceKlass* ik = oop_type->klass()->as_instance_klass();
3921 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3922 ciField* field = ik->nonstatic_field_at(i);
3923 if (field->offset() >= TrackedInitializationLimit * HeapWordSize)
3924 continue; // do not bother to track really large numbers of fields
3925 // Find (or create) the alias category for this field:
3926 int fieldidx = C->alias_type(field)->index();
3927 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3928 }
3929 }
3930 }
3931
3932 // Cast raw oop to the real thing...
3933 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3934 javaoop = _gvn.transform(javaoop);
3935 C->set_recent_alloc(control(), javaoop);
3936 assert(just_allocated_object(control()) == javaoop, "just allocated");
3937
3938 #ifdef ASSERT
3939 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
3940 assert(AllocateNode::Ideal_allocation(rawoop, &_gvn) == alloc,
3941 "Ideal_allocation works");
3942 assert(AllocateNode::Ideal_allocation(javaoop, &_gvn) == alloc,
3943 "Ideal_allocation works");
3944 if (alloc->is_AllocateArray()) {
3945 assert(AllocateArrayNode::Ideal_array_allocation(rawoop, &_gvn) == alloc->as_AllocateArray(),
3946 "Ideal_allocation works");
3947 assert(AllocateArrayNode::Ideal_array_allocation(javaoop, &_gvn) == alloc->as_AllocateArray(),
3948 "Ideal_allocation works");
3949 } else {
3950 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3951 }
3952 }
3953 #endif //ASSERT
3954
3955 return javaoop;
3956 }
3957
3958 //---------------------------new_instance--------------------------------------
3959 // This routine takes a klass_node which may be constant (for a static type)
3960 // or may be non-constant (for reflective code). It will work equally well
3961 // for either, and the graph will fold nicely if the optimizer later reduces
3962 // the type to a constant.
3963 // The optional arguments are for specialized use by intrinsics:
3964 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3965 // - If 'return_size_val', report the the total object size to the caller.
3966 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3967 Node* GraphKit::new_instance(Node* klass_node,
3968 Node* extra_slow_test,
3969 Node* *return_size_val,
3970 bool deoptimize_on_exception,
3971 InlineTypeBaseNode* inline_type_node) {
3972 // Compute size in doublewords
3973 // The size is always an integral number of doublewords, represented
3974 // as a positive bytewise size stored in the klass's layout_helper.
3975 // The layout_helper also encodes (in a low bit) the need for a slow path.
3976 jint layout_con = Klass::_lh_neutral_value;
3977 Node* layout_val = get_layout_helper(klass_node, layout_con);
3978 bool layout_is_con = (layout_val == NULL);
3979
3980 if (extra_slow_test == NULL) extra_slow_test = intcon(0);
3981 // Generate the initial go-slow test. It's either ALWAYS (return a
3982 // Node for 1) or NEVER (return a NULL) or perhaps (in the reflective
3983 // case) a computed value derived from the layout_helper.
3984 Node* initial_slow_test = NULL;
3985 if (layout_is_con) {
3986 assert(!StressReflectiveCode, "stress mode does not use these paths");
3987 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3988 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3989 } else { // reflective case
3990 // This reflective path is used by Unsafe.allocateInstance.
3991 // (It may be stress-tested by specifying StressReflectiveCode.)
3992 // Basically, we want to get into the VM is there's an illegal argument.
3993 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3994 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3995 if (extra_slow_test != intcon(0)) {
3996 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3997 }
3998 // (Macro-expander will further convert this to a Bool, if necessary.)
3999 }
4000
4001 // Find the size in bytes. This is easy; it's the layout_helper.
4002 // The size value must be valid even if the slow path is taken.
4003 Node* size = NULL;
4004 if (layout_is_con) {
4005 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4006 } else { // reflective case
4007 // This reflective path is used by clone and Unsafe.allocateInstance.
4008 size = ConvI2X(layout_val);
4009
4010 // Clear the low bits to extract layout_helper_size_in_bytes:
4011 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4012 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4013 size = _gvn.transform( new AndXNode(size, mask) );
4014 }
4015 if (return_size_val != NULL) {
4016 (*return_size_val) = size;
4017 }
4018
4019 // This is a precise notnull oop of the klass.
4020 // (Actually, it need not be precise if this is a reflective allocation.)
4021 // It's what we cast the result to.
4022 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4023 if (!tklass) tklass = TypeKlassPtr::OBJECT;
4024 const TypeOopPtr* oop_type = tklass->as_instance_type();
4025
4026 // Now generate allocation code
4027
4028 // The entire memory state is needed for slow path of the allocation
4029 // since GC and deoptimization can happen.
4030 Node *mem = reset_memory();
4031 set_all_memory(mem); // Create new memory state
4032
4033 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4034 control(), mem, i_o(),
4035 size, klass_node,
4036 initial_slow_test, inline_type_node);
4037
4038 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4039 }
4040
4041 // With compressed oops, the 64 bit init value for non flattened value
4042 // arrays is built from 2 32 bit compressed oops
4043 static Node* raw_default_for_coops(Node* default_value, GraphKit& kit) {
4044 Node* lower = kit.gvn().transform(new CastP2XNode(kit.control(), default_value));
4045 Node* upper = kit.gvn().transform(new LShiftLNode(lower, kit.intcon(32)));
4046 return kit.gvn().transform(new OrLNode(lower, upper));
4047 }
4048
4049 //-------------------------------new_array-------------------------------------
4050 // helper for newarray and anewarray
4051 // The 'length' parameter is (obviously) the length of the array.
4052 // See comments on new_instance for the meaning of the other arguments.
4053 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4054 Node* length, // number of array elements
4055 int nargs, // number of arguments to push back for uncommon trap
4056 Node* *return_size_val,
4057 bool deoptimize_on_exception) {
4058 jint layout_con = Klass::_lh_neutral_value;
4059 Node* layout_val = get_layout_helper(klass_node, layout_con);
4060 bool layout_is_con = (layout_val == NULL);
4061
4062 if (!layout_is_con && !StressReflectiveCode &&
4063 !too_many_traps(Deoptimization::Reason_class_check)) {
4064 // This is a reflective array creation site.
4065 // Optimistically assume that it is a subtype of Object[],
4066 // so that we can fold up all the address arithmetic.
4067 layout_con = Klass::array_layout_helper(T_OBJECT);
4068 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4069 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4070 { BuildCutout unless(this, bol_lh, PROB_MAX);
4071 inc_sp(nargs);
4072 uncommon_trap(Deoptimization::Reason_class_check,
4073 Deoptimization::Action_maybe_recompile);
4074 }
4075 layout_val = NULL;
4076 layout_is_con = true;
4077 }
4078
4079 // Generate the initial go-slow test. Make sure we do not overflow
4080 // if length is huge (near 2Gig) or negative! We do not need
4081 // exact double-words here, just a close approximation of needed
4082 // double-words. We can't add any offset or rounding bits, lest we
4083 // take a size -1 of bytes and make it positive. Use an unsigned
4084 // compare, so negative sizes look hugely positive.
4085 int fast_size_limit = FastAllocateSizeLimit;
4086 if (layout_is_con) {
4087 assert(!StressReflectiveCode, "stress mode does not use these paths");
4088 // Increase the size limit if we have exact knowledge of array type.
4089 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4090 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4091 }
4092
4093 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4094 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4095
4096 // --- Size Computation ---
4097 // array_size = round_to_heap(array_header + (length << elem_shift));
4098 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4099 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4100 // The rounding mask is strength-reduced, if possible.
4101 int round_mask = MinObjAlignmentInBytes - 1;
4102 Node* header_size = NULL;
4103 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4104 // (T_BYTE has the weakest alignment and size restrictions...)
4105 if (layout_is_con) {
4106 int hsize = Klass::layout_helper_header_size(layout_con);
4107 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4108 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4109 if ((round_mask & ~right_n_bits(eshift)) == 0)
4110 round_mask = 0; // strength-reduce it if it goes away completely
4111 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4112 assert(header_size_min <= hsize, "generic minimum is smallest");
4113 header_size_min = hsize;
4114 header_size = intcon(hsize + round_mask);
4115 } else {
4116 Node* hss = intcon(Klass::_lh_header_size_shift);
4117 Node* hsm = intcon(Klass::_lh_header_size_mask);
4118 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
4119 hsize = _gvn.transform( new AndINode(hsize, hsm) );
4120 Node* mask = intcon(round_mask);
4121 header_size = _gvn.transform( new AddINode(hsize, mask) );
4122 }
4123
4124 Node* elem_shift = NULL;
4125 if (layout_is_con) {
4126 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4127 if (eshift != 0)
4128 elem_shift = intcon(eshift);
4129 } else {
4130 // There is no need to mask or shift this value.
4131 // The semantics of LShiftINode include an implicit mask to 0x1F.
4132 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4133 elem_shift = layout_val;
4134 }
4135
4136 // Transition to native address size for all offset calculations:
4137 Node* lengthx = ConvI2X(length);
4138 Node* headerx = ConvI2X(header_size);
4139 #ifdef _LP64
4140 { const TypeInt* tilen = _gvn.find_int_type(length);
4141 if (tilen != NULL && tilen->_lo < 0) {
4142 // Add a manual constraint to a positive range. Cf. array_element_address.
4143 jint size_max = fast_size_limit;
4144 if (size_max > tilen->_hi) size_max = tilen->_hi;
4145 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4146
4147 // Only do a narrow I2L conversion if the range check passed.
4148 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4149 _gvn.transform(iff);
4150 RegionNode* region = new RegionNode(3);
4151 _gvn.set_type(region, Type::CONTROL);
4152 lengthx = new PhiNode(region, TypeLong::LONG);
4153 _gvn.set_type(lengthx, TypeLong::LONG);
4154
4155 // Range check passed. Use ConvI2L node with narrow type.
4156 Node* passed = IfFalse(iff);
4157 region->init_req(1, passed);
4158 // Make I2L conversion control dependent to prevent it from
4159 // floating above the range check during loop optimizations.
4160 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4161
4162 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4163 region->init_req(2, IfTrue(iff));
4164 lengthx->init_req(2, ConvI2X(length));
4165
4166 set_control(region);
4167 record_for_igvn(region);
4168 record_for_igvn(lengthx);
4169 }
4170 }
4171 #endif
4172
4173 // Combine header size (plus rounding) and body size. Then round down.
4174 // This computation cannot overflow, because it is used only in two
4175 // places, one where the length is sharply limited, and the other
4176 // after a successful allocation.
4177 Node* abody = lengthx;
4178 if (elem_shift != NULL)
4179 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
4180 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
4181 if (round_mask != 0) {
4182 Node* mask = MakeConX(~round_mask);
4183 size = _gvn.transform( new AndXNode(size, mask) );
4184 }
4185 // else if round_mask == 0, the size computation is self-rounding
4186
4187 if (return_size_val != NULL) {
4188 // This is the size
4189 (*return_size_val) = size;
4190 }
4191
4192 // Now generate allocation code
4193
4194 // The entire memory state is needed for slow path of the allocation
4195 // since GC and deoptimization can happen.
4196 Node *mem = reset_memory();
4197 set_all_memory(mem); // Create new memory state
4198
4199 if (initial_slow_test->is_Bool()) {
4200 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4201 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4202 }
4203
4204 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4205 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4206 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4207
4208 // Inline type array variants:
4209 // - null-ok: MyValue.ref[] (ciObjArrayKlass "[LMyValue$ref")
4210 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue$val")
4211 // - null-free, flattened: MyValue.val[] (ciFlatArrayKlass "[QMyValue$val")
4212 // Check if array is a null-free, non-flattened inline type array
4213 // that needs to be initialized with the default inline type.
4214 Node* default_value = NULL;
4215 Node* raw_default_value = NULL;
4216 if (ary_ptr != NULL && ary_ptr->klass_is_exact()) {
4217 // Array type is known
4218 ciKlass* elem_klass = ary_ptr->klass()->as_array_klass()->element_klass();
4219 if (elem_klass != NULL && elem_klass->is_inlinetype()) {
4220 ciInlineKlass* vk = elem_klass->as_inline_klass();
4221 if (!vk->flatten_array()) {
4222 default_value = InlineTypeNode::default_oop(gvn(), vk);
4223 if (UseCompressedOops) {
4224 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4225 raw_default_value = raw_default_for_coops(default_value, *this);
4226 } else {
4227 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4228 }
4229 }
4230 }
4231 } else if (ary_klass->klass()->can_be_inline_array_klass()) {
4232 // Array type is not known, add runtime checks
4233 assert(!ary_klass->klass_is_exact(), "unexpected exact type");
4234 Node* r = new RegionNode(4);
4235 default_value = new PhiNode(r, TypeInstPtr::BOTTOM);
4236
4237 // Check if array is an object array
4238 Node* cmp = gen_lh_array_test(klass_node, Klass::_lh_array_tag_obj_value);
4239 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4240 IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4241
4242 // Not an object array, initialize with all zero
4243 r->init_req(1, _gvn.transform(new IfFalseNode(iff)));
4244 default_value->init_req(1, null());
4245
4246 // Object array, check if null-free
4247 set_control(_gvn.transform(new IfTrueNode(iff)));
4248 Node* lhp = basic_plus_adr(klass_node, in_bytes(Klass::layout_helper_offset()));
4249 Node* layout_val = _gvn.transform(LoadNode::make(_gvn, NULL, immutable_memory(), lhp, lhp->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
4250 Node* null_free = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_null_free_shift)));
4251 null_free = _gvn.transform(new AndINode(null_free, intcon(Klass::_lh_null_free_mask)));
4252 cmp = _gvn.transform(new CmpINode(null_free, intcon(0)));
4253 bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
4254 iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4255
4256 // Not null-free, initialize with all zero
4257 r->init_req(2, _gvn.transform(new IfFalseNode(iff)));
4258 default_value->init_req(2, null());
4259
4260 // Null-free, non-flattened inline type array, initialize with the default value
4261 set_control(_gvn.transform(new IfTrueNode(iff)));
4262 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset()));
4263 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS));
4264 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()));
4265 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4266 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(InlineKlass::default_value_offset_offset()));
4267 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
4268 Node* elem_mirror = load_mirror_from_klass(eklass);
4269 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset));
4270 Node* val = access_load_at(elem_mirror, default_value_addr, _gvn.type(default_value_addr)->is_ptr(), TypeInstPtr::BOTTOM, T_OBJECT, IN_HEAP);
4271 r->init_req(3, control());
4272 default_value->init_req(3, val);
4273
4274 set_control(_gvn.transform(r));
4275 default_value = _gvn.transform(default_value);
4276 if (UseCompressedOops) {
4277 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4278 raw_default_value = raw_default_for_coops(default_value, *this);
4279 } else {
4280 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4281 }
4282 }
4283
4284 // Create the AllocateArrayNode and its result projections
4285 AllocateArrayNode* alloc = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4286 control(), mem, i_o(),
4287 size, klass_node,
4288 initial_slow_test,
4289 length, default_value,
4290 raw_default_value);
4291
4292 // Cast to correct type. Note that the klass_node may be constant or not,
4293 // and in the latter case the actual array type will be inexact also.
4294 // (This happens via a non-constant argument to inline_native_newArray.)
4295 // In any case, the value of klass_node provides the desired array type.
4296 const TypeInt* length_type = _gvn.find_int_type(length);
4297 if (ary_type->isa_aryptr() && length_type != NULL) {
4298 // Try to get a better type than POS for the size
4299 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4300 }
4301
4302 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4303
4304 // Cast length on remaining path to be as narrow as possible
4305 if (map()->find_edge(length) >= 0) {
4306 Node* ccast = alloc->make_ideal_length(ary_type, &_gvn);
4307 if (ccast != length) {
4308 _gvn.set_type_bottom(ccast);
4309 record_for_igvn(ccast);
4310 replace_in_map(length, ccast);
4311 }
4312 }
4313
4314 return javaoop;
4315 }
4316
4317 // The following "Ideal_foo" functions are placed here because they recognize
4318 // the graph shapes created by the functions immediately above.
4319
4320 //---------------------------Ideal_allocation----------------------------------
4321 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4322 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase) {
4323 if (ptr == NULL) { // reduce dumb test in callers
4324 return NULL;
4325 }
4326
4327 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4328 ptr = bs->step_over_gc_barrier(ptr);
4329
4330 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4331 ptr = ptr->in(1);
4332 if (ptr == NULL) return NULL;
4333 }
4334 // Return NULL for allocations with several casts:
4335 // j.l.reflect.Array.newInstance(jobject, jint)
4336 // Object.clone()
4337 // to keep more precise type from last cast.
4338 if (ptr->is_Proj()) {
4339 Node* allo = ptr->in(0);
4340 if (allo != NULL && allo->is_Allocate()) {
4341 return allo->as_Allocate();
4342 }
4343 }
4344 // Report failure to match.
4345 return NULL;
4346 }
4347
4348 // Fancy version which also strips off an offset (and reports it to caller).
4349 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseTransform* phase,
4350 intptr_t& offset) {
4351 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4352 if (base == NULL) return NULL;
4353 return Ideal_allocation(base, phase);
4354 }
4355
4356 // Trace Initialize <- Proj[Parm] <- Allocate
4357 AllocateNode* InitializeNode::allocation() {
4358 Node* rawoop = in(InitializeNode::RawAddress);
4359 if (rawoop->is_Proj()) {
4360 Node* alloc = rawoop->in(0);
4361 if (alloc->is_Allocate()) {
4362 return alloc->as_Allocate();
4363 }
4364 }
4365 return NULL;
4366 }
4367
4368 // Trace Allocate -> Proj[Parm] -> Initialize
4369 InitializeNode* AllocateNode::initialization() {
4370 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4371 if (rawoop == NULL) return NULL;
4372 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4373 Node* init = rawoop->fast_out(i);
4374 if (init->is_Initialize()) {
4375 assert(init->as_Initialize()->allocation() == this, "2-way link");
4376 return init->as_Initialize();
4377 }
4378 }
4379 return NULL;
4380 }
4381
4382 //----------------------------- loop predicates ---------------------------
4383
4384 //------------------------------add_predicate_impl----------------------------
4385 void GraphKit::add_empty_predicate_impl(Deoptimization::DeoptReason reason, int nargs) {
4386 // Too many traps seen?
4387 if (too_many_traps(reason)) {
4388 #ifdef ASSERT
4389 if (TraceLoopPredicate) {
4390 int tc = C->trap_count(reason);
4391 tty->print("too many traps=%s tcount=%d in ",
4392 Deoptimization::trap_reason_name(reason), tc);
4393 method()->print(); // which method has too many predicate traps
4394 tty->cr();
4395 }
4396 #endif
4397 // We cannot afford to take more traps here,
4398 // do not generate predicate.
4399 return;
4400 }
4401
4402 Node *cont = _gvn.intcon(1);
4403 Node* opq = _gvn.transform(new Opaque1Node(C, cont));
4404 Node *bol = _gvn.transform(new Conv2BNode(opq));
4405 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4406 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
4407 C->add_predicate_opaq(opq);
4408 {
4409 PreserveJVMState pjvms(this);
4410 set_control(iffalse);
4411 inc_sp(nargs);
4412 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4413 }
4414 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
4415 set_control(iftrue);
4416 }
4417
4418 //------------------------------add_predicate---------------------------------
4419 void GraphKit::add_empty_predicates(int nargs) {
4420 // These loop predicates remain empty. All concrete loop predicates are inserted above the corresponding
4421 // empty loop predicate later by 'PhaseIdealLoop::create_new_if_for_predicate'. All concrete loop predicates of
4422 // a specific kind (normal, profile or limit check) share the same uncommon trap as the empty loop predicate.
4423 if (UseLoopPredicate) {
4424 add_empty_predicate_impl(Deoptimization::Reason_predicate, nargs);
4425 }
4426 if (UseProfiledLoopPredicate) {
4427 add_empty_predicate_impl(Deoptimization::Reason_profile_predicate, nargs);
4428 }
4429 // loop's limit check predicate should be near the loop.
4430 add_empty_predicate_impl(Deoptimization::Reason_loop_limit_check, nargs);
4431 }
4432
4433 void GraphKit::sync_kit(IdealKit& ideal) {
4434 set_all_memory(ideal.merged_memory());
4435 set_i_o(ideal.i_o());
4436 set_control(ideal.ctrl());
4437 }
4438
4439 void GraphKit::final_sync(IdealKit& ideal) {
4440 // Final sync IdealKit and graphKit.
4441 sync_kit(ideal);
4442 }
4443
4444 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4445 Node* len = load_array_length(load_String_value(str, set_ctrl));
4446 Node* coder = load_String_coder(str, set_ctrl);
4447 // Divide length by 2 if coder is UTF16
4448 return _gvn.transform(new RShiftINode(len, coder));
4449 }
4450
4451 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4452 int value_offset = java_lang_String::value_offset();
4453 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4454 false, NULL, Type::Offset(0));
4455 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4456 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4457 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, true, true),
4458 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4459 Node* p = basic_plus_adr(str, str, value_offset);
4460 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4461 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4462 return load;
4463 }
4464
4465 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4466 if (!CompactStrings) {
4467 return intcon(java_lang_String::CODER_UTF16);
4468 }
4469 int coder_offset = java_lang_String::coder_offset();
4470 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4471 false, NULL, Type::Offset(0));
4472 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4473
4474 Node* p = basic_plus_adr(str, str, coder_offset);
4475 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4476 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4477 return load;
4478 }
4479
4480 void GraphKit::store_String_value(Node* str, Node* value) {
4481 int value_offset = java_lang_String::value_offset();
4482 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4483 false, NULL, Type::Offset(0));
4484 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4485
4486 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4487 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4488 }
4489
4490 void GraphKit::store_String_coder(Node* str, Node* value) {
4491 int coder_offset = java_lang_String::coder_offset();
4492 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4493 false, NULL, Type::Offset(0));
4494 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4495
4496 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4497 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4498 }
4499
4500 // Capture src and dst memory state with a MergeMemNode
4501 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4502 if (src_type == dst_type) {
4503 // Types are equal, we don't need a MergeMemNode
4504 return memory(src_type);
4505 }
4506 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4507 record_for_igvn(merge); // fold it up later, if possible
4508 int src_idx = C->get_alias_index(src_type);
4509 int dst_idx = C->get_alias_index(dst_type);
4510 merge->set_memory_at(src_idx, memory(src_idx));
4511 merge->set_memory_at(dst_idx, memory(dst_idx));
4512 return merge;
4513 }
4514
4515 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4516 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4517 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4518 // If input and output memory types differ, capture both states to preserve
4519 // the dependency between preceding and subsequent loads/stores.
4520 // For example, the following program:
4521 // StoreB
4522 // compress_string
4523 // LoadB
4524 // has this memory graph (use->def):
4525 // LoadB -> compress_string -> CharMem
4526 // ... -> StoreB -> ByteMem
4527 // The intrinsic hides the dependency between LoadB and StoreB, causing
4528 // the load to read from memory not containing the result of the StoreB.
4529 // The correct memory graph should look like this:
4530 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4531 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4532 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4533 Node* res_mem = _gvn.transform(new SCMemProjNode(str));
4534 set_memory(res_mem, TypeAryPtr::BYTES);
4535 return str;
4536 }
4537
4538 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4539 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4540 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4541 // Capture src and dst memory (see comment in 'compress_string').
4542 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4543 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4544 set_memory(_gvn.transform(str), dst_type);
4545 }
4546
4547 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4548 /**
4549 * int i_char = start;
4550 * for (int i_byte = 0; i_byte < count; i_byte++) {
4551 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4552 * }
4553 */
4554 add_empty_predicates();
4555 RegionNode* head = new RegionNode(3);
4556 head->init_req(1, control());
4557 gvn().set_type(head, Type::CONTROL);
4558 record_for_igvn(head);
4559
4560 Node* i_byte = new PhiNode(head, TypeInt::INT);
4561 i_byte->init_req(1, intcon(0));
4562 gvn().set_type(i_byte, TypeInt::INT);
4563 record_for_igvn(i_byte);
4564
4565 Node* i_char = new PhiNode(head, TypeInt::INT);
4566 i_char->init_req(1, start);
4567 gvn().set_type(i_char, TypeInt::INT);
4568 record_for_igvn(i_char);
4569
4570 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4571 gvn().set_type(mem, Type::MEMORY);
4572 record_for_igvn(mem);
4573 set_control(head);
4574 set_memory(mem, TypeAryPtr::BYTES);
4575 Node* ch = load_array_element(control(), src, i_byte, TypeAryPtr::BYTES);
4576 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4577 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4578 false, false, true /* mismatched */);
4579
4580 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4581 head->init_req(2, IfTrue(iff));
4582 mem->init_req(2, st);
4583 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4584 i_char->init_req(2, AddI(i_char, intcon(2)));
4585
4586 set_control(IfFalse(iff));
4587 set_memory(st, TypeAryPtr::BYTES);
4588 }
4589
4590 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4591 if (!field->is_constant()) {
4592 return NULL; // Field not marked as constant.
4593 }
4594 ciInstance* holder = NULL;
4595 if (!field->is_static()) {
4596 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4597 if (const_oop != NULL && const_oop->is_instance()) {
4598 holder = const_oop->as_instance();
4599 }
4600 }
4601 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4602 /*is_unsigned_load=*/false);
4603 if (con_type != NULL) {
4604 Node* con = makecon(con_type);
4605 assert(!field->type()->is_inlinetype() || (field->is_static() && !con_type->is_zero_type()), "sanity");
4606 // Check type of constant which might be more precise
4607 if (con_type->is_inlinetypeptr() && con_type->inline_klass()->is_scalarizable()) {
4608 // Load inline type from constant oop
4609 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4610 }
4611 return con;
4612 }
4613 return NULL;
4614 }
4615
4616 //---------------------------load_mirror_from_klass----------------------------
4617 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4618 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4619 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4620 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4621 // mirror = ((OopHandle)mirror)->resolve();
4622 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4623 }