1 /*
2 * Copyright (c) 2001, 2024, 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 "ci/ciObjArray.hpp"
31 #include "asm/register.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/inlinetypenode.hpp"
43 #include "opto/intrinsicnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "runtime/deoptimization.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "utilities/bitMap.inline.hpp"
55 #include "utilities/powerOfTwo.hpp"
56 #include "utilities/growableArray.hpp"
57
58 //----------------------------GraphKit-----------------------------------------
59 // Main utility constructor.
60 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
61 : Phase(Phase::Parser),
62 _env(C->env()),
63 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
64 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
65 {
66 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
67 _exceptions = jvms->map()->next_exception();
68 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
69 set_jvms(jvms);
70 #ifdef ASSERT
71 if (_gvn.is_IterGVN() != nullptr) {
72 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
73 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
74 _worklist_size = _gvn.C->igvn_worklist()->size();
75 }
76 #endif
77 }
78
79 // Private constructor for parser.
80 GraphKit::GraphKit()
81 : Phase(Phase::Parser),
82 _env(C->env()),
83 _gvn(*C->initial_gvn()),
84 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
85 {
86 _exceptions = nullptr;
87 set_map(nullptr);
88 debug_only(_sp = -99);
89 debug_only(set_bci(-99));
90 }
91
92
93
94 //---------------------------clean_stack---------------------------------------
95 // Clear away rubbish from the stack area of the JVM state.
96 // This destroys any arguments that may be waiting on the stack.
97 void GraphKit::clean_stack(int from_sp) {
98 SafePointNode* map = this->map();
99 JVMState* jvms = this->jvms();
100 int stk_size = jvms->stk_size();
101 int stkoff = jvms->stkoff();
102 Node* top = this->top();
103 for (int i = from_sp; i < stk_size; i++) {
104 if (map->in(stkoff + i) != top) {
105 map->set_req(stkoff + i, top);
106 }
107 }
108 }
109
110
111 //--------------------------------sync_jvms-----------------------------------
112 // Make sure our current jvms agrees with our parse state.
113 JVMState* GraphKit::sync_jvms() const {
114 JVMState* jvms = this->jvms();
115 jvms->set_bci(bci()); // Record the new bci in the JVMState
116 jvms->set_sp(sp()); // Record the new sp in the JVMState
117 assert(jvms_in_sync(), "jvms is now in sync");
118 return jvms;
119 }
120
121 //--------------------------------sync_jvms_for_reexecute---------------------
122 // Make sure our current jvms agrees with our parse state. This version
123 // uses the reexecute_sp for reexecuting bytecodes.
124 JVMState* GraphKit::sync_jvms_for_reexecute() {
125 JVMState* jvms = this->jvms();
126 jvms->set_bci(bci()); // Record the new bci in the JVMState
127 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
128 return jvms;
129 }
130
131 #ifdef ASSERT
132 bool GraphKit::jvms_in_sync() const {
133 Parse* parse = is_Parse();
134 if (parse == nullptr) {
135 if (bci() != jvms()->bci()) return false;
136 if (sp() != (int)jvms()->sp()) return false;
137 return true;
138 }
139 if (jvms()->method() != parse->method()) return false;
140 if (jvms()->bci() != parse->bci()) return false;
141 int jvms_sp = jvms()->sp();
142 if (jvms_sp != parse->sp()) return false;
143 int jvms_depth = jvms()->depth();
144 if (jvms_depth != parse->depth()) return false;
145 return true;
146 }
147
148 // Local helper checks for special internal merge points
149 // used to accumulate and merge exception states.
150 // They are marked by the region's in(0) edge being the map itself.
151 // Such merge points must never "escape" into the parser at large,
152 // until they have been handed to gvn.transform.
153 static bool is_hidden_merge(Node* reg) {
154 if (reg == nullptr) return false;
155 if (reg->is_Phi()) {
156 reg = reg->in(0);
157 if (reg == nullptr) return false;
158 }
159 return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root();
160 }
161
162 void GraphKit::verify_map() const {
163 if (map() == nullptr) return; // null map is OK
164 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
165 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
166 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
167 }
168
169 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
170 assert(ex_map->next_exception() == nullptr, "not already part of a chain");
171 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
172 }
173 #endif
174
175 //---------------------------stop_and_kill_map---------------------------------
176 // Set _map to null, signalling a stop to further bytecode execution.
177 // First smash the current map's control to a constant, to mark it dead.
178 void GraphKit::stop_and_kill_map() {
179 SafePointNode* dead_map = stop();
180 if (dead_map != nullptr) {
181 dead_map->disconnect_inputs(C); // Mark the map as killed.
182 assert(dead_map->is_killed(), "must be so marked");
183 }
184 }
185
186
187 //--------------------------------stopped--------------------------------------
188 // Tell if _map is null, or control is top.
189 bool GraphKit::stopped() {
190 if (map() == nullptr) return true;
191 else if (control() == top()) return true;
192 else return false;
193 }
194
195
196 //-----------------------------has_exception_handler----------------------------------
197 // Tell if this method or any caller method has exception handlers.
198 bool GraphKit::has_exception_handler() {
199 for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) {
200 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
201 return true;
202 }
203 }
204 return false;
205 }
206
207 //------------------------------save_ex_oop------------------------------------
208 // Save an exception without blowing stack contents or other JVM state.
209 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
210 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
211 ex_map->add_req(ex_oop);
212 debug_only(verify_exception_state(ex_map));
213 }
214
215 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
216 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
217 Node* ex_oop = ex_map->in(ex_map->req()-1);
218 if (clear_it) ex_map->del_req(ex_map->req()-1);
219 return ex_oop;
220 }
221
222 //-----------------------------saved_ex_oop------------------------------------
223 // Recover a saved exception from its map.
224 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
225 return common_saved_ex_oop(ex_map, false);
226 }
227
228 //--------------------------clear_saved_ex_oop---------------------------------
229 // Erase a previously saved exception from its map.
230 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
231 return common_saved_ex_oop(ex_map, true);
232 }
233
234 #ifdef ASSERT
235 //---------------------------has_saved_ex_oop----------------------------------
236 // Erase a previously saved exception from its map.
237 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
238 return ex_map->req() == ex_map->jvms()->endoff()+1;
239 }
240 #endif
241
242 //-------------------------make_exception_state--------------------------------
243 // Turn the current JVM state into an exception state, appending the ex_oop.
244 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
245 sync_jvms();
246 SafePointNode* ex_map = stop(); // do not manipulate this map any more
247 set_saved_ex_oop(ex_map, ex_oop);
248 return ex_map;
249 }
250
251
252 //--------------------------add_exception_state--------------------------------
253 // Add an exception to my list of exceptions.
254 void GraphKit::add_exception_state(SafePointNode* ex_map) {
255 if (ex_map == nullptr || ex_map->control() == top()) {
256 return;
257 }
258 #ifdef ASSERT
259 verify_exception_state(ex_map);
260 if (has_exceptions()) {
261 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
262 }
263 #endif
264
265 // If there is already an exception of exactly this type, merge with it.
266 // In particular, null-checks and other low-level exceptions common up here.
267 Node* ex_oop = saved_ex_oop(ex_map);
268 const Type* ex_type = _gvn.type(ex_oop);
269 if (ex_oop == top()) {
270 // No action needed.
271 return;
272 }
273 assert(ex_type->isa_instptr(), "exception must be an instance");
274 for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) {
275 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
276 // We check sp also because call bytecodes can generate exceptions
277 // both before and after arguments are popped!
278 if (ex_type2 == ex_type
279 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
280 combine_exception_states(ex_map, e2);
281 return;
282 }
283 }
284
285 // No pre-existing exception of the same type. Chain it on the list.
286 push_exception_state(ex_map);
287 }
288
289 //-----------------------add_exception_states_from-----------------------------
290 void GraphKit::add_exception_states_from(JVMState* jvms) {
291 SafePointNode* ex_map = jvms->map()->next_exception();
292 if (ex_map != nullptr) {
293 jvms->map()->set_next_exception(nullptr);
294 for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) {
295 next_map = ex_map->next_exception();
296 ex_map->set_next_exception(nullptr);
297 add_exception_state(ex_map);
298 }
299 }
300 }
301
302 //-----------------------transfer_exceptions_into_jvms-------------------------
303 JVMState* GraphKit::transfer_exceptions_into_jvms() {
304 if (map() == nullptr) {
305 // We need a JVMS to carry the exceptions, but the map has gone away.
306 // Create a scratch JVMS, cloned from any of the exception states...
307 if (has_exceptions()) {
308 _map = _exceptions;
309 _map = clone_map();
310 _map->set_next_exception(nullptr);
311 clear_saved_ex_oop(_map);
312 debug_only(verify_map());
313 } else {
314 // ...or created from scratch
315 JVMState* jvms = new (C) JVMState(_method, nullptr);
316 jvms->set_bci(_bci);
317 jvms->set_sp(_sp);
318 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
319 set_jvms(jvms);
320 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
321 set_all_memory(top());
322 while (map()->req() < jvms->endoff()) map()->add_req(top());
323 }
324 // (This is a kludge, in case you didn't notice.)
325 set_control(top());
326 }
327 JVMState* jvms = sync_jvms();
328 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
329 jvms->map()->set_next_exception(_exceptions);
330 _exceptions = nullptr; // done with this set of exceptions
331 return jvms;
332 }
333
334 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
335 assert(is_hidden_merge(dstphi), "must be a special merge node");
336 assert(is_hidden_merge(srcphi), "must be a special merge node");
337 uint limit = srcphi->req();
338 for (uint i = PhiNode::Input; i < limit; i++) {
339 dstphi->add_req(srcphi->in(i));
340 }
341 }
342 static inline void add_one_req(Node* dstphi, Node* src) {
343 assert(is_hidden_merge(dstphi), "must be a special merge node");
344 assert(!is_hidden_merge(src), "must not be a special merge node");
345 dstphi->add_req(src);
346 }
347
348 //-----------------------combine_exception_states------------------------------
349 // This helper function combines exception states by building phis on a
350 // specially marked state-merging region. These regions and phis are
351 // untransformed, and can build up gradually. The region is marked by
352 // having a control input of its exception map, rather than null. Such
353 // regions do not appear except in this function, and in use_exception_state.
354 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
355 if (failing()) return; // dying anyway...
356 JVMState* ex_jvms = ex_map->_jvms;
357 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
358 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
359 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
360 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
361 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
362 assert(ex_map->req() == phi_map->req(), "matching maps");
363 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
364 Node* hidden_merge_mark = root();
365 Node* region = phi_map->control();
366 MergeMemNode* phi_mem = phi_map->merged_memory();
367 MergeMemNode* ex_mem = ex_map->merged_memory();
368 if (region->in(0) != hidden_merge_mark) {
369 // The control input is not (yet) a specially-marked region in phi_map.
370 // Make it so, and build some phis.
371 region = new RegionNode(2);
372 _gvn.set_type(region, Type::CONTROL);
373 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
374 region->init_req(1, phi_map->control());
375 phi_map->set_control(region);
376 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
377 record_for_igvn(io_phi);
378 _gvn.set_type(io_phi, Type::ABIO);
379 phi_map->set_i_o(io_phi);
380 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
381 Node* m = mms.memory();
382 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
383 record_for_igvn(m_phi);
384 _gvn.set_type(m_phi, Type::MEMORY);
385 mms.set_memory(m_phi);
386 }
387 }
388
389 // Either or both of phi_map and ex_map might already be converted into phis.
390 Node* ex_control = ex_map->control();
391 // if there is special marking on ex_map also, we add multiple edges from src
392 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
393 // how wide was the destination phi_map, originally?
394 uint orig_width = region->req();
395
396 if (add_multiple) {
397 add_n_reqs(region, ex_control);
398 add_n_reqs(phi_map->i_o(), ex_map->i_o());
399 } else {
400 // ex_map has no merges, so we just add single edges everywhere
401 add_one_req(region, ex_control);
402 add_one_req(phi_map->i_o(), ex_map->i_o());
403 }
404 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
405 if (mms.is_empty()) {
406 // get a copy of the base memory, and patch some inputs into it
407 const TypePtr* adr_type = mms.adr_type(C);
408 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
409 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
410 mms.set_memory(phi);
411 // Prepare to append interesting stuff onto the newly sliced phi:
412 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
413 }
414 // Append stuff from ex_map:
415 if (add_multiple) {
416 add_n_reqs(mms.memory(), mms.memory2());
417 } else {
418 add_one_req(mms.memory(), mms.memory2());
419 }
420 }
421 uint limit = ex_map->req();
422 for (uint i = TypeFunc::Parms; i < limit; i++) {
423 // Skip everything in the JVMS after tos. (The ex_oop follows.)
424 if (i == tos) i = ex_jvms->monoff();
425 Node* src = ex_map->in(i);
426 Node* dst = phi_map->in(i);
427 if (src != dst) {
428 PhiNode* phi;
429 if (dst->in(0) != region) {
430 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
431 record_for_igvn(phi);
432 _gvn.set_type(phi, phi->type());
433 phi_map->set_req(i, dst);
434 // Prepare to append interesting stuff onto the new phi:
435 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
436 } else {
437 assert(dst->is_Phi(), "nobody else uses a hidden region");
438 phi = dst->as_Phi();
439 }
440 if (add_multiple && src->in(0) == ex_control) {
441 // Both are phis.
442 add_n_reqs(dst, src);
443 } else {
444 while (dst->req() < region->req()) add_one_req(dst, src);
445 }
446 const Type* srctype = _gvn.type(src);
447 if (phi->type() != srctype) {
448 const Type* dsttype = phi->type()->meet_speculative(srctype);
449 if (phi->type() != dsttype) {
450 phi->set_type(dsttype);
451 _gvn.set_type(phi, dsttype);
452 }
453 }
454 }
455 }
456 phi_map->merge_replaced_nodes_with(ex_map);
457 }
458
459 //--------------------------use_exception_state--------------------------------
460 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
461 if (failing()) { stop(); return top(); }
462 Node* region = phi_map->control();
463 Node* hidden_merge_mark = root();
464 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
465 Node* ex_oop = clear_saved_ex_oop(phi_map);
466 if (region->in(0) == hidden_merge_mark) {
467 // Special marking for internal ex-states. Process the phis now.
468 region->set_req(0, region); // now it's an ordinary region
469 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
470 // Note: Setting the jvms also sets the bci and sp.
471 set_control(_gvn.transform(region));
472 uint tos = jvms()->stkoff() + sp();
473 for (uint i = 1; i < tos; i++) {
474 Node* x = phi_map->in(i);
475 if (x->in(0) == region) {
476 assert(x->is_Phi(), "expected a special phi");
477 phi_map->set_req(i, _gvn.transform(x));
478 }
479 }
480 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
481 Node* x = mms.memory();
482 if (x->in(0) == region) {
483 assert(x->is_Phi(), "nobody else uses a hidden region");
484 mms.set_memory(_gvn.transform(x));
485 }
486 }
487 if (ex_oop->in(0) == region) {
488 assert(ex_oop->is_Phi(), "expected a special phi");
489 ex_oop = _gvn.transform(ex_oop);
490 }
491 } else {
492 set_jvms(phi_map->jvms());
493 }
494
495 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
496 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
497 return ex_oop;
498 }
499
500 //---------------------------------java_bc-------------------------------------
501 Bytecodes::Code GraphKit::java_bc() const {
502 ciMethod* method = this->method();
503 int bci = this->bci();
504 if (method != nullptr && bci != InvocationEntryBci)
505 return method->java_code_at_bci(bci);
506 else
507 return Bytecodes::_illegal;
508 }
509
510 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
511 bool must_throw) {
512 // if the exception capability is set, then we will generate code
513 // to check the JavaThread.should_post_on_exceptions flag to see
514 // if we actually need to report exception events (for this
515 // thread). If we don't need to report exception events, we will
516 // take the normal fast path provided by add_exception_events. If
517 // exception event reporting is enabled for this thread, we will
518 // take the uncommon_trap in the BuildCutout below.
519
520 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
521 Node* jthread = _gvn.transform(new ThreadLocalNode());
522 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
523 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
524
525 // Test the should_post_on_exceptions_flag vs. 0
526 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
527 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
528
529 // Branch to slow_path if should_post_on_exceptions_flag was true
530 { BuildCutout unless(this, tst, PROB_MAX);
531 // Do not try anything fancy if we're notifying the VM on every throw.
532 // Cf. case Bytecodes::_athrow in parse2.cpp.
533 uncommon_trap(reason, Deoptimization::Action_none,
534 (ciKlass*)nullptr, (char*)nullptr, must_throw);
535 }
536
537 }
538
539 //------------------------------builtin_throw----------------------------------
540 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
541 bool must_throw = true;
542
543 // If this particular condition has not yet happened at this
544 // bytecode, then use the uncommon trap mechanism, and allow for
545 // a future recompilation if several traps occur here.
546 // If the throw is hot, try to use a more complicated inline mechanism
547 // which keeps execution inside the compiled code.
548 bool treat_throw_as_hot = false;
549 ciMethodData* md = method()->method_data();
550
551 if (ProfileTraps) {
552 if (too_many_traps(reason)) {
553 treat_throw_as_hot = true;
554 }
555 // (If there is no MDO at all, assume it is early in
556 // execution, and that any deopts are part of the
557 // startup transient, and don't need to be remembered.)
558
559 // Also, if there is a local exception handler, treat all throws
560 // as hot if there has been at least one in this method.
561 if (C->trap_count(reason) != 0
562 && method()->method_data()->trap_count(reason) != 0
563 && has_exception_handler()) {
564 treat_throw_as_hot = true;
565 }
566 }
567
568 // If this throw happens frequently, an uncommon trap might cause
569 // a performance pothole. If there is a local exception handler,
570 // and if this particular bytecode appears to be deoptimizing often,
571 // let us handle the throw inline, with a preconstructed instance.
572 // Note: If the deopt count has blown up, the uncommon trap
573 // runtime is going to flush this nmethod, not matter what.
574 if (treat_throw_as_hot && method()->can_omit_stack_trace()) {
575 // If the throw is local, we use a pre-existing instance and
576 // punt on the backtrace. This would lead to a missing backtrace
577 // (a repeat of 4292742) if the backtrace object is ever asked
578 // for its backtrace.
579 // Fixing this remaining case of 4292742 requires some flavor of
580 // escape analysis. Leave that for the future.
581 ciInstance* ex_obj = nullptr;
582 switch (reason) {
583 case Deoptimization::Reason_null_check:
584 ex_obj = env()->NullPointerException_instance();
585 break;
586 case Deoptimization::Reason_div0_check:
587 ex_obj = env()->ArithmeticException_instance();
588 break;
589 case Deoptimization::Reason_range_check:
590 ex_obj = env()->ArrayIndexOutOfBoundsException_instance();
591 break;
592 case Deoptimization::Reason_class_check:
593 ex_obj = env()->ClassCastException_instance();
594 break;
595 case Deoptimization::Reason_array_check:
596 ex_obj = env()->ArrayStoreException_instance();
597 break;
598 default:
599 break;
600 }
601 if (failing()) { stop(); return; } // exception allocation might fail
602 if (ex_obj != nullptr) {
603 if (env()->jvmti_can_post_on_exceptions()) {
604 // check if we must post exception events, take uncommon trap if so
605 uncommon_trap_if_should_post_on_exceptions(reason, must_throw);
606 // here if should_post_on_exceptions is false
607 // continue on with the normal codegen
608 }
609
610 // Cheat with a preallocated exception object.
611 if (C->log() != nullptr)
612 C->log()->elem("hot_throw preallocated='1' reason='%s'",
613 Deoptimization::trap_reason_name(reason));
614 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
615 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
616
617 // Clear the detail message of the preallocated exception object.
618 // Weblogic sometimes mutates the detail message of exceptions
619 // using reflection.
620 int offset = java_lang_Throwable::get_detailMessage_offset();
621 const TypePtr* adr_typ = ex_con->add_offset(offset);
622
623 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
624 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
625 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
626
627 if (!method()->has_exception_handlers()) {
628 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
629 // This prevents issues with exception handling and late inlining.
630 set_sp(0);
631 clean_stack(0);
632 }
633
634 add_exception_state(make_exception_state(ex_node));
635 return;
636 }
637 }
638
639 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
640 // It won't be much cheaper than bailing to the interp., since we'll
641 // have to pass up all the debug-info, and the runtime will have to
642 // create the stack trace.
643
644 // Usual case: Bail to interpreter.
645 // Reserve the right to recompile if we haven't seen anything yet.
646
647 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr;
648 Deoptimization::DeoptAction action = Deoptimization::Action_maybe_recompile;
649 if (treat_throw_as_hot
650 && (method()->method_data()->trap_recompiled_at(bci(), m)
651 || C->too_many_traps(reason))) {
652 // We cannot afford to take more traps here. Suffer in the interpreter.
653 if (C->log() != nullptr)
654 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
655 Deoptimization::trap_reason_name(reason),
656 C->trap_count(reason));
657 action = Deoptimization::Action_none;
658 }
659
660 // "must_throw" prunes the JVM state to include only the stack, if there
661 // are no local exception handlers. This should cut down on register
662 // allocation time and code size, by drastically reducing the number
663 // of in-edges on the call to the uncommon trap.
664
665 uncommon_trap(reason, action, (ciKlass*)nullptr, (char*)nullptr, must_throw);
666 }
667
668
669 //----------------------------PreserveJVMState---------------------------------
670 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
671 debug_only(kit->verify_map());
672 _kit = kit;
673 _map = kit->map(); // preserve the map
674 _sp = kit->sp();
675 kit->set_map(clone_map ? kit->clone_map() : nullptr);
676 #ifdef ASSERT
677 _bci = kit->bci();
678 Parse* parser = kit->is_Parse();
679 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
680 _block = block;
681 #endif
682 }
683 PreserveJVMState::~PreserveJVMState() {
684 GraphKit* kit = _kit;
685 #ifdef ASSERT
686 assert(kit->bci() == _bci, "bci must not shift");
687 Parse* parser = kit->is_Parse();
688 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
689 assert(block == _block, "block must not shift");
690 #endif
691 kit->set_map(_map);
692 kit->set_sp(_sp);
693 }
694
695
696 //-----------------------------BuildCutout-------------------------------------
697 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
698 : PreserveJVMState(kit)
699 {
700 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
701 SafePointNode* outer_map = _map; // preserved map is caller's
702 SafePointNode* inner_map = kit->map();
703 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
704 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
705 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
706 }
707 BuildCutout::~BuildCutout() {
708 GraphKit* kit = _kit;
709 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
710 }
711
712 //---------------------------PreserveReexecuteState----------------------------
713 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
714 assert(!kit->stopped(), "must call stopped() before");
715 _kit = kit;
716 _sp = kit->sp();
717 _reexecute = kit->jvms()->_reexecute;
718 }
719 PreserveReexecuteState::~PreserveReexecuteState() {
720 if (_kit->stopped()) return;
721 _kit->jvms()->_reexecute = _reexecute;
722 _kit->set_sp(_sp);
723 }
724
725 //------------------------------clone_map--------------------------------------
726 // Implementation of PreserveJVMState
727 //
728 // Only clone_map(...) here. If this function is only used in the
729 // PreserveJVMState class we may want to get rid of this extra
730 // function eventually and do it all there.
731
732 SafePointNode* GraphKit::clone_map() {
733 if (map() == nullptr) return nullptr;
734
735 // Clone the memory edge first
736 Node* mem = MergeMemNode::make(map()->memory());
737 gvn().set_type_bottom(mem);
738
739 SafePointNode *clonemap = (SafePointNode*)map()->clone();
740 JVMState* jvms = this->jvms();
741 JVMState* clonejvms = jvms->clone_shallow(C);
742 clonemap->set_memory(mem);
743 clonemap->set_jvms(clonejvms);
744 clonejvms->set_map(clonemap);
745 record_for_igvn(clonemap);
746 gvn().set_type_bottom(clonemap);
747 return clonemap;
748 }
749
750 //-----------------------------destruct_map_clone------------------------------
751 //
752 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need
753 // to destruct/free/delete in the exact opposite order as clone_map().
754 void GraphKit::destruct_map_clone(SafePointNode* sfp) {
755 if (sfp == nullptr) return;
756
757 Node* mem = sfp->memory();
758 JVMState* jvms = sfp->jvms();
759
760 if (jvms != nullptr) {
761 delete jvms;
762 }
763
764 remove_for_igvn(sfp);
765 gvn().clear_type(sfp);
766 sfp->destruct(&_gvn);
767
768 if (mem != nullptr) {
769 gvn().clear_type(mem);
770 mem->destruct(&_gvn);
771 }
772 }
773
774 //-----------------------------set_map_clone-----------------------------------
775 void GraphKit::set_map_clone(SafePointNode* m) {
776 _map = m;
777 _map = clone_map();
778 _map->set_next_exception(nullptr);
779 debug_only(verify_map());
780 }
781
782
783 //----------------------------kill_dead_locals---------------------------------
784 // Detect any locals which are known to be dead, and force them to top.
785 void GraphKit::kill_dead_locals() {
786 // Consult the liveness information for the locals. If any
787 // of them are unused, then they can be replaced by top(). This
788 // should help register allocation time and cut down on the size
789 // of the deoptimization information.
790
791 // This call is made from many of the bytecode handling
792 // subroutines called from the Big Switch in do_one_bytecode.
793 // Every bytecode which might include a slow path is responsible
794 // for killing its dead locals. The more consistent we
795 // are about killing deads, the fewer useless phis will be
796 // constructed for them at various merge points.
797
798 // bci can be -1 (InvocationEntryBci). We return the entry
799 // liveness for the method.
800
801 if (method() == nullptr || method()->code_size() == 0) {
802 // We are building a graph for a call to a native method.
803 // All locals are live.
804 return;
805 }
806
807 ResourceMark rm;
808
809 // Consult the liveness information for the locals. If any
810 // of them are unused, then they can be replaced by top(). This
811 // should help register allocation time and cut down on the size
812 // of the deoptimization information.
813 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
814
815 int len = (int)live_locals.size();
816 assert(len <= jvms()->loc_size(), "too many live locals");
817 for (int local = 0; local < len; local++) {
818 if (!live_locals.at(local)) {
819 set_local(local, top());
820 }
821 }
822 }
823
824 #ifdef ASSERT
825 //-------------------------dead_locals_are_killed------------------------------
826 // Return true if all dead locals are set to top in the map.
827 // Used to assert "clean" debug info at various points.
828 bool GraphKit::dead_locals_are_killed() {
829 if (method() == nullptr || method()->code_size() == 0) {
830 // No locals need to be dead, so all is as it should be.
831 return true;
832 }
833
834 // Make sure somebody called kill_dead_locals upstream.
835 ResourceMark rm;
836 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
837 if (jvms->loc_size() == 0) continue; // no locals to consult
838 SafePointNode* map = jvms->map();
839 ciMethod* method = jvms->method();
840 int bci = jvms->bci();
841 if (jvms == this->jvms()) {
842 bci = this->bci(); // it might not yet be synched
843 }
844 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
845 int len = (int)live_locals.size();
846 if (!live_locals.is_valid() || len == 0)
847 // This method is trivial, or is poisoned by a breakpoint.
848 return true;
849 assert(len == jvms->loc_size(), "live map consistent with locals map");
850 for (int local = 0; local < len; local++) {
851 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
852 if (PrintMiscellaneous && (Verbose || WizardMode)) {
853 tty->print_cr("Zombie local %d: ", local);
854 jvms->dump();
855 }
856 return false;
857 }
858 }
859 }
860 return true;
861 }
862
863 #endif //ASSERT
864
865 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
866 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
867 ciMethod* cur_method = jvms->method();
868 int cur_bci = jvms->bci();
869 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
870 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
871 return Interpreter::bytecode_should_reexecute(code) ||
872 (is_anewarray && (code == Bytecodes::_multianewarray));
873 // Reexecute _multianewarray bytecode which was replaced with
874 // sequence of [a]newarray. See Parse::do_multianewarray().
875 //
876 // Note: interpreter should not have it set since this optimization
877 // is limited by dimensions and guarded by flag so in some cases
878 // multianewarray() runtime calls will be generated and
879 // the bytecode should not be reexecutes (stack will not be reset).
880 } else {
881 return false;
882 }
883 }
884
885 // Helper function for adding JVMState and debug information to node
886 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
887 // Add the safepoint edges to the call (or other safepoint).
888
889 // Make sure dead locals are set to top. This
890 // should help register allocation time and cut down on the size
891 // of the deoptimization information.
892 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
893
894 // Walk the inline list to fill in the correct set of JVMState's
895 // Also fill in the associated edges for each JVMState.
896
897 // If the bytecode needs to be reexecuted we need to put
898 // the arguments back on the stack.
899 const bool should_reexecute = jvms()->should_reexecute();
900 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
901
902 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
903 // undefined if the bci is different. This is normal for Parse but it
904 // should not happen for LibraryCallKit because only one bci is processed.
905 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
906 "in LibraryCallKit the reexecute bit should not change");
907
908 // If we are guaranteed to throw, we can prune everything but the
909 // input to the current bytecode.
910 bool can_prune_locals = false;
911 uint stack_slots_not_pruned = 0;
912 int inputs = 0, depth = 0;
913 if (must_throw) {
914 assert(method() == youngest_jvms->method(), "sanity");
915 if (compute_stack_effects(inputs, depth)) {
916 can_prune_locals = true;
917 stack_slots_not_pruned = inputs;
918 }
919 }
920
921 if (env()->should_retain_local_variables()) {
922 // At any safepoint, this method can get breakpointed, which would
923 // then require an immediate deoptimization.
924 can_prune_locals = false; // do not prune locals
925 stack_slots_not_pruned = 0;
926 }
927
928 // do not scribble on the input jvms
929 JVMState* out_jvms = youngest_jvms->clone_deep(C);
930 call->set_jvms(out_jvms); // Start jvms list for call node
931
932 // For a known set of bytecodes, the interpreter should reexecute them if
933 // deoptimization happens. We set the reexecute state for them here
934 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
935 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
936 #ifdef ASSERT
937 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
938 assert(method() == youngest_jvms->method(), "sanity");
939 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
940 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
941 #endif // ASSERT
942 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
943 }
944
945 // Presize the call:
946 DEBUG_ONLY(uint non_debug_edges = call->req());
947 call->add_req_batch(top(), youngest_jvms->debug_depth());
948 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
949
950 // Set up edges so that the call looks like this:
951 // Call [state:] ctl io mem fptr retadr
952 // [parms:] parm0 ... parmN
953 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
954 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
955 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
956 // Note that caller debug info precedes callee debug info.
957
958 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
959 uint debug_ptr = call->req();
960
961 // Loop over the map input edges associated with jvms, add them
962 // to the call node, & reset all offsets to match call node array.
963
964 JVMState* callee_jvms = nullptr;
965 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
966 uint debug_end = debug_ptr;
967 uint debug_start = debug_ptr - in_jvms->debug_size();
968 debug_ptr = debug_start; // back up the ptr
969
970 uint p = debug_start; // walks forward in [debug_start, debug_end)
971 uint j, k, l;
972 SafePointNode* in_map = in_jvms->map();
973 out_jvms->set_map(call);
974
975 if (can_prune_locals) {
976 assert(in_jvms->method() == out_jvms->method(), "sanity");
977 // If the current throw can reach an exception handler in this JVMS,
978 // then we must keep everything live that can reach that handler.
979 // As a quick and dirty approximation, we look for any handlers at all.
980 if (in_jvms->method()->has_exception_handlers()) {
981 can_prune_locals = false;
982 }
983 }
984
985 // Add the Locals
986 k = in_jvms->locoff();
987 l = in_jvms->loc_size();
988 out_jvms->set_locoff(p);
989 if (!can_prune_locals) {
990 for (j = 0; j < l; j++) {
991 Node* val = in_map->in(k + j);
992 // Check if there's a larval that has been written in the callee state (constructor) and update it in the caller state
993 if (callee_jvms != nullptr && val->is_InlineType() && val->as_InlineType()->is_larval() &&
994 callee_jvms->method()->is_object_constructor() && val == in_map->argument(in_jvms, 0) &&
995 val->bottom_type()->is_inlinetypeptr()) {
996 val = callee_jvms->map()->local(callee_jvms, 0); // Receiver
997 }
998 call->set_req(p++, val);
999 }
1000 } else {
1001 p += l; // already set to top above by add_req_batch
1002 }
1003
1004 // Add the Expression Stack
1005 k = in_jvms->stkoff();
1006 l = in_jvms->sp();
1007 out_jvms->set_stkoff(p);
1008 if (!can_prune_locals) {
1009 for (j = 0; j < l; j++) {
1010 Node* val = in_map->in(k + j);
1011 // Check if there's a larval that has been written in the callee state (constructor) and update it in the caller state
1012 if (callee_jvms != nullptr && val->is_InlineType() && val->as_InlineType()->is_larval() &&
1013 callee_jvms->method()->is_object_constructor() && val == in_map->argument(in_jvms, 0) &&
1014 val->bottom_type()->is_inlinetypeptr()) {
1015 val = callee_jvms->map()->local(callee_jvms, 0); // Receiver
1016 }
1017 call->set_req(p++, val);
1018 }
1019 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1020 // Divide stack into {S0,...,S1}, where S0 is set to top.
1021 uint s1 = stack_slots_not_pruned;
1022 stack_slots_not_pruned = 0; // for next iteration
1023 if (s1 > l) s1 = l;
1024 uint s0 = l - s1;
1025 p += s0; // skip the tops preinstalled by add_req_batch
1026 for (j = s0; j < l; j++)
1027 call->set_req(p++, in_map->in(k+j));
1028 } else {
1029 p += l; // already set to top above by add_req_batch
1030 }
1031
1032 // Add the Monitors
1033 k = in_jvms->monoff();
1034 l = in_jvms->mon_size();
1035 out_jvms->set_monoff(p);
1036 for (j = 0; j < l; j++)
1037 call->set_req(p++, in_map->in(k+j));
1038
1039 // Copy any scalar object fields.
1040 k = in_jvms->scloff();
1041 l = in_jvms->scl_size();
1042 out_jvms->set_scloff(p);
1043 for (j = 0; j < l; j++)
1044 call->set_req(p++, in_map->in(k+j));
1045
1046 // Finish the new jvms.
1047 out_jvms->set_endoff(p);
1048
1049 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1050 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1051 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1052 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1053 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1054 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1055
1056 // Update the two tail pointers in parallel.
1057 callee_jvms = out_jvms;
1058 out_jvms = out_jvms->caller();
1059 in_jvms = in_jvms->caller();
1060 }
1061
1062 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1063
1064 // Test the correctness of JVMState::debug_xxx accessors:
1065 assert(call->jvms()->debug_start() == non_debug_edges, "");
1066 assert(call->jvms()->debug_end() == call->req(), "");
1067 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1068 }
1069
1070 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1071 Bytecodes::Code code = java_bc();
1072 if (code == Bytecodes::_wide) {
1073 code = method()->java_code_at_bci(bci() + 1);
1074 }
1075
1076 if (code != Bytecodes::_illegal) {
1077 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1078 }
1079
1080 auto rsize = [&]() {
1081 assert(code != Bytecodes::_illegal, "code is illegal!");
1082 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1083 return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1084 };
1085
1086 switch (code) {
1087 case Bytecodes::_illegal:
1088 return false;
1089
1090 case Bytecodes::_ldc:
1091 case Bytecodes::_ldc_w:
1092 case Bytecodes::_ldc2_w:
1093 inputs = 0;
1094 break;
1095
1096 case Bytecodes::_dup: inputs = 1; break;
1097 case Bytecodes::_dup_x1: inputs = 2; break;
1098 case Bytecodes::_dup_x2: inputs = 3; break;
1099 case Bytecodes::_dup2: inputs = 2; break;
1100 case Bytecodes::_dup2_x1: inputs = 3; break;
1101 case Bytecodes::_dup2_x2: inputs = 4; break;
1102 case Bytecodes::_swap: inputs = 2; break;
1103 case Bytecodes::_arraylength: inputs = 1; break;
1104
1105 case Bytecodes::_getstatic:
1106 case Bytecodes::_putstatic:
1107 case Bytecodes::_getfield:
1108 case Bytecodes::_putfield:
1109 {
1110 bool ignored_will_link;
1111 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1112 int size = field->type()->size();
1113 bool is_get = (depth >= 0), is_static = (depth & 1);
1114 inputs = (is_static ? 0 : 1);
1115 if (is_get) {
1116 depth = size - inputs;
1117 } else {
1118 inputs += size; // putxxx pops the value from the stack
1119 depth = - inputs;
1120 }
1121 }
1122 break;
1123
1124 case Bytecodes::_invokevirtual:
1125 case Bytecodes::_invokespecial:
1126 case Bytecodes::_invokestatic:
1127 case Bytecodes::_invokedynamic:
1128 case Bytecodes::_invokeinterface:
1129 {
1130 bool ignored_will_link;
1131 ciSignature* declared_signature = nullptr;
1132 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1133 assert(declared_signature != nullptr, "cannot be null");
1134 inputs = declared_signature->arg_size_for_bc(code);
1135 int size = declared_signature->return_type()->size();
1136 depth = size - inputs;
1137 }
1138 break;
1139
1140 case Bytecodes::_multianewarray:
1141 {
1142 ciBytecodeStream iter(method());
1143 iter.reset_to_bci(bci());
1144 iter.next();
1145 inputs = iter.get_dimensions();
1146 assert(rsize() == 1, "");
1147 depth = 1 - inputs;
1148 }
1149 break;
1150
1151 case Bytecodes::_ireturn:
1152 case Bytecodes::_lreturn:
1153 case Bytecodes::_freturn:
1154 case Bytecodes::_dreturn:
1155 case Bytecodes::_areturn:
1156 assert(rsize() == -depth, "");
1157 inputs = -depth;
1158 break;
1159
1160 case Bytecodes::_jsr:
1161 case Bytecodes::_jsr_w:
1162 inputs = 0;
1163 depth = 1; // S.B. depth=1, not zero
1164 break;
1165
1166 default:
1167 // bytecode produces a typed result
1168 inputs = rsize() - depth;
1169 assert(inputs >= 0, "");
1170 break;
1171 }
1172
1173 #ifdef ASSERT
1174 // spot check
1175 int outputs = depth + inputs;
1176 assert(outputs >= 0, "sanity");
1177 switch (code) {
1178 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1179 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1180 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1181 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1182 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1183 default: break;
1184 }
1185 #endif //ASSERT
1186
1187 return true;
1188 }
1189
1190
1191
1192 //------------------------------basic_plus_adr---------------------------------
1193 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1194 // short-circuit a common case
1195 if (offset == intcon(0)) return ptr;
1196 return _gvn.transform( new AddPNode(base, ptr, offset) );
1197 }
1198
1199 Node* GraphKit::ConvI2L(Node* offset) {
1200 // short-circuit a common case
1201 jint offset_con = find_int_con(offset, Type::OffsetBot);
1202 if (offset_con != Type::OffsetBot) {
1203 return longcon((jlong) offset_con);
1204 }
1205 return _gvn.transform( new ConvI2LNode(offset));
1206 }
1207
1208 Node* GraphKit::ConvI2UL(Node* offset) {
1209 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1210 if (offset_con != (juint) Type::OffsetBot) {
1211 return longcon((julong) offset_con);
1212 }
1213 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1214 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1215 return _gvn.transform( new AndLNode(conv, mask) );
1216 }
1217
1218 Node* GraphKit::ConvL2I(Node* offset) {
1219 // short-circuit a common case
1220 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1221 if (offset_con != (jlong)Type::OffsetBot) {
1222 return intcon((int) offset_con);
1223 }
1224 return _gvn.transform( new ConvL2INode(offset));
1225 }
1226
1227 //-------------------------load_object_klass-----------------------------------
1228 Node* GraphKit::load_object_klass(Node* obj) {
1229 // Special-case a fresh allocation to avoid building nodes:
1230 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1231 if (akls != nullptr) return akls;
1232 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1233 return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1234 }
1235
1236 //-------------------------load_array_length-----------------------------------
1237 Node* GraphKit::load_array_length(Node* array) {
1238 // Special-case a fresh allocation to avoid building nodes:
1239 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1240 Node *alen;
1241 if (alloc == nullptr) {
1242 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1243 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1244 } else {
1245 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1246 }
1247 return alen;
1248 }
1249
1250 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1251 const TypeOopPtr* oop_type,
1252 bool replace_length_in_map) {
1253 Node* length = alloc->Ideal_length();
1254 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1255 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1256 if (ccast != length) {
1257 // do not transform ccast here, it might convert to top node for
1258 // negative array length and break assumptions in parsing stage.
1259 _gvn.set_type_bottom(ccast);
1260 record_for_igvn(ccast);
1261 if (replace_length_in_map) {
1262 replace_in_map(length, ccast);
1263 }
1264 return ccast;
1265 }
1266 }
1267 return length;
1268 }
1269
1270 //------------------------------do_null_check----------------------------------
1271 // Helper function to do a null pointer check. Returned value is
1272 // the incoming address with null casted away. You are allowed to use the
1273 // not-null value only if you are control dependent on the test.
1274 #ifndef PRODUCT
1275 extern uint explicit_null_checks_inserted,
1276 explicit_null_checks_elided;
1277 #endif
1278 Node* GraphKit::null_check_common(Node* value, BasicType type,
1279 // optional arguments for variations:
1280 bool assert_null,
1281 Node* *null_control,
1282 bool speculative,
1283 bool is_init_check) {
1284 assert(!assert_null || null_control == nullptr, "not both at once");
1285 if (stopped()) return top();
1286 NOT_PRODUCT(explicit_null_checks_inserted++);
1287
1288 if (value->is_InlineType()) {
1289 // Null checking a scalarized but nullable inline type. Check the IsInit
1290 // input instead of the oop input to avoid keeping buffer allocations alive.
1291 InlineTypeNode* vtptr = value->as_InlineType();
1292 while (vtptr->get_oop()->is_InlineType()) {
1293 vtptr = vtptr->get_oop()->as_InlineType();
1294 }
1295 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true);
1296 if (stopped()) {
1297 return top();
1298 }
1299 if (assert_null) {
1300 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1301 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1302 // replace_in_map(value, vtptr);
1303 // return vtptr;
1304 replace_in_map(value, null());
1305 return null();
1306 }
1307 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1308 return cast_not_null(value, do_replace_in_map);
1309 }
1310
1311 // Construct null check
1312 Node *chk = nullptr;
1313 switch(type) {
1314 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1315 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1316 case T_ARRAY : // fall through
1317 type = T_OBJECT; // simplify further tests
1318 case T_OBJECT : {
1319 const Type *t = _gvn.type( value );
1320
1321 const TypeOopPtr* tp = t->isa_oopptr();
1322 if (tp != nullptr && !tp->is_loaded()
1323 // Only for do_null_check, not any of its siblings:
1324 && !assert_null && null_control == nullptr) {
1325 // Usually, any field access or invocation on an unloaded oop type
1326 // will simply fail to link, since the statically linked class is
1327 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1328 // the static class is loaded but the sharper oop type is not.
1329 // Rather than checking for this obscure case in lots of places,
1330 // we simply observe that a null check on an unloaded class
1331 // will always be followed by a nonsense operation, so we
1332 // can just issue the uncommon trap here.
1333 // Our access to the unloaded class will only be correct
1334 // after it has been loaded and initialized, which requires
1335 // a trip through the interpreter.
1336 ciKlass* klass = tp->unloaded_klass();
1337 #ifndef PRODUCT
1338 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1339 #endif
1340 uncommon_trap(Deoptimization::Reason_unloaded,
1341 Deoptimization::Action_reinterpret,
1342 klass, "!loaded");
1343 return top();
1344 }
1345
1346 if (assert_null) {
1347 // See if the type is contained in NULL_PTR.
1348 // If so, then the value is already null.
1349 if (t->higher_equal(TypePtr::NULL_PTR)) {
1350 NOT_PRODUCT(explicit_null_checks_elided++);
1351 return value; // Elided null assert quickly!
1352 }
1353 } else {
1354 // See if mixing in the null pointer changes type.
1355 // If so, then the null pointer was not allowed in the original
1356 // type. In other words, "value" was not-null.
1357 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1358 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1359 NOT_PRODUCT(explicit_null_checks_elided++);
1360 return value; // Elided null check quickly!
1361 }
1362 }
1363 chk = new CmpPNode( value, null() );
1364 break;
1365 }
1366
1367 default:
1368 fatal("unexpected type: %s", type2name(type));
1369 }
1370 assert(chk != nullptr, "sanity check");
1371 chk = _gvn.transform(chk);
1372
1373 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1374 BoolNode *btst = new BoolNode( chk, btest);
1375 Node *tst = _gvn.transform( btst );
1376
1377 //-----------
1378 // if peephole optimizations occurred, a prior test existed.
1379 // If a prior test existed, maybe it dominates as we can avoid this test.
1380 if (tst != btst && type == T_OBJECT) {
1381 // At this point we want to scan up the CFG to see if we can
1382 // find an identical test (and so avoid this test altogether).
1383 Node *cfg = control();
1384 int depth = 0;
1385 while( depth < 16 ) { // Limit search depth for speed
1386 if( cfg->Opcode() == Op_IfTrue &&
1387 cfg->in(0)->in(1) == tst ) {
1388 // Found prior test. Use "cast_not_null" to construct an identical
1389 // CastPP (and hence hash to) as already exists for the prior test.
1390 // Return that casted value.
1391 if (assert_null) {
1392 replace_in_map(value, null());
1393 return null(); // do not issue the redundant test
1394 }
1395 Node *oldcontrol = control();
1396 set_control(cfg);
1397 Node *res = cast_not_null(value);
1398 set_control(oldcontrol);
1399 NOT_PRODUCT(explicit_null_checks_elided++);
1400 return res;
1401 }
1402 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1403 if (cfg == nullptr) break; // Quit at region nodes
1404 depth++;
1405 }
1406 }
1407
1408 //-----------
1409 // Branch to failure if null
1410 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1411 Deoptimization::DeoptReason reason;
1412 if (assert_null) {
1413 reason = Deoptimization::reason_null_assert(speculative);
1414 } else if (type == T_OBJECT || is_init_check) {
1415 reason = Deoptimization::reason_null_check(speculative);
1416 } else {
1417 reason = Deoptimization::Reason_div0_check;
1418 }
1419 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1420 // ciMethodData::has_trap_at will return a conservative -1 if any
1421 // must-be-null assertion has failed. This could cause performance
1422 // problems for a method after its first do_null_assert failure.
1423 // Consider using 'Reason_class_check' instead?
1424
1425 // To cause an implicit null check, we set the not-null probability
1426 // to the maximum (PROB_MAX). For an explicit check the probability
1427 // is set to a smaller value.
1428 if (null_control != nullptr || too_many_traps(reason)) {
1429 // probability is less likely
1430 ok_prob = PROB_LIKELY_MAG(3);
1431 } else if (!assert_null &&
1432 (ImplicitNullCheckThreshold > 0) &&
1433 method() != nullptr &&
1434 (method()->method_data()->trap_count(reason)
1435 >= (uint)ImplicitNullCheckThreshold)) {
1436 ok_prob = PROB_LIKELY_MAG(3);
1437 }
1438
1439 if (null_control != nullptr) {
1440 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1441 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1442 set_control( _gvn.transform( new IfTrueNode(iff)));
1443 #ifndef PRODUCT
1444 if (null_true == top()) {
1445 explicit_null_checks_elided++;
1446 }
1447 #endif
1448 (*null_control) = null_true;
1449 } else {
1450 BuildCutout unless(this, tst, ok_prob);
1451 // Check for optimizer eliding test at parse time
1452 if (stopped()) {
1453 // Failure not possible; do not bother making uncommon trap.
1454 NOT_PRODUCT(explicit_null_checks_elided++);
1455 } else if (assert_null) {
1456 uncommon_trap(reason,
1457 Deoptimization::Action_make_not_entrant,
1458 nullptr, "assert_null");
1459 } else {
1460 replace_in_map(value, zerocon(type));
1461 builtin_throw(reason);
1462 }
1463 }
1464
1465 // Must throw exception, fall-thru not possible?
1466 if (stopped()) {
1467 return top(); // No result
1468 }
1469
1470 if (assert_null) {
1471 // Cast obj to null on this path.
1472 replace_in_map(value, zerocon(type));
1473 return zerocon(type);
1474 }
1475
1476 // Cast obj to not-null on this path, if there is no null_control.
1477 // (If there is a null_control, a non-null value may come back to haunt us.)
1478 if (type == T_OBJECT) {
1479 Node* cast = cast_not_null(value, false);
1480 if (null_control == nullptr || (*null_control) == top())
1481 replace_in_map(value, cast);
1482 value = cast;
1483 }
1484
1485 return value;
1486 }
1487
1488 //------------------------------cast_not_null----------------------------------
1489 // Cast obj to not-null on this path
1490 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1491 if (obj->is_InlineType()) {
1492 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1493 vt->as_InlineType()->set_is_init(_gvn);
1494 vt = _gvn.transform(vt);
1495 if (do_replace_in_map) {
1496 replace_in_map(obj, vt);
1497 }
1498 return vt;
1499 }
1500 const Type *t = _gvn.type(obj);
1501 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1502 // Object is already not-null?
1503 if( t == t_not_null ) return obj;
1504
1505 Node* cast = new CastPPNode(control(), obj,t_not_null);
1506 cast = _gvn.transform( cast );
1507
1508 // Scan for instances of 'obj' in the current JVM mapping.
1509 // These instances are known to be not-null after the test.
1510 if (do_replace_in_map)
1511 replace_in_map(obj, cast);
1512
1513 return cast; // Return casted value
1514 }
1515
1516 // Sometimes in intrinsics, we implicitly know an object is not null
1517 // (there's no actual null check) so we can cast it to not null. In
1518 // the course of optimizations, the input to the cast can become null.
1519 // In that case that data path will die and we need the control path
1520 // to become dead as well to keep the graph consistent. So we have to
1521 // add a check for null for which one branch can't be taken. It uses
1522 // an Opaque4 node that will cause the check to be removed after loop
1523 // opts so the test goes away and the compiled code doesn't execute a
1524 // useless check.
1525 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1526 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1527 return value;
1528 }
1529 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1530 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1531 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1532 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1533 _gvn.set_type(iff, iff->Value(&_gvn));
1534 if (!tst->is_Con()) {
1535 record_for_igvn(iff);
1536 }
1537 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1538 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1539 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1540 C->root()->add_req(halt);
1541 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1542 set_control(if_t);
1543 return cast_not_null(value, do_replace_in_map);
1544 }
1545
1546
1547 //--------------------------replace_in_map-------------------------------------
1548 void GraphKit::replace_in_map(Node* old, Node* neww) {
1549 if (old == neww) {
1550 return;
1551 }
1552
1553 map()->replace_edge(old, neww);
1554
1555 // Note: This operation potentially replaces any edge
1556 // on the map. This includes locals, stack, and monitors
1557 // of the current (innermost) JVM state.
1558
1559 // don't let inconsistent types from profiling escape this
1560 // method
1561
1562 const Type* told = _gvn.type(old);
1563 const Type* tnew = _gvn.type(neww);
1564
1565 if (!tnew->higher_equal(told)) {
1566 return;
1567 }
1568
1569 map()->record_replaced_node(old, neww);
1570 }
1571
1572
1573 //=============================================================================
1574 //--------------------------------memory---------------------------------------
1575 Node* GraphKit::memory(uint alias_idx) {
1576 MergeMemNode* mem = merged_memory();
1577 Node* p = mem->memory_at(alias_idx);
1578 assert(p != mem->empty_memory(), "empty");
1579 _gvn.set_type(p, Type::MEMORY); // must be mapped
1580 return p;
1581 }
1582
1583 //-----------------------------reset_memory------------------------------------
1584 Node* GraphKit::reset_memory() {
1585 Node* mem = map()->memory();
1586 // do not use this node for any more parsing!
1587 debug_only( map()->set_memory((Node*)nullptr) );
1588 return _gvn.transform( mem );
1589 }
1590
1591 //------------------------------set_all_memory---------------------------------
1592 void GraphKit::set_all_memory(Node* newmem) {
1593 Node* mergemem = MergeMemNode::make(newmem);
1594 gvn().set_type_bottom(mergemem);
1595 map()->set_memory(mergemem);
1596 }
1597
1598 //------------------------------set_all_memory_call----------------------------
1599 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1600 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1601 set_all_memory(newmem);
1602 }
1603
1604 //=============================================================================
1605 //
1606 // parser factory methods for MemNodes
1607 //
1608 // These are layered on top of the factory methods in LoadNode and StoreNode,
1609 // and integrate with the parser's memory state and _gvn engine.
1610 //
1611
1612 // factory methods in "int adr_idx"
1613 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1614 int adr_idx,
1615 MemNode::MemOrd mo,
1616 LoadNode::ControlDependency control_dependency,
1617 bool require_atomic_access,
1618 bool unaligned,
1619 bool mismatched,
1620 bool unsafe,
1621 uint8_t barrier_data) {
1622 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1623 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1624 debug_only(adr_type = C->get_adr_type(adr_idx));
1625 Node* mem = memory(adr_idx);
1626 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1627 ld = _gvn.transform(ld);
1628
1629 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1630 // Improve graph before escape analysis and boxing elimination.
1631 record_for_igvn(ld);
1632 if (ld->is_DecodeN()) {
1633 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1634 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1635 // a Phi). Recording such cases is still perfectly sound, but may be
1636 // unnecessary and result in some minor IGVN overhead.
1637 record_for_igvn(ld->in(1));
1638 }
1639 }
1640 return ld;
1641 }
1642
1643 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1644 int adr_idx,
1645 MemNode::MemOrd mo,
1646 bool require_atomic_access,
1647 bool unaligned,
1648 bool mismatched,
1649 bool unsafe,
1650 int barrier_data) {
1651 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1652 const TypePtr* adr_type = nullptr;
1653 debug_only(adr_type = C->get_adr_type(adr_idx));
1654 Node *mem = memory(adr_idx);
1655 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1656 if (unaligned) {
1657 st->as_Store()->set_unaligned_access();
1658 }
1659 if (mismatched) {
1660 st->as_Store()->set_mismatched_access();
1661 }
1662 if (unsafe) {
1663 st->as_Store()->set_unsafe_access();
1664 }
1665 st->as_Store()->set_barrier_data(barrier_data);
1666 st = _gvn.transform(st);
1667 set_memory(st, adr_idx);
1668 // Back-to-back stores can only remove intermediate store with DU info
1669 // so push on worklist for optimizer.
1670 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1671 record_for_igvn(st);
1672
1673 return st;
1674 }
1675
1676 Node* GraphKit::access_store_at(Node* obj,
1677 Node* adr,
1678 const TypePtr* adr_type,
1679 Node* val,
1680 const Type* val_type,
1681 BasicType bt,
1682 DecoratorSet decorators,
1683 bool safe_for_replace) {
1684 // Transformation of a value which could be null pointer (CastPP #null)
1685 // could be delayed during Parse (for example, in adjust_map_after_if()).
1686 // Execute transformation here to avoid barrier generation in such case.
1687 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1688 val = _gvn.makecon(TypePtr::NULL_PTR);
1689 }
1690
1691 if (stopped()) {
1692 return top(); // Dead path ?
1693 }
1694
1695 assert(val != nullptr, "not dead path");
1696 if (val->is_InlineType()) {
1697 // Store to non-flat field. Buffer the inline type and make sure
1698 // the store is re-executed if the allocation triggers deoptimization.
1699 PreserveReexecuteState preexecs(this);
1700 jvms()->set_should_reexecute(true);
1701 val = val->as_InlineType()->buffer(this, safe_for_replace);
1702 }
1703
1704 C2AccessValuePtr addr(adr, adr_type);
1705 C2AccessValue value(val, val_type);
1706 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1707 if (access.is_raw()) {
1708 return _barrier_set->BarrierSetC2::store_at(access, value);
1709 } else {
1710 return _barrier_set->store_at(access, value);
1711 }
1712 }
1713
1714 Node* GraphKit::access_load_at(Node* obj, // containing obj
1715 Node* adr, // actual address to store val at
1716 const TypePtr* adr_type,
1717 const Type* val_type,
1718 BasicType bt,
1719 DecoratorSet decorators,
1720 Node* ctl) {
1721 if (stopped()) {
1722 return top(); // Dead path ?
1723 }
1724
1725 C2AccessValuePtr addr(adr, adr_type);
1726 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1727 if (access.is_raw()) {
1728 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1729 } else {
1730 return _barrier_set->load_at(access, val_type);
1731 }
1732 }
1733
1734 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1735 const Type* val_type,
1736 BasicType bt,
1737 DecoratorSet decorators) {
1738 if (stopped()) {
1739 return top(); // Dead path ?
1740 }
1741
1742 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1743 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1744 if (access.is_raw()) {
1745 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1746 } else {
1747 return _barrier_set->load_at(access, val_type);
1748 }
1749 }
1750
1751 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1752 Node* adr,
1753 const TypePtr* adr_type,
1754 int alias_idx,
1755 Node* expected_val,
1756 Node* new_val,
1757 const Type* value_type,
1758 BasicType bt,
1759 DecoratorSet decorators) {
1760 C2AccessValuePtr addr(adr, adr_type);
1761 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1762 bt, obj, addr, alias_idx);
1763 if (access.is_raw()) {
1764 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1765 } else {
1766 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1767 }
1768 }
1769
1770 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1771 Node* adr,
1772 const TypePtr* adr_type,
1773 int alias_idx,
1774 Node* expected_val,
1775 Node* new_val,
1776 const Type* value_type,
1777 BasicType bt,
1778 DecoratorSet decorators) {
1779 C2AccessValuePtr addr(adr, adr_type);
1780 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1781 bt, obj, addr, alias_idx);
1782 if (access.is_raw()) {
1783 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1784 } else {
1785 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1786 }
1787 }
1788
1789 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1790 Node* adr,
1791 const TypePtr* adr_type,
1792 int alias_idx,
1793 Node* new_val,
1794 const Type* value_type,
1795 BasicType bt,
1796 DecoratorSet decorators) {
1797 C2AccessValuePtr addr(adr, adr_type);
1798 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1799 bt, obj, addr, alias_idx);
1800 if (access.is_raw()) {
1801 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1802 } else {
1803 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1804 }
1805 }
1806
1807 Node* GraphKit::access_atomic_add_at(Node* obj,
1808 Node* adr,
1809 const TypePtr* adr_type,
1810 int alias_idx,
1811 Node* new_val,
1812 const Type* value_type,
1813 BasicType bt,
1814 DecoratorSet decorators) {
1815 C2AccessValuePtr addr(adr, adr_type);
1816 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1817 if (access.is_raw()) {
1818 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1819 } else {
1820 return _barrier_set->atomic_add_at(access, new_val, value_type);
1821 }
1822 }
1823
1824 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1825 return _barrier_set->clone(this, src, dst, size, is_array);
1826 }
1827
1828 //-------------------------array_element_address-------------------------
1829 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1830 const TypeInt* sizetype, Node* ctrl) {
1831 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1832 uint shift = arytype->is_flat() ? arytype->flat_log_elem_size() : exact_log2(type2aelembytes(elembt));
1833 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1834
1835 // short-circuit a common case (saves lots of confusing waste motion)
1836 jint idx_con = find_int_con(idx, -1);
1837 if (idx_con >= 0) {
1838 intptr_t offset = header + ((intptr_t)idx_con << shift);
1839 return basic_plus_adr(ary, offset);
1840 }
1841
1842 // must be correct type for alignment purposes
1843 Node* base = basic_plus_adr(ary, header);
1844 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1845 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1846 return basic_plus_adr(ary, base, scale);
1847 }
1848
1849 //-------------------------load_array_element-------------------------
1850 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1851 const Type* elemtype = arytype->elem();
1852 BasicType elembt = elemtype->array_element_basic_type();
1853 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1854 if (elembt == T_NARROWOOP) {
1855 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1856 }
1857 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1858 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1859 return ld;
1860 }
1861
1862 //-------------------------set_arguments_for_java_call-------------------------
1863 // Arguments (pre-popped from the stack) are taken from the JVMS.
1864 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1865 PreserveReexecuteState preexecs(this);
1866 if (EnableValhalla) {
1867 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1868 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1869 jvms()->set_should_reexecute(true);
1870 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1871 inc_sp(arg_size);
1872 }
1873 // Add the call arguments
1874 const TypeTuple* domain = call->tf()->domain_sig();
1875 uint nargs = domain->cnt();
1876 int arg_num = 0;
1877 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1878 Node* arg = argument(i-TypeFunc::Parms);
1879 const Type* t = domain->field_at(i);
1880 // TODO 8284443 A static call to a mismatched method should still be scalarized
1881 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1882 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1883 if (!arg->is_InlineType()) {
1884 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1885 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass(), t->inline_klass()->is_null_free());
1886 }
1887 InlineTypeNode* vt = arg->as_InlineType();
1888 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1889 // If an inline type argument is passed as fields, attach the Method* to the call site
1890 // to be able to access the extended signature later via attached_method_before_pc().
1891 // For example, see CompiledMethod::preserve_callee_argument_oops().
1892 call->set_override_symbolic_info(true);
1893 // Register an evol dependency on the callee method to make sure that this method is deoptimized and
1894 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1895 C->dependencies()->assert_evol_method(call->method());
1896 arg_num++;
1897 continue;
1898 } else if (arg->is_InlineType()) {
1899 // Pass inline type argument via oop to callee
1900 InlineTypeNode* inline_type = arg->as_InlineType();
1901 const ciMethod* method = call->method();
1902 ciInstanceKlass* holder = method->holder();
1903 const bool is_receiver = (i == TypeFunc::Parms);
1904 const bool is_abstract_or_object_klass_constructor = method->is_object_constructor() &&
1905 (holder->is_abstract() || holder->is_java_lang_Object());
1906 const bool is_larval_receiver_on_super_constructor = is_receiver && is_abstract_or_object_klass_constructor;
1907 bool must_init_buffer = true;
1908 // We always need to buffer inline types when they are escaping. However, we can skip the actual initialization
1909 // of the buffer if the inline type is a larval because we are going to update the buffer anyway which requires
1910 // us to create a new one. But there is one special case where we are still required to initialize the buffer:
1911 // When we have a larval receiver invoked on an abstract (value class) constructor or the Object constructor (that
1912 // is not going to be inlined). After this call, the larval is completely initialized and thus not a larval anymore.
1913 // We therefore need to force an initialization of the buffer to not lose all the field writes so far in case the
1914 // buffer needs to be used (e.g. to read from when deoptimizing at runtime) or further updated in abstract super
1915 // value class constructors which could have more fields to be initialized. Note that we do not need to
1916 // initialize the buffer when invoking another constructor in the same class on a larval receiver because we
1917 // have not initialized any fields, yet (this is done completely by the other constructor call).
1918 if (inline_type->is_larval() && !is_larval_receiver_on_super_constructor) {
1919 must_init_buffer = false;
1920 }
1921 arg = inline_type->buffer(this, true, must_init_buffer);
1922 }
1923 if (t != Type::HALF) {
1924 arg_num++;
1925 }
1926 call->init_req(idx++, arg);
1927 }
1928 }
1929
1930 //---------------------------set_edges_for_java_call---------------------------
1931 // Connect a newly created call into the current JVMS.
1932 // A return value node (if any) is returned from set_edges_for_java_call.
1933 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1934
1935 // Add the predefined inputs:
1936 call->init_req( TypeFunc::Control, control() );
1937 call->init_req( TypeFunc::I_O , i_o() );
1938 call->init_req( TypeFunc::Memory , reset_memory() );
1939 call->init_req( TypeFunc::FramePtr, frameptr() );
1940 call->init_req( TypeFunc::ReturnAdr, top() );
1941
1942 add_safepoint_edges(call, must_throw);
1943
1944 Node* xcall = _gvn.transform(call);
1945
1946 if (xcall == top()) {
1947 set_control(top());
1948 return;
1949 }
1950 assert(xcall == call, "call identity is stable");
1951
1952 // Re-use the current map to produce the result.
1953
1954 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1955 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1956 set_all_memory_call(xcall, separate_io_proj);
1957
1958 //return xcall; // no need, caller already has it
1959 }
1960
1961 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1962 if (stopped()) return top(); // maybe the call folded up?
1963
1964 // Note: Since any out-of-line call can produce an exception,
1965 // we always insert an I_O projection from the call into the result.
1966
1967 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1968
1969 if (separate_io_proj) {
1970 // The caller requested separate projections be used by the fall
1971 // through and exceptional paths, so replace the projections for
1972 // the fall through path.
1973 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1974 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1975 }
1976
1977 // Capture the return value, if any.
1978 Node* ret;
1979 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
1980 ret = top();
1981 } else if (call->tf()->returns_inline_type_as_fields()) {
1982 // Return of multiple values (inline type fields): we create a
1983 // InlineType node, each field is a projection from the call.
1984 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
1985 uint base_input = TypeFunc::Parms;
1986 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
1987 } else {
1988 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1989 ciType* t = call->method()->return_type();
1990 if (t->is_klass()) {
1991 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
1992 if (type->is_inlinetypeptr()) {
1993 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass(), type->inline_klass()->is_null_free());
1994 }
1995 }
1996 }
1997
1998 // We just called the constructor on a value type receiver. Reload it from the buffer
1999 ciMethod* method = call->method();
2000 if (method->is_object_constructor() && !method->holder()->is_java_lang_Object()) {
2001 InlineTypeNode* inline_type_receiver = call->in(TypeFunc::Parms)->isa_InlineType();
2002 if (inline_type_receiver != nullptr) {
2003 assert(inline_type_receiver->is_larval(), "must be larval");
2004 assert(inline_type_receiver->is_allocated(&gvn()), "larval must be buffered");
2005 InlineTypeNode* reloaded = InlineTypeNode::make_from_oop(this, inline_type_receiver->get_oop(),
2006 inline_type_receiver->bottom_type()->inline_klass(), true);
2007 assert(!reloaded->is_larval(), "should not be larval anymore");
2008 replace_in_map(inline_type_receiver, reloaded);
2009 }
2010 }
2011
2012 return ret;
2013 }
2014
2015 //--------------------set_predefined_input_for_runtime_call--------------------
2016 // Reading and setting the memory state is way conservative here.
2017 // The real problem is that I am not doing real Type analysis on memory,
2018 // so I cannot distinguish card mark stores from other stores. Across a GC
2019 // point the Store Barrier and the card mark memory has to agree. I cannot
2020 // have a card mark store and its barrier split across the GC point from
2021 // either above or below. Here I get that to happen by reading ALL of memory.
2022 // A better answer would be to separate out card marks from other memory.
2023 // For now, return the input memory state, so that it can be reused
2024 // after the call, if this call has restricted memory effects.
2025 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2026 // Set fixed predefined input arguments
2027 Node* memory = reset_memory();
2028 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2029 call->init_req( TypeFunc::Control, control() );
2030 call->init_req( TypeFunc::I_O, top() ); // does no i/o
2031 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
2032 call->init_req( TypeFunc::FramePtr, frameptr() );
2033 call->init_req( TypeFunc::ReturnAdr, top() );
2034 return memory;
2035 }
2036
2037 //-------------------set_predefined_output_for_runtime_call--------------------
2038 // Set control and memory (not i_o) from the call.
2039 // If keep_mem is not null, use it for the output state,
2040 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
2041 // If hook_mem is null, this call produces no memory effects at all.
2042 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
2043 // then only that memory slice is taken from the call.
2044 // In the last case, we must put an appropriate memory barrier before
2045 // the call, so as to create the correct anti-dependencies on loads
2046 // preceding the call.
2047 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
2048 Node* keep_mem,
2049 const TypePtr* hook_mem) {
2050 // no i/o
2051 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
2052 if (keep_mem) {
2053 // First clone the existing memory state
2054 set_all_memory(keep_mem);
2055 if (hook_mem != nullptr) {
2056 // Make memory for the call
2057 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
2058 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
2059 // We also use hook_mem to extract specific effects from arraycopy stubs.
2060 set_memory(mem, hook_mem);
2061 }
2062 // ...else the call has NO memory effects.
2063
2064 // Make sure the call advertises its memory effects precisely.
2065 // This lets us build accurate anti-dependences in gcm.cpp.
2066 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
2067 "call node must be constructed correctly");
2068 } else {
2069 assert(hook_mem == nullptr, "");
2070 // This is not a "slow path" call; all memory comes from the call.
2071 set_all_memory_call(call);
2072 }
2073 }
2074
2075 // Keep track of MergeMems feeding into other MergeMems
2076 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2077 if (!mem->is_MergeMem()) {
2078 return;
2079 }
2080 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2081 Node* use = i.get();
2082 if (use->is_MergeMem()) {
2083 wl.push(use);
2084 }
2085 }
2086 }
2087
2088 // Replace the call with the current state of the kit.
2089 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
2090 JVMState* ejvms = nullptr;
2091 if (has_exceptions()) {
2092 ejvms = transfer_exceptions_into_jvms();
2093 }
2094
2095 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2096 ReplacedNodes replaced_nodes_exception;
2097 Node* ex_ctl = top();
2098
2099 SafePointNode* final_state = stop();
2100
2101 // Find all the needed outputs of this call
2102 CallProjections* callprojs = call->extract_projections(true);
2103
2104 Unique_Node_List wl;
2105 Node* init_mem = call->in(TypeFunc::Memory);
2106 Node* final_mem = final_state->in(TypeFunc::Memory);
2107 Node* final_ctl = final_state->in(TypeFunc::Control);
2108 Node* final_io = final_state->in(TypeFunc::I_O);
2109
2110 // Replace all the old call edges with the edges from the inlining result
2111 if (callprojs->fallthrough_catchproj != nullptr) {
2112 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2113 }
2114 if (callprojs->fallthrough_memproj != nullptr) {
2115 if (final_mem->is_MergeMem()) {
2116 // Parser's exits MergeMem was not transformed but may be optimized
2117 final_mem = _gvn.transform(final_mem);
2118 }
2119 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2120 add_mergemem_users_to_worklist(wl, final_mem);
2121 }
2122 if (callprojs->fallthrough_ioproj != nullptr) {
2123 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2124 }
2125
2126 // Replace the result with the new result if it exists and is used
2127 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2128 // If the inlined code is dead, the result projections for an inline type returned as
2129 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2130 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2131 "unexpected number of results");
2132 C->gvn_replace_by(callprojs->resproj[0], result);
2133 }
2134
2135 if (ejvms == nullptr) {
2136 // No exception edges to simply kill off those paths
2137 if (callprojs->catchall_catchproj != nullptr) {
2138 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2139 }
2140 if (callprojs->catchall_memproj != nullptr) {
2141 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2142 }
2143 if (callprojs->catchall_ioproj != nullptr) {
2144 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2145 }
2146 // Replace the old exception object with top
2147 if (callprojs->exobj != nullptr) {
2148 C->gvn_replace_by(callprojs->exobj, C->top());
2149 }
2150 } else {
2151 GraphKit ekit(ejvms);
2152
2153 // Load my combined exception state into the kit, with all phis transformed:
2154 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2155 replaced_nodes_exception = ex_map->replaced_nodes();
2156
2157 Node* ex_oop = ekit.use_exception_state(ex_map);
2158
2159 if (callprojs->catchall_catchproj != nullptr) {
2160 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2161 ex_ctl = ekit.control();
2162 }
2163 if (callprojs->catchall_memproj != nullptr) {
2164 Node* ex_mem = ekit.reset_memory();
2165 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2166 add_mergemem_users_to_worklist(wl, ex_mem);
2167 }
2168 if (callprojs->catchall_ioproj != nullptr) {
2169 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2170 }
2171
2172 // Replace the old exception object with the newly created one
2173 if (callprojs->exobj != nullptr) {
2174 C->gvn_replace_by(callprojs->exobj, ex_oop);
2175 }
2176 }
2177
2178 // Disconnect the call from the graph
2179 call->disconnect_inputs(C);
2180 C->gvn_replace_by(call, C->top());
2181
2182 // Clean up any MergeMems that feed other MergeMems since the
2183 // optimizer doesn't like that.
2184 while (wl.size() > 0) {
2185 _gvn.transform(wl.pop());
2186 }
2187
2188 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2189 replaced_nodes.apply(C, final_ctl);
2190 }
2191 if (!ex_ctl->is_top() && do_replaced_nodes) {
2192 replaced_nodes_exception.apply(C, ex_ctl);
2193 }
2194 }
2195
2196
2197 //------------------------------increment_counter------------------------------
2198 // for statistics: increment a VM counter by 1
2199
2200 void GraphKit::increment_counter(address counter_addr) {
2201 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2202 increment_counter(adr1);
2203 }
2204
2205 void GraphKit::increment_counter(Node* counter_addr) {
2206 int adr_type = Compile::AliasIdxRaw;
2207 Node* ctrl = control();
2208 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2209 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2210 store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered);
2211 }
2212
2213
2214 //------------------------------uncommon_trap----------------------------------
2215 // Bail out to the interpreter in mid-method. Implemented by calling the
2216 // uncommon_trap blob. This helper function inserts a runtime call with the
2217 // right debug info.
2218 Node* GraphKit::uncommon_trap(int trap_request,
2219 ciKlass* klass, const char* comment,
2220 bool must_throw,
2221 bool keep_exact_action) {
2222 if (failing()) stop();
2223 if (stopped()) return nullptr; // trap reachable?
2224
2225 // Note: If ProfileTraps is true, and if a deopt. actually
2226 // occurs here, the runtime will make sure an MDO exists. There is
2227 // no need to call method()->ensure_method_data() at this point.
2228
2229 // Set the stack pointer to the right value for reexecution:
2230 set_sp(reexecute_sp());
2231
2232 #ifdef ASSERT
2233 if (!must_throw) {
2234 // Make sure the stack has at least enough depth to execute
2235 // the current bytecode.
2236 int inputs, ignored_depth;
2237 if (compute_stack_effects(inputs, ignored_depth)) {
2238 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2239 Bytecodes::name(java_bc()), sp(), inputs);
2240 }
2241 }
2242 #endif
2243
2244 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2245 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2246
2247 switch (action) {
2248 case Deoptimization::Action_maybe_recompile:
2249 case Deoptimization::Action_reinterpret:
2250 // Temporary fix for 6529811 to allow virtual calls to be sure they
2251 // get the chance to go from mono->bi->mega
2252 if (!keep_exact_action &&
2253 Deoptimization::trap_request_index(trap_request) < 0 &&
2254 too_many_recompiles(reason)) {
2255 // This BCI is causing too many recompilations.
2256 if (C->log() != nullptr) {
2257 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2258 Deoptimization::trap_reason_name(reason),
2259 Deoptimization::trap_action_name(action));
2260 }
2261 action = Deoptimization::Action_none;
2262 trap_request = Deoptimization::make_trap_request(reason, action);
2263 } else {
2264 C->set_trap_can_recompile(true);
2265 }
2266 break;
2267 case Deoptimization::Action_make_not_entrant:
2268 C->set_trap_can_recompile(true);
2269 break;
2270 case Deoptimization::Action_none:
2271 case Deoptimization::Action_make_not_compilable:
2272 break;
2273 default:
2274 #ifdef ASSERT
2275 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2276 #endif
2277 break;
2278 }
2279
2280 if (TraceOptoParse) {
2281 char buf[100];
2282 tty->print_cr("Uncommon trap %s at bci:%d",
2283 Deoptimization::format_trap_request(buf, sizeof(buf),
2284 trap_request), bci());
2285 }
2286
2287 CompileLog* log = C->log();
2288 if (log != nullptr) {
2289 int kid = (klass == nullptr)? -1: log->identify(klass);
2290 log->begin_elem("uncommon_trap bci='%d'", bci());
2291 char buf[100];
2292 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2293 trap_request));
2294 if (kid >= 0) log->print(" klass='%d'", kid);
2295 if (comment != nullptr) log->print(" comment='%s'", comment);
2296 log->end_elem();
2297 }
2298
2299 // Make sure any guarding test views this path as very unlikely
2300 Node *i0 = control()->in(0);
2301 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2302 IfNode *iff = i0->as_If();
2303 float f = iff->_prob; // Get prob
2304 if (control()->Opcode() == Op_IfTrue) {
2305 if (f > PROB_UNLIKELY_MAG(4))
2306 iff->_prob = PROB_MIN;
2307 } else {
2308 if (f < PROB_LIKELY_MAG(4))
2309 iff->_prob = PROB_MAX;
2310 }
2311 }
2312
2313 // Clear out dead values from the debug info.
2314 kill_dead_locals();
2315
2316 // Now insert the uncommon trap subroutine call
2317 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
2318 const TypePtr* no_memory_effects = nullptr;
2319 // Pass the index of the class to be loaded
2320 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2321 (must_throw ? RC_MUST_THROW : 0),
2322 OptoRuntime::uncommon_trap_Type(),
2323 call_addr, "uncommon_trap", no_memory_effects,
2324 intcon(trap_request));
2325 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2326 "must extract request correctly from the graph");
2327 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2328
2329 call->set_req(TypeFunc::ReturnAdr, returnadr());
2330 // The debug info is the only real input to this call.
2331
2332 // Halt-and-catch fire here. The above call should never return!
2333 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2334 PRODUCT_ONLY(COMMA /*reachable*/false));
2335 _gvn.set_type_bottom(halt);
2336 root()->add_req(halt);
2337
2338 stop_and_kill_map();
2339 return call;
2340 }
2341
2342
2343 //--------------------------just_allocated_object------------------------------
2344 // Report the object that was just allocated.
2345 // It must be the case that there are no intervening safepoints.
2346 // We use this to determine if an object is so "fresh" that
2347 // it does not require card marks.
2348 Node* GraphKit::just_allocated_object(Node* current_control) {
2349 Node* ctrl = current_control;
2350 // Object::<init> is invoked after allocation, most of invoke nodes
2351 // will be reduced, but a region node is kept in parse time, we check
2352 // the pattern and skip the region node if it degraded to a copy.
2353 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2354 ctrl->as_Region()->is_copy()) {
2355 ctrl = ctrl->as_Region()->is_copy();
2356 }
2357 if (C->recent_alloc_ctl() == ctrl) {
2358 return C->recent_alloc_obj();
2359 }
2360 return nullptr;
2361 }
2362
2363
2364 /**
2365 * Record profiling data exact_kls for Node n with the type system so
2366 * that it can propagate it (speculation)
2367 *
2368 * @param n node that the type applies to
2369 * @param exact_kls type from profiling
2370 * @param maybe_null did profiling see null?
2371 *
2372 * @return node with improved type
2373 */
2374 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2375 const Type* current_type = _gvn.type(n);
2376 assert(UseTypeSpeculation, "type speculation must be on");
2377
2378 const TypePtr* speculative = current_type->speculative();
2379
2380 // Should the klass from the profile be recorded in the speculative type?
2381 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2382 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2383 const TypeOopPtr* xtype = tklass->as_instance_type();
2384 assert(xtype->klass_is_exact(), "Should be exact");
2385 // Any reason to believe n is not null (from this profiling or a previous one)?
2386 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2387 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2388 // record the new speculative type's depth
2389 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2390 speculative = speculative->with_inline_depth(jvms()->depth());
2391 } else if (current_type->would_improve_ptr(ptr_kind)) {
2392 // Profiling report that null was never seen so we can change the
2393 // speculative type to non null ptr.
2394 if (ptr_kind == ProfileAlwaysNull) {
2395 speculative = TypePtr::NULL_PTR;
2396 } else {
2397 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2398 const TypePtr* ptr = TypePtr::NOTNULL;
2399 if (speculative != nullptr) {
2400 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2401 } else {
2402 speculative = ptr;
2403 }
2404 }
2405 }
2406
2407 if (speculative != current_type->speculative()) {
2408 // Build a type with a speculative type (what we think we know
2409 // about the type but will need a guard when we use it)
2410 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2411 // We're changing the type, we need a new CheckCast node to carry
2412 // the new type. The new type depends on the control: what
2413 // profiling tells us is only valid from here as far as we can
2414 // tell.
2415 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2416 cast = _gvn.transform(cast);
2417 replace_in_map(n, cast);
2418 n = cast;
2419 }
2420
2421 return n;
2422 }
2423
2424 /**
2425 * Record profiling data from receiver profiling at an invoke with the
2426 * type system so that it can propagate it (speculation)
2427 *
2428 * @param n receiver node
2429 *
2430 * @return node with improved type
2431 */
2432 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2433 if (!UseTypeSpeculation) {
2434 return n;
2435 }
2436 ciKlass* exact_kls = profile_has_unique_klass();
2437 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2438 if ((java_bc() == Bytecodes::_checkcast ||
2439 java_bc() == Bytecodes::_instanceof ||
2440 java_bc() == Bytecodes::_aastore) &&
2441 method()->method_data()->is_mature()) {
2442 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2443 if (data != nullptr) {
2444 if (java_bc() == Bytecodes::_aastore) {
2445 ciKlass* array_type = nullptr;
2446 ciKlass* element_type = nullptr;
2447 ProfilePtrKind element_ptr = ProfileMaybeNull;
2448 bool flat_array = true;
2449 bool null_free_array = true;
2450 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2451 exact_kls = element_type;
2452 ptr_kind = element_ptr;
2453 } else {
2454 if (!data->as_BitData()->null_seen()) {
2455 ptr_kind = ProfileNeverNull;
2456 } else {
2457 assert(data->is_ReceiverTypeData(), "bad profile data type");
2458 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2459 uint i = 0;
2460 for (; i < call->row_limit(); i++) {
2461 ciKlass* receiver = call->receiver(i);
2462 if (receiver != nullptr) {
2463 break;
2464 }
2465 }
2466 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2467 }
2468 }
2469 }
2470 }
2471 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2472 }
2473
2474 /**
2475 * Record profiling data from argument profiling at an invoke with the
2476 * type system so that it can propagate it (speculation)
2477 *
2478 * @param dest_method target method for the call
2479 * @param bc what invoke bytecode is this?
2480 */
2481 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2482 if (!UseTypeSpeculation) {
2483 return;
2484 }
2485 const TypeFunc* tf = TypeFunc::make(dest_method);
2486 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2487 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2488 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2489 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2490 if (is_reference_type(targ->basic_type())) {
2491 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2492 ciKlass* better_type = nullptr;
2493 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2494 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2495 }
2496 i++;
2497 }
2498 }
2499 }
2500
2501 /**
2502 * Record profiling data from parameter profiling at an invoke with
2503 * the type system so that it can propagate it (speculation)
2504 */
2505 void GraphKit::record_profiled_parameters_for_speculation() {
2506 if (!UseTypeSpeculation) {
2507 return;
2508 }
2509 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2510 if (_gvn.type(local(i))->isa_oopptr()) {
2511 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2512 ciKlass* better_type = nullptr;
2513 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2514 record_profile_for_speculation(local(i), better_type, ptr_kind);
2515 }
2516 j++;
2517 }
2518 }
2519 }
2520
2521 /**
2522 * Record profiling data from return value profiling at an invoke with
2523 * the type system so that it can propagate it (speculation)
2524 */
2525 void GraphKit::record_profiled_return_for_speculation() {
2526 if (!UseTypeSpeculation) {
2527 return;
2528 }
2529 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2530 ciKlass* better_type = nullptr;
2531 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2532 // If profiling reports a single type for the return value,
2533 // feed it to the type system so it can propagate it as a
2534 // speculative type
2535 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2536 }
2537 }
2538
2539 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2540 if (Matcher::strict_fp_requires_explicit_rounding) {
2541 // (Note: TypeFunc::make has a cache that makes this fast.)
2542 const TypeFunc* tf = TypeFunc::make(dest_method);
2543 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2544 for (int j = 0; j < nargs; j++) {
2545 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2546 if (targ->basic_type() == T_DOUBLE) {
2547 // If any parameters are doubles, they must be rounded before
2548 // the call, dprecision_rounding does gvn.transform
2549 Node *arg = argument(j);
2550 arg = dprecision_rounding(arg);
2551 set_argument(j, arg);
2552 }
2553 }
2554 }
2555 }
2556
2557 // rounding for strict float precision conformance
2558 Node* GraphKit::precision_rounding(Node* n) {
2559 if (Matcher::strict_fp_requires_explicit_rounding) {
2560 #ifdef IA32
2561 if (UseSSE == 0) {
2562 return _gvn.transform(new RoundFloatNode(nullptr, n));
2563 }
2564 #else
2565 Unimplemented();
2566 #endif // IA32
2567 }
2568 return n;
2569 }
2570
2571 // rounding for strict double precision conformance
2572 Node* GraphKit::dprecision_rounding(Node *n) {
2573 if (Matcher::strict_fp_requires_explicit_rounding) {
2574 #ifdef IA32
2575 if (UseSSE < 2) {
2576 return _gvn.transform(new RoundDoubleNode(nullptr, n));
2577 }
2578 #else
2579 Unimplemented();
2580 #endif // IA32
2581 }
2582 return n;
2583 }
2584
2585 //=============================================================================
2586 // Generate a fast path/slow path idiom. Graph looks like:
2587 // [foo] indicates that 'foo' is a parameter
2588 //
2589 // [in] null
2590 // \ /
2591 // CmpP
2592 // Bool ne
2593 // If
2594 // / \
2595 // True False-<2>
2596 // / |
2597 // / cast_not_null
2598 // Load | | ^
2599 // [fast_test] | |
2600 // gvn to opt_test | |
2601 // / \ | <1>
2602 // True False |
2603 // | \\ |
2604 // [slow_call] \[fast_result]
2605 // Ctl Val \ \
2606 // | \ \
2607 // Catch <1> \ \
2608 // / \ ^ \ \
2609 // Ex No_Ex | \ \
2610 // | \ \ | \ <2> \
2611 // ... \ [slow_res] | | \ [null_result]
2612 // \ \--+--+--- | |
2613 // \ | / \ | /
2614 // --------Region Phi
2615 //
2616 //=============================================================================
2617 // Code is structured as a series of driver functions all called 'do_XXX' that
2618 // call a set of helper functions. Helper functions first, then drivers.
2619
2620 //------------------------------null_check_oop---------------------------------
2621 // Null check oop. Set null-path control into Region in slot 3.
2622 // Make a cast-not-nullness use the other not-null control. Return cast.
2623 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2624 bool never_see_null,
2625 bool safe_for_replace,
2626 bool speculative) {
2627 // Initial null check taken path
2628 (*null_control) = top();
2629 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2630
2631 // Generate uncommon_trap:
2632 if (never_see_null && (*null_control) != top()) {
2633 // If we see an unexpected null at a check-cast we record it and force a
2634 // recompile; the offending check-cast will be compiled to handle nulls.
2635 // If we see more than one offending BCI, then all checkcasts in the
2636 // method will be compiled to handle nulls.
2637 PreserveJVMState pjvms(this);
2638 set_control(*null_control);
2639 replace_in_map(value, null());
2640 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2641 uncommon_trap(reason,
2642 Deoptimization::Action_make_not_entrant);
2643 (*null_control) = top(); // null path is dead
2644 }
2645 if ((*null_control) == top() && safe_for_replace) {
2646 replace_in_map(value, cast);
2647 }
2648
2649 // Cast away null-ness on the result
2650 return cast;
2651 }
2652
2653 //------------------------------opt_iff----------------------------------------
2654 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2655 // Return slow-path control.
2656 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2657 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2658
2659 // Fast path taken; set region slot 2
2660 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2661 region->init_req(2,fast_taken); // Capture fast-control
2662
2663 // Fast path not-taken, i.e. slow path
2664 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2665 return slow_taken;
2666 }
2667
2668 //-----------------------------make_runtime_call-------------------------------
2669 Node* GraphKit::make_runtime_call(int flags,
2670 const TypeFunc* call_type, address call_addr,
2671 const char* call_name,
2672 const TypePtr* adr_type,
2673 // The following parms are all optional.
2674 // The first null ends the list.
2675 Node* parm0, Node* parm1,
2676 Node* parm2, Node* parm3,
2677 Node* parm4, Node* parm5,
2678 Node* parm6, Node* parm7) {
2679 assert(call_addr != nullptr, "must not call null targets");
2680
2681 // Slow-path call
2682 bool is_leaf = !(flags & RC_NO_LEAF);
2683 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2684 if (call_name == nullptr) {
2685 assert(!is_leaf, "must supply name for leaf");
2686 call_name = OptoRuntime::stub_name(call_addr);
2687 }
2688 CallNode* call;
2689 if (!is_leaf) {
2690 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2691 } else if (flags & RC_NO_FP) {
2692 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2693 } else if (flags & RC_VECTOR){
2694 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2695 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2696 } else {
2697 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2698 }
2699
2700 // The following is similar to set_edges_for_java_call,
2701 // except that the memory effects of the call are restricted to AliasIdxRaw.
2702
2703 // Slow path call has no side-effects, uses few values
2704 bool wide_in = !(flags & RC_NARROW_MEM);
2705 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2706
2707 Node* prev_mem = nullptr;
2708 if (wide_in) {
2709 prev_mem = set_predefined_input_for_runtime_call(call);
2710 } else {
2711 assert(!wide_out, "narrow in => narrow out");
2712 Node* narrow_mem = memory(adr_type);
2713 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2714 }
2715
2716 // Hook each parm in order. Stop looking at the first null.
2717 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2718 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2719 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2720 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2721 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2722 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2723 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2724 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2725 /* close each nested if ===> */ } } } } } } } }
2726 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2727
2728 if (!is_leaf) {
2729 // Non-leaves can block and take safepoints:
2730 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2731 }
2732 // Non-leaves can throw exceptions:
2733 if (has_io) {
2734 call->set_req(TypeFunc::I_O, i_o());
2735 }
2736
2737 if (flags & RC_UNCOMMON) {
2738 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2739 // (An "if" probability corresponds roughly to an unconditional count.
2740 // Sort of.)
2741 call->set_cnt(PROB_UNLIKELY_MAG(4));
2742 }
2743
2744 Node* c = _gvn.transform(call);
2745 assert(c == call, "cannot disappear");
2746
2747 if (wide_out) {
2748 // Slow path call has full side-effects.
2749 set_predefined_output_for_runtime_call(call);
2750 } else {
2751 // Slow path call has few side-effects, and/or sets few values.
2752 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2753 }
2754
2755 if (has_io) {
2756 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2757 }
2758 return call;
2759
2760 }
2761
2762 // i2b
2763 Node* GraphKit::sign_extend_byte(Node* in) {
2764 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2765 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2766 }
2767
2768 // i2s
2769 Node* GraphKit::sign_extend_short(Node* in) {
2770 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2771 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2772 }
2773
2774
2775 //------------------------------merge_memory-----------------------------------
2776 // Merge memory from one path into the current memory state.
2777 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2778 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2779 Node* old_slice = mms.force_memory();
2780 Node* new_slice = mms.memory2();
2781 if (old_slice != new_slice) {
2782 PhiNode* phi;
2783 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2784 if (mms.is_empty()) {
2785 // clone base memory Phi's inputs for this memory slice
2786 assert(old_slice == mms.base_memory(), "sanity");
2787 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2788 _gvn.set_type(phi, Type::MEMORY);
2789 for (uint i = 1; i < phi->req(); i++) {
2790 phi->init_req(i, old_slice->in(i));
2791 }
2792 } else {
2793 phi = old_slice->as_Phi(); // Phi was generated already
2794 }
2795 } else {
2796 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2797 _gvn.set_type(phi, Type::MEMORY);
2798 }
2799 phi->set_req(new_path, new_slice);
2800 mms.set_memory(phi);
2801 }
2802 }
2803 }
2804
2805 //------------------------------make_slow_call_ex------------------------------
2806 // Make the exception handler hookups for the slow call
2807 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2808 if (stopped()) return;
2809
2810 // Make a catch node with just two handlers: fall-through and catch-all
2811 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2812 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2813 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2814 _gvn.set_type_bottom(norm);
2815 C->record_for_igvn(norm);
2816 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2817
2818 { PreserveJVMState pjvms(this);
2819 set_control(excp);
2820 set_i_o(i_o);
2821
2822 if (excp != top()) {
2823 if (deoptimize) {
2824 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2825 uncommon_trap(Deoptimization::Reason_unhandled,
2826 Deoptimization::Action_none);
2827 } else {
2828 // Create an exception state also.
2829 // Use an exact type if the caller has a specific exception.
2830 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2831 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2832 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2833 }
2834 }
2835 }
2836
2837 // Get the no-exception control from the CatchNode.
2838 set_control(norm);
2839 }
2840
2841 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2842 Node* cmp = nullptr;
2843 switch(bt) {
2844 case T_INT: cmp = new CmpINode(in1, in2); break;
2845 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2846 default: fatal("unexpected comparison type %s", type2name(bt));
2847 }
2848 cmp = gvn.transform(cmp);
2849 Node* bol = gvn.transform(new BoolNode(cmp, test));
2850 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2851 gvn.transform(iff);
2852 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2853 return iff;
2854 }
2855
2856 //-------------------------------gen_subtype_check-----------------------------
2857 // Generate a subtyping check. Takes as input the subtype and supertype.
2858 // Returns 2 values: sets the default control() to the true path and returns
2859 // the false path. Only reads invariant memory; sets no (visible) memory.
2860 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2861 // but that's not exposed to the optimizer. This call also doesn't take in an
2862 // Object; if you wish to check an Object you need to load the Object's class
2863 // prior to coming here.
2864 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2865 ciMethod* method, int bci) {
2866 Compile* C = gvn.C;
2867 if ((*ctrl)->is_top()) {
2868 return C->top();
2869 }
2870
2871 // Fast check for identical types, perhaps identical constants.
2872 // The types can even be identical non-constants, in cases
2873 // involving Array.newInstance, Object.clone, etc.
2874 if (subklass == superklass)
2875 return C->top(); // false path is dead; no test needed.
2876
2877 if (gvn.type(superklass)->singleton()) {
2878 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2879 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2880
2881 // In the common case of an exact superklass, try to fold up the
2882 // test before generating code. You may ask, why not just generate
2883 // the code and then let it fold up? The answer is that the generated
2884 // code will necessarily include null checks, which do not always
2885 // completely fold away. If they are also needless, then they turn
2886 // into a performance loss. Example:
2887 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2888 // Here, the type of 'fa' is often exact, so the store check
2889 // of fa[1]=x will fold up, without testing the nullness of x.
2890 //
2891 // At macro expansion, we would have already folded the SubTypeCheckNode
2892 // being expanded here because we always perform the static sub type
2893 // check in SubTypeCheckNode::sub() regardless of whether
2894 // StressReflectiveCode is set or not. We can therefore skip this
2895 // static check when StressReflectiveCode is on.
2896 switch (C->static_subtype_check(superk, subk)) {
2897 case Compile::SSC_always_false:
2898 {
2899 Node* always_fail = *ctrl;
2900 *ctrl = gvn.C->top();
2901 return always_fail;
2902 }
2903 case Compile::SSC_always_true:
2904 return C->top();
2905 case Compile::SSC_easy_test:
2906 {
2907 // Just do a direct pointer compare and be done.
2908 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2909 *ctrl = gvn.transform(new IfTrueNode(iff));
2910 return gvn.transform(new IfFalseNode(iff));
2911 }
2912 case Compile::SSC_full_test:
2913 break;
2914 default:
2915 ShouldNotReachHere();
2916 }
2917 }
2918
2919 // %%% Possible further optimization: Even if the superklass is not exact,
2920 // if the subklass is the unique subtype of the superklass, the check
2921 // will always succeed. We could leave a dependency behind to ensure this.
2922
2923 // First load the super-klass's check-offset
2924 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2925 Node* m = C->immutable_memory();
2926 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2927 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2928 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2929 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
2930 bool might_be_cache = (chk_off_con == cacheoff_con);
2931
2932 // Load from the sub-klass's super-class display list, or a 1-word cache of
2933 // the secondary superclass list, or a failing value with a sentinel offset
2934 // if the super-klass is an interface or exceptionally deep in the Java
2935 // hierarchy and we have to scan the secondary superclass list the hard way.
2936 // Worst-case type is a little odd: null is allowed as a result (usually
2937 // klass loads can never produce a null).
2938 Node *chk_off_X = chk_off;
2939 #ifdef _LP64
2940 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2941 #endif
2942 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2943 // For some types like interfaces the following loadKlass is from a 1-word
2944 // cache which is mutable so can't use immutable memory. Other
2945 // types load from the super-class display table which is immutable.
2946 Node *kmem = C->immutable_memory();
2947 // secondary_super_cache is not immutable but can be treated as such because:
2948 // - no ideal node writes to it in a way that could cause an
2949 // incorrect/missed optimization of the following Load.
2950 // - it's a cache so, worse case, not reading the latest value
2951 // wouldn't cause incorrect execution
2952 if (might_be_cache && mem != nullptr) {
2953 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2954 }
2955 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, nullptr, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
2956
2957 // Compile speed common case: ARE a subtype and we canNOT fail
2958 if (superklass == nkls) {
2959 return C->top(); // false path is dead; no test needed.
2960 }
2961
2962 // Gather the various success & failures here
2963 RegionNode* r_not_subtype = new RegionNode(3);
2964 gvn.record_for_igvn(r_not_subtype);
2965 RegionNode* r_ok_subtype = new RegionNode(4);
2966 gvn.record_for_igvn(r_ok_subtype);
2967
2968 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2969 // SubTypeCheck node
2970 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2971 ciCallProfile profile = method->call_profile_at_bci(bci);
2972 float total_prob = 0;
2973 for (int i = 0; profile.has_receiver(i); ++i) {
2974 float prob = profile.receiver_prob(i);
2975 total_prob += prob;
2976 }
2977 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2978 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2979 for (int i = 0; profile.has_receiver(i); ++i) {
2980 ciKlass* klass = profile.receiver(i);
2981 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2982 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2983 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2984 continue;
2985 }
2986 float prob = profile.receiver_prob(i);
2987 ConNode* klass_node = gvn.makecon(klass_t);
2988 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2989 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2990
2991 if (result == Compile::SSC_always_true) {
2992 r_ok_subtype->add_req(iftrue);
2993 } else {
2994 assert(result == Compile::SSC_always_false, "");
2995 r_not_subtype->add_req(iftrue);
2996 }
2997 *ctrl = gvn.transform(new IfFalseNode(iff));
2998 }
2999 }
3000 }
3001
3002 // See if we get an immediate positive hit. Happens roughly 83% of the
3003 // time. Test to see if the value loaded just previously from the subklass
3004 // is exactly the superklass.
3005 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3006 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
3007 *ctrl = gvn.transform(new IfFalseNode(iff1));
3008
3009 // Compile speed common case: Check for being deterministic right now. If
3010 // chk_off is a constant and not equal to cacheoff then we are NOT a
3011 // subklass. In this case we need exactly the 1 test above and we can
3012 // return those results immediately.
3013 if (!might_be_cache) {
3014 Node* not_subtype_ctrl = *ctrl;
3015 *ctrl = iftrue1; // We need exactly the 1 test above
3016 PhaseIterGVN* igvn = gvn.is_IterGVN();
3017 if (igvn != nullptr) {
3018 igvn->remove_globally_dead_node(r_ok_subtype);
3019 igvn->remove_globally_dead_node(r_not_subtype);
3020 }
3021 return not_subtype_ctrl;
3022 }
3023
3024 r_ok_subtype->init_req(1, iftrue1);
3025
3026 // Check for immediate negative hit. Happens roughly 11% of the time (which
3027 // is roughly 63% of the remaining cases). Test to see if the loaded
3028 // check-offset points into the subklass display list or the 1-element
3029 // cache. If it points to the display (and NOT the cache) and the display
3030 // missed then it's not a subtype.
3031 Node *cacheoff = gvn.intcon(cacheoff_con);
3032 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3033 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3034 *ctrl = gvn.transform(new IfFalseNode(iff2));
3035
3036 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3037 // No performance impact (too rare) but allows sharing of secondary arrays
3038 // which has some footprint reduction.
3039 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3040 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3041 *ctrl = gvn.transform(new IfFalseNode(iff3));
3042
3043 // -- Roads not taken here: --
3044 // We could also have chosen to perform the self-check at the beginning
3045 // of this code sequence, as the assembler does. This would not pay off
3046 // the same way, since the optimizer, unlike the assembler, can perform
3047 // static type analysis to fold away many successful self-checks.
3048 // Non-foldable self checks work better here in second position, because
3049 // the initial primary superclass check subsumes a self-check for most
3050 // types. An exception would be a secondary type like array-of-interface,
3051 // which does not appear in its own primary supertype display.
3052 // Finally, we could have chosen to move the self-check into the
3053 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3054 // dependent manner. But it is worthwhile to have the check here,
3055 // where it can be perhaps be optimized. The cost in code space is
3056 // small (register compare, branch).
3057
3058 // Now do a linear scan of the secondary super-klass array. Again, no real
3059 // performance impact (too rare) but it's gotta be done.
3060 // Since the code is rarely used, there is no penalty for moving it
3061 // out of line, and it can only improve I-cache density.
3062 // The decision to inline or out-of-line this final check is platform
3063 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3064 Node* psc = gvn.transform(
3065 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3066
3067 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3068 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3069 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3070
3071 // Return false path; set default control to true path.
3072 *ctrl = gvn.transform(r_ok_subtype);
3073 return gvn.transform(r_not_subtype);
3074 }
3075
3076 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3077 const Type* sub_t = _gvn.type(obj_or_subklass);
3078 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3079 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3080 obj_or_subklass = makecon(sub_t);
3081 }
3082 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3083 if (expand_subtype_check) {
3084 MergeMemNode* mem = merged_memory();
3085 Node* ctrl = control();
3086 Node* subklass = obj_or_subklass;
3087 if (!sub_t->isa_klassptr()) {
3088 subklass = load_object_klass(obj_or_subklass);
3089 }
3090
3091 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3092 set_control(ctrl);
3093 return n;
3094 }
3095
3096 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3097 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3098 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3099 set_control(_gvn.transform(new IfTrueNode(iff)));
3100 return _gvn.transform(new IfFalseNode(iff));
3101 }
3102
3103 // Profile-driven exact type check:
3104 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3105 float prob, Node* *casted_receiver) {
3106 assert(!klass->is_interface(), "no exact type check on interfaces");
3107 Node* fail = top();
3108 const Type* rec_t = _gvn.type(receiver);
3109 if (rec_t->is_inlinetypeptr()) {
3110 if (klass->equals(rec_t->inline_klass())) {
3111 (*casted_receiver) = receiver; // Always passes
3112 } else {
3113 (*casted_receiver) = top(); // Always fails
3114 fail = control();
3115 set_control(top());
3116 }
3117 return fail;
3118 }
3119 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3120 Node* recv_klass = load_object_klass(receiver);
3121 fail = type_check(recv_klass, tklass, prob);
3122
3123 if (!stopped()) {
3124 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3125 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3126 assert(recv_xtype->klass_is_exact(), "");
3127
3128 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3129 // Subsume downstream occurrences of receiver with a cast to
3130 // recv_xtype, since now we know what the type will be.
3131 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3132 Node* res = _gvn.transform(cast);
3133 if (recv_xtype->is_inlinetypeptr()) {
3134 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3135 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3136 }
3137 (*casted_receiver) = res;
3138 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3139 // (User must make the replace_in_map call.)
3140 }
3141 }
3142
3143 return fail;
3144 }
3145
3146 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3147 float prob) {
3148 Node* want_klass = makecon(tklass);
3149 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3150 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3151 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3152 set_control(_gvn.transform(new IfTrueNode (iff)));
3153 Node* fail = _gvn.transform(new IfFalseNode(iff));
3154 return fail;
3155 }
3156
3157 //------------------------------subtype_check_receiver-------------------------
3158 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3159 Node** casted_receiver) {
3160 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3161 Node* want_klass = makecon(tklass);
3162
3163 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3164
3165 // Ignore interface type information until interface types are properly tracked.
3166 if (!stopped() && !klass->is_interface()) {
3167 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3168 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3169 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3170 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3171 if (recv_type->is_inlinetypeptr()) {
3172 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3173 }
3174 (*casted_receiver) = cast;
3175 }
3176 }
3177
3178 return slow_ctl;
3179 }
3180
3181 //------------------------------seems_never_null-------------------------------
3182 // Use null_seen information if it is available from the profile.
3183 // If we see an unexpected null at a type check we record it and force a
3184 // recompile; the offending check will be recompiled to handle nulls.
3185 // If we see several offending BCIs, then all checks in the
3186 // method will be recompiled.
3187 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3188 speculating = !_gvn.type(obj)->speculative_maybe_null();
3189 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3190 if (UncommonNullCast // Cutout for this technique
3191 && obj != null() // And not the -Xcomp stupid case?
3192 && !too_many_traps(reason)
3193 ) {
3194 if (speculating) {
3195 return true;
3196 }
3197 if (data == nullptr)
3198 // Edge case: no mature data. Be optimistic here.
3199 return true;
3200 // If the profile has not seen a null, assume it won't happen.
3201 assert(java_bc() == Bytecodes::_checkcast ||
3202 java_bc() == Bytecodes::_instanceof ||
3203 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3204 return !data->as_BitData()->null_seen();
3205 }
3206 speculating = false;
3207 return false;
3208 }
3209
3210 void GraphKit::guard_klass_being_initialized(Node* klass) {
3211 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3212 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3213 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3214 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3215 T_BYTE, MemNode::unordered);
3216 init_state = _gvn.transform(init_state);
3217
3218 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3219
3220 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3221 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3222
3223 { BuildCutout unless(this, tst, PROB_MAX);
3224 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3225 }
3226 }
3227
3228 void GraphKit::guard_init_thread(Node* klass) {
3229 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3230 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3231
3232 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3233 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3234 T_ADDRESS, MemNode::unordered);
3235 init_thread = _gvn.transform(init_thread);
3236
3237 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3238
3239 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3240 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3241
3242 { BuildCutout unless(this, tst, PROB_MAX);
3243 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3244 }
3245 }
3246
3247 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3248 if (ik->is_being_initialized()) {
3249 if (C->needs_clinit_barrier(ik, context)) {
3250 Node* klass = makecon(TypeKlassPtr::make(ik));
3251 guard_klass_being_initialized(klass);
3252 guard_init_thread(klass);
3253 insert_mem_bar(Op_MemBarCPUOrder);
3254 }
3255 } else if (ik->is_initialized()) {
3256 return; // no barrier needed
3257 } else {
3258 uncommon_trap(Deoptimization::Reason_uninitialized,
3259 Deoptimization::Action_reinterpret,
3260 nullptr);
3261 }
3262 }
3263
3264 //------------------------maybe_cast_profiled_receiver-------------------------
3265 // If the profile has seen exactly one type, narrow to exactly that type.
3266 // Subsequent type checks will always fold up.
3267 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3268 const TypeKlassPtr* require_klass,
3269 ciKlass* spec_klass,
3270 bool safe_for_replace) {
3271 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3272
3273 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3274
3275 // Make sure we haven't already deoptimized from this tactic.
3276 if (too_many_traps_or_recompiles(reason))
3277 return nullptr;
3278
3279 // (No, this isn't a call, but it's enough like a virtual call
3280 // to use the same ciMethod accessor to get the profile info...)
3281 // If we have a speculative type use it instead of profiling (which
3282 // may not help us)
3283 ciKlass* exact_kls = spec_klass;
3284 if (exact_kls == nullptr) {
3285 if (java_bc() == Bytecodes::_aastore) {
3286 ciKlass* array_type = nullptr;
3287 ciKlass* element_type = nullptr;
3288 ProfilePtrKind element_ptr = ProfileMaybeNull;
3289 bool flat_array = true;
3290 bool null_free_array = true;
3291 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3292 exact_kls = element_type;
3293 } else {
3294 exact_kls = profile_has_unique_klass();
3295 }
3296 }
3297 if (exact_kls != nullptr) {// no cast failures here
3298 if (require_klass == nullptr ||
3299 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3300 // If we narrow the type to match what the type profile sees or
3301 // the speculative type, we can then remove the rest of the
3302 // cast.
3303 // This is a win, even if the exact_kls is very specific,
3304 // because downstream operations, such as method calls,
3305 // will often benefit from the sharper type.
3306 Node* exact_obj = not_null_obj; // will get updated in place...
3307 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3308 &exact_obj);
3309 { PreserveJVMState pjvms(this);
3310 set_control(slow_ctl);
3311 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3312 }
3313 if (safe_for_replace) {
3314 replace_in_map(not_null_obj, exact_obj);
3315 }
3316 return exact_obj;
3317 }
3318 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3319 }
3320
3321 return nullptr;
3322 }
3323
3324 /**
3325 * Cast obj to type and emit guard unless we had too many traps here
3326 * already
3327 *
3328 * @param obj node being casted
3329 * @param type type to cast the node to
3330 * @param not_null true if we know node cannot be null
3331 */
3332 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3333 ciKlass* type,
3334 bool not_null) {
3335 if (stopped()) {
3336 return obj;
3337 }
3338
3339 // type is null if profiling tells us this object is always null
3340 if (type != nullptr) {
3341 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3342 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3343
3344 if (!too_many_traps_or_recompiles(null_reason) &&
3345 !too_many_traps_or_recompiles(class_reason)) {
3346 Node* not_null_obj = nullptr;
3347 // not_null is true if we know the object is not null and
3348 // there's no need for a null check
3349 if (!not_null) {
3350 Node* null_ctl = top();
3351 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3352 assert(null_ctl->is_top(), "no null control here");
3353 } else {
3354 not_null_obj = obj;
3355 }
3356
3357 Node* exact_obj = not_null_obj;
3358 ciKlass* exact_kls = type;
3359 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3360 &exact_obj);
3361 {
3362 PreserveJVMState pjvms(this);
3363 set_control(slow_ctl);
3364 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3365 }
3366 replace_in_map(not_null_obj, exact_obj);
3367 obj = exact_obj;
3368 }
3369 } else {
3370 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3371 Node* exact_obj = null_assert(obj);
3372 replace_in_map(obj, exact_obj);
3373 obj = exact_obj;
3374 }
3375 }
3376 return obj;
3377 }
3378
3379 //-------------------------------gen_instanceof--------------------------------
3380 // Generate an instance-of idiom. Used by both the instance-of bytecode
3381 // and the reflective instance-of call.
3382 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3383 kill_dead_locals(); // Benefit all the uncommon traps
3384 assert( !stopped(), "dead parse path should be checked in callers" );
3385 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3386 "must check for not-null not-dead klass in callers");
3387
3388 // Make the merge point
3389 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3390 RegionNode* region = new RegionNode(PATH_LIMIT);
3391 Node* phi = new PhiNode(region, TypeInt::BOOL);
3392 C->set_has_split_ifs(true); // Has chance for split-if optimization
3393
3394 ciProfileData* data = nullptr;
3395 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3396 data = method()->method_data()->bci_to_data(bci());
3397 }
3398 bool speculative_not_null = false;
3399 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3400 && seems_never_null(obj, data, speculative_not_null));
3401
3402 // Null check; get casted pointer; set region slot 3
3403 Node* null_ctl = top();
3404 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3405
3406 // If not_null_obj is dead, only null-path is taken
3407 if (stopped()) { // Doing instance-of on a null?
3408 set_control(null_ctl);
3409 return intcon(0);
3410 }
3411 region->init_req(_null_path, null_ctl);
3412 phi ->init_req(_null_path, intcon(0)); // Set null path value
3413 if (null_ctl == top()) {
3414 // Do this eagerly, so that pattern matches like is_diamond_phi
3415 // will work even during parsing.
3416 assert(_null_path == PATH_LIMIT-1, "delete last");
3417 region->del_req(_null_path);
3418 phi ->del_req(_null_path);
3419 }
3420
3421 // Do we know the type check always succeed?
3422 bool known_statically = false;
3423 if (_gvn.type(superklass)->singleton()) {
3424 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3425 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3426 if (subk != nullptr && subk->is_loaded()) {
3427 int static_res = C->static_subtype_check(superk, subk);
3428 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3429 }
3430 }
3431
3432 if (!known_statically) {
3433 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3434 // We may not have profiling here or it may not help us. If we
3435 // have a speculative type use it to perform an exact cast.
3436 ciKlass* spec_obj_type = obj_type->speculative_type();
3437 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3438 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3439 if (stopped()) { // Profile disagrees with this path.
3440 set_control(null_ctl); // Null is the only remaining possibility.
3441 return intcon(0);
3442 }
3443 if (cast_obj != nullptr) {
3444 not_null_obj = cast_obj;
3445 }
3446 }
3447 }
3448
3449 // Generate the subtype check
3450 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3451
3452 // Plug in the success path to the general merge in slot 1.
3453 region->init_req(_obj_path, control());
3454 phi ->init_req(_obj_path, intcon(1));
3455
3456 // Plug in the failing path to the general merge in slot 2.
3457 region->init_req(_fail_path, not_subtype_ctrl);
3458 phi ->init_req(_fail_path, intcon(0));
3459
3460 // Return final merged results
3461 set_control( _gvn.transform(region) );
3462 record_for_igvn(region);
3463
3464 // If we know the type check always succeeds then we don't use the
3465 // profiling data at this bytecode. Don't lose it, feed it to the
3466 // type system as a speculative type.
3467 if (safe_for_replace) {
3468 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3469 replace_in_map(obj, casted_obj);
3470 }
3471
3472 return _gvn.transform(phi);
3473 }
3474
3475 //-------------------------------gen_checkcast---------------------------------
3476 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3477 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3478 // uncommon-trap paths work. Adjust stack after this call.
3479 // If failure_control is supplied and not null, it is filled in with
3480 // the control edge for the cast failure. Otherwise, an appropriate
3481 // uncommon trap or exception is thrown.
3482 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) {
3483 kill_dead_locals(); // Benefit all the uncommon traps
3484 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3485 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3486 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3487 bool safe_for_replace = (failure_control == nullptr);
3488 assert(!null_free || toop->is_inlinetypeptr(), "must be an inline type pointer");
3489
3490 // Fast cutout: Check the case that the cast is vacuously true.
3491 // This detects the common cases where the test will short-circuit
3492 // away completely. We do this before we perform the null check,
3493 // because if the test is going to turn into zero code, we don't
3494 // want a residual null check left around. (Causes a slowdown,
3495 // for example, in some objArray manipulations, such as a[i]=a[j].)
3496 if (improved_klass_ptr_type->singleton()) {
3497 const TypeKlassPtr* kptr = nullptr;
3498 const Type* t = _gvn.type(obj);
3499 if (t->isa_oop_ptr()) {
3500 kptr = t->is_oopptr()->as_klass_type();
3501 } else if (obj->is_InlineType()) {
3502 ciInlineKlass* vk = t->inline_klass();
3503 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3504 }
3505 if (kptr != nullptr) {
3506 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3507 case Compile::SSC_always_true:
3508 // If we know the type check always succeed then we don't use
3509 // the profiling data at this bytecode. Don't lose it, feed it
3510 // to the type system as a speculative type.
3511 obj = record_profiled_receiver_for_speculation(obj);
3512 if (null_free) {
3513 assert(safe_for_replace, "must be");
3514 obj = null_check(obj);
3515 }
3516 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3517 return obj;
3518 case Compile::SSC_always_false:
3519 if (null_free) {
3520 assert(safe_for_replace, "must be");
3521 obj = null_check(obj);
3522 }
3523 // It needs a null check because a null will *pass* the cast check.
3524 if (t->isa_oopptr() != nullptr && !t->is_oopptr()->maybe_null()) {
3525 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3526 Deoptimization::DeoptReason reason = is_aastore ?
3527 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3528 builtin_throw(reason);
3529 return top();
3530 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3531 return null_assert(obj);
3532 }
3533 break; // Fall through to full check
3534 default:
3535 break;
3536 }
3537 }
3538 }
3539
3540 ciProfileData* data = nullptr;
3541 if (failure_control == nullptr) { // use MDO in regular case only
3542 assert(java_bc() == Bytecodes::_aastore ||
3543 java_bc() == Bytecodes::_checkcast,
3544 "interpreter profiles type checks only for these BCs");
3545 if (method()->method_data()->is_mature()) {
3546 data = method()->method_data()->bci_to_data(bci());
3547 }
3548 }
3549
3550 // Make the merge point
3551 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3552 RegionNode* region = new RegionNode(PATH_LIMIT);
3553 Node* phi = new PhiNode(region, toop);
3554 _gvn.set_type(region, Type::CONTROL);
3555 _gvn.set_type(phi, toop);
3556
3557 C->set_has_split_ifs(true); // Has chance for split-if optimization
3558
3559 // Use null-cast information if it is available
3560 bool speculative_not_null = false;
3561 bool never_see_null = ((failure_control == nullptr) // regular case only
3562 && seems_never_null(obj, data, speculative_not_null));
3563
3564 if (obj->is_InlineType()) {
3565 // Re-execute if buffering during triggers deoptimization
3566 PreserveReexecuteState preexecs(this);
3567 jvms()->set_should_reexecute(true);
3568 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3569 }
3570
3571 // Null check; get casted pointer; set region slot 3
3572 Node* null_ctl = top();
3573 Node* not_null_obj = nullptr;
3574 if (null_free) {
3575 assert(safe_for_replace, "must be");
3576 not_null_obj = null_check(obj);
3577 } else {
3578 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3579 }
3580
3581 // If not_null_obj is dead, only null-path is taken
3582 if (stopped()) { // Doing instance-of on a null?
3583 set_control(null_ctl);
3584 if (toop->is_inlinetypeptr()) {
3585 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3586 }
3587 return null();
3588 }
3589 region->init_req(_null_path, null_ctl);
3590 phi ->init_req(_null_path, null()); // Set null path value
3591 if (null_ctl == top()) {
3592 // Do this eagerly, so that pattern matches like is_diamond_phi
3593 // will work even during parsing.
3594 assert(_null_path == PATH_LIMIT-1, "delete last");
3595 region->del_req(_null_path);
3596 phi ->del_req(_null_path);
3597 }
3598
3599 Node* cast_obj = nullptr;
3600 if (improved_klass_ptr_type->klass_is_exact()) {
3601 // The following optimization tries to statically cast the speculative type of the object
3602 // (for example obtained during profiling) to the type of the superklass and then do a
3603 // dynamic check that the type of the object is what we expect. To work correctly
3604 // for checkcast and aastore the type of superklass should be exact.
3605 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3606 // We may not have profiling here or it may not help us. If we have
3607 // a speculative type use it to perform an exact cast.
3608 ciKlass* spec_obj_type = obj_type->speculative_type();
3609 if (spec_obj_type != nullptr || data != nullptr) {
3610 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3611 if (cast_obj != nullptr) {
3612 if (failure_control != nullptr) // failure is now impossible
3613 (*failure_control) = top();
3614 // adjust the type of the phi to the exact klass:
3615 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3616 }
3617 }
3618 }
3619
3620 if (cast_obj == nullptr) {
3621 // Generate the subtype check
3622 Node* improved_superklass = superklass;
3623 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3624 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3625 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3626 // Additionally, the benefit would only be minor in non-constant cases.
3627 improved_superklass = makecon(improved_klass_ptr_type);
3628 }
3629 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3630 // Plug in success path into the merge
3631 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3632 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3633 if (failure_control == nullptr) {
3634 if (not_subtype_ctrl != top()) { // If failure is possible
3635 PreserveJVMState pjvms(this);
3636 set_control(not_subtype_ctrl);
3637 Node* obj_klass = nullptr;
3638 if (not_null_obj->is_InlineType()) {
3639 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3640 } else {
3641 obj_klass = load_object_klass(not_null_obj);
3642 }
3643 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3644 Deoptimization::DeoptReason reason = is_aastore ?
3645 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3646 builtin_throw(reason);
3647 }
3648 } else {
3649 (*failure_control) = not_subtype_ctrl;
3650 }
3651 }
3652
3653 region->init_req(_obj_path, control());
3654 phi ->init_req(_obj_path, cast_obj);
3655
3656 // A merge of null or Casted-NotNull obj
3657 Node* res = _gvn.transform(phi);
3658
3659 // Note I do NOT always 'replace_in_map(obj,result)' here.
3660 // if( tk->klass()->can_be_primary_super() )
3661 // This means that if I successfully store an Object into an array-of-String
3662 // I 'forget' that the Object is really now known to be a String. I have to
3663 // do this because we don't have true union types for interfaces - if I store
3664 // a Baz into an array-of-Interface and then tell the optimizer it's an
3665 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3666 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3667 // replace_in_map( obj, res );
3668
3669 // Return final merged results
3670 set_control( _gvn.transform(region) );
3671 record_for_igvn(region);
3672
3673 bool not_inline = !toop->can_be_inline_type();
3674 bool not_flat_in_array = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flat_in_array());
3675 if (EnableValhalla && not_flat_in_array) {
3676 // Check if obj has been loaded from an array
3677 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3678 Node* array = nullptr;
3679 if (obj->isa_Load()) {
3680 Node* address = obj->in(MemNode::Address);
3681 if (address->isa_AddP()) {
3682 array = address->as_AddP()->in(AddPNode::Base);
3683 }
3684 } else if (obj->is_Phi()) {
3685 Node* region = obj->in(0);
3686 // TODO make this more robust (see JDK-8231346)
3687 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3688 IfNode* iff = region->in(2)->in(0)->isa_If();
3689 if (iff != nullptr) {
3690 iff->is_flat_array_check(&_gvn, &array);
3691 }
3692 }
3693 }
3694 if (array != nullptr) {
3695 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3696 if (ary_t != nullptr && !ary_t->is_flat()) {
3697 if (!ary_t->is_not_null_free() && not_inline) {
3698 // Casting array element to a non-inline-type, mark array as not null-free.
3699 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3700 replace_in_map(array, cast);
3701 } else if (!ary_t->is_not_flat()) {
3702 // Casting array element to a non-flat type, mark array as not flat.
3703 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3704 replace_in_map(array, cast);
3705 }
3706 }
3707 }
3708 }
3709
3710 if (!stopped() && !res->is_InlineType()) {
3711 res = record_profiled_receiver_for_speculation(res);
3712 if (toop->is_inlinetypeptr()) {
3713 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null());
3714 res = vt;
3715 if (safe_for_replace) {
3716 replace_in_map(obj, vt);
3717 replace_in_map(not_null_obj, vt);
3718 replace_in_map(res, vt);
3719 }
3720 }
3721 }
3722 return res;
3723 }
3724
3725 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3726 // Load markword
3727 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3728 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3729 if (check_lock) {
3730 // Check if obj is locked
3731 Node* locked_bit = MakeConX(markWord::unlocked_value);
3732 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3733 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3734 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3735 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3736 _gvn.transform(iff);
3737 Node* locked_region = new RegionNode(3);
3738 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3739
3740 // Unlocked: Use bits from mark word
3741 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3742 mark_phi->init_req(1, mark);
3743
3744 // Locked: Load prototype header from klass
3745 set_control(_gvn.transform(new IfFalseNode(iff)));
3746 // Make loads control dependent to make sure they are only executed if array is locked
3747 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3748 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, control(), C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3749 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
3750 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3751
3752 locked_region->init_req(2, control());
3753 mark_phi->init_req(2, proto);
3754 set_control(_gvn.transform(locked_region));
3755 record_for_igvn(locked_region);
3756
3757 mark = mark_phi;
3758 }
3759
3760 // Now check if mark word bits are set
3761 Node* mask = MakeConX(mask_val);
3762 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3763 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3764 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3765 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3766 }
3767
3768 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3769 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3770 }
3771
3772 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3773 // We can't use immutable memory here because the mark word is mutable.
3774 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3775 // check is moved out of loops (mainly to enable loop unswitching).
3776 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3777 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3778 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3779 }
3780
3781 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3782 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3783 }
3784
3785 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3786 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3787 RegionNode* region = new RegionNode(3);
3788 Node* null_ctl = top();
3789 null_check_oop(val, &null_ctl);
3790 if (null_ctl != top()) {
3791 PreserveJVMState pjvms(this);
3792 set_control(null_ctl);
3793 {
3794 // Deoptimize if null-free array
3795 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3796 inc_sp(nargs);
3797 uncommon_trap(Deoptimization::Reason_null_check,
3798 Deoptimization::Action_none);
3799 }
3800 region->init_req(1, control());
3801 }
3802 region->init_req(2, control());
3803 set_control(_gvn.transform(region));
3804 record_for_igvn(region);
3805 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3806 // Since we were just successfully storing null, the array can't be null free.
3807 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3808 ary_t = ary_t->cast_to_not_null_free();
3809 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3810 if (safe_for_replace) {
3811 replace_in_map(ary, cast);
3812 }
3813 ary = cast;
3814 }
3815 return ary;
3816 }
3817
3818 //------------------------------next_monitor-----------------------------------
3819 // What number should be given to the next monitor?
3820 int GraphKit::next_monitor() {
3821 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3822 int next = current + C->sync_stack_slots();
3823 // Keep the toplevel high water mark current:
3824 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3825 return current;
3826 }
3827
3828 //------------------------------insert_mem_bar---------------------------------
3829 // Memory barrier to avoid floating things around
3830 // The membar serves as a pinch point between both control and all memory slices.
3831 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3832 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3833 mb->init_req(TypeFunc::Control, control());
3834 mb->init_req(TypeFunc::Memory, reset_memory());
3835 Node* membar = _gvn.transform(mb);
3836 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3837 set_all_memory_call(membar);
3838 return membar;
3839 }
3840
3841 //-------------------------insert_mem_bar_volatile----------------------------
3842 // Memory barrier to avoid floating things around
3843 // The membar serves as a pinch point between both control and memory(alias_idx).
3844 // If you want to make a pinch point on all memory slices, do not use this
3845 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3846 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3847 // When Parse::do_put_xxx updates a volatile field, it appends a series
3848 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3849 // The first membar is on the same memory slice as the field store opcode.
3850 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3851 // All the other membars (for other volatile slices, including AliasIdxBot,
3852 // which stands for all unknown volatile slices) are control-dependent
3853 // on the first membar. This prevents later volatile loads or stores
3854 // from sliding up past the just-emitted store.
3855
3856 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3857 mb->set_req(TypeFunc::Control,control());
3858 if (alias_idx == Compile::AliasIdxBot) {
3859 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3860 } else {
3861 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3862 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3863 }
3864 Node* membar = _gvn.transform(mb);
3865 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3866 if (alias_idx == Compile::AliasIdxBot) {
3867 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3868 } else {
3869 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3870 }
3871 return membar;
3872 }
3873
3874 //------------------------------shared_lock------------------------------------
3875 // Emit locking code.
3876 FastLockNode* GraphKit::shared_lock(Node* obj) {
3877 // bci is either a monitorenter bc or InvocationEntryBci
3878 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3879 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3880
3881 if( !GenerateSynchronizationCode )
3882 return nullptr; // Not locking things?
3883
3884 if (stopped()) // Dead monitor?
3885 return nullptr;
3886
3887 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3888
3889 // Box the stack location
3890 Node* box = new BoxLockNode(next_monitor());
3891 // Check for bailout after new BoxLockNode
3892 if (failing()) { return nullptr; }
3893 box = _gvn.transform(box);
3894 Node* mem = reset_memory();
3895
3896 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3897
3898 // Add monitor to debug info for the slow path. If we block inside the
3899 // slow path and de-opt, we need the monitor hanging around
3900 map()->push_monitor( flock );
3901
3902 const TypeFunc *tf = LockNode::lock_type();
3903 LockNode *lock = new LockNode(C, tf);
3904
3905 lock->init_req( TypeFunc::Control, control() );
3906 lock->init_req( TypeFunc::Memory , mem );
3907 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3908 lock->init_req( TypeFunc::FramePtr, frameptr() );
3909 lock->init_req( TypeFunc::ReturnAdr, top() );
3910
3911 lock->init_req(TypeFunc::Parms + 0, obj);
3912 lock->init_req(TypeFunc::Parms + 1, box);
3913 lock->init_req(TypeFunc::Parms + 2, flock);
3914 add_safepoint_edges(lock);
3915
3916 lock = _gvn.transform( lock )->as_Lock();
3917
3918 // lock has no side-effects, sets few values
3919 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3920
3921 insert_mem_bar(Op_MemBarAcquireLock);
3922
3923 // Add this to the worklist so that the lock can be eliminated
3924 record_for_igvn(lock);
3925
3926 #ifndef PRODUCT
3927 if (PrintLockStatistics) {
3928 // Update the counter for this lock. Don't bother using an atomic
3929 // operation since we don't require absolute accuracy.
3930 lock->create_lock_counter(map()->jvms());
3931 increment_counter(lock->counter()->addr());
3932 }
3933 #endif
3934
3935 return flock;
3936 }
3937
3938
3939 //------------------------------shared_unlock----------------------------------
3940 // Emit unlocking code.
3941 void GraphKit::shared_unlock(Node* box, Node* obj) {
3942 // bci is either a monitorenter bc or InvocationEntryBci
3943 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3944 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3945
3946 if( !GenerateSynchronizationCode )
3947 return;
3948 if (stopped()) { // Dead monitor?
3949 map()->pop_monitor(); // Kill monitor from debug info
3950 return;
3951 }
3952 assert(!obj->is_InlineType(), "should not unlock on inline type");
3953
3954 // Memory barrier to avoid floating things down past the locked region
3955 insert_mem_bar(Op_MemBarReleaseLock);
3956
3957 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3958 UnlockNode *unlock = new UnlockNode(C, tf);
3959 #ifdef ASSERT
3960 unlock->set_dbg_jvms(sync_jvms());
3961 #endif
3962 uint raw_idx = Compile::AliasIdxRaw;
3963 unlock->init_req( TypeFunc::Control, control() );
3964 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3965 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3966 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3967 unlock->init_req( TypeFunc::ReturnAdr, top() );
3968
3969 unlock->init_req(TypeFunc::Parms + 0, obj);
3970 unlock->init_req(TypeFunc::Parms + 1, box);
3971 unlock = _gvn.transform(unlock)->as_Unlock();
3972
3973 Node* mem = reset_memory();
3974
3975 // unlock has no side-effects, sets few values
3976 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3977
3978 // Kill monitor from debug info
3979 map()->pop_monitor( );
3980 }
3981
3982 //-------------------------------get_layout_helper-----------------------------
3983 // If the given klass is a constant or known to be an array,
3984 // fetch the constant layout helper value into constant_value
3985 // and return null. Otherwise, load the non-constant
3986 // layout helper value, and return the node which represents it.
3987 // This two-faced routine is useful because allocation sites
3988 // almost always feature constant types.
3989 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3990 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3991 if (!StressReflectiveCode && klass_t != nullptr) {
3992 bool xklass = klass_t->klass_is_exact();
3993 bool can_be_flat = false;
3994 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
3995 if (UseFlatArray && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
3996 // Don't constant fold if the runtime type might be a flat array but the static type is not.
3997 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
3998 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->flat_in_array());
3999 }
4000 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4001 jint lhelper;
4002 if (klass_t->is_flat()) {
4003 lhelper = ary_type->flat_layout_helper();
4004 } else if (klass_t->isa_aryklassptr()) {
4005 BasicType elem = ary_type->elem()->array_element_basic_type();
4006 if (is_reference_type(elem, true)) {
4007 elem = T_OBJECT;
4008 }
4009 lhelper = Klass::array_layout_helper(elem);
4010 } else {
4011 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4012 }
4013 if (lhelper != Klass::_lh_neutral_value) {
4014 constant_value = lhelper;
4015 return (Node*) nullptr;
4016 }
4017 }
4018 }
4019 constant_value = Klass::_lh_neutral_value; // put in a known value
4020 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
4021 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4022 }
4023
4024 // We just put in an allocate/initialize with a big raw-memory effect.
4025 // Hook selected additional alias categories on the initialization.
4026 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4027 MergeMemNode* init_in_merge,
4028 Node* init_out_raw) {
4029 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4030 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4031
4032 Node* prevmem = kit.memory(alias_idx);
4033 init_in_merge->set_memory_at(alias_idx, prevmem);
4034 if (init_out_raw != nullptr) {
4035 kit.set_memory(init_out_raw, alias_idx);
4036 }
4037 }
4038
4039 //---------------------------set_output_for_allocation-------------------------
4040 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4041 const TypeOopPtr* oop_type,
4042 bool deoptimize_on_exception) {
4043 int rawidx = Compile::AliasIdxRaw;
4044 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4045 add_safepoint_edges(alloc);
4046 Node* allocx = _gvn.transform(alloc);
4047 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4048 // create memory projection for i_o
4049 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4050 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4051
4052 // create a memory projection as for the normal control path
4053 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4054 set_memory(malloc, rawidx);
4055
4056 // a normal slow-call doesn't change i_o, but an allocation does
4057 // we create a separate i_o projection for the normal control path
4058 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4059 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4060
4061 // put in an initialization barrier
4062 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4063 rawoop)->as_Initialize();
4064 assert(alloc->initialization() == init, "2-way macro link must work");
4065 assert(init ->allocation() == alloc, "2-way macro link must work");
4066 {
4067 // Extract memory strands which may participate in the new object's
4068 // initialization, and source them from the new InitializeNode.
4069 // This will allow us to observe initializations when they occur,
4070 // and link them properly (as a group) to the InitializeNode.
4071 assert(init->in(InitializeNode::Memory) == malloc, "");
4072 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4073 init->set_req(InitializeNode::Memory, minit_in);
4074 record_for_igvn(minit_in); // fold it up later, if possible
4075 _gvn.set_type(minit_in, Type::MEMORY);
4076 Node* minit_out = memory(rawidx);
4077 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4078 // Add an edge in the MergeMem for the header fields so an access
4079 // to one of those has correct memory state
4080 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
4081 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
4082 if (oop_type->isa_aryptr()) {
4083 const TypeAryPtr* arytype = oop_type->is_aryptr();
4084 if (arytype->is_flat()) {
4085 // Initially all flat array accesses share a single slice
4086 // but that changes after parsing. Prepare the memory graph so
4087 // it can optimize flat array accesses properly once they
4088 // don't share a single slice.
4089 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4090 C->set_flat_accesses_share_alias(false);
4091 ciInlineKlass* vk = arytype->elem()->inline_klass();
4092 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
4093 ciField* field = vk->nonstatic_field_at(i);
4094 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4095 continue; // do not bother to track really large numbers of fields
4096 int off_in_vt = field->offset_in_bytes() - vk->first_field_offset();
4097 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
4098 int fieldidx = C->get_alias_index(adr_type, true);
4099 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node
4100 // can result in per flat array field Phis to be created which confuses the logic of
4101 // Compile::adjust_flat_array_access_aliases().
4102 hook_memory_on_init(*this, fieldidx, minit_in, nullptr);
4103 }
4104 C->set_flat_accesses_share_alias(true);
4105 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
4106 } else {
4107 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4108 int elemidx = C->get_alias_index(telemref);
4109 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
4110 }
4111 } else if (oop_type->isa_instptr()) {
4112 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
4113 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4114 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4115 ciField* field = ik->nonstatic_field_at(i);
4116 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4117 continue; // do not bother to track really large numbers of fields
4118 // Find (or create) the alias category for this field:
4119 int fieldidx = C->alias_type(field)->index();
4120 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
4121 }
4122 }
4123 }
4124
4125 // Cast raw oop to the real thing...
4126 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4127 javaoop = _gvn.transform(javaoop);
4128 C->set_recent_alloc(control(), javaoop);
4129 assert(just_allocated_object(control()) == javaoop, "just allocated");
4130
4131 #ifdef ASSERT
4132 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4133 assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
4134 "Ideal_allocation works");
4135 assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
4136 "Ideal_allocation works");
4137 if (alloc->is_AllocateArray()) {
4138 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
4139 "Ideal_allocation works");
4140 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
4141 "Ideal_allocation works");
4142 } else {
4143 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4144 }
4145 }
4146 #endif //ASSERT
4147
4148 return javaoop;
4149 }
4150
4151 //---------------------------new_instance--------------------------------------
4152 // This routine takes a klass_node which may be constant (for a static type)
4153 // or may be non-constant (for reflective code). It will work equally well
4154 // for either, and the graph will fold nicely if the optimizer later reduces
4155 // the type to a constant.
4156 // The optional arguments are for specialized use by intrinsics:
4157 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4158 // - If 'return_size_val', report the total object size to the caller.
4159 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4160 Node* GraphKit::new_instance(Node* klass_node,
4161 Node* extra_slow_test,
4162 Node* *return_size_val,
4163 bool deoptimize_on_exception,
4164 InlineTypeNode* inline_type_node) {
4165 // Compute size in doublewords
4166 // The size is always an integral number of doublewords, represented
4167 // as a positive bytewise size stored in the klass's layout_helper.
4168 // The layout_helper also encodes (in a low bit) the need for a slow path.
4169 jint layout_con = Klass::_lh_neutral_value;
4170 Node* layout_val = get_layout_helper(klass_node, layout_con);
4171 bool layout_is_con = (layout_val == nullptr);
4172
4173 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4174 // Generate the initial go-slow test. It's either ALWAYS (return a
4175 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4176 // case) a computed value derived from the layout_helper.
4177 Node* initial_slow_test = nullptr;
4178 if (layout_is_con) {
4179 assert(!StressReflectiveCode, "stress mode does not use these paths");
4180 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4181 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4182 } else { // reflective case
4183 // This reflective path is used by Unsafe.allocateInstance.
4184 // (It may be stress-tested by specifying StressReflectiveCode.)
4185 // Basically, we want to get into the VM is there's an illegal argument.
4186 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4187 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4188 if (extra_slow_test != intcon(0)) {
4189 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4190 }
4191 // (Macro-expander will further convert this to a Bool, if necessary.)
4192 }
4193
4194 // Find the size in bytes. This is easy; it's the layout_helper.
4195 // The size value must be valid even if the slow path is taken.
4196 Node* size = nullptr;
4197 if (layout_is_con) {
4198 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4199 } else { // reflective case
4200 // This reflective path is used by clone and Unsafe.allocateInstance.
4201 size = ConvI2X(layout_val);
4202
4203 // Clear the low bits to extract layout_helper_size_in_bytes:
4204 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4205 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4206 size = _gvn.transform( new AndXNode(size, mask) );
4207 }
4208 if (return_size_val != nullptr) {
4209 (*return_size_val) = size;
4210 }
4211
4212 // This is a precise notnull oop of the klass.
4213 // (Actually, it need not be precise if this is a reflective allocation.)
4214 // It's what we cast the result to.
4215 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4216 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4217 const TypeOopPtr* oop_type = tklass->as_instance_type();
4218
4219 // Now generate allocation code
4220
4221 // The entire memory state is needed for slow path of the allocation
4222 // since GC and deoptimization can happen.
4223 Node *mem = reset_memory();
4224 set_all_memory(mem); // Create new memory state
4225
4226 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4227 control(), mem, i_o(),
4228 size, klass_node,
4229 initial_slow_test, inline_type_node);
4230
4231 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4232 }
4233
4234 //-------------------------------new_array-------------------------------------
4235 // helper for newarray and anewarray
4236 // The 'length' parameter is (obviously) the length of the array.
4237 // The optional arguments are for specialized use by intrinsics:
4238 // - If 'return_size_val', report the non-padded array size (sum of header size
4239 // and array body) to the caller.
4240 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4241 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4242 Node* length, // number of array elements
4243 int nargs, // number of arguments to push back for uncommon trap
4244 Node* *return_size_val,
4245 bool deoptimize_on_exception) {
4246 jint layout_con = Klass::_lh_neutral_value;
4247 Node* layout_val = get_layout_helper(klass_node, layout_con);
4248 bool layout_is_con = (layout_val == nullptr);
4249
4250 if (!layout_is_con && !StressReflectiveCode &&
4251 !too_many_traps(Deoptimization::Reason_class_check)) {
4252 // This is a reflective array creation site.
4253 // Optimistically assume that it is a subtype of Object[],
4254 // so that we can fold up all the address arithmetic.
4255 layout_con = Klass::array_layout_helper(T_OBJECT);
4256 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4257 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4258 { BuildCutout unless(this, bol_lh, PROB_MAX);
4259 inc_sp(nargs);
4260 uncommon_trap(Deoptimization::Reason_class_check,
4261 Deoptimization::Action_maybe_recompile);
4262 }
4263 layout_val = nullptr;
4264 layout_is_con = true;
4265 }
4266
4267 // Generate the initial go-slow test. Make sure we do not overflow
4268 // if length is huge (near 2Gig) or negative! We do not need
4269 // exact double-words here, just a close approximation of needed
4270 // double-words. We can't add any offset or rounding bits, lest we
4271 // take a size -1 of bytes and make it positive. Use an unsigned
4272 // compare, so negative sizes look hugely positive.
4273 int fast_size_limit = FastAllocateSizeLimit;
4274 if (layout_is_con) {
4275 assert(!StressReflectiveCode, "stress mode does not use these paths");
4276 // Increase the size limit if we have exact knowledge of array type.
4277 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4278 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4279 }
4280
4281 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4282 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4283
4284 // --- Size Computation ---
4285 // array_size = round_to_heap(array_header + (length << elem_shift));
4286 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4287 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4288 // The rounding mask is strength-reduced, if possible.
4289 int round_mask = MinObjAlignmentInBytes - 1;
4290 Node* header_size = nullptr;
4291 // (T_BYTE has the weakest alignment and size restrictions...)
4292 if (layout_is_con) {
4293 int hsize = Klass::layout_helper_header_size(layout_con);
4294 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4295 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4296 if ((round_mask & ~right_n_bits(eshift)) == 0)
4297 round_mask = 0; // strength-reduce it if it goes away completely
4298 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4299 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4300 assert(header_size_min <= hsize, "generic minimum is smallest");
4301 header_size = intcon(hsize);
4302 } else {
4303 Node* hss = intcon(Klass::_lh_header_size_shift);
4304 Node* hsm = intcon(Klass::_lh_header_size_mask);
4305 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4306 header_size = _gvn.transform(new AndINode(header_size, hsm));
4307 }
4308
4309 Node* elem_shift = nullptr;
4310 if (layout_is_con) {
4311 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4312 if (eshift != 0)
4313 elem_shift = intcon(eshift);
4314 } else {
4315 // There is no need to mask or shift this value.
4316 // The semantics of LShiftINode include an implicit mask to 0x1F.
4317 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4318 elem_shift = layout_val;
4319 }
4320
4321 // Transition to native address size for all offset calculations:
4322 Node* lengthx = ConvI2X(length);
4323 Node* headerx = ConvI2X(header_size);
4324 #ifdef _LP64
4325 { const TypeInt* tilen = _gvn.find_int_type(length);
4326 if (tilen != nullptr && tilen->_lo < 0) {
4327 // Add a manual constraint to a positive range. Cf. array_element_address.
4328 jint size_max = fast_size_limit;
4329 if (size_max > tilen->_hi) size_max = tilen->_hi;
4330 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4331
4332 // Only do a narrow I2L conversion if the range check passed.
4333 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4334 _gvn.transform(iff);
4335 RegionNode* region = new RegionNode(3);
4336 _gvn.set_type(region, Type::CONTROL);
4337 lengthx = new PhiNode(region, TypeLong::LONG);
4338 _gvn.set_type(lengthx, TypeLong::LONG);
4339
4340 // Range check passed. Use ConvI2L node with narrow type.
4341 Node* passed = IfFalse(iff);
4342 region->init_req(1, passed);
4343 // Make I2L conversion control dependent to prevent it from
4344 // floating above the range check during loop optimizations.
4345 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4346
4347 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4348 region->init_req(2, IfTrue(iff));
4349 lengthx->init_req(2, ConvI2X(length));
4350
4351 set_control(region);
4352 record_for_igvn(region);
4353 record_for_igvn(lengthx);
4354 }
4355 }
4356 #endif
4357
4358 // Combine header size and body size for the array copy part, then align (if
4359 // necessary) for the allocation part. This computation cannot overflow,
4360 // because it is used only in two places, one where the length is sharply
4361 // limited, and the other after a successful allocation.
4362 Node* abody = lengthx;
4363 if (elem_shift != nullptr) {
4364 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
4365 }
4366 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4367
4368 if (return_size_val != nullptr) {
4369 // This is the size
4370 (*return_size_val) = non_rounded_size;
4371 }
4372
4373 Node* size = non_rounded_size;
4374 if (round_mask != 0) {
4375 Node* mask1 = MakeConX(round_mask);
4376 size = _gvn.transform(new AddXNode(size, mask1));
4377 Node* mask2 = MakeConX(~round_mask);
4378 size = _gvn.transform(new AndXNode(size, mask2));
4379 }
4380 // else if round_mask == 0, the size computation is self-rounding
4381
4382 // Now generate allocation code
4383
4384 // The entire memory state is needed for slow path of the allocation
4385 // since GC and deoptimization can happen.
4386 Node *mem = reset_memory();
4387 set_all_memory(mem); // Create new memory state
4388
4389 if (initial_slow_test->is_Bool()) {
4390 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4391 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4392 }
4393
4394 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4395 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4396 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4397
4398 // Inline type array variants:
4399 // - null-ok: ciObjArrayKlass with is_elem_null_free() = false
4400 // - null-free: ciObjArrayKlass with is_elem_null_free() = true
4401 // - null-free, flat: ciFlatArrayKlass with is_elem_null_free() = true
4402 // Check if array is a null-free, non-flat inline type array
4403 // that needs to be initialized with the default inline type.
4404 Node* default_value = nullptr;
4405 Node* raw_default_value = nullptr;
4406 if (ary_ptr != nullptr && ary_ptr->klass_is_exact()) {
4407 // Array type is known
4408 if (ary_ptr->is_null_free() && !ary_ptr->is_flat()) {
4409 ciInlineKlass* vk = ary_ptr->elem()->inline_klass();
4410 default_value = InlineTypeNode::default_oop(gvn(), vk);
4411 if (UseCompressedOops) {
4412 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4413 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4414 Node* lower = _gvn.transform(new CastP2XNode(control(), default_value));
4415 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4416 raw_default_value = _gvn.transform(new OrLNode(lower, upper));
4417 } else {
4418 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4419 }
4420 }
4421 }
4422
4423 Node* valid_length_test = _gvn.intcon(1);
4424 if (ary_type->isa_aryptr()) {
4425 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4426 jint max = TypeAryPtr::max_array_length(bt);
4427 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4428 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4429 }
4430
4431 // Create the AllocateArrayNode and its result projections
4432 AllocateArrayNode* alloc
4433 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4434 control(), mem, i_o(),
4435 size, klass_node,
4436 initial_slow_test,
4437 length, valid_length_test,
4438 default_value, raw_default_value);
4439 // Cast to correct type. Note that the klass_node may be constant or not,
4440 // and in the latter case the actual array type will be inexact also.
4441 // (This happens via a non-constant argument to inline_native_newArray.)
4442 // In any case, the value of klass_node provides the desired array type.
4443 const TypeInt* length_type = _gvn.find_int_type(length);
4444 if (ary_type->isa_aryptr() && length_type != nullptr) {
4445 // Try to get a better type than POS for the size
4446 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4447 }
4448
4449 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4450
4451 array_ideal_length(alloc, ary_type, true);
4452 return javaoop;
4453 }
4454
4455 // The following "Ideal_foo" functions are placed here because they recognize
4456 // the graph shapes created by the functions immediately above.
4457
4458 //---------------------------Ideal_allocation----------------------------------
4459 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4460 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
4461 if (ptr == nullptr) { // reduce dumb test in callers
4462 return nullptr;
4463 }
4464
4465 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4466 ptr = bs->step_over_gc_barrier(ptr);
4467
4468 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4469 ptr = ptr->in(1);
4470 if (ptr == nullptr) return nullptr;
4471 }
4472 // Return null for allocations with several casts:
4473 // j.l.reflect.Array.newInstance(jobject, jint)
4474 // Object.clone()
4475 // to keep more precise type from last cast.
4476 if (ptr->is_Proj()) {
4477 Node* allo = ptr->in(0);
4478 if (allo != nullptr && allo->is_Allocate()) {
4479 return allo->as_Allocate();
4480 }
4481 }
4482 // Report failure to match.
4483 return nullptr;
4484 }
4485
4486 // Fancy version which also strips off an offset (and reports it to caller).
4487 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4488 intptr_t& offset) {
4489 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4490 if (base == nullptr) return nullptr;
4491 return Ideal_allocation(base);
4492 }
4493
4494 // Trace Initialize <- Proj[Parm] <- Allocate
4495 AllocateNode* InitializeNode::allocation() {
4496 Node* rawoop = in(InitializeNode::RawAddress);
4497 if (rawoop->is_Proj()) {
4498 Node* alloc = rawoop->in(0);
4499 if (alloc->is_Allocate()) {
4500 return alloc->as_Allocate();
4501 }
4502 }
4503 return nullptr;
4504 }
4505
4506 // Trace Allocate -> Proj[Parm] -> Initialize
4507 InitializeNode* AllocateNode::initialization() {
4508 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4509 if (rawoop == nullptr) return nullptr;
4510 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4511 Node* init = rawoop->fast_out(i);
4512 if (init->is_Initialize()) {
4513 assert(init->as_Initialize()->allocation() == this, "2-way link");
4514 return init->as_Initialize();
4515 }
4516 }
4517 return nullptr;
4518 }
4519
4520 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4521 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4522 // Too many traps seen?
4523 if (too_many_traps(reason)) {
4524 #ifdef ASSERT
4525 if (TraceLoopPredicate) {
4526 int tc = C->trap_count(reason);
4527 tty->print("too many traps=%s tcount=%d in ",
4528 Deoptimization::trap_reason_name(reason), tc);
4529 method()->print(); // which method has too many predicate traps
4530 tty->cr();
4531 }
4532 #endif
4533 // We cannot afford to take more traps here,
4534 // do not generate Parse Predicate.
4535 return;
4536 }
4537
4538 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4539 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4540 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4541 {
4542 PreserveJVMState pjvms(this);
4543 set_control(if_false);
4544 inc_sp(nargs);
4545 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4546 }
4547 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4548 set_control(if_true);
4549 }
4550
4551 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4552 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4553 void GraphKit::add_parse_predicates(int nargs) {
4554 if (UseLoopPredicate) {
4555 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4556 }
4557 if (UseProfiledLoopPredicate) {
4558 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4559 }
4560 // Loop Limit Check Predicate should be near the loop.
4561 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4562 }
4563
4564 void GraphKit::sync_kit(IdealKit& ideal) {
4565 set_all_memory(ideal.merged_memory());
4566 set_i_o(ideal.i_o());
4567 set_control(ideal.ctrl());
4568 }
4569
4570 void GraphKit::final_sync(IdealKit& ideal) {
4571 // Final sync IdealKit and graphKit.
4572 sync_kit(ideal);
4573 }
4574
4575 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4576 Node* len = load_array_length(load_String_value(str, set_ctrl));
4577 Node* coder = load_String_coder(str, set_ctrl);
4578 // Divide length by 2 if coder is UTF16
4579 return _gvn.transform(new RShiftINode(len, coder));
4580 }
4581
4582 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4583 int value_offset = java_lang_String::value_offset();
4584 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4585 false, nullptr, Type::Offset(0));
4586 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4587 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4588 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true),
4589 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4590 Node* p = basic_plus_adr(str, str, value_offset);
4591 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4592 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4593 return load;
4594 }
4595
4596 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4597 if (!CompactStrings) {
4598 return intcon(java_lang_String::CODER_UTF16);
4599 }
4600 int coder_offset = java_lang_String::coder_offset();
4601 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4602 false, nullptr, Type::Offset(0));
4603 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4604
4605 Node* p = basic_plus_adr(str, str, coder_offset);
4606 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4607 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4608 return load;
4609 }
4610
4611 void GraphKit::store_String_value(Node* str, Node* value) {
4612 int value_offset = java_lang_String::value_offset();
4613 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4614 false, nullptr, Type::Offset(0));
4615 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4616
4617 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4618 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4619 }
4620
4621 void GraphKit::store_String_coder(Node* str, Node* value) {
4622 int coder_offset = java_lang_String::coder_offset();
4623 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4624 false, nullptr, Type::Offset(0));
4625 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4626
4627 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4628 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4629 }
4630
4631 // Capture src and dst memory state with a MergeMemNode
4632 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4633 if (src_type == dst_type) {
4634 // Types are equal, we don't need a MergeMemNode
4635 return memory(src_type);
4636 }
4637 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4638 record_for_igvn(merge); // fold it up later, if possible
4639 int src_idx = C->get_alias_index(src_type);
4640 int dst_idx = C->get_alias_index(dst_type);
4641 merge->set_memory_at(src_idx, memory(src_idx));
4642 merge->set_memory_at(dst_idx, memory(dst_idx));
4643 return merge;
4644 }
4645
4646 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4647 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4648 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4649 // If input and output memory types differ, capture both states to preserve
4650 // the dependency between preceding and subsequent loads/stores.
4651 // For example, the following program:
4652 // StoreB
4653 // compress_string
4654 // LoadB
4655 // has this memory graph (use->def):
4656 // LoadB -> compress_string -> CharMem
4657 // ... -> StoreB -> ByteMem
4658 // The intrinsic hides the dependency between LoadB and StoreB, causing
4659 // the load to read from memory not containing the result of the StoreB.
4660 // The correct memory graph should look like this:
4661 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4662 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4663 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4664 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4665 set_memory(res_mem, TypeAryPtr::BYTES);
4666 return str;
4667 }
4668
4669 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4670 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4671 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4672 // Capture src and dst memory (see comment in 'compress_string').
4673 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4674 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4675 set_memory(_gvn.transform(str), dst_type);
4676 }
4677
4678 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4679 /**
4680 * int i_char = start;
4681 * for (int i_byte = 0; i_byte < count; i_byte++) {
4682 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4683 * }
4684 */
4685 add_parse_predicates();
4686 C->set_has_loops(true);
4687
4688 RegionNode* head = new RegionNode(3);
4689 head->init_req(1, control());
4690 gvn().set_type(head, Type::CONTROL);
4691 record_for_igvn(head);
4692
4693 Node* i_byte = new PhiNode(head, TypeInt::INT);
4694 i_byte->init_req(1, intcon(0));
4695 gvn().set_type(i_byte, TypeInt::INT);
4696 record_for_igvn(i_byte);
4697
4698 Node* i_char = new PhiNode(head, TypeInt::INT);
4699 i_char->init_req(1, start);
4700 gvn().set_type(i_char, TypeInt::INT);
4701 record_for_igvn(i_char);
4702
4703 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4704 gvn().set_type(mem, Type::MEMORY);
4705 record_for_igvn(mem);
4706 set_control(head);
4707 set_memory(mem, TypeAryPtr::BYTES);
4708 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4709 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4710 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4711 false, false, true /* mismatched */);
4712
4713 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4714 head->init_req(2, IfTrue(iff));
4715 mem->init_req(2, st);
4716 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4717 i_char->init_req(2, AddI(i_char, intcon(2)));
4718
4719 set_control(IfFalse(iff));
4720 set_memory(st, TypeAryPtr::BYTES);
4721 }
4722
4723 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4724 if (!field->is_constant()) {
4725 return nullptr; // Field not marked as constant.
4726 }
4727 ciInstance* holder = nullptr;
4728 if (!field->is_static()) {
4729 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4730 if (const_oop != nullptr && const_oop->is_instance()) {
4731 holder = const_oop->as_instance();
4732 }
4733 }
4734 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4735 /*is_unsigned_load=*/false);
4736 if (con_type != nullptr) {
4737 Node* con = makecon(con_type);
4738 if (field->type()->is_inlinetype()) {
4739 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free());
4740 } else if (con_type->is_inlinetypeptr()) {
4741 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free());
4742 }
4743 return con;
4744 }
4745 return nullptr;
4746 }
4747
4748 //---------------------------load_mirror_from_klass----------------------------
4749 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4750 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4751 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4752 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4753 // mirror = ((OopHandle)mirror)->resolve();
4754 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4755 }
4756
4757 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4758 const Type* obj_type = obj->bottom_type();
4759 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4760 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4761 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4762 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4763 obj = casted_obj;
4764 }
4765 if (sig_type->is_inlinetypeptr()) {
4766 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass(), !gvn().type(obj)->maybe_null());
4767 }
4768 return obj;
4769 }