1 /*
2 * Copyright (c) 2001, 2023, 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_ex_handler----------------------------------
197 // Tell if this method or any caller method has exception handlers.
198 bool GraphKit::has_ex_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_ex_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 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
964 uint debug_end = debug_ptr;
965 uint debug_start = debug_ptr - in_jvms->debug_size();
966 debug_ptr = debug_start; // back up the ptr
967
968 uint p = debug_start; // walks forward in [debug_start, debug_end)
969 uint j, k, l;
970 SafePointNode* in_map = in_jvms->map();
971 out_jvms->set_map(call);
972
973 if (can_prune_locals) {
974 assert(in_jvms->method() == out_jvms->method(), "sanity");
975 // If the current throw can reach an exception handler in this JVMS,
976 // then we must keep everything live that can reach that handler.
977 // As a quick and dirty approximation, we look for any handlers at all.
978 if (in_jvms->method()->has_exception_handlers()) {
979 can_prune_locals = false;
980 }
981 }
982
983 // Add the Locals
984 k = in_jvms->locoff();
985 l = in_jvms->loc_size();
986 out_jvms->set_locoff(p);
987 if (!can_prune_locals) {
988 for (j = 0; j < l; j++)
989 call->set_req(p++, in_map->in(k+j));
990 } else {
991 p += l; // already set to top above by add_req_batch
992 }
993
994 // Add the Expression Stack
995 k = in_jvms->stkoff();
996 l = in_jvms->sp();
997 out_jvms->set_stkoff(p);
998 if (!can_prune_locals) {
999 for (j = 0; j < l; j++)
1000 call->set_req(p++, in_map->in(k+j));
1001 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1002 // Divide stack into {S0,...,S1}, where S0 is set to top.
1003 uint s1 = stack_slots_not_pruned;
1004 stack_slots_not_pruned = 0; // for next iteration
1005 if (s1 > l) s1 = l;
1006 uint s0 = l - s1;
1007 p += s0; // skip the tops preinstalled by add_req_batch
1008 for (j = s0; j < l; j++)
1009 call->set_req(p++, in_map->in(k+j));
1010 } else {
1011 p += l; // already set to top above by add_req_batch
1012 }
1013
1014 // Add the Monitors
1015 k = in_jvms->monoff();
1016 l = in_jvms->mon_size();
1017 out_jvms->set_monoff(p);
1018 for (j = 0; j < l; j++)
1019 call->set_req(p++, in_map->in(k+j));
1020
1021 // Copy any scalar object fields.
1022 k = in_jvms->scloff();
1023 l = in_jvms->scl_size();
1024 out_jvms->set_scloff(p);
1025 for (j = 0; j < l; j++)
1026 call->set_req(p++, in_map->in(k+j));
1027
1028 // Finish the new jvms.
1029 out_jvms->set_endoff(p);
1030
1031 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1032 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1033 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1034 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1035 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1036 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1037
1038 // Update the two tail pointers in parallel.
1039 out_jvms = out_jvms->caller();
1040 in_jvms = in_jvms->caller();
1041 }
1042
1043 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1044
1045 // Test the correctness of JVMState::debug_xxx accessors:
1046 assert(call->jvms()->debug_start() == non_debug_edges, "");
1047 assert(call->jvms()->debug_end() == call->req(), "");
1048 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1049 }
1050
1051 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1052 Bytecodes::Code code = java_bc();
1053 if (code == Bytecodes::_wide) {
1054 code = method()->java_code_at_bci(bci() + 1);
1055 }
1056
1057 if (code != Bytecodes::_illegal) {
1058 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1059 }
1060
1061 auto rsize = [&]() {
1062 assert(code != Bytecodes::_illegal, "code is illegal!");
1063 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1064 return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1065 };
1066
1067 switch (code) {
1068 case Bytecodes::_illegal:
1069 return false;
1070
1071 case Bytecodes::_ldc:
1072 case Bytecodes::_ldc_w:
1073 case Bytecodes::_ldc2_w:
1074 inputs = 0;
1075 break;
1076
1077 case Bytecodes::_dup: inputs = 1; break;
1078 case Bytecodes::_dup_x1: inputs = 2; break;
1079 case Bytecodes::_dup_x2: inputs = 3; break;
1080 case Bytecodes::_dup2: inputs = 2; break;
1081 case Bytecodes::_dup2_x1: inputs = 3; break;
1082 case Bytecodes::_dup2_x2: inputs = 4; break;
1083 case Bytecodes::_swap: inputs = 2; break;
1084 case Bytecodes::_arraylength: inputs = 1; break;
1085
1086 case Bytecodes::_getstatic:
1087 case Bytecodes::_putstatic:
1088 case Bytecodes::_getfield:
1089 case Bytecodes::_putfield:
1090 {
1091 bool ignored_will_link;
1092 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1093 int size = field->type()->size();
1094 bool is_get = (depth >= 0), is_static = (depth & 1);
1095 inputs = (is_static ? 0 : 1);
1096 if (is_get) {
1097 depth = size - inputs;
1098 } else {
1099 inputs += size; // putxxx pops the value from the stack
1100 depth = - inputs;
1101 }
1102 }
1103 break;
1104
1105 case Bytecodes::_invokevirtual:
1106 case Bytecodes::_invokespecial:
1107 case Bytecodes::_invokestatic:
1108 case Bytecodes::_invokedynamic:
1109 case Bytecodes::_invokeinterface:
1110 {
1111 bool ignored_will_link;
1112 ciSignature* declared_signature = nullptr;
1113 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1114 assert(declared_signature != nullptr, "cannot be null");
1115 inputs = declared_signature->arg_size_for_bc(code);
1116 int size = declared_signature->return_type()->size();
1117 depth = size - inputs;
1118 }
1119 break;
1120
1121 case Bytecodes::_multianewarray:
1122 {
1123 ciBytecodeStream iter(method());
1124 iter.reset_to_bci(bci());
1125 iter.next();
1126 inputs = iter.get_dimensions();
1127 assert(rsize() == 1, "");
1128 depth = 1 - inputs;
1129 }
1130 break;
1131
1132 case Bytecodes::_withfield: {
1133 bool ignored_will_link;
1134 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1135 int size = field->type()->size();
1136 inputs = size+1;
1137 depth = rsize() - inputs;
1138 break;
1139 }
1140
1141 case Bytecodes::_ireturn:
1142 case Bytecodes::_lreturn:
1143 case Bytecodes::_freturn:
1144 case Bytecodes::_dreturn:
1145 case Bytecodes::_areturn:
1146 assert(rsize() == -depth, "");
1147 inputs = -depth;
1148 break;
1149
1150 case Bytecodes::_jsr:
1151 case Bytecodes::_jsr_w:
1152 inputs = 0;
1153 depth = 1; // S.B. depth=1, not zero
1154 break;
1155
1156 default:
1157 // bytecode produces a typed result
1158 inputs = rsize() - depth;
1159 assert(inputs >= 0, "");
1160 break;
1161 }
1162
1163 #ifdef ASSERT
1164 // spot check
1165 int outputs = depth + inputs;
1166 assert(outputs >= 0, "sanity");
1167 switch (code) {
1168 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1169 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1170 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1171 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1172 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1173 default: break;
1174 }
1175 #endif //ASSERT
1176
1177 return true;
1178 }
1179
1180
1181
1182 //------------------------------basic_plus_adr---------------------------------
1183 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1184 // short-circuit a common case
1185 if (offset == intcon(0)) return ptr;
1186 return _gvn.transform( new AddPNode(base, ptr, offset) );
1187 }
1188
1189 Node* GraphKit::ConvI2L(Node* offset) {
1190 // short-circuit a common case
1191 jint offset_con = find_int_con(offset, Type::OffsetBot);
1192 if (offset_con != Type::OffsetBot) {
1193 return longcon((jlong) offset_con);
1194 }
1195 return _gvn.transform( new ConvI2LNode(offset));
1196 }
1197
1198 Node* GraphKit::ConvI2UL(Node* offset) {
1199 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1200 if (offset_con != (juint) Type::OffsetBot) {
1201 return longcon((julong) offset_con);
1202 }
1203 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1204 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1205 return _gvn.transform( new AndLNode(conv, mask) );
1206 }
1207
1208 Node* GraphKit::ConvL2I(Node* offset) {
1209 // short-circuit a common case
1210 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1211 if (offset_con != (jlong)Type::OffsetBot) {
1212 return intcon((int) offset_con);
1213 }
1214 return _gvn.transform( new ConvL2INode(offset));
1215 }
1216
1217 //-------------------------load_object_klass-----------------------------------
1218 Node* GraphKit::load_object_klass(Node* obj) {
1219 // Special-case a fresh allocation to avoid building nodes:
1220 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1221 if (akls != nullptr) return akls;
1222 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1223 return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1224 }
1225
1226 //-------------------------load_array_length-----------------------------------
1227 Node* GraphKit::load_array_length(Node* array) {
1228 // Special-case a fresh allocation to avoid building nodes:
1229 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1230 Node *alen;
1231 if (alloc == nullptr) {
1232 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1233 alen = _gvn.transform( new LoadRangeNode(0, immutable_memory(), r_adr, TypeInt::POS));
1234 } else {
1235 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1236 }
1237 return alen;
1238 }
1239
1240 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1241 const TypeOopPtr* oop_type,
1242 bool replace_length_in_map) {
1243 Node* length = alloc->Ideal_length();
1244 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1245 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1246 if (ccast != length) {
1247 // do not transform ccast here, it might convert to top node for
1248 // negative array length and break assumptions in parsing stage.
1249 _gvn.set_type_bottom(ccast);
1250 record_for_igvn(ccast);
1251 if (replace_length_in_map) {
1252 replace_in_map(length, ccast);
1253 }
1254 return ccast;
1255 }
1256 }
1257 return length;
1258 }
1259
1260 //------------------------------do_null_check----------------------------------
1261 // Helper function to do a null pointer check. Returned value is
1262 // the incoming address with null casted away. You are allowed to use the
1263 // not-null value only if you are control dependent on the test.
1264 #ifndef PRODUCT
1265 extern uint explicit_null_checks_inserted,
1266 explicit_null_checks_elided;
1267 #endif
1268 Node* GraphKit::null_check_common(Node* value, BasicType type,
1269 // optional arguments for variations:
1270 bool assert_null,
1271 Node* *null_control,
1272 bool speculative,
1273 bool is_init_check) {
1274 assert(!assert_null || null_control == nullptr, "not both at once");
1275 if (stopped()) return top();
1276 NOT_PRODUCT(explicit_null_checks_inserted++);
1277
1278 if (value->is_InlineType()) {
1279 // Null checking a scalarized but nullable inline type. Check the IsInit
1280 // input instead of the oop input to avoid keeping buffer allocations alive.
1281 InlineTypeNode* vtptr = value->as_InlineType();
1282 while (vtptr->get_oop()->is_InlineType()) {
1283 vtptr = vtptr->get_oop()->as_InlineType();
1284 }
1285 null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true);
1286 if (stopped()) {
1287 return top();
1288 }
1289 if (assert_null) {
1290 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1291 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1292 // replace_in_map(value, vtptr);
1293 // return vtptr;
1294 return null();
1295 }
1296 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1297 return cast_not_null(value, do_replace_in_map);
1298 }
1299
1300 // Construct null check
1301 Node *chk = nullptr;
1302 switch(type) {
1303 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1304 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1305 case T_PRIMITIVE_OBJECT : // fall through
1306 case T_ARRAY : // fall through
1307 type = T_OBJECT; // simplify further tests
1308 case T_OBJECT : {
1309 const Type *t = _gvn.type( value );
1310
1311 const TypeOopPtr* tp = t->isa_oopptr();
1312 if (tp != nullptr && !tp->is_loaded()
1313 // Only for do_null_check, not any of its siblings:
1314 && !assert_null && null_control == nullptr) {
1315 // Usually, any field access or invocation on an unloaded oop type
1316 // will simply fail to link, since the statically linked class is
1317 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1318 // the static class is loaded but the sharper oop type is not.
1319 // Rather than checking for this obscure case in lots of places,
1320 // we simply observe that a null check on an unloaded class
1321 // will always be followed by a nonsense operation, so we
1322 // can just issue the uncommon trap here.
1323 // Our access to the unloaded class will only be correct
1324 // after it has been loaded and initialized, which requires
1325 // a trip through the interpreter.
1326 ciKlass* klass = tp->unloaded_klass();
1327 #ifndef PRODUCT
1328 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1329 #endif
1330 uncommon_trap(Deoptimization::Reason_unloaded,
1331 Deoptimization::Action_reinterpret,
1332 klass, "!loaded");
1333 return top();
1334 }
1335
1336 if (assert_null) {
1337 // See if the type is contained in NULL_PTR.
1338 // If so, then the value is already null.
1339 if (t->higher_equal(TypePtr::NULL_PTR)) {
1340 NOT_PRODUCT(explicit_null_checks_elided++);
1341 return value; // Elided null assert quickly!
1342 }
1343 } else {
1344 // See if mixing in the null pointer changes type.
1345 // If so, then the null pointer was not allowed in the original
1346 // type. In other words, "value" was not-null.
1347 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1348 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1349 NOT_PRODUCT(explicit_null_checks_elided++);
1350 return value; // Elided null check quickly!
1351 }
1352 }
1353 chk = new CmpPNode( value, null() );
1354 break;
1355 }
1356
1357 default:
1358 fatal("unexpected type: %s", type2name(type));
1359 }
1360 assert(chk != nullptr, "sanity check");
1361 chk = _gvn.transform(chk);
1362
1363 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1364 BoolNode *btst = new BoolNode( chk, btest);
1365 Node *tst = _gvn.transform( btst );
1366
1367 //-----------
1368 // if peephole optimizations occurred, a prior test existed.
1369 // If a prior test existed, maybe it dominates as we can avoid this test.
1370 if (tst != btst && type == T_OBJECT) {
1371 // At this point we want to scan up the CFG to see if we can
1372 // find an identical test (and so avoid this test altogether).
1373 Node *cfg = control();
1374 int depth = 0;
1375 while( depth < 16 ) { // Limit search depth for speed
1376 if( cfg->Opcode() == Op_IfTrue &&
1377 cfg->in(0)->in(1) == tst ) {
1378 // Found prior test. Use "cast_not_null" to construct an identical
1379 // CastPP (and hence hash to) as already exists for the prior test.
1380 // Return that casted value.
1381 if (assert_null) {
1382 replace_in_map(value, null());
1383 return null(); // do not issue the redundant test
1384 }
1385 Node *oldcontrol = control();
1386 set_control(cfg);
1387 Node *res = cast_not_null(value);
1388 set_control(oldcontrol);
1389 NOT_PRODUCT(explicit_null_checks_elided++);
1390 return res;
1391 }
1392 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1393 if (cfg == nullptr) break; // Quit at region nodes
1394 depth++;
1395 }
1396 }
1397
1398 //-----------
1399 // Branch to failure if null
1400 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1401 Deoptimization::DeoptReason reason;
1402 if (assert_null) {
1403 reason = Deoptimization::reason_null_assert(speculative);
1404 } else if (type == T_OBJECT || is_init_check) {
1405 reason = Deoptimization::reason_null_check(speculative);
1406 } else {
1407 reason = Deoptimization::Reason_div0_check;
1408 }
1409 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1410 // ciMethodData::has_trap_at will return a conservative -1 if any
1411 // must-be-null assertion has failed. This could cause performance
1412 // problems for a method after its first do_null_assert failure.
1413 // Consider using 'Reason_class_check' instead?
1414
1415 // To cause an implicit null check, we set the not-null probability
1416 // to the maximum (PROB_MAX). For an explicit check the probability
1417 // is set to a smaller value.
1418 if (null_control != nullptr || too_many_traps(reason)) {
1419 // probability is less likely
1420 ok_prob = PROB_LIKELY_MAG(3);
1421 } else if (!assert_null &&
1422 (ImplicitNullCheckThreshold > 0) &&
1423 method() != nullptr &&
1424 (method()->method_data()->trap_count(reason)
1425 >= (uint)ImplicitNullCheckThreshold)) {
1426 ok_prob = PROB_LIKELY_MAG(3);
1427 }
1428
1429 if (null_control != nullptr) {
1430 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1431 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1432 set_control( _gvn.transform( new IfTrueNode(iff)));
1433 #ifndef PRODUCT
1434 if (null_true == top()) {
1435 explicit_null_checks_elided++;
1436 }
1437 #endif
1438 (*null_control) = null_true;
1439 } else {
1440 BuildCutout unless(this, tst, ok_prob);
1441 // Check for optimizer eliding test at parse time
1442 if (stopped()) {
1443 // Failure not possible; do not bother making uncommon trap.
1444 NOT_PRODUCT(explicit_null_checks_elided++);
1445 } else if (assert_null) {
1446 uncommon_trap(reason,
1447 Deoptimization::Action_make_not_entrant,
1448 nullptr, "assert_null");
1449 } else {
1450 replace_in_map(value, zerocon(type));
1451 builtin_throw(reason);
1452 }
1453 }
1454
1455 // Must throw exception, fall-thru not possible?
1456 if (stopped()) {
1457 return top(); // No result
1458 }
1459
1460 if (assert_null) {
1461 // Cast obj to null on this path.
1462 replace_in_map(value, zerocon(type));
1463 return zerocon(type);
1464 }
1465
1466 // Cast obj to not-null on this path, if there is no null_control.
1467 // (If there is a null_control, a non-null value may come back to haunt us.)
1468 if (type == T_OBJECT) {
1469 Node* cast = cast_not_null(value, false);
1470 if (null_control == nullptr || (*null_control) == top())
1471 replace_in_map(value, cast);
1472 value = cast;
1473 }
1474
1475 return value;
1476 }
1477
1478 //------------------------------cast_not_null----------------------------------
1479 // Cast obj to not-null on this path
1480 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1481 if (obj->is_InlineType()) {
1482 Node* vt = obj->clone();
1483 vt->as_InlineType()->set_is_init(_gvn);
1484 vt = _gvn.transform(vt);
1485 if (do_replace_in_map) {
1486 replace_in_map(obj, vt);
1487 }
1488 return vt;
1489 }
1490 const Type *t = _gvn.type(obj);
1491 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1492 // Object is already not-null?
1493 if( t == t_not_null ) return obj;
1494
1495 Node *cast = new CastPPNode(obj,t_not_null);
1496 cast->init_req(0, control());
1497 cast = _gvn.transform( cast );
1498
1499 // Scan for instances of 'obj' in the current JVM mapping.
1500 // These instances are known to be not-null after the test.
1501 if (do_replace_in_map)
1502 replace_in_map(obj, cast);
1503
1504 return cast; // Return casted value
1505 }
1506
1507 // Sometimes in intrinsics, we implicitly know an object is not null
1508 // (there's no actual null check) so we can cast it to not null. In
1509 // the course of optimizations, the input to the cast can become null.
1510 // In that case that data path will die and we need the control path
1511 // to become dead as well to keep the graph consistent. So we have to
1512 // add a check for null for which one branch can't be taken. It uses
1513 // an Opaque4 node that will cause the check to be removed after loop
1514 // opts so the test goes away and the compiled code doesn't execute a
1515 // useless check.
1516 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1517 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1518 return value;
1519 }
1520 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1521 Node *tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1522 Node* opaq = _gvn.transform(new Opaque4Node(C, tst, intcon(1)));
1523 IfNode *iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1524 _gvn.set_type(iff, iff->Value(&_gvn));
1525 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1526 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1527 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1528 C->root()->add_req(halt);
1529 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1530 set_control(if_t);
1531 return cast_not_null(value, do_replace_in_map);
1532 }
1533
1534
1535 //--------------------------replace_in_map-------------------------------------
1536 void GraphKit::replace_in_map(Node* old, Node* neww) {
1537 if (old == neww) {
1538 return;
1539 }
1540
1541 map()->replace_edge(old, neww);
1542
1543 // Note: This operation potentially replaces any edge
1544 // on the map. This includes locals, stack, and monitors
1545 // of the current (innermost) JVM state.
1546
1547 // don't let inconsistent types from profiling escape this
1548 // method
1549
1550 const Type* told = _gvn.type(old);
1551 const Type* tnew = _gvn.type(neww);
1552
1553 if (!tnew->higher_equal(told)) {
1554 return;
1555 }
1556
1557 map()->record_replaced_node(old, neww);
1558 }
1559
1560
1561 //=============================================================================
1562 //--------------------------------memory---------------------------------------
1563 Node* GraphKit::memory(uint alias_idx) {
1564 MergeMemNode* mem = merged_memory();
1565 Node* p = mem->memory_at(alias_idx);
1566 assert(p != mem->empty_memory(), "empty");
1567 _gvn.set_type(p, Type::MEMORY); // must be mapped
1568 return p;
1569 }
1570
1571 //-----------------------------reset_memory------------------------------------
1572 Node* GraphKit::reset_memory() {
1573 Node* mem = map()->memory();
1574 // do not use this node for any more parsing!
1575 debug_only( map()->set_memory((Node*)nullptr) );
1576 return _gvn.transform( mem );
1577 }
1578
1579 //------------------------------set_all_memory---------------------------------
1580 void GraphKit::set_all_memory(Node* newmem) {
1581 Node* mergemem = MergeMemNode::make(newmem);
1582 gvn().set_type_bottom(mergemem);
1583 map()->set_memory(mergemem);
1584 }
1585
1586 //------------------------------set_all_memory_call----------------------------
1587 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1588 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1589 set_all_memory(newmem);
1590 }
1591
1592 //=============================================================================
1593 //
1594 // parser factory methods for MemNodes
1595 //
1596 // These are layered on top of the factory methods in LoadNode and StoreNode,
1597 // and integrate with the parser's memory state and _gvn engine.
1598 //
1599
1600 // factory methods in "int adr_idx"
1601 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1602 int adr_idx,
1603 MemNode::MemOrd mo,
1604 LoadNode::ControlDependency control_dependency,
1605 bool require_atomic_access,
1606 bool unaligned,
1607 bool mismatched,
1608 bool unsafe,
1609 uint8_t barrier_data) {
1610 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1611 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1612 debug_only(adr_type = C->get_adr_type(adr_idx));
1613 Node* mem = memory(adr_idx);
1614 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1615 ld = _gvn.transform(ld);
1616
1617 if (((bt == T_OBJECT || bt == T_PRIMITIVE_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1618 // Improve graph before escape analysis and boxing elimination.
1619 record_for_igvn(ld);
1620 }
1621 return ld;
1622 }
1623
1624 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1625 int adr_idx,
1626 MemNode::MemOrd mo,
1627 bool require_atomic_access,
1628 bool unaligned,
1629 bool mismatched,
1630 bool unsafe,
1631 int barrier_data) {
1632 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1633 const TypePtr* adr_type = nullptr;
1634 debug_only(adr_type = C->get_adr_type(adr_idx));
1635 Node *mem = memory(adr_idx);
1636 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1637 if (unaligned) {
1638 st->as_Store()->set_unaligned_access();
1639 }
1640 if (mismatched) {
1641 st->as_Store()->set_mismatched_access();
1642 }
1643 if (unsafe) {
1644 st->as_Store()->set_unsafe_access();
1645 }
1646 st->as_Store()->set_barrier_data(barrier_data);
1647 st = _gvn.transform(st);
1648 set_memory(st, adr_idx);
1649 // Back-to-back stores can only remove intermediate store with DU info
1650 // so push on worklist for optimizer.
1651 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1652 record_for_igvn(st);
1653
1654 return st;
1655 }
1656
1657 Node* GraphKit::access_store_at(Node* obj,
1658 Node* adr,
1659 const TypePtr* adr_type,
1660 Node* val,
1661 const Type* val_type,
1662 BasicType bt,
1663 DecoratorSet decorators,
1664 bool safe_for_replace) {
1665 // Transformation of a value which could be null pointer (CastPP #null)
1666 // could be delayed during Parse (for example, in adjust_map_after_if()).
1667 // Execute transformation here to avoid barrier generation in such case.
1668 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1669 val = _gvn.makecon(TypePtr::NULL_PTR);
1670 }
1671
1672 if (stopped()) {
1673 return top(); // Dead path ?
1674 }
1675
1676 assert(val != nullptr, "not dead path");
1677 if (val->is_InlineType()) {
1678 // Store to non-flat field. Buffer the inline type and make sure
1679 // the store is re-executed if the allocation triggers deoptimization.
1680 PreserveReexecuteState preexecs(this);
1681 jvms()->set_should_reexecute(true);
1682 val = val->as_InlineType()->buffer(this, safe_for_replace);
1683 }
1684
1685 C2AccessValuePtr addr(adr, adr_type);
1686 C2AccessValue value(val, val_type);
1687 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1688 if (access.is_raw()) {
1689 return _barrier_set->BarrierSetC2::store_at(access, value);
1690 } else {
1691 return _barrier_set->store_at(access, value);
1692 }
1693 }
1694
1695 Node* GraphKit::access_load_at(Node* obj, // containing obj
1696 Node* adr, // actual address to store val at
1697 const TypePtr* adr_type,
1698 const Type* val_type,
1699 BasicType bt,
1700 DecoratorSet decorators,
1701 Node* ctl) {
1702 if (stopped()) {
1703 return top(); // Dead path ?
1704 }
1705
1706 C2AccessValuePtr addr(adr, adr_type);
1707 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1708 if (access.is_raw()) {
1709 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1710 } else {
1711 return _barrier_set->load_at(access, val_type);
1712 }
1713 }
1714
1715 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1716 const Type* val_type,
1717 BasicType bt,
1718 DecoratorSet decorators) {
1719 if (stopped()) {
1720 return top(); // Dead path ?
1721 }
1722
1723 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1724 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1725 if (access.is_raw()) {
1726 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1727 } else {
1728 return _barrier_set->load_at(access, val_type);
1729 }
1730 }
1731
1732 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1733 Node* adr,
1734 const TypePtr* adr_type,
1735 int alias_idx,
1736 Node* expected_val,
1737 Node* new_val,
1738 const Type* value_type,
1739 BasicType bt,
1740 DecoratorSet decorators) {
1741 C2AccessValuePtr addr(adr, adr_type);
1742 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1743 bt, obj, addr, alias_idx);
1744 if (access.is_raw()) {
1745 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1746 } else {
1747 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1748 }
1749 }
1750
1751 Node* GraphKit::access_atomic_cmpxchg_bool_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_bool_at(access, expected_val, new_val, value_type);
1765 } else {
1766 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1767 }
1768 }
1769
1770 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1771 Node* adr,
1772 const TypePtr* adr_type,
1773 int alias_idx,
1774 Node* new_val,
1775 const Type* value_type,
1776 BasicType bt,
1777 DecoratorSet decorators) {
1778 C2AccessValuePtr addr(adr, adr_type);
1779 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1780 bt, obj, addr, alias_idx);
1781 if (access.is_raw()) {
1782 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1783 } else {
1784 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1785 }
1786 }
1787
1788 Node* GraphKit::access_atomic_add_at(Node* obj,
1789 Node* adr,
1790 const TypePtr* adr_type,
1791 int alias_idx,
1792 Node* new_val,
1793 const Type* value_type,
1794 BasicType bt,
1795 DecoratorSet decorators) {
1796 C2AccessValuePtr addr(adr, adr_type);
1797 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1798 if (access.is_raw()) {
1799 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1800 } else {
1801 return _barrier_set->atomic_add_at(access, new_val, value_type);
1802 }
1803 }
1804
1805 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1806 return _barrier_set->clone(this, src, dst, size, is_array);
1807 }
1808
1809 //-------------------------array_element_address-------------------------
1810 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1811 const TypeInt* sizetype, Node* ctrl) {
1812 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1813 uint shift = arytype->is_flat() ? arytype->flat_log_elem_size() : exact_log2(type2aelembytes(elembt));
1814 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1815
1816 // short-circuit a common case (saves lots of confusing waste motion)
1817 jint idx_con = find_int_con(idx, -1);
1818 if (idx_con >= 0) {
1819 intptr_t offset = header + ((intptr_t)idx_con << shift);
1820 return basic_plus_adr(ary, offset);
1821 }
1822
1823 // must be correct type for alignment purposes
1824 Node* base = basic_plus_adr(ary, header);
1825 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1826 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1827 return basic_plus_adr(ary, base, scale);
1828 }
1829
1830 //-------------------------load_array_element-------------------------
1831 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1832 const Type* elemtype = arytype->elem();
1833 BasicType elembt = elemtype->array_element_basic_type();
1834 assert(elembt != T_PRIMITIVE_OBJECT, "inline types are not supported by this method");
1835 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1836 if (elembt == T_NARROWOOP) {
1837 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1838 }
1839 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1840 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1841 return ld;
1842 }
1843
1844 //-------------------------set_arguments_for_java_call-------------------------
1845 // Arguments (pre-popped from the stack) are taken from the JVMS.
1846 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1847 PreserveReexecuteState preexecs(this);
1848 if (EnableValhalla) {
1849 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1850 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1851 jvms()->set_should_reexecute(true);
1852 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1853 inc_sp(arg_size);
1854 }
1855 // Add the call arguments
1856 const TypeTuple* domain = call->tf()->domain_sig();
1857 uint nargs = domain->cnt();
1858 int arg_num = 0;
1859 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1860 Node* arg = argument(i-TypeFunc::Parms);
1861 const Type* t = domain->field_at(i);
1862 // TODO 8284443 A static call to a mismatched method should still be scalarized
1863 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1864 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1865 if (!arg->is_InlineType()) {
1866 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1867 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass(), t->inline_klass()->is_null_free());
1868 }
1869 InlineTypeNode* vt = arg->as_InlineType();
1870 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1871 // If an inline type argument is passed as fields, attach the Method* to the call site
1872 // to be able to access the extended signature later via attached_method_before_pc().
1873 // For example, see CompiledMethod::preserve_callee_argument_oops().
1874 call->set_override_symbolic_info(true);
1875 // Register an evol dependency on the callee method to make sure that this method is deoptimized and
1876 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1877 C->dependencies()->assert_evol_method(call->method());
1878 arg_num++;
1879 continue;
1880 } else if (arg->is_InlineType()) {
1881 // Pass inline type argument via oop to callee
1882 arg = arg->as_InlineType()->buffer(this);
1883 if (!is_late_inline) {
1884 arg = arg->as_InlineType()->get_oop();
1885 }
1886 }
1887 if (t != Type::HALF) {
1888 arg_num++;
1889 }
1890 call->init_req(idx++, arg);
1891 }
1892 }
1893
1894 //---------------------------set_edges_for_java_call---------------------------
1895 // Connect a newly created call into the current JVMS.
1896 // A return value node (if any) is returned from set_edges_for_java_call.
1897 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1898
1899 // Add the predefined inputs:
1900 call->init_req( TypeFunc::Control, control() );
1901 call->init_req( TypeFunc::I_O , i_o() );
1902 call->init_req( TypeFunc::Memory , reset_memory() );
1903 call->init_req( TypeFunc::FramePtr, frameptr() );
1904 call->init_req( TypeFunc::ReturnAdr, top() );
1905
1906 add_safepoint_edges(call, must_throw);
1907
1908 Node* xcall = _gvn.transform(call);
1909
1910 if (xcall == top()) {
1911 set_control(top());
1912 return;
1913 }
1914 assert(xcall == call, "call identity is stable");
1915
1916 // Re-use the current map to produce the result.
1917
1918 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1919 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1920 set_all_memory_call(xcall, separate_io_proj);
1921
1922 //return xcall; // no need, caller already has it
1923 }
1924
1925 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1926 if (stopped()) return top(); // maybe the call folded up?
1927
1928 // Note: Since any out-of-line call can produce an exception,
1929 // we always insert an I_O projection from the call into the result.
1930
1931 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1932
1933 if (separate_io_proj) {
1934 // The caller requested separate projections be used by the fall
1935 // through and exceptional paths, so replace the projections for
1936 // the fall through path.
1937 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1938 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1939 }
1940
1941 // Capture the return value, if any.
1942 Node* ret;
1943 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
1944 ret = top();
1945 } else if (call->tf()->returns_inline_type_as_fields()) {
1946 // Return of multiple values (inline type fields): we create a
1947 // InlineType node, each field is a projection from the call.
1948 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
1949 uint base_input = TypeFunc::Parms;
1950 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, call->method()->signature()->returns_null_free_inline_type());
1951 } else {
1952 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1953 ciType* t = call->method()->return_type();
1954 if (t->is_klass()) {
1955 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
1956 if (type->is_inlinetypeptr()) {
1957 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass(), type->inline_klass()->is_null_free());
1958 }
1959 }
1960 }
1961
1962 return ret;
1963 }
1964
1965 //--------------------set_predefined_input_for_runtime_call--------------------
1966 // Reading and setting the memory state is way conservative here.
1967 // The real problem is that I am not doing real Type analysis on memory,
1968 // so I cannot distinguish card mark stores from other stores. Across a GC
1969 // point the Store Barrier and the card mark memory has to agree. I cannot
1970 // have a card mark store and its barrier split across the GC point from
1971 // either above or below. Here I get that to happen by reading ALL of memory.
1972 // A better answer would be to separate out card marks from other memory.
1973 // For now, return the input memory state, so that it can be reused
1974 // after the call, if this call has restricted memory effects.
1975 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1976 // Set fixed predefined input arguments
1977 Node* memory = reset_memory();
1978 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
1979 call->init_req( TypeFunc::Control, control() );
1980 call->init_req( TypeFunc::I_O, top() ); // does no i/o
1981 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
1982 call->init_req( TypeFunc::FramePtr, frameptr() );
1983 call->init_req( TypeFunc::ReturnAdr, top() );
1984 return memory;
1985 }
1986
1987 //-------------------set_predefined_output_for_runtime_call--------------------
1988 // Set control and memory (not i_o) from the call.
1989 // If keep_mem is not null, use it for the output state,
1990 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
1991 // If hook_mem is null, this call produces no memory effects at all.
1992 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
1993 // then only that memory slice is taken from the call.
1994 // In the last case, we must put an appropriate memory barrier before
1995 // the call, so as to create the correct anti-dependencies on loads
1996 // preceding the call.
1997 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
1998 Node* keep_mem,
1999 const TypePtr* hook_mem) {
2000 // no i/o
2001 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
2002 if (keep_mem) {
2003 // First clone the existing memory state
2004 set_all_memory(keep_mem);
2005 if (hook_mem != nullptr) {
2006 // Make memory for the call
2007 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
2008 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
2009 // We also use hook_mem to extract specific effects from arraycopy stubs.
2010 set_memory(mem, hook_mem);
2011 }
2012 // ...else the call has NO memory effects.
2013
2014 // Make sure the call advertises its memory effects precisely.
2015 // This lets us build accurate anti-dependences in gcm.cpp.
2016 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
2017 "call node must be constructed correctly");
2018 } else {
2019 assert(hook_mem == nullptr, "");
2020 // This is not a "slow path" call; all memory comes from the call.
2021 set_all_memory_call(call);
2022 }
2023 }
2024
2025 // Keep track of MergeMems feeding into other MergeMems
2026 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2027 if (!mem->is_MergeMem()) {
2028 return;
2029 }
2030 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2031 Node* use = i.get();
2032 if (use->is_MergeMem()) {
2033 wl.push(use);
2034 }
2035 }
2036 }
2037
2038 // Replace the call with the current state of the kit.
2039 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes) {
2040 JVMState* ejvms = nullptr;
2041 if (has_exceptions()) {
2042 ejvms = transfer_exceptions_into_jvms();
2043 }
2044
2045 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2046 ReplacedNodes replaced_nodes_exception;
2047 Node* ex_ctl = top();
2048
2049 SafePointNode* final_state = stop();
2050
2051 // Find all the needed outputs of this call
2052 CallProjections* callprojs = call->extract_projections(true);
2053
2054 Unique_Node_List wl;
2055 Node* init_mem = call->in(TypeFunc::Memory);
2056 Node* final_mem = final_state->in(TypeFunc::Memory);
2057 Node* final_ctl = final_state->in(TypeFunc::Control);
2058 Node* final_io = final_state->in(TypeFunc::I_O);
2059
2060 // Replace all the old call edges with the edges from the inlining result
2061 if (callprojs->fallthrough_catchproj != nullptr) {
2062 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2063 }
2064 if (callprojs->fallthrough_memproj != nullptr) {
2065 if (final_mem->is_MergeMem()) {
2066 // Parser's exits MergeMem was not transformed but may be optimized
2067 final_mem = _gvn.transform(final_mem);
2068 }
2069 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2070 add_mergemem_users_to_worklist(wl, final_mem);
2071 }
2072 if (callprojs->fallthrough_ioproj != nullptr) {
2073 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2074 }
2075
2076 // Replace the result with the new result if it exists and is used
2077 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2078 // If the inlined code is dead, the result projections for an inline type returned as
2079 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2080 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2081 "unexpected number of results");
2082 C->gvn_replace_by(callprojs->resproj[0], result);
2083 }
2084
2085 if (ejvms == nullptr) {
2086 // No exception edges to simply kill off those paths
2087 if (callprojs->catchall_catchproj != nullptr) {
2088 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2089 }
2090 if (callprojs->catchall_memproj != nullptr) {
2091 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2092 }
2093 if (callprojs->catchall_ioproj != nullptr) {
2094 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2095 }
2096 // Replace the old exception object with top
2097 if (callprojs->exobj != nullptr) {
2098 C->gvn_replace_by(callprojs->exobj, C->top());
2099 }
2100 } else {
2101 GraphKit ekit(ejvms);
2102
2103 // Load my combined exception state into the kit, with all phis transformed:
2104 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2105 replaced_nodes_exception = ex_map->replaced_nodes();
2106
2107 Node* ex_oop = ekit.use_exception_state(ex_map);
2108
2109 if (callprojs->catchall_catchproj != nullptr) {
2110 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2111 ex_ctl = ekit.control();
2112 }
2113 if (callprojs->catchall_memproj != nullptr) {
2114 Node* ex_mem = ekit.reset_memory();
2115 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2116 add_mergemem_users_to_worklist(wl, ex_mem);
2117 }
2118 if (callprojs->catchall_ioproj != nullptr) {
2119 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2120 }
2121
2122 // Replace the old exception object with the newly created one
2123 if (callprojs->exobj != nullptr) {
2124 C->gvn_replace_by(callprojs->exobj, ex_oop);
2125 }
2126 }
2127
2128 // Disconnect the call from the graph
2129 call->disconnect_inputs(C);
2130 C->gvn_replace_by(call, C->top());
2131
2132 // Clean up any MergeMems that feed other MergeMems since the
2133 // optimizer doesn't like that.
2134 while (wl.size() > 0) {
2135 _gvn.transform(wl.pop());
2136 }
2137
2138 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2139 replaced_nodes.apply(C, final_ctl);
2140 }
2141 if (!ex_ctl->is_top() && do_replaced_nodes) {
2142 replaced_nodes_exception.apply(C, ex_ctl);
2143 }
2144 }
2145
2146
2147 //------------------------------increment_counter------------------------------
2148 // for statistics: increment a VM counter by 1
2149
2150 void GraphKit::increment_counter(address counter_addr) {
2151 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2152 increment_counter(adr1);
2153 }
2154
2155 void GraphKit::increment_counter(Node* counter_addr) {
2156 int adr_type = Compile::AliasIdxRaw;
2157 Node* ctrl = control();
2158 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, adr_type, MemNode::unordered);
2159 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2160 store_to_memory(ctrl, counter_addr, incr, T_LONG, adr_type, MemNode::unordered);
2161 }
2162
2163
2164 //------------------------------uncommon_trap----------------------------------
2165 // Bail out to the interpreter in mid-method. Implemented by calling the
2166 // uncommon_trap blob. This helper function inserts a runtime call with the
2167 // right debug info.
2168 Node* GraphKit::uncommon_trap(int trap_request,
2169 ciKlass* klass, const char* comment,
2170 bool must_throw,
2171 bool keep_exact_action) {
2172 if (failing()) stop();
2173 if (stopped()) return nullptr; // trap reachable?
2174
2175 // Note: If ProfileTraps is true, and if a deopt. actually
2176 // occurs here, the runtime will make sure an MDO exists. There is
2177 // no need to call method()->ensure_method_data() at this point.
2178
2179 // Set the stack pointer to the right value for reexecution:
2180 set_sp(reexecute_sp());
2181
2182 #ifdef ASSERT
2183 if (!must_throw) {
2184 // Make sure the stack has at least enough depth to execute
2185 // the current bytecode.
2186 int inputs, ignored_depth;
2187 if (compute_stack_effects(inputs, ignored_depth)) {
2188 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2189 Bytecodes::name(java_bc()), sp(), inputs);
2190 }
2191 }
2192 #endif
2193
2194 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2195 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2196
2197 switch (action) {
2198 case Deoptimization::Action_maybe_recompile:
2199 case Deoptimization::Action_reinterpret:
2200 // Temporary fix for 6529811 to allow virtual calls to be sure they
2201 // get the chance to go from mono->bi->mega
2202 if (!keep_exact_action &&
2203 Deoptimization::trap_request_index(trap_request) < 0 &&
2204 too_many_recompiles(reason)) {
2205 // This BCI is causing too many recompilations.
2206 if (C->log() != nullptr) {
2207 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2208 Deoptimization::trap_reason_name(reason),
2209 Deoptimization::trap_action_name(action));
2210 }
2211 action = Deoptimization::Action_none;
2212 trap_request = Deoptimization::make_trap_request(reason, action);
2213 } else {
2214 C->set_trap_can_recompile(true);
2215 }
2216 break;
2217 case Deoptimization::Action_make_not_entrant:
2218 C->set_trap_can_recompile(true);
2219 break;
2220 case Deoptimization::Action_none:
2221 case Deoptimization::Action_make_not_compilable:
2222 break;
2223 default:
2224 #ifdef ASSERT
2225 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2226 #endif
2227 break;
2228 }
2229
2230 if (TraceOptoParse) {
2231 char buf[100];
2232 tty->print_cr("Uncommon trap %s at bci:%d",
2233 Deoptimization::format_trap_request(buf, sizeof(buf),
2234 trap_request), bci());
2235 }
2236
2237 CompileLog* log = C->log();
2238 if (log != nullptr) {
2239 int kid = (klass == nullptr)? -1: log->identify(klass);
2240 log->begin_elem("uncommon_trap bci='%d'", bci());
2241 char buf[100];
2242 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2243 trap_request));
2244 if (kid >= 0) log->print(" klass='%d'", kid);
2245 if (comment != nullptr) log->print(" comment='%s'", comment);
2246 log->end_elem();
2247 }
2248
2249 // Make sure any guarding test views this path as very unlikely
2250 Node *i0 = control()->in(0);
2251 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2252 IfNode *iff = i0->as_If();
2253 float f = iff->_prob; // Get prob
2254 if (control()->Opcode() == Op_IfTrue) {
2255 if (f > PROB_UNLIKELY_MAG(4))
2256 iff->_prob = PROB_MIN;
2257 } else {
2258 if (f < PROB_LIKELY_MAG(4))
2259 iff->_prob = PROB_MAX;
2260 }
2261 }
2262
2263 // Clear out dead values from the debug info.
2264 kill_dead_locals();
2265
2266 // Now insert the uncommon trap subroutine call
2267 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2268 const TypePtr* no_memory_effects = nullptr;
2269 // Pass the index of the class to be loaded
2270 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2271 (must_throw ? RC_MUST_THROW : 0),
2272 OptoRuntime::uncommon_trap_Type(),
2273 call_addr, "uncommon_trap", no_memory_effects,
2274 intcon(trap_request));
2275 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2276 "must extract request correctly from the graph");
2277 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2278
2279 call->set_req(TypeFunc::ReturnAdr, returnadr());
2280 // The debug info is the only real input to this call.
2281
2282 // Halt-and-catch fire here. The above call should never return!
2283 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2284 PRODUCT_ONLY(COMMA /*reachable*/false));
2285 _gvn.set_type_bottom(halt);
2286 root()->add_req(halt);
2287
2288 stop_and_kill_map();
2289 return call;
2290 }
2291
2292
2293 //--------------------------just_allocated_object------------------------------
2294 // Report the object that was just allocated.
2295 // It must be the case that there are no intervening safepoints.
2296 // We use this to determine if an object is so "fresh" that
2297 // it does not require card marks.
2298 Node* GraphKit::just_allocated_object(Node* current_control) {
2299 Node* ctrl = current_control;
2300 // Object::<init> is invoked after allocation, most of invoke nodes
2301 // will be reduced, but a region node is kept in parse time, we check
2302 // the pattern and skip the region node if it degraded to a copy.
2303 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2304 ctrl->as_Region()->is_copy()) {
2305 ctrl = ctrl->as_Region()->is_copy();
2306 }
2307 if (C->recent_alloc_ctl() == ctrl) {
2308 return C->recent_alloc_obj();
2309 }
2310 return nullptr;
2311 }
2312
2313
2314 /**
2315 * Record profiling data exact_kls for Node n with the type system so
2316 * that it can propagate it (speculation)
2317 *
2318 * @param n node that the type applies to
2319 * @param exact_kls type from profiling
2320 * @param maybe_null did profiling see null?
2321 *
2322 * @return node with improved type
2323 */
2324 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2325 const Type* current_type = _gvn.type(n);
2326 assert(UseTypeSpeculation, "type speculation must be on");
2327
2328 const TypePtr* speculative = current_type->speculative();
2329
2330 // Should the klass from the profile be recorded in the speculative type?
2331 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2332 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2333 const TypeOopPtr* xtype = tklass->as_instance_type();
2334 assert(xtype->klass_is_exact(), "Should be exact");
2335 // Any reason to believe n is not null (from this profiling or a previous one)?
2336 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2337 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2338 // record the new speculative type's depth
2339 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2340 speculative = speculative->with_inline_depth(jvms()->depth());
2341 } else if (current_type->would_improve_ptr(ptr_kind)) {
2342 // Profiling report that null was never seen so we can change the
2343 // speculative type to non null ptr.
2344 if (ptr_kind == ProfileAlwaysNull) {
2345 speculative = TypePtr::NULL_PTR;
2346 } else {
2347 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2348 const TypePtr* ptr = TypePtr::NOTNULL;
2349 if (speculative != nullptr) {
2350 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2351 } else {
2352 speculative = ptr;
2353 }
2354 }
2355 }
2356
2357 if (speculative != current_type->speculative()) {
2358 // Build a type with a speculative type (what we think we know
2359 // about the type but will need a guard when we use it)
2360 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2361 // We're changing the type, we need a new CheckCast node to carry
2362 // the new type. The new type depends on the control: what
2363 // profiling tells us is only valid from here as far as we can
2364 // tell.
2365 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2366 cast = _gvn.transform(cast);
2367 replace_in_map(n, cast);
2368 n = cast;
2369 }
2370
2371 return n;
2372 }
2373
2374 /**
2375 * Record profiling data from receiver profiling at an invoke with the
2376 * type system so that it can propagate it (speculation)
2377 *
2378 * @param n receiver node
2379 *
2380 * @return node with improved type
2381 */
2382 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2383 if (!UseTypeSpeculation) {
2384 return n;
2385 }
2386 ciKlass* exact_kls = profile_has_unique_klass();
2387 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2388 if ((java_bc() == Bytecodes::_checkcast ||
2389 java_bc() == Bytecodes::_instanceof ||
2390 java_bc() == Bytecodes::_aastore) &&
2391 method()->method_data()->is_mature()) {
2392 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2393 if (data != nullptr) {
2394 if (java_bc() == Bytecodes::_aastore) {
2395 ciKlass* array_type = nullptr;
2396 ciKlass* element_type = nullptr;
2397 ProfilePtrKind element_ptr = ProfileMaybeNull;
2398 bool flat_array = true;
2399 bool null_free_array = true;
2400 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2401 exact_kls = element_type;
2402 ptr_kind = element_ptr;
2403 } else {
2404 if (!data->as_BitData()->null_seen()) {
2405 ptr_kind = ProfileNeverNull;
2406 } else {
2407 assert(data->is_ReceiverTypeData(), "bad profile data type");
2408 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2409 uint i = 0;
2410 for (; i < call->row_limit(); i++) {
2411 ciKlass* receiver = call->receiver(i);
2412 if (receiver != nullptr) {
2413 break;
2414 }
2415 }
2416 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2417 }
2418 }
2419 }
2420 }
2421 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2422 }
2423
2424 /**
2425 * Record profiling data from argument profiling at an invoke with the
2426 * type system so that it can propagate it (speculation)
2427 *
2428 * @param dest_method target method for the call
2429 * @param bc what invoke bytecode is this?
2430 */
2431 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2432 if (!UseTypeSpeculation) {
2433 return;
2434 }
2435 const TypeFunc* tf = TypeFunc::make(dest_method);
2436 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2437 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2438 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2439 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2440 if (is_reference_type(targ->basic_type())) {
2441 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2442 ciKlass* better_type = nullptr;
2443 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2444 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2445 }
2446 i++;
2447 }
2448 }
2449 }
2450
2451 /**
2452 * Record profiling data from parameter profiling at an invoke with
2453 * the type system so that it can propagate it (speculation)
2454 */
2455 void GraphKit::record_profiled_parameters_for_speculation() {
2456 if (!UseTypeSpeculation) {
2457 return;
2458 }
2459 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2460 if (_gvn.type(local(i))->isa_oopptr()) {
2461 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2462 ciKlass* better_type = nullptr;
2463 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2464 record_profile_for_speculation(local(i), better_type, ptr_kind);
2465 }
2466 j++;
2467 }
2468 }
2469 }
2470
2471 /**
2472 * Record profiling data from return value profiling at an invoke with
2473 * the type system so that it can propagate it (speculation)
2474 */
2475 void GraphKit::record_profiled_return_for_speculation() {
2476 if (!UseTypeSpeculation) {
2477 return;
2478 }
2479 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2480 ciKlass* better_type = nullptr;
2481 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2482 // If profiling reports a single type for the return value,
2483 // feed it to the type system so it can propagate it as a
2484 // speculative type
2485 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2486 }
2487 }
2488
2489 void GraphKit::round_double_arguments(ciMethod* dest_method) {
2490 if (Matcher::strict_fp_requires_explicit_rounding) {
2491 // (Note: TypeFunc::make has a cache that makes this fast.)
2492 const TypeFunc* tf = TypeFunc::make(dest_method);
2493 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2494 for (int j = 0; j < nargs; j++) {
2495 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2496 if (targ->basic_type() == T_DOUBLE) {
2497 // If any parameters are doubles, they must be rounded before
2498 // the call, dprecision_rounding does gvn.transform
2499 Node *arg = argument(j);
2500 arg = dprecision_rounding(arg);
2501 set_argument(j, arg);
2502 }
2503 }
2504 }
2505 }
2506
2507 // rounding for strict float precision conformance
2508 Node* GraphKit::precision_rounding(Node* n) {
2509 if (Matcher::strict_fp_requires_explicit_rounding) {
2510 #ifdef IA32
2511 if (UseSSE == 0) {
2512 return _gvn.transform(new RoundFloatNode(0, n));
2513 }
2514 #else
2515 Unimplemented();
2516 #endif // IA32
2517 }
2518 return n;
2519 }
2520
2521 // rounding for strict double precision conformance
2522 Node* GraphKit::dprecision_rounding(Node *n) {
2523 if (Matcher::strict_fp_requires_explicit_rounding) {
2524 #ifdef IA32
2525 if (UseSSE < 2) {
2526 return _gvn.transform(new RoundDoubleNode(0, n));
2527 }
2528 #else
2529 Unimplemented();
2530 #endif // IA32
2531 }
2532 return n;
2533 }
2534
2535 //=============================================================================
2536 // Generate a fast path/slow path idiom. Graph looks like:
2537 // [foo] indicates that 'foo' is a parameter
2538 //
2539 // [in] null
2540 // \ /
2541 // CmpP
2542 // Bool ne
2543 // If
2544 // / \
2545 // True False-<2>
2546 // / |
2547 // / cast_not_null
2548 // Load | | ^
2549 // [fast_test] | |
2550 // gvn to opt_test | |
2551 // / \ | <1>
2552 // True False |
2553 // | \\ |
2554 // [slow_call] \[fast_result]
2555 // Ctl Val \ \
2556 // | \ \
2557 // Catch <1> \ \
2558 // / \ ^ \ \
2559 // Ex No_Ex | \ \
2560 // | \ \ | \ <2> \
2561 // ... \ [slow_res] | | \ [null_result]
2562 // \ \--+--+--- | |
2563 // \ | / \ | /
2564 // --------Region Phi
2565 //
2566 //=============================================================================
2567 // Code is structured as a series of driver functions all called 'do_XXX' that
2568 // call a set of helper functions. Helper functions first, then drivers.
2569
2570 //------------------------------null_check_oop---------------------------------
2571 // Null check oop. Set null-path control into Region in slot 3.
2572 // Make a cast-not-nullness use the other not-null control. Return cast.
2573 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2574 bool never_see_null,
2575 bool safe_for_replace,
2576 bool speculative) {
2577 // Initial null check taken path
2578 (*null_control) = top();
2579 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2580
2581 // Generate uncommon_trap:
2582 if (never_see_null && (*null_control) != top()) {
2583 // If we see an unexpected null at a check-cast we record it and force a
2584 // recompile; the offending check-cast will be compiled to handle nulls.
2585 // If we see more than one offending BCI, then all checkcasts in the
2586 // method will be compiled to handle nulls.
2587 PreserveJVMState pjvms(this);
2588 set_control(*null_control);
2589 replace_in_map(value, null());
2590 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2591 uncommon_trap(reason,
2592 Deoptimization::Action_make_not_entrant);
2593 (*null_control) = top(); // null path is dead
2594 }
2595 if ((*null_control) == top() && safe_for_replace) {
2596 replace_in_map(value, cast);
2597 }
2598
2599 // Cast away null-ness on the result
2600 return cast;
2601 }
2602
2603 //------------------------------opt_iff----------------------------------------
2604 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2605 // Return slow-path control.
2606 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2607 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2608
2609 // Fast path taken; set region slot 2
2610 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2611 region->init_req(2,fast_taken); // Capture fast-control
2612
2613 // Fast path not-taken, i.e. slow path
2614 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2615 return slow_taken;
2616 }
2617
2618 //-----------------------------make_runtime_call-------------------------------
2619 Node* GraphKit::make_runtime_call(int flags,
2620 const TypeFunc* call_type, address call_addr,
2621 const char* call_name,
2622 const TypePtr* adr_type,
2623 // The following parms are all optional.
2624 // The first null ends the list.
2625 Node* parm0, Node* parm1,
2626 Node* parm2, Node* parm3,
2627 Node* parm4, Node* parm5,
2628 Node* parm6, Node* parm7) {
2629 assert(call_addr != nullptr, "must not call null targets");
2630
2631 // Slow-path call
2632 bool is_leaf = !(flags & RC_NO_LEAF);
2633 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2634 if (call_name == nullptr) {
2635 assert(!is_leaf, "must supply name for leaf");
2636 call_name = OptoRuntime::stub_name(call_addr);
2637 }
2638 CallNode* call;
2639 if (!is_leaf) {
2640 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2641 } else if (flags & RC_NO_FP) {
2642 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2643 } else if (flags & RC_VECTOR){
2644 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2645 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2646 } else {
2647 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2648 }
2649
2650 // The following is similar to set_edges_for_java_call,
2651 // except that the memory effects of the call are restricted to AliasIdxRaw.
2652
2653 // Slow path call has no side-effects, uses few values
2654 bool wide_in = !(flags & RC_NARROW_MEM);
2655 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2656
2657 Node* prev_mem = nullptr;
2658 if (wide_in) {
2659 prev_mem = set_predefined_input_for_runtime_call(call);
2660 } else {
2661 assert(!wide_out, "narrow in => narrow out");
2662 Node* narrow_mem = memory(adr_type);
2663 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2664 }
2665
2666 // Hook each parm in order. Stop looking at the first null.
2667 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2668 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2669 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2670 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2671 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2672 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2673 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2674 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2675 /* close each nested if ===> */ } } } } } } } }
2676 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2677
2678 if (!is_leaf) {
2679 // Non-leaves can block and take safepoints:
2680 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2681 }
2682 // Non-leaves can throw exceptions:
2683 if (has_io) {
2684 call->set_req(TypeFunc::I_O, i_o());
2685 }
2686
2687 if (flags & RC_UNCOMMON) {
2688 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2689 // (An "if" probability corresponds roughly to an unconditional count.
2690 // Sort of.)
2691 call->set_cnt(PROB_UNLIKELY_MAG(4));
2692 }
2693
2694 Node* c = _gvn.transform(call);
2695 assert(c == call, "cannot disappear");
2696
2697 if (wide_out) {
2698 // Slow path call has full side-effects.
2699 set_predefined_output_for_runtime_call(call);
2700 } else {
2701 // Slow path call has few side-effects, and/or sets few values.
2702 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2703 }
2704
2705 if (has_io) {
2706 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2707 }
2708 return call;
2709
2710 }
2711
2712 // i2b
2713 Node* GraphKit::sign_extend_byte(Node* in) {
2714 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2715 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2716 }
2717
2718 // i2s
2719 Node* GraphKit::sign_extend_short(Node* in) {
2720 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2721 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2722 }
2723
2724
2725 //------------------------------merge_memory-----------------------------------
2726 // Merge memory from one path into the current memory state.
2727 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2728 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2729 Node* old_slice = mms.force_memory();
2730 Node* new_slice = mms.memory2();
2731 if (old_slice != new_slice) {
2732 PhiNode* phi;
2733 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2734 if (mms.is_empty()) {
2735 // clone base memory Phi's inputs for this memory slice
2736 assert(old_slice == mms.base_memory(), "sanity");
2737 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2738 _gvn.set_type(phi, Type::MEMORY);
2739 for (uint i = 1; i < phi->req(); i++) {
2740 phi->init_req(i, old_slice->in(i));
2741 }
2742 } else {
2743 phi = old_slice->as_Phi(); // Phi was generated already
2744 }
2745 } else {
2746 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2747 _gvn.set_type(phi, Type::MEMORY);
2748 }
2749 phi->set_req(new_path, new_slice);
2750 mms.set_memory(phi);
2751 }
2752 }
2753 }
2754
2755 //------------------------------make_slow_call_ex------------------------------
2756 // Make the exception handler hookups for the slow call
2757 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2758 if (stopped()) return;
2759
2760 // Make a catch node with just two handlers: fall-through and catch-all
2761 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2762 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2763 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2764 _gvn.set_type_bottom(norm);
2765 C->record_for_igvn(norm);
2766 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2767
2768 { PreserveJVMState pjvms(this);
2769 set_control(excp);
2770 set_i_o(i_o);
2771
2772 if (excp != top()) {
2773 if (deoptimize) {
2774 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2775 uncommon_trap(Deoptimization::Reason_unhandled,
2776 Deoptimization::Action_none);
2777 } else {
2778 // Create an exception state also.
2779 // Use an exact type if the caller has a specific exception.
2780 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2781 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2782 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2783 }
2784 }
2785 }
2786
2787 // Get the no-exception control from the CatchNode.
2788 set_control(norm);
2789 }
2790
2791 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2792 Node* cmp = nullptr;
2793 switch(bt) {
2794 case T_INT: cmp = new CmpINode(in1, in2); break;
2795 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2796 default: fatal("unexpected comparison type %s", type2name(bt));
2797 }
2798 cmp = gvn.transform(cmp);
2799 Node* bol = gvn.transform(new BoolNode(cmp, test));
2800 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2801 gvn.transform(iff);
2802 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2803 return iff;
2804 }
2805
2806 //-------------------------------gen_subtype_check-----------------------------
2807 // Generate a subtyping check. Takes as input the subtype and supertype.
2808 // Returns 2 values: sets the default control() to the true path and returns
2809 // the false path. Only reads invariant memory; sets no (visible) memory.
2810 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2811 // but that's not exposed to the optimizer. This call also doesn't take in an
2812 // Object; if you wish to check an Object you need to load the Object's class
2813 // prior to coming here.
2814 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn) {
2815 Compile* C = gvn.C;
2816 if ((*ctrl)->is_top()) {
2817 return C->top();
2818 }
2819
2820 // Fast check for identical types, perhaps identical constants.
2821 // The types can even be identical non-constants, in cases
2822 // involving Array.newInstance, Object.clone, etc.
2823 if (subklass == superklass)
2824 return C->top(); // false path is dead; no test needed.
2825
2826 if (gvn.type(superklass)->singleton()) {
2827 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2828 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2829
2830 // In the common case of an exact superklass, try to fold up the
2831 // test before generating code. You may ask, why not just generate
2832 // the code and then let it fold up? The answer is that the generated
2833 // code will necessarily include null checks, which do not always
2834 // completely fold away. If they are also needless, then they turn
2835 // into a performance loss. Example:
2836 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2837 // Here, the type of 'fa' is often exact, so the store check
2838 // of fa[1]=x will fold up, without testing the nullness of x.
2839 switch (C->static_subtype_check(superk, subk)) {
2840 case Compile::SSC_always_false:
2841 {
2842 Node* always_fail = *ctrl;
2843 *ctrl = gvn.C->top();
2844 return always_fail;
2845 }
2846 case Compile::SSC_always_true:
2847 return C->top();
2848 case Compile::SSC_easy_test:
2849 {
2850 // Just do a direct pointer compare and be done.
2851 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2852 *ctrl = gvn.transform(new IfTrueNode(iff));
2853 return gvn.transform(new IfFalseNode(iff));
2854 }
2855 case Compile::SSC_full_test:
2856 break;
2857 default:
2858 ShouldNotReachHere();
2859 }
2860 }
2861
2862 // %%% Possible further optimization: Even if the superklass is not exact,
2863 // if the subklass is the unique subtype of the superklass, the check
2864 // will always succeed. We could leave a dependency behind to ensure this.
2865
2866 // First load the super-klass's check-offset
2867 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2868 Node* m = C->immutable_memory();
2869 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2870 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2871 bool might_be_cache = (gvn.find_int_con(chk_off, cacheoff_con) == cacheoff_con);
2872
2873 // Load from the sub-klass's super-class display list, or a 1-word cache of
2874 // the secondary superclass list, or a failing value with a sentinel offset
2875 // if the super-klass is an interface or exceptionally deep in the Java
2876 // hierarchy and we have to scan the secondary superclass list the hard way.
2877 // Worst-case type is a little odd: null is allowed as a result (usually
2878 // klass loads can never produce a null).
2879 Node *chk_off_X = chk_off;
2880 #ifdef _LP64
2881 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2882 #endif
2883 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2884 // For some types like interfaces the following loadKlass is from a 1-word
2885 // cache which is mutable so can't use immutable memory. Other
2886 // types load from the super-class display table which is immutable.
2887 Node *kmem = C->immutable_memory();
2888 // secondary_super_cache is not immutable but can be treated as such because:
2889 // - no ideal node writes to it in a way that could cause an
2890 // incorrect/missed optimization of the following Load.
2891 // - it's a cache so, worse case, not reading the latest value
2892 // wouldn't cause incorrect execution
2893 if (might_be_cache && mem != nullptr) {
2894 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2895 }
2896 Node *nkls = gvn.transform(LoadKlassNode::make(gvn, nullptr, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
2897
2898 // Compile speed common case: ARE a subtype and we canNOT fail
2899 if( superklass == nkls )
2900 return C->top(); // false path is dead; no test needed.
2901
2902 // See if we get an immediate positive hit. Happens roughly 83% of the
2903 // time. Test to see if the value loaded just previously from the subklass
2904 // is exactly the superklass.
2905 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2906 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
2907 *ctrl = gvn.transform(new IfFalseNode(iff1));
2908
2909 // Compile speed common case: Check for being deterministic right now. If
2910 // chk_off is a constant and not equal to cacheoff then we are NOT a
2911 // subklass. In this case we need exactly the 1 test above and we can
2912 // return those results immediately.
2913 if (!might_be_cache) {
2914 Node* not_subtype_ctrl = *ctrl;
2915 *ctrl = iftrue1; // We need exactly the 1 test above
2916 return not_subtype_ctrl;
2917 }
2918
2919 // Gather the various success & failures here
2920 RegionNode *r_ok_subtype = new RegionNode(4);
2921 gvn.record_for_igvn(r_ok_subtype);
2922 RegionNode *r_not_subtype = new RegionNode(3);
2923 gvn.record_for_igvn(r_not_subtype);
2924
2925 r_ok_subtype->init_req(1, iftrue1);
2926
2927 // Check for immediate negative hit. Happens roughly 11% of the time (which
2928 // is roughly 63% of the remaining cases). Test to see if the loaded
2929 // check-offset points into the subklass display list or the 1-element
2930 // cache. If it points to the display (and NOT the cache) and the display
2931 // missed then it's not a subtype.
2932 Node *cacheoff = gvn.intcon(cacheoff_con);
2933 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2934 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2935 *ctrl = gvn.transform(new IfFalseNode(iff2));
2936
2937 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2938 // No performance impact (too rare) but allows sharing of secondary arrays
2939 // which has some footprint reduction.
2940 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2941 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2942 *ctrl = gvn.transform(new IfFalseNode(iff3));
2943
2944 // -- Roads not taken here: --
2945 // We could also have chosen to perform the self-check at the beginning
2946 // of this code sequence, as the assembler does. This would not pay off
2947 // the same way, since the optimizer, unlike the assembler, can perform
2948 // static type analysis to fold away many successful self-checks.
2949 // Non-foldable self checks work better here in second position, because
2950 // the initial primary superclass check subsumes a self-check for most
2951 // types. An exception would be a secondary type like array-of-interface,
2952 // which does not appear in its own primary supertype display.
2953 // Finally, we could have chosen to move the self-check into the
2954 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2955 // dependent manner. But it is worthwhile to have the check here,
2956 // where it can be perhaps be optimized. The cost in code space is
2957 // small (register compare, branch).
2958
2959 // Now do a linear scan of the secondary super-klass array. Again, no real
2960 // performance impact (too rare) but it's gotta be done.
2961 // Since the code is rarely used, there is no penalty for moving it
2962 // out of line, and it can only improve I-cache density.
2963 // The decision to inline or out-of-line this final check is platform
2964 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2965 Node* psc = gvn.transform(
2966 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2967
2968 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2969 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2970 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2971
2972 // Return false path; set default control to true path.
2973 *ctrl = gvn.transform(r_ok_subtype);
2974 return gvn.transform(r_not_subtype);
2975 }
2976
2977 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2978 const Type* sub_t = _gvn.type(obj_or_subklass);
2979 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
2980 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
2981 obj_or_subklass = makecon(sub_t);
2982 }
2983 bool expand_subtype_check = C->post_loop_opts_phase() || // macro node expansion is over
2984 ExpandSubTypeCheckAtParseTime; // forced expansion
2985 if (expand_subtype_check) {
2986 MergeMemNode* mem = merged_memory();
2987 Node* ctrl = control();
2988 Node* subklass = obj_or_subklass;
2989 if (!sub_t->isa_klassptr()) {
2990 subklass = load_object_klass(obj_or_subklass);
2991 }
2992 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn);
2993 set_control(ctrl);
2994 return n;
2995 }
2996
2997 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass));
2998 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2999 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3000 set_control(_gvn.transform(new IfTrueNode(iff)));
3001 return _gvn.transform(new IfFalseNode(iff));
3002 }
3003
3004 // Profile-driven exact type check:
3005 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3006 float prob, Node* *casted_receiver) {
3007 assert(!klass->is_interface(), "no exact type check on interfaces");
3008 Node* fail = top();
3009 const Type* rec_t = _gvn.type(receiver);
3010 if (rec_t->is_inlinetypeptr()) {
3011 if (klass->equals(rec_t->inline_klass())) {
3012 (*casted_receiver) = receiver; // Always passes
3013 } else {
3014 (*casted_receiver) = top(); // Always fails
3015 fail = control();
3016 set_control(top());
3017 }
3018 return fail;
3019 }
3020 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3021 Node* recv_klass = load_object_klass(receiver);
3022 fail = type_check(recv_klass, tklass, prob);
3023
3024 if (!stopped()) {
3025 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3026 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3027 assert(recv_xtype->klass_is_exact(), "");
3028
3029 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3030 // Subsume downstream occurrences of receiver with a cast to
3031 // recv_xtype, since now we know what the type will be.
3032 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3033 Node* res = _gvn.transform(cast);
3034 if (recv_xtype->is_inlinetypeptr()) {
3035 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3036 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3037 }
3038 (*casted_receiver) = res;
3039 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3040 // (User must make the replace_in_map call.)
3041 }
3042 }
3043
3044 return fail;
3045 }
3046
3047 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3048 float prob) {
3049 Node* want_klass = makecon(tklass);
3050 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3051 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3052 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3053 set_control(_gvn.transform(new IfTrueNode (iff)));
3054 Node* fail = _gvn.transform(new IfFalseNode(iff));
3055 return fail;
3056 }
3057
3058 //------------------------------subtype_check_receiver-------------------------
3059 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3060 Node** casted_receiver) {
3061 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3062 Node* want_klass = makecon(tklass);
3063
3064 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3065
3066 // Ignore interface type information until interface types are properly tracked.
3067 if (!stopped() && !klass->is_interface()) {
3068 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3069 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3070 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3071 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3072 if (recv_type->is_inlinetypeptr()) {
3073 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3074 }
3075 (*casted_receiver) = cast;
3076 }
3077 }
3078
3079 return slow_ctl;
3080 }
3081
3082 //------------------------------seems_never_null-------------------------------
3083 // Use null_seen information if it is available from the profile.
3084 // If we see an unexpected null at a type check we record it and force a
3085 // recompile; the offending check will be recompiled to handle nulls.
3086 // If we see several offending BCIs, then all checks in the
3087 // method will be recompiled.
3088 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3089 speculating = !_gvn.type(obj)->speculative_maybe_null();
3090 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3091 if (UncommonNullCast // Cutout for this technique
3092 && obj != null() // And not the -Xcomp stupid case?
3093 && !too_many_traps(reason)
3094 ) {
3095 if (speculating) {
3096 return true;
3097 }
3098 if (data == nullptr)
3099 // Edge case: no mature data. Be optimistic here.
3100 return true;
3101 // If the profile has not seen a null, assume it won't happen.
3102 assert(java_bc() == Bytecodes::_checkcast ||
3103 java_bc() == Bytecodes::_instanceof ||
3104 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3105 if (java_bc() == Bytecodes::_aastore) {
3106 return ((ciArrayLoadStoreData*)data->as_ArrayLoadStoreData())->element()->ptr_kind() == ProfileNeverNull;
3107 }
3108 return !data->as_BitData()->null_seen();
3109 }
3110 speculating = false;
3111 return false;
3112 }
3113
3114 void GraphKit::guard_klass_being_initialized(Node* klass) {
3115 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3116 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3117 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3118 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3119 T_BYTE, MemNode::unordered);
3120 init_state = _gvn.transform(init_state);
3121
3122 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3123
3124 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3125 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3126
3127 { BuildCutout unless(this, tst, PROB_MAX);
3128 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3129 }
3130 }
3131
3132 void GraphKit::guard_init_thread(Node* klass) {
3133 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3134 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3135
3136 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3137 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3138 T_ADDRESS, MemNode::unordered);
3139 init_thread = _gvn.transform(init_thread);
3140
3141 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3142
3143 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3144 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3145
3146 { BuildCutout unless(this, tst, PROB_MAX);
3147 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3148 }
3149 }
3150
3151 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3152 if (ik->is_being_initialized()) {
3153 if (C->needs_clinit_barrier(ik, context)) {
3154 Node* klass = makecon(TypeKlassPtr::make(ik));
3155 guard_klass_being_initialized(klass);
3156 guard_init_thread(klass);
3157 insert_mem_bar(Op_MemBarCPUOrder);
3158 }
3159 } else if (ik->is_initialized()) {
3160 return; // no barrier needed
3161 } else {
3162 uncommon_trap(Deoptimization::Reason_uninitialized,
3163 Deoptimization::Action_reinterpret,
3164 nullptr);
3165 }
3166 }
3167
3168 //------------------------maybe_cast_profiled_receiver-------------------------
3169 // If the profile has seen exactly one type, narrow to exactly that type.
3170 // Subsequent type checks will always fold up.
3171 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3172 const TypeKlassPtr* require_klass,
3173 ciKlass* spec_klass,
3174 bool safe_for_replace) {
3175 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3176
3177 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3178
3179 // Make sure we haven't already deoptimized from this tactic.
3180 if (too_many_traps_or_recompiles(reason))
3181 return nullptr;
3182
3183 // (No, this isn't a call, but it's enough like a virtual call
3184 // to use the same ciMethod accessor to get the profile info...)
3185 // If we have a speculative type use it instead of profiling (which
3186 // may not help us)
3187 ciKlass* exact_kls = spec_klass;
3188 if (exact_kls == nullptr) {
3189 if (java_bc() == Bytecodes::_aastore) {
3190 ciKlass* array_type = nullptr;
3191 ciKlass* element_type = nullptr;
3192 ProfilePtrKind element_ptr = ProfileMaybeNull;
3193 bool flat_array = true;
3194 bool null_free_array = true;
3195 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3196 exact_kls = element_type;
3197 } else {
3198 exact_kls = profile_has_unique_klass();
3199 }
3200 }
3201 if (exact_kls != nullptr) {// no cast failures here
3202 if (require_klass == nullptr ||
3203 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3204 // If we narrow the type to match what the type profile sees or
3205 // the speculative type, we can then remove the rest of the
3206 // cast.
3207 // This is a win, even if the exact_kls is very specific,
3208 // because downstream operations, such as method calls,
3209 // will often benefit from the sharper type.
3210 Node* exact_obj = not_null_obj; // will get updated in place...
3211 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3212 &exact_obj);
3213 { PreserveJVMState pjvms(this);
3214 set_control(slow_ctl);
3215 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3216 }
3217 if (safe_for_replace) {
3218 replace_in_map(not_null_obj, exact_obj);
3219 }
3220 return exact_obj;
3221 }
3222 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3223 }
3224
3225 return nullptr;
3226 }
3227
3228 /**
3229 * Cast obj to type and emit guard unless we had too many traps here
3230 * already
3231 *
3232 * @param obj node being casted
3233 * @param type type to cast the node to
3234 * @param not_null true if we know node cannot be null
3235 */
3236 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3237 ciKlass* type,
3238 bool not_null) {
3239 if (stopped()) {
3240 return obj;
3241 }
3242
3243 // type is null if profiling tells us this object is always null
3244 if (type != nullptr) {
3245 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3246 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3247
3248 if (!too_many_traps_or_recompiles(null_reason) &&
3249 !too_many_traps_or_recompiles(class_reason)) {
3250 Node* not_null_obj = nullptr;
3251 // not_null is true if we know the object is not null and
3252 // there's no need for a null check
3253 if (!not_null) {
3254 Node* null_ctl = top();
3255 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3256 assert(null_ctl->is_top(), "no null control here");
3257 } else {
3258 not_null_obj = obj;
3259 }
3260
3261 Node* exact_obj = not_null_obj;
3262 ciKlass* exact_kls = type;
3263 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3264 &exact_obj);
3265 {
3266 PreserveJVMState pjvms(this);
3267 set_control(slow_ctl);
3268 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3269 }
3270 replace_in_map(not_null_obj, exact_obj);
3271 obj = exact_obj;
3272 }
3273 } else {
3274 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3275 Node* exact_obj = null_assert(obj);
3276 replace_in_map(obj, exact_obj);
3277 obj = exact_obj;
3278 }
3279 }
3280 return obj;
3281 }
3282
3283 //-------------------------------gen_instanceof--------------------------------
3284 // Generate an instance-of idiom. Used by both the instance-of bytecode
3285 // and the reflective instance-of call.
3286 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3287 kill_dead_locals(); // Benefit all the uncommon traps
3288 assert( !stopped(), "dead parse path should be checked in callers" );
3289 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3290 "must check for not-null not-dead klass in callers");
3291
3292 // Make the merge point
3293 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3294 RegionNode* region = new RegionNode(PATH_LIMIT);
3295 Node* phi = new PhiNode(region, TypeInt::BOOL);
3296 C->set_has_split_ifs(true); // Has chance for split-if optimization
3297
3298 ciProfileData* data = nullptr;
3299 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3300 data = method()->method_data()->bci_to_data(bci());
3301 }
3302 bool speculative_not_null = false;
3303 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3304 && seems_never_null(obj, data, speculative_not_null));
3305
3306 // Null check; get casted pointer; set region slot 3
3307 Node* null_ctl = top();
3308 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3309
3310 // If not_null_obj is dead, only null-path is taken
3311 if (stopped()) { // Doing instance-of on a null?
3312 set_control(null_ctl);
3313 return intcon(0);
3314 }
3315 region->init_req(_null_path, null_ctl);
3316 phi ->init_req(_null_path, intcon(0)); // Set null path value
3317 if (null_ctl == top()) {
3318 // Do this eagerly, so that pattern matches like is_diamond_phi
3319 // will work even during parsing.
3320 assert(_null_path == PATH_LIMIT-1, "delete last");
3321 region->del_req(_null_path);
3322 phi ->del_req(_null_path);
3323 }
3324
3325 // Do we know the type check always succeed?
3326 bool known_statically = false;
3327 if (_gvn.type(superklass)->singleton()) {
3328 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3329 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3330 if (subk != nullptr && subk->is_loaded()) {
3331 int static_res = C->static_subtype_check(superk, subk);
3332 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3333 }
3334 }
3335
3336 if (!known_statically) {
3337 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3338 // We may not have profiling here or it may not help us. If we
3339 // have a speculative type use it to perform an exact cast.
3340 ciKlass* spec_obj_type = obj_type->speculative_type();
3341 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3342 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3343 if (stopped()) { // Profile disagrees with this path.
3344 set_control(null_ctl); // Null is the only remaining possibility.
3345 return intcon(0);
3346 }
3347 if (cast_obj != nullptr) {
3348 not_null_obj = cast_obj;
3349 }
3350 }
3351 }
3352
3353 // Generate the subtype check
3354 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3355
3356 // Plug in the success path to the general merge in slot 1.
3357 region->init_req(_obj_path, control());
3358 phi ->init_req(_obj_path, intcon(1));
3359
3360 // Plug in the failing path to the general merge in slot 2.
3361 region->init_req(_fail_path, not_subtype_ctrl);
3362 phi ->init_req(_fail_path, intcon(0));
3363
3364 // Return final merged results
3365 set_control( _gvn.transform(region) );
3366 record_for_igvn(region);
3367
3368 // If we know the type check always succeeds then we don't use the
3369 // profiling data at this bytecode. Don't lose it, feed it to the
3370 // type system as a speculative type.
3371 if (safe_for_replace) {
3372 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3373 replace_in_map(obj, casted_obj);
3374 }
3375
3376 return _gvn.transform(phi);
3377 }
3378
3379 //-------------------------------gen_checkcast---------------------------------
3380 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3381 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3382 // uncommon-trap paths work. Adjust stack after this call.
3383 // If failure_control is supplied and not null, it is filled in with
3384 // the control edge for the cast failure. Otherwise, an appropriate
3385 // uncommon trap or exception is thrown.
3386 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) {
3387 kill_dead_locals(); // Benefit all the uncommon traps
3388 const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr()->try_improve();
3389 const TypeOopPtr *toop = tk->cast_to_exactness(false)->as_instance_type();
3390 bool safe_for_replace = (failure_control == nullptr);
3391 assert(!null_free || toop->is_inlinetypeptr(), "must be an inline type pointer");
3392
3393 // Fast cutout: Check the case that the cast is vacuously true.
3394 // This detects the common cases where the test will short-circuit
3395 // away completely. We do this before we perform the null check,
3396 // because if the test is going to turn into zero code, we don't
3397 // want a residual null check left around. (Causes a slowdown,
3398 // for example, in some objArray manipulations, such as a[i]=a[j].)
3399 if (tk->singleton()) {
3400 const TypeKlassPtr* kptr = nullptr;
3401 const Type* t = _gvn.type(obj);
3402 if (t->isa_oop_ptr()) {
3403 kptr = t->is_oopptr()->as_klass_type();
3404 } else if (obj->is_InlineType()) {
3405 ciInlineKlass* vk = t->inline_klass();
3406 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3407 }
3408 if (kptr != nullptr) {
3409 switch (C->static_subtype_check(tk, kptr)) {
3410 case Compile::SSC_always_true:
3411 // If we know the type check always succeed then we don't use
3412 // the profiling data at this bytecode. Don't lose it, feed it
3413 // to the type system as a speculative type.
3414 obj = record_profiled_receiver_for_speculation(obj);
3415 if (null_free) {
3416 assert(safe_for_replace, "must be");
3417 obj = null_check(obj);
3418 }
3419 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3420 return obj;
3421 case Compile::SSC_always_false:
3422 if (null_free) {
3423 assert(safe_for_replace, "must be");
3424 obj = null_check(obj);
3425 }
3426 // It needs a null check because a null will *pass* the cast check.
3427 if (t->isa_oopptr() != nullptr && !t->is_oopptr()->maybe_null()) {
3428 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3429 Deoptimization::DeoptReason reason = is_aastore ?
3430 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3431 builtin_throw(reason);
3432 return top();
3433 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3434 return null_assert(obj);
3435 }
3436 break; // Fall through to full check
3437 default:
3438 break;
3439 }
3440 }
3441 }
3442
3443 ciProfileData* data = nullptr;
3444 if (failure_control == nullptr) { // use MDO in regular case only
3445 assert(java_bc() == Bytecodes::_aastore ||
3446 java_bc() == Bytecodes::_checkcast,
3447 "interpreter profiles type checks only for these BCs");
3448 if (method()->method_data()->is_mature()) {
3449 data = method()->method_data()->bci_to_data(bci());
3450 }
3451 }
3452
3453 // Make the merge point
3454 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3455 RegionNode* region = new RegionNode(PATH_LIMIT);
3456 Node* phi = new PhiNode(region, toop);
3457 _gvn.set_type(region, Type::CONTROL);
3458 _gvn.set_type(phi, toop);
3459
3460 C->set_has_split_ifs(true); // Has chance for split-if optimization
3461
3462 // Use null-cast information if it is available
3463 bool speculative_not_null = false;
3464 bool never_see_null = ((failure_control == nullptr) // regular case only
3465 && seems_never_null(obj, data, speculative_not_null));
3466
3467 if (obj->is_InlineType()) {
3468 // Re-execute if buffering during triggers deoptimization
3469 PreserveReexecuteState preexecs(this);
3470 jvms()->set_should_reexecute(true);
3471 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3472 }
3473
3474 // Null check; get casted pointer; set region slot 3
3475 Node* null_ctl = top();
3476 Node* not_null_obj = nullptr;
3477 if (null_free) {
3478 assert(safe_for_replace, "must be");
3479 not_null_obj = null_check(obj);
3480 } else {
3481 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3482 }
3483
3484 // If not_null_obj is dead, only null-path is taken
3485 if (stopped()) { // Doing instance-of on a null?
3486 set_control(null_ctl);
3487 if (toop->is_inlinetypeptr()) {
3488 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3489 }
3490 return null();
3491 }
3492 region->init_req(_null_path, null_ctl);
3493 phi ->init_req(_null_path, null()); // Set null path value
3494 if (null_ctl == top()) {
3495 // Do this eagerly, so that pattern matches like is_diamond_phi
3496 // will work even during parsing.
3497 assert(_null_path == PATH_LIMIT-1, "delete last");
3498 region->del_req(_null_path);
3499 phi ->del_req(_null_path);
3500 }
3501
3502 Node* cast_obj = nullptr;
3503 if (tk->klass_is_exact()) {
3504 // The following optimization tries to statically cast the speculative type of the object
3505 // (for example obtained during profiling) to the type of the superklass and then do a
3506 // dynamic check that the type of the object is what we expect. To work correctly
3507 // for checkcast and aastore the type of superklass should be exact.
3508 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3509 // We may not have profiling here or it may not help us. If we have
3510 // a speculative type use it to perform an exact cast.
3511 ciKlass* spec_obj_type = obj_type->speculative_type();
3512 if (spec_obj_type != nullptr || data != nullptr) {
3513 cast_obj = maybe_cast_profiled_receiver(not_null_obj, tk, spec_obj_type, safe_for_replace);
3514 if (cast_obj != nullptr) {
3515 if (failure_control != nullptr) // failure is now impossible
3516 (*failure_control) = top();
3517 // adjust the type of the phi to the exact klass:
3518 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3519 }
3520 }
3521 }
3522
3523 if (cast_obj == nullptr) {
3524 // Generate the subtype check
3525 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3526
3527 // Plug in success path into the merge
3528 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3529 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3530 if (failure_control == nullptr) {
3531 if (not_subtype_ctrl != top()) { // If failure is possible
3532 PreserveJVMState pjvms(this);
3533 set_control(not_subtype_ctrl);
3534 Node* obj_klass = nullptr;
3535 if (not_null_obj->is_InlineType()) {
3536 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3537 } else {
3538 obj_klass = load_object_klass(not_null_obj);
3539 }
3540 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3541 Deoptimization::DeoptReason reason = is_aastore ?
3542 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3543 builtin_throw(reason);
3544 }
3545 } else {
3546 (*failure_control) = not_subtype_ctrl;
3547 }
3548 }
3549
3550 region->init_req(_obj_path, control());
3551 phi ->init_req(_obj_path, cast_obj);
3552
3553 // A merge of null or Casted-NotNull obj
3554 Node* res = _gvn.transform(phi);
3555
3556 // Note I do NOT always 'replace_in_map(obj,result)' here.
3557 // if( tk->klass()->can_be_primary_super() )
3558 // This means that if I successfully store an Object into an array-of-String
3559 // I 'forget' that the Object is really now known to be a String. I have to
3560 // do this because we don't have true union types for interfaces - if I store
3561 // a Baz into an array-of-Interface and then tell the optimizer it's an
3562 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3563 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3564 // replace_in_map( obj, res );
3565
3566 // Return final merged results
3567 set_control( _gvn.transform(region) );
3568 record_for_igvn(region);
3569
3570 bool not_inline = !toop->can_be_inline_type();
3571 bool not_flat = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flat_array());
3572 if (EnableValhalla && not_flat) {
3573 // Check if obj has been loaded from an array
3574 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3575 Node* array = nullptr;
3576 if (obj->isa_Load()) {
3577 Node* address = obj->in(MemNode::Address);
3578 if (address->isa_AddP()) {
3579 array = address->as_AddP()->in(AddPNode::Base);
3580 }
3581 } else if (obj->is_Phi()) {
3582 Node* region = obj->in(0);
3583 // TODO make this more robust (see JDK-8231346)
3584 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3585 IfNode* iff = region->in(2)->in(0)->isa_If();
3586 if (iff != nullptr) {
3587 iff->is_flat_array_check(&_gvn, &array);
3588 }
3589 }
3590 }
3591 if (array != nullptr) {
3592 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3593 if (ary_t != nullptr) {
3594 if (!ary_t->is_not_null_free() && not_inline) {
3595 // Casting array element to a non-inline-type, mark array as not null-free.
3596 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3597 replace_in_map(array, cast);
3598 } else if (!ary_t->is_not_flat()) {
3599 // Casting array element to a non-flat type, mark array as not flat.
3600 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3601 replace_in_map(array, cast);
3602 }
3603 }
3604 }
3605 }
3606
3607 if (!stopped() && !res->is_InlineType()) {
3608 res = record_profiled_receiver_for_speculation(res);
3609 if (toop->is_inlinetypeptr()) {
3610 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null());
3611 res = vt;
3612 if (safe_for_replace) {
3613 replace_in_map(obj, vt);
3614 replace_in_map(not_null_obj, vt);
3615 replace_in_map(res, vt);
3616 }
3617 }
3618 }
3619 return res;
3620 }
3621
3622 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3623 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3624 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3625 Node* mask = MakeConX(markWord::inline_type_pattern);
3626 Node* masked = _gvn.transform(new AndXNode(mark, mask));
3627 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3628 return _gvn.transform(new BoolNode(cmp, is_inline ? BoolTest::eq : BoolTest::ne));
3629 }
3630
3631 Node* GraphKit::is_val_mirror(Node* mirror) {
3632 Node* p = basic_plus_adr(mirror, java_lang_Class::secondary_mirror_offset());
3633 Node* secondary_mirror = access_load_at(mirror, p, _gvn.type(p)->is_ptr(), TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR), T_OBJECT, IN_HEAP);
3634 Node* cmp = _gvn.transform(new CmpPNode(mirror, secondary_mirror));
3635 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3636 }
3637
3638 Node* GraphKit::array_lh_test(Node* klass, jint mask, jint val, bool eq) {
3639 Node* lh_adr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
3640 // Make sure to use immutable memory here to enable hoisting the check out of loops
3641 Node* lh_val = _gvn.transform(LoadNode::make(_gvn, nullptr, immutable_memory(), lh_adr, lh_adr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
3642 Node* masked = _gvn.transform(new AndINode(lh_val, intcon(mask)));
3643 Node* cmp = _gvn.transform(new CmpINode(masked, intcon(val)));
3644 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3645 }
3646
3647 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3648 // We can't use immutable memory here because the mark word is mutable.
3649 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3650 // check is moved out of loops (mainly to enable loop unswitching).
3651 Node* mem = UseArrayMarkWordCheck ? memory(Compile::AliasIdxRaw) : immutable_memory();
3652 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, mem, array_or_klass));
3653 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3654 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3655 }
3656
3657 Node* GraphKit::null_free_array_test(Node* klass, bool null_free) {
3658 return array_lh_test(klass, Klass::_lh_null_free_array_bit_inplace, 0, !null_free);
3659 }
3660
3661 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3662 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3663 RegionNode* region = new RegionNode(3);
3664 Node* null_ctl = top();
3665 null_check_oop(val, &null_ctl);
3666 if (null_ctl != top()) {
3667 PreserveJVMState pjvms(this);
3668 set_control(null_ctl);
3669 {
3670 // Deoptimize if null-free array
3671 BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX);
3672 inc_sp(nargs);
3673 uncommon_trap(Deoptimization::Reason_null_check,
3674 Deoptimization::Action_none);
3675 }
3676 region->init_req(1, control());
3677 }
3678 region->init_req(2, control());
3679 set_control(_gvn.transform(region));
3680 record_for_igvn(region);
3681 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3682 // Since we were just successfully storing null, the array can't be null free.
3683 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3684 ary_t = ary_t->cast_to_not_null_free();
3685 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3686 if (safe_for_replace) {
3687 replace_in_map(ary, cast);
3688 }
3689 ary = cast;
3690 }
3691 return ary;
3692 }
3693
3694 //------------------------------next_monitor-----------------------------------
3695 // What number should be given to the next monitor?
3696 int GraphKit::next_monitor() {
3697 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3698 int next = current + C->sync_stack_slots();
3699 // Keep the toplevel high water mark current:
3700 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3701 return current;
3702 }
3703
3704 //------------------------------insert_mem_bar---------------------------------
3705 // Memory barrier to avoid floating things around
3706 // The membar serves as a pinch point between both control and all memory slices.
3707 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3708 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3709 mb->init_req(TypeFunc::Control, control());
3710 mb->init_req(TypeFunc::Memory, reset_memory());
3711 Node* membar = _gvn.transform(mb);
3712 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3713 set_all_memory_call(membar);
3714 return membar;
3715 }
3716
3717 //-------------------------insert_mem_bar_volatile----------------------------
3718 // Memory barrier to avoid floating things around
3719 // The membar serves as a pinch point between both control and memory(alias_idx).
3720 // If you want to make a pinch point on all memory slices, do not use this
3721 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3722 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3723 // When Parse::do_put_xxx updates a volatile field, it appends a series
3724 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3725 // The first membar is on the same memory slice as the field store opcode.
3726 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3727 // All the other membars (for other volatile slices, including AliasIdxBot,
3728 // which stands for all unknown volatile slices) are control-dependent
3729 // on the first membar. This prevents later volatile loads or stores
3730 // from sliding up past the just-emitted store.
3731
3732 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3733 mb->set_req(TypeFunc::Control,control());
3734 if (alias_idx == Compile::AliasIdxBot) {
3735 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3736 } else {
3737 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3738 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3739 }
3740 Node* membar = _gvn.transform(mb);
3741 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3742 if (alias_idx == Compile::AliasIdxBot) {
3743 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3744 } else {
3745 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3746 }
3747 return membar;
3748 }
3749
3750 //------------------------------shared_lock------------------------------------
3751 // Emit locking code.
3752 FastLockNode* GraphKit::shared_lock(Node* obj) {
3753 // bci is either a monitorenter bc or InvocationEntryBci
3754 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3755 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3756
3757 if( !GenerateSynchronizationCode )
3758 return nullptr; // Not locking things?
3759
3760 if (stopped()) // Dead monitor?
3761 return nullptr;
3762
3763 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3764
3765 // Box the stack location
3766 Node* box = _gvn.transform(new BoxLockNode(next_monitor()));
3767 Node* mem = reset_memory();
3768
3769 FastLockNode * flock = _gvn.transform(new FastLockNode(0, obj, box) )->as_FastLock();
3770
3771 // Create the rtm counters for this fast lock if needed.
3772 flock->create_rtm_lock_counter(sync_jvms()); // sync_jvms used to get current bci
3773
3774 // Add monitor to debug info for the slow path. If we block inside the
3775 // slow path and de-opt, we need the monitor hanging around
3776 map()->push_monitor( flock );
3777
3778 const TypeFunc *tf = LockNode::lock_type();
3779 LockNode *lock = new LockNode(C, tf);
3780
3781 lock->init_req( TypeFunc::Control, control() );
3782 lock->init_req( TypeFunc::Memory , mem );
3783 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3784 lock->init_req( TypeFunc::FramePtr, frameptr() );
3785 lock->init_req( TypeFunc::ReturnAdr, top() );
3786
3787 lock->init_req(TypeFunc::Parms + 0, obj);
3788 lock->init_req(TypeFunc::Parms + 1, box);
3789 lock->init_req(TypeFunc::Parms + 2, flock);
3790 add_safepoint_edges(lock);
3791
3792 lock = _gvn.transform( lock )->as_Lock();
3793
3794 // lock has no side-effects, sets few values
3795 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3796
3797 insert_mem_bar(Op_MemBarAcquireLock);
3798
3799 // Add this to the worklist so that the lock can be eliminated
3800 record_for_igvn(lock);
3801
3802 #ifndef PRODUCT
3803 if (PrintLockStatistics) {
3804 // Update the counter for this lock. Don't bother using an atomic
3805 // operation since we don't require absolute accuracy.
3806 lock->create_lock_counter(map()->jvms());
3807 increment_counter(lock->counter()->addr());
3808 }
3809 #endif
3810
3811 return flock;
3812 }
3813
3814
3815 //------------------------------shared_unlock----------------------------------
3816 // Emit unlocking code.
3817 void GraphKit::shared_unlock(Node* box, Node* obj) {
3818 // bci is either a monitorenter bc or InvocationEntryBci
3819 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3820 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3821
3822 if( !GenerateSynchronizationCode )
3823 return;
3824 if (stopped()) { // Dead monitor?
3825 map()->pop_monitor(); // Kill monitor from debug info
3826 return;
3827 }
3828 assert(!obj->is_InlineType(), "should not unlock on inline type");
3829
3830 // Memory barrier to avoid floating things down past the locked region
3831 insert_mem_bar(Op_MemBarReleaseLock);
3832
3833 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3834 UnlockNode *unlock = new UnlockNode(C, tf);
3835 #ifdef ASSERT
3836 unlock->set_dbg_jvms(sync_jvms());
3837 #endif
3838 uint raw_idx = Compile::AliasIdxRaw;
3839 unlock->init_req( TypeFunc::Control, control() );
3840 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3841 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3842 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3843 unlock->init_req( TypeFunc::ReturnAdr, top() );
3844
3845 unlock->init_req(TypeFunc::Parms + 0, obj);
3846 unlock->init_req(TypeFunc::Parms + 1, box);
3847 unlock = _gvn.transform(unlock)->as_Unlock();
3848
3849 Node* mem = reset_memory();
3850
3851 // unlock has no side-effects, sets few values
3852 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3853
3854 // Kill monitor from debug info
3855 map()->pop_monitor( );
3856 }
3857
3858 //-------------------------------get_layout_helper-----------------------------
3859 // If the given klass is a constant or known to be an array,
3860 // fetch the constant layout helper value into constant_value
3861 // and return null. Otherwise, load the non-constant
3862 // layout helper value, and return the node which represents it.
3863 // This two-faced routine is useful because allocation sites
3864 // almost always feature constant types.
3865 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3866 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3867 if (!StressReflectiveCode && klass_t != nullptr) {
3868 bool xklass = klass_t->klass_is_exact();
3869 bool can_be_flat = false;
3870 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
3871 if (UseFlatArray && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
3872 // The runtime type of [LMyValue might be [QMyValue due to [QMyValue <: [LMyValue. Don't constant fold.
3873 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
3874 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->flat_array());
3875 }
3876 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
3877 jint lhelper;
3878 if (klass_t->is_flat()) {
3879 lhelper = ary_type->flat_layout_helper();
3880 } else if (klass_t->isa_aryklassptr()) {
3881 BasicType elem = ary_type->elem()->array_element_basic_type();
3882 if (is_reference_type(elem, true)) {
3883 elem = T_OBJECT;
3884 }
3885 lhelper = Klass::array_layout_helper(elem);
3886 } else {
3887 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3888 }
3889 if (lhelper != Klass::_lh_neutral_value) {
3890 constant_value = lhelper;
3891 return (Node*) nullptr;
3892 }
3893 }
3894 }
3895 constant_value = Klass::_lh_neutral_value; // put in a known value
3896 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
3897 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3898 }
3899
3900 // We just put in an allocate/initialize with a big raw-memory effect.
3901 // Hook selected additional alias categories on the initialization.
3902 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3903 MergeMemNode* init_in_merge,
3904 Node* init_out_raw) {
3905 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3906 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3907
3908 Node* prevmem = kit.memory(alias_idx);
3909 init_in_merge->set_memory_at(alias_idx, prevmem);
3910 if (init_out_raw != nullptr) {
3911 kit.set_memory(init_out_raw, alias_idx);
3912 }
3913 }
3914
3915 //---------------------------set_output_for_allocation-------------------------
3916 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3917 const TypeOopPtr* oop_type,
3918 bool deoptimize_on_exception) {
3919 int rawidx = Compile::AliasIdxRaw;
3920 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3921 add_safepoint_edges(alloc);
3922 Node* allocx = _gvn.transform(alloc);
3923 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3924 // create memory projection for i_o
3925 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3926 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3927
3928 // create a memory projection as for the normal control path
3929 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3930 set_memory(malloc, rawidx);
3931
3932 // a normal slow-call doesn't change i_o, but an allocation does
3933 // we create a separate i_o projection for the normal control path
3934 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3935 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3936
3937 // put in an initialization barrier
3938 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3939 rawoop)->as_Initialize();
3940 assert(alloc->initialization() == init, "2-way macro link must work");
3941 assert(init ->allocation() == alloc, "2-way macro link must work");
3942 {
3943 // Extract memory strands which may participate in the new object's
3944 // initialization, and source them from the new InitializeNode.
3945 // This will allow us to observe initializations when they occur,
3946 // and link them properly (as a group) to the InitializeNode.
3947 assert(init->in(InitializeNode::Memory) == malloc, "");
3948 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3949 init->set_req(InitializeNode::Memory, minit_in);
3950 record_for_igvn(minit_in); // fold it up later, if possible
3951 _gvn.set_type(minit_in, Type::MEMORY);
3952 Node* minit_out = memory(rawidx);
3953 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3954 // Add an edge in the MergeMem for the header fields so an access
3955 // to one of those has correct memory state
3956 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
3957 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
3958 if (oop_type->isa_aryptr()) {
3959 const TypeAryPtr* arytype = oop_type->is_aryptr();
3960 if (arytype->is_flat()) {
3961 // Initially all flat array accesses share a single slice
3962 // but that changes after parsing. Prepare the memory graph so
3963 // it can optimize flat array accesses properly once they
3964 // don't share a single slice.
3965 assert(C->flat_accesses_share_alias(), "should be set at parse time");
3966 C->set_flat_accesses_share_alias(false);
3967 ciInlineKlass* vk = arytype->elem()->inline_klass();
3968 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
3969 ciField* field = vk->nonstatic_field_at(i);
3970 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3971 continue; // do not bother to track really large numbers of fields
3972 int off_in_vt = field->offset_in_bytes() - vk->first_field_offset();
3973 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
3974 int fieldidx = C->get_alias_index(adr_type, true);
3975 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node
3976 // can result in per flat array field Phis to be created which confuses the logic of
3977 // Compile::adjust_flat_array_access_aliases().
3978 hook_memory_on_init(*this, fieldidx, minit_in, nullptr);
3979 }
3980 C->set_flat_accesses_share_alias(true);
3981 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
3982 } else {
3983 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3984 int elemidx = C->get_alias_index(telemref);
3985 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
3986 }
3987 } else if (oop_type->isa_instptr()) {
3988 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
3989 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3990 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3991 ciField* field = ik->nonstatic_field_at(i);
3992 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3993 continue; // do not bother to track really large numbers of fields
3994 // Find (or create) the alias category for this field:
3995 int fieldidx = C->alias_type(field)->index();
3996 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
3997 }
3998 }
3999 }
4000
4001 // Cast raw oop to the real thing...
4002 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4003 javaoop = _gvn.transform(javaoop);
4004 C->set_recent_alloc(control(), javaoop);
4005 assert(just_allocated_object(control()) == javaoop, "just allocated");
4006
4007 #ifdef ASSERT
4008 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4009 assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
4010 "Ideal_allocation works");
4011 assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
4012 "Ideal_allocation works");
4013 if (alloc->is_AllocateArray()) {
4014 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
4015 "Ideal_allocation works");
4016 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
4017 "Ideal_allocation works");
4018 } else {
4019 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4020 }
4021 }
4022 #endif //ASSERT
4023
4024 return javaoop;
4025 }
4026
4027 //---------------------------new_instance--------------------------------------
4028 // This routine takes a klass_node which may be constant (for a static type)
4029 // or may be non-constant (for reflective code). It will work equally well
4030 // for either, and the graph will fold nicely if the optimizer later reduces
4031 // the type to a constant.
4032 // The optional arguments are for specialized use by intrinsics:
4033 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4034 // - If 'return_size_val', report the total object size to the caller.
4035 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4036 Node* GraphKit::new_instance(Node* klass_node,
4037 Node* extra_slow_test,
4038 Node* *return_size_val,
4039 bool deoptimize_on_exception,
4040 InlineTypeNode* inline_type_node) {
4041 // Compute size in doublewords
4042 // The size is always an integral number of doublewords, represented
4043 // as a positive bytewise size stored in the klass's layout_helper.
4044 // The layout_helper also encodes (in a low bit) the need for a slow path.
4045 jint layout_con = Klass::_lh_neutral_value;
4046 Node* layout_val = get_layout_helper(klass_node, layout_con);
4047 bool layout_is_con = (layout_val == nullptr);
4048
4049 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4050 // Generate the initial go-slow test. It's either ALWAYS (return a
4051 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4052 // case) a computed value derived from the layout_helper.
4053 Node* initial_slow_test = nullptr;
4054 if (layout_is_con) {
4055 assert(!StressReflectiveCode, "stress mode does not use these paths");
4056 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4057 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4058 } else { // reflective case
4059 // This reflective path is used by Unsafe.allocateInstance.
4060 // (It may be stress-tested by specifying StressReflectiveCode.)
4061 // Basically, we want to get into the VM is there's an illegal argument.
4062 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4063 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4064 if (extra_slow_test != intcon(0)) {
4065 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4066 }
4067 // (Macro-expander will further convert this to a Bool, if necessary.)
4068 }
4069
4070 // Find the size in bytes. This is easy; it's the layout_helper.
4071 // The size value must be valid even if the slow path is taken.
4072 Node* size = nullptr;
4073 if (layout_is_con) {
4074 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4075 } else { // reflective case
4076 // This reflective path is used by clone and Unsafe.allocateInstance.
4077 size = ConvI2X(layout_val);
4078
4079 // Clear the low bits to extract layout_helper_size_in_bytes:
4080 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4081 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4082 size = _gvn.transform( new AndXNode(size, mask) );
4083 }
4084 if (return_size_val != nullptr) {
4085 (*return_size_val) = size;
4086 }
4087
4088 // This is a precise notnull oop of the klass.
4089 // (Actually, it need not be precise if this is a reflective allocation.)
4090 // It's what we cast the result to.
4091 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4092 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4093 const TypeOopPtr* oop_type = tklass->as_instance_type();
4094
4095 // Now generate allocation code
4096
4097 // The entire memory state is needed for slow path of the allocation
4098 // since GC and deoptimization can happen.
4099 Node *mem = reset_memory();
4100 set_all_memory(mem); // Create new memory state
4101
4102 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4103 control(), mem, i_o(),
4104 size, klass_node,
4105 initial_slow_test, inline_type_node);
4106
4107 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4108 }
4109
4110 //-------------------------------new_array-------------------------------------
4111 // helper for newarray and anewarray
4112 // The 'length' parameter is (obviously) the length of the array.
4113 // See comments on new_instance for the meaning of the other arguments.
4114 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4115 Node* length, // number of array elements
4116 int nargs, // number of arguments to push back for uncommon trap
4117 Node* *return_size_val,
4118 bool deoptimize_on_exception) {
4119 jint layout_con = Klass::_lh_neutral_value;
4120 Node* layout_val = get_layout_helper(klass_node, layout_con);
4121 bool layout_is_con = (layout_val == nullptr);
4122
4123 if (!layout_is_con && !StressReflectiveCode &&
4124 !too_many_traps(Deoptimization::Reason_class_check)) {
4125 // This is a reflective array creation site.
4126 // Optimistically assume that it is a subtype of Object[],
4127 // so that we can fold up all the address arithmetic.
4128 layout_con = Klass::array_layout_helper(T_OBJECT);
4129 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4130 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4131 { BuildCutout unless(this, bol_lh, PROB_MAX);
4132 inc_sp(nargs);
4133 uncommon_trap(Deoptimization::Reason_class_check,
4134 Deoptimization::Action_maybe_recompile);
4135 }
4136 layout_val = nullptr;
4137 layout_is_con = true;
4138 }
4139
4140 // Generate the initial go-slow test. Make sure we do not overflow
4141 // if length is huge (near 2Gig) or negative! We do not need
4142 // exact double-words here, just a close approximation of needed
4143 // double-words. We can't add any offset or rounding bits, lest we
4144 // take a size -1 of bytes and make it positive. Use an unsigned
4145 // compare, so negative sizes look hugely positive.
4146 int fast_size_limit = FastAllocateSizeLimit;
4147 if (layout_is_con) {
4148 assert(!StressReflectiveCode, "stress mode does not use these paths");
4149 // Increase the size limit if we have exact knowledge of array type.
4150 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4151 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4152 }
4153
4154 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4155 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4156
4157 // --- Size Computation ---
4158 // array_size = round_to_heap(array_header + (length << elem_shift));
4159 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4160 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4161 // The rounding mask is strength-reduced, if possible.
4162 int round_mask = MinObjAlignmentInBytes - 1;
4163 Node* header_size = nullptr;
4164 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4165 // (T_BYTE has the weakest alignment and size restrictions...)
4166 if (layout_is_con) {
4167 int hsize = Klass::layout_helper_header_size(layout_con);
4168 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4169 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4170 if ((round_mask & ~right_n_bits(eshift)) == 0)
4171 round_mask = 0; // strength-reduce it if it goes away completely
4172 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4173 assert(header_size_min <= hsize, "generic minimum is smallest");
4174 header_size_min = hsize;
4175 header_size = intcon(hsize + round_mask);
4176 } else {
4177 Node* hss = intcon(Klass::_lh_header_size_shift);
4178 Node* hsm = intcon(Klass::_lh_header_size_mask);
4179 Node* hsize = _gvn.transform( new URShiftINode(layout_val, hss) );
4180 hsize = _gvn.transform( new AndINode(hsize, hsm) );
4181 Node* mask = intcon(round_mask);
4182 header_size = _gvn.transform( new AddINode(hsize, mask) );
4183 }
4184
4185 Node* elem_shift = nullptr;
4186 if (layout_is_con) {
4187 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4188 if (eshift != 0)
4189 elem_shift = intcon(eshift);
4190 } else {
4191 // There is no need to mask or shift this value.
4192 // The semantics of LShiftINode include an implicit mask to 0x1F.
4193 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4194 elem_shift = layout_val;
4195 }
4196
4197 // Transition to native address size for all offset calculations:
4198 Node* lengthx = ConvI2X(length);
4199 Node* headerx = ConvI2X(header_size);
4200 #ifdef _LP64
4201 { const TypeInt* tilen = _gvn.find_int_type(length);
4202 if (tilen != nullptr && tilen->_lo < 0) {
4203 // Add a manual constraint to a positive range. Cf. array_element_address.
4204 jint size_max = fast_size_limit;
4205 if (size_max > tilen->_hi) size_max = tilen->_hi;
4206 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4207
4208 // Only do a narrow I2L conversion if the range check passed.
4209 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4210 _gvn.transform(iff);
4211 RegionNode* region = new RegionNode(3);
4212 _gvn.set_type(region, Type::CONTROL);
4213 lengthx = new PhiNode(region, TypeLong::LONG);
4214 _gvn.set_type(lengthx, TypeLong::LONG);
4215
4216 // Range check passed. Use ConvI2L node with narrow type.
4217 Node* passed = IfFalse(iff);
4218 region->init_req(1, passed);
4219 // Make I2L conversion control dependent to prevent it from
4220 // floating above the range check during loop optimizations.
4221 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4222
4223 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4224 region->init_req(2, IfTrue(iff));
4225 lengthx->init_req(2, ConvI2X(length));
4226
4227 set_control(region);
4228 record_for_igvn(region);
4229 record_for_igvn(lengthx);
4230 }
4231 }
4232 #endif
4233
4234 // Combine header size (plus rounding) and body size. Then round down.
4235 // This computation cannot overflow, because it is used only in two
4236 // places, one where the length is sharply limited, and the other
4237 // after a successful allocation.
4238 Node* abody = lengthx;
4239 if (elem_shift != nullptr)
4240 abody = _gvn.transform( new LShiftXNode(lengthx, elem_shift) );
4241 Node* size = _gvn.transform( new AddXNode(headerx, abody) );
4242 if (round_mask != 0) {
4243 Node* mask = MakeConX(~round_mask);
4244 size = _gvn.transform( new AndXNode(size, mask) );
4245 }
4246 // else if round_mask == 0, the size computation is self-rounding
4247
4248 if (return_size_val != nullptr) {
4249 // This is the size
4250 (*return_size_val) = size;
4251 }
4252
4253 // Now generate allocation code
4254
4255 // The entire memory state is needed for slow path of the allocation
4256 // since GC and deoptimization can happen.
4257 Node *mem = reset_memory();
4258 set_all_memory(mem); // Create new memory state
4259
4260 if (initial_slow_test->is_Bool()) {
4261 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4262 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4263 }
4264
4265 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4266 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4267 const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
4268
4269 // Inline type array variants:
4270 // - null-ok: MyValue.ref[] (ciObjArrayKlass "[LMyValue")
4271 // - null-free: MyValue.val[] (ciObjArrayKlass "[QMyValue")
4272 // - null-free, flat : MyValue.val[] (ciFlatArrayKlass "[QMyValue")
4273 // Check if array is a null-free, non-flat inline type array
4274 // that needs to be initialized with the default inline type.
4275 Node* default_value = nullptr;
4276 Node* raw_default_value = nullptr;
4277 if (ary_ptr != nullptr && ary_ptr->klass_is_exact()) {
4278 // Array type is known
4279 if (ary_ptr->is_null_free() && !ary_ptr->is_flat()) {
4280 ciInlineKlass* vk = ary_ptr->elem()->inline_klass();
4281 default_value = InlineTypeNode::default_oop(gvn(), vk);
4282 }
4283 } else if (ary_type->can_be_inline_array()) {
4284 // Array type is not known, add runtime checks
4285 assert(!ary_klass->klass_is_exact(), "unexpected exact type");
4286 Node* r = new RegionNode(3);
4287 default_value = new PhiNode(r, TypeInstPtr::BOTTOM);
4288
4289 Node* bol = array_lh_test(klass_node, Klass::_lh_array_tag_flat_value_bit_inplace | Klass::_lh_null_free_array_bit_inplace, Klass::_lh_null_free_array_bit_inplace);
4290 IfNode* iff = create_and_map_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4291
4292 // Null-free, non-flat inline type array, initialize with the default value
4293 set_control(_gvn.transform(new IfTrueNode(iff)));
4294 Node* p = basic_plus_adr(klass_node, in_bytes(ArrayKlass::element_klass_offset()));
4295 Node* eklass = _gvn.transform(LoadKlassNode::make(_gvn, control(), immutable_memory(), p, TypeInstPtr::KLASS));
4296 Node* adr_fixed_block_addr = basic_plus_adr(eklass, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()));
4297 Node* adr_fixed_block = make_load(control(), adr_fixed_block_addr, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4298 Node* default_value_offset_addr = basic_plus_adr(adr_fixed_block, in_bytes(InlineKlass::default_value_offset_offset()));
4299 Node* default_value_offset = make_load(control(), default_value_offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
4300 Node* elem_mirror = load_mirror_from_klass(eklass);
4301 Node* default_value_addr = basic_plus_adr(elem_mirror, ConvI2X(default_value_offset));
4302 Node* val = access_load_at(elem_mirror, default_value_addr, TypeInstPtr::MIRROR, TypeInstPtr::NOTNULL, T_OBJECT, IN_HEAP);
4303 r->init_req(1, control());
4304 default_value->init_req(1, val);
4305
4306 // Otherwise initialize with all zero
4307 r->init_req(2, _gvn.transform(new IfFalseNode(iff)));
4308 default_value->init_req(2, null());
4309
4310 set_control(_gvn.transform(r));
4311 default_value = _gvn.transform(default_value);
4312 }
4313 if (default_value != nullptr) {
4314 if (UseCompressedOops) {
4315 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4316 default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
4317 Node* lower = _gvn.transform(new CastP2XNode(control(), default_value));
4318 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4319 raw_default_value = _gvn.transform(new OrLNode(lower, upper));
4320 } else {
4321 raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
4322 }
4323 }
4324
4325 Node* valid_length_test = _gvn.intcon(1);
4326 if (ary_type->isa_aryptr()) {
4327 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4328 jint max = TypeAryPtr::max_array_length(bt);
4329 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4330 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4331 }
4332
4333 // Create the AllocateArrayNode and its result projections
4334 AllocateArrayNode* alloc
4335 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4336 control(), mem, i_o(),
4337 size, klass_node,
4338 initial_slow_test,
4339 length, valid_length_test,
4340 default_value, raw_default_value);
4341 // Cast to correct type. Note that the klass_node may be constant or not,
4342 // and in the latter case the actual array type will be inexact also.
4343 // (This happens via a non-constant argument to inline_native_newArray.)
4344 // In any case, the value of klass_node provides the desired array type.
4345 const TypeInt* length_type = _gvn.find_int_type(length);
4346 if (ary_type->isa_aryptr() && length_type != nullptr) {
4347 // Try to get a better type than POS for the size
4348 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4349 }
4350
4351 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4352
4353 array_ideal_length(alloc, ary_type, true);
4354 return javaoop;
4355 }
4356
4357 // The following "Ideal_foo" functions are placed here because they recognize
4358 // the graph shapes created by the functions immediately above.
4359
4360 //---------------------------Ideal_allocation----------------------------------
4361 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4362 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
4363 if (ptr == nullptr) { // reduce dumb test in callers
4364 return nullptr;
4365 }
4366
4367 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4368 ptr = bs->step_over_gc_barrier(ptr);
4369
4370 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4371 ptr = ptr->in(1);
4372 if (ptr == nullptr) return nullptr;
4373 }
4374 // Return null for allocations with several casts:
4375 // j.l.reflect.Array.newInstance(jobject, jint)
4376 // Object.clone()
4377 // to keep more precise type from last cast.
4378 if (ptr->is_Proj()) {
4379 Node* allo = ptr->in(0);
4380 if (allo != nullptr && allo->is_Allocate()) {
4381 return allo->as_Allocate();
4382 }
4383 }
4384 // Report failure to match.
4385 return nullptr;
4386 }
4387
4388 // Fancy version which also strips off an offset (and reports it to caller).
4389 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4390 intptr_t& offset) {
4391 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4392 if (base == nullptr) return nullptr;
4393 return Ideal_allocation(base);
4394 }
4395
4396 // Trace Initialize <- Proj[Parm] <- Allocate
4397 AllocateNode* InitializeNode::allocation() {
4398 Node* rawoop = in(InitializeNode::RawAddress);
4399 if (rawoop->is_Proj()) {
4400 Node* alloc = rawoop->in(0);
4401 if (alloc->is_Allocate()) {
4402 return alloc->as_Allocate();
4403 }
4404 }
4405 return nullptr;
4406 }
4407
4408 // Trace Allocate -> Proj[Parm] -> Initialize
4409 InitializeNode* AllocateNode::initialization() {
4410 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4411 if (rawoop == nullptr) return nullptr;
4412 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4413 Node* init = rawoop->fast_out(i);
4414 if (init->is_Initialize()) {
4415 assert(init->as_Initialize()->allocation() == this, "2-way link");
4416 return init->as_Initialize();
4417 }
4418 }
4419 return nullptr;
4420 }
4421
4422 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4423 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4424 // Too many traps seen?
4425 if (too_many_traps(reason)) {
4426 #ifdef ASSERT
4427 if (TraceLoopPredicate) {
4428 int tc = C->trap_count(reason);
4429 tty->print("too many traps=%s tcount=%d in ",
4430 Deoptimization::trap_reason_name(reason), tc);
4431 method()->print(); // which method has too many predicate traps
4432 tty->cr();
4433 }
4434 #endif
4435 // We cannot afford to take more traps here,
4436 // do not generate Parse Predicate.
4437 return;
4438 }
4439
4440 Node* cont = _gvn.intcon(1);
4441 Node* opaq = _gvn.transform(new Opaque1Node(C, cont));
4442 C->add_parse_predicate_opaq(opaq);
4443 Node* bol = _gvn.transform(new Conv2BNode(opaq));
4444 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), bol, reason);
4445 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4446 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4447 {
4448 PreserveJVMState pjvms(this);
4449 set_control(if_false);
4450 inc_sp(nargs);
4451 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4452 }
4453 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4454 set_control(if_true);
4455 }
4456
4457 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4458 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4459 void GraphKit::add_parse_predicates(int nargs) {
4460 if (UseLoopPredicate) {
4461 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4462 }
4463 if (UseProfiledLoopPredicate) {
4464 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4465 }
4466 // Loop Limit Check Predicate should be near the loop.
4467 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4468 }
4469
4470 void GraphKit::sync_kit(IdealKit& ideal) {
4471 set_all_memory(ideal.merged_memory());
4472 set_i_o(ideal.i_o());
4473 set_control(ideal.ctrl());
4474 }
4475
4476 void GraphKit::final_sync(IdealKit& ideal) {
4477 // Final sync IdealKit and graphKit.
4478 sync_kit(ideal);
4479 }
4480
4481 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4482 Node* len = load_array_length(load_String_value(str, set_ctrl));
4483 Node* coder = load_String_coder(str, set_ctrl);
4484 // Divide length by 2 if coder is UTF16
4485 return _gvn.transform(new RShiftINode(len, coder));
4486 }
4487
4488 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4489 int value_offset = java_lang_String::value_offset();
4490 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4491 false, nullptr, Type::Offset(0));
4492 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4493 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4494 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true),
4495 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4496 Node* p = basic_plus_adr(str, str, value_offset);
4497 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4498 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4499 return load;
4500 }
4501
4502 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4503 if (!CompactStrings) {
4504 return intcon(java_lang_String::CODER_UTF16);
4505 }
4506 int coder_offset = java_lang_String::coder_offset();
4507 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4508 false, nullptr, Type::Offset(0));
4509 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4510
4511 Node* p = basic_plus_adr(str, str, coder_offset);
4512 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4513 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4514 return load;
4515 }
4516
4517 void GraphKit::store_String_value(Node* str, Node* value) {
4518 int value_offset = java_lang_String::value_offset();
4519 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4520 false, nullptr, Type::Offset(0));
4521 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4522
4523 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4524 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4525 }
4526
4527 void GraphKit::store_String_coder(Node* str, Node* value) {
4528 int coder_offset = java_lang_String::coder_offset();
4529 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4530 false, nullptr, Type::Offset(0));
4531 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4532
4533 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4534 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4535 }
4536
4537 // Capture src and dst memory state with a MergeMemNode
4538 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4539 if (src_type == dst_type) {
4540 // Types are equal, we don't need a MergeMemNode
4541 return memory(src_type);
4542 }
4543 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4544 record_for_igvn(merge); // fold it up later, if possible
4545 int src_idx = C->get_alias_index(src_type);
4546 int dst_idx = C->get_alias_index(dst_type);
4547 merge->set_memory_at(src_idx, memory(src_idx));
4548 merge->set_memory_at(dst_idx, memory(dst_idx));
4549 return merge;
4550 }
4551
4552 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4553 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4554 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4555 // If input and output memory types differ, capture both states to preserve
4556 // the dependency between preceding and subsequent loads/stores.
4557 // For example, the following program:
4558 // StoreB
4559 // compress_string
4560 // LoadB
4561 // has this memory graph (use->def):
4562 // LoadB -> compress_string -> CharMem
4563 // ... -> StoreB -> ByteMem
4564 // The intrinsic hides the dependency between LoadB and StoreB, causing
4565 // the load to read from memory not containing the result of the StoreB.
4566 // The correct memory graph should look like this:
4567 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4568 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4569 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4570 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4571 set_memory(res_mem, TypeAryPtr::BYTES);
4572 return str;
4573 }
4574
4575 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4576 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4577 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4578 // Capture src and dst memory (see comment in 'compress_string').
4579 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4580 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4581 set_memory(_gvn.transform(str), dst_type);
4582 }
4583
4584 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4585 /**
4586 * int i_char = start;
4587 * for (int i_byte = 0; i_byte < count; i_byte++) {
4588 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4589 * }
4590 */
4591 add_parse_predicates();
4592 C->set_has_loops(true);
4593
4594 RegionNode* head = new RegionNode(3);
4595 head->init_req(1, control());
4596 gvn().set_type(head, Type::CONTROL);
4597 record_for_igvn(head);
4598
4599 Node* i_byte = new PhiNode(head, TypeInt::INT);
4600 i_byte->init_req(1, intcon(0));
4601 gvn().set_type(i_byte, TypeInt::INT);
4602 record_for_igvn(i_byte);
4603
4604 Node* i_char = new PhiNode(head, TypeInt::INT);
4605 i_char->init_req(1, start);
4606 gvn().set_type(i_char, TypeInt::INT);
4607 record_for_igvn(i_char);
4608
4609 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4610 gvn().set_type(mem, Type::MEMORY);
4611 record_for_igvn(mem);
4612 set_control(head);
4613 set_memory(mem, TypeAryPtr::BYTES);
4614 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4615 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4616 AndI(ch, intcon(0xff)), T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
4617 false, false, true /* mismatched */);
4618
4619 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4620 head->init_req(2, IfTrue(iff));
4621 mem->init_req(2, st);
4622 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4623 i_char->init_req(2, AddI(i_char, intcon(2)));
4624
4625 set_control(IfFalse(iff));
4626 set_memory(st, TypeAryPtr::BYTES);
4627 }
4628
4629 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4630 if (!field->is_constant()) {
4631 return nullptr; // Field not marked as constant.
4632 }
4633 ciInstance* holder = nullptr;
4634 if (!field->is_static()) {
4635 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4636 if (const_oop != nullptr && const_oop->is_instance()) {
4637 holder = const_oop->as_instance();
4638 }
4639 }
4640 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4641 /*is_unsigned_load=*/false);
4642 if (con_type != nullptr) {
4643 Node* con = makecon(con_type);
4644 if (field->type()->is_inlinetype()) {
4645 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free());
4646 } else if (con_type->is_inlinetypeptr()) {
4647 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free());
4648 }
4649 return con;
4650 }
4651 return nullptr;
4652 }
4653
4654 //---------------------------load_mirror_from_klass----------------------------
4655 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4656 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4657 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4658 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4659 // mirror = ((OopHandle)mirror)->resolve();
4660 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4661 }