1 /*
2 * Copyright (c) 2001, 2025, 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 "ci/ciFlatArrayKlass.hpp"
26 #include "ci/ciInlineKlass.hpp"
27 #include "ci/ciUtilities.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "asm/register.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "gc/shared/barrierSet.hpp"
33 #include "gc/shared/c2/barrierSetC2.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "oops/flatArrayKlass.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/graphKit.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/inlinetypenode.hpp"
43 #include "opto/intrinsicnode.hpp"
44 #include "opto/locknode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/runtime.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "runtime/deoptimization.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "utilities/bitMap.inline.hpp"
55 #include "utilities/powerOfTwo.hpp"
56 #include "utilities/growableArray.hpp"
57
58 //----------------------------GraphKit-----------------------------------------
59 // Main utility constructor.
60 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
61 : Phase(Phase::Parser),
62 _env(C->env()),
63 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
64 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
65 {
66 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
67 _exceptions = jvms->map()->next_exception();
68 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
69 set_jvms(jvms);
70 #ifdef ASSERT
71 if (_gvn.is_IterGVN() != nullptr) {
72 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
73 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
74 _worklist_size = _gvn.C->igvn_worklist()->size();
75 }
76 #endif
77 }
78
79 // Private constructor for parser.
80 GraphKit::GraphKit()
81 : Phase(Phase::Parser),
82 _env(C->env()),
83 _gvn(*C->initial_gvn()),
84 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
85 {
86 _exceptions = nullptr;
87 set_map(nullptr);
88 DEBUG_ONLY(_sp = -99);
89 DEBUG_ONLY(set_bci(-99));
90 }
91
92
93
94 //---------------------------clean_stack---------------------------------------
95 // Clear away rubbish from the stack area of the JVM state.
96 // This destroys any arguments that may be waiting on the stack.
97 void GraphKit::clean_stack(int from_sp) {
98 SafePointNode* map = this->map();
99 JVMState* jvms = this->jvms();
100 int stk_size = jvms->stk_size();
101 int stkoff = jvms->stkoff();
102 Node* top = this->top();
103 for (int i = from_sp; i < stk_size; i++) {
104 if (map->in(stkoff + i) != top) {
105 map->set_req(stkoff + i, top);
106 }
107 }
108 }
109
110
111 //--------------------------------sync_jvms-----------------------------------
112 // Make sure our current jvms agrees with our parse state.
113 JVMState* GraphKit::sync_jvms() const {
114 JVMState* jvms = this->jvms();
115 jvms->set_bci(bci()); // Record the new bci in the JVMState
116 jvms->set_sp(sp()); // Record the new sp in the JVMState
117 assert(jvms_in_sync(), "jvms is now in sync");
118 return jvms;
119 }
120
121 //--------------------------------sync_jvms_for_reexecute---------------------
122 // Make sure our current jvms agrees with our parse state. This version
123 // uses the reexecute_sp for reexecuting bytecodes.
124 JVMState* GraphKit::sync_jvms_for_reexecute() {
125 JVMState* jvms = this->jvms();
126 jvms->set_bci(bci()); // Record the new bci in the JVMState
127 jvms->set_sp(reexecute_sp()); // Record the new sp in the JVMState
128 return jvms;
129 }
130
131 #ifdef ASSERT
132 bool GraphKit::jvms_in_sync() const {
133 Parse* parse = is_Parse();
134 if (parse == nullptr) {
135 if (bci() != jvms()->bci()) return false;
136 if (sp() != (int)jvms()->sp()) return false;
137 return true;
138 }
139 if (jvms()->method() != parse->method()) return false;
140 if (jvms()->bci() != parse->bci()) return false;
141 int jvms_sp = jvms()->sp();
142 if (jvms_sp != parse->sp()) return false;
143 int jvms_depth = jvms()->depth();
144 if (jvms_depth != parse->depth()) return false;
145 return true;
146 }
147
148 // Local helper checks for special internal merge points
149 // used to accumulate and merge exception states.
150 // They are marked by the region's in(0) edge being the map itself.
151 // Such merge points must never "escape" into the parser at large,
152 // until they have been handed to gvn.transform.
153 static bool is_hidden_merge(Node* reg) {
154 if (reg == nullptr) return false;
155 if (reg->is_Phi()) {
156 reg = reg->in(0);
157 if (reg == nullptr) return false;
158 }
159 return reg->is_Region() && reg->in(0) != nullptr && reg->in(0)->is_Root();
160 }
161
162 void GraphKit::verify_map() const {
163 if (map() == nullptr) return; // null map is OK
164 assert(map()->req() <= jvms()->endoff(), "no extra garbage on map");
165 assert(!map()->has_exceptions(), "call add_exception_states_from 1st");
166 assert(!is_hidden_merge(control()), "call use_exception_state, not set_map");
167 }
168
169 void GraphKit::verify_exception_state(SafePointNode* ex_map) {
170 assert(ex_map->next_exception() == nullptr, "not already part of a chain");
171 assert(has_saved_ex_oop(ex_map), "every exception state has an ex_oop");
172 }
173 #endif
174
175 //---------------------------stop_and_kill_map---------------------------------
176 // Set _map to null, signalling a stop to further bytecode execution.
177 // First smash the current map's control to a constant, to mark it dead.
178 void GraphKit::stop_and_kill_map() {
179 SafePointNode* dead_map = stop();
180 if (dead_map != nullptr) {
181 dead_map->disconnect_inputs(C); // Mark the map as killed.
182 assert(dead_map->is_killed(), "must be so marked");
183 }
184 }
185
186
187 //--------------------------------stopped--------------------------------------
188 // Tell if _map is null, or control is top.
189 bool GraphKit::stopped() {
190 if (map() == nullptr) return true;
191 else if (control() == top()) return true;
192 else return false;
193 }
194
195
196 //-----------------------------has_exception_handler----------------------------------
197 // Tell if this method or any caller method has exception handlers.
198 bool GraphKit::has_exception_handler() {
199 for (JVMState* jvmsp = jvms(); jvmsp != nullptr; jvmsp = jvmsp->caller()) {
200 if (jvmsp->has_method() && jvmsp->method()->has_exception_handlers()) {
201 return true;
202 }
203 }
204 return false;
205 }
206
207 //------------------------------save_ex_oop------------------------------------
208 // Save an exception without blowing stack contents or other JVM state.
209 void GraphKit::set_saved_ex_oop(SafePointNode* ex_map, Node* ex_oop) {
210 assert(!has_saved_ex_oop(ex_map), "clear ex-oop before setting again");
211 ex_map->add_req(ex_oop);
212 DEBUG_ONLY(verify_exception_state(ex_map));
213 }
214
215 inline static Node* common_saved_ex_oop(SafePointNode* ex_map, bool clear_it) {
216 assert(GraphKit::has_saved_ex_oop(ex_map), "ex_oop must be there");
217 Node* ex_oop = ex_map->in(ex_map->req()-1);
218 if (clear_it) ex_map->del_req(ex_map->req()-1);
219 return ex_oop;
220 }
221
222 //-----------------------------saved_ex_oop------------------------------------
223 // Recover a saved exception from its map.
224 Node* GraphKit::saved_ex_oop(SafePointNode* ex_map) {
225 return common_saved_ex_oop(ex_map, false);
226 }
227
228 //--------------------------clear_saved_ex_oop---------------------------------
229 // Erase a previously saved exception from its map.
230 Node* GraphKit::clear_saved_ex_oop(SafePointNode* ex_map) {
231 return common_saved_ex_oop(ex_map, true);
232 }
233
234 #ifdef ASSERT
235 //---------------------------has_saved_ex_oop----------------------------------
236 // Erase a previously saved exception from its map.
237 bool GraphKit::has_saved_ex_oop(SafePointNode* ex_map) {
238 return ex_map->req() == ex_map->jvms()->endoff()+1;
239 }
240 #endif
241
242 //-------------------------make_exception_state--------------------------------
243 // Turn the current JVM state into an exception state, appending the ex_oop.
244 SafePointNode* GraphKit::make_exception_state(Node* ex_oop) {
245 sync_jvms();
246 SafePointNode* ex_map = stop(); // do not manipulate this map any more
247 set_saved_ex_oop(ex_map, ex_oop);
248 return ex_map;
249 }
250
251
252 //--------------------------add_exception_state--------------------------------
253 // Add an exception to my list of exceptions.
254 void GraphKit::add_exception_state(SafePointNode* ex_map) {
255 if (ex_map == nullptr || ex_map->control() == top()) {
256 return;
257 }
258 #ifdef ASSERT
259 verify_exception_state(ex_map);
260 if (has_exceptions()) {
261 assert(ex_map->jvms()->same_calls_as(_exceptions->jvms()), "all collected exceptions must come from the same place");
262 }
263 #endif
264
265 // If there is already an exception of exactly this type, merge with it.
266 // In particular, null-checks and other low-level exceptions common up here.
267 Node* ex_oop = saved_ex_oop(ex_map);
268 const Type* ex_type = _gvn.type(ex_oop);
269 if (ex_oop == top()) {
270 // No action needed.
271 return;
272 }
273 assert(ex_type->isa_instptr(), "exception must be an instance");
274 for (SafePointNode* e2 = _exceptions; e2 != nullptr; e2 = e2->next_exception()) {
275 const Type* ex_type2 = _gvn.type(saved_ex_oop(e2));
276 // We check sp also because call bytecodes can generate exceptions
277 // both before and after arguments are popped!
278 if (ex_type2 == ex_type
279 && e2->_jvms->sp() == ex_map->_jvms->sp()) {
280 combine_exception_states(ex_map, e2);
281 return;
282 }
283 }
284
285 // No pre-existing exception of the same type. Chain it on the list.
286 push_exception_state(ex_map);
287 }
288
289 //-----------------------add_exception_states_from-----------------------------
290 void GraphKit::add_exception_states_from(JVMState* jvms) {
291 SafePointNode* ex_map = jvms->map()->next_exception();
292 if (ex_map != nullptr) {
293 jvms->map()->set_next_exception(nullptr);
294 for (SafePointNode* next_map; ex_map != nullptr; ex_map = next_map) {
295 next_map = ex_map->next_exception();
296 ex_map->set_next_exception(nullptr);
297 add_exception_state(ex_map);
298 }
299 }
300 }
301
302 //-----------------------transfer_exceptions_into_jvms-------------------------
303 JVMState* GraphKit::transfer_exceptions_into_jvms() {
304 if (map() == nullptr) {
305 // We need a JVMS to carry the exceptions, but the map has gone away.
306 // Create a scratch JVMS, cloned from any of the exception states...
307 if (has_exceptions()) {
308 _map = _exceptions;
309 _map = clone_map();
310 _map->set_next_exception(nullptr);
311 clear_saved_ex_oop(_map);
312 DEBUG_ONLY(verify_map());
313 } else {
314 // ...or created from scratch
315 JVMState* jvms = new (C) JVMState(_method, nullptr);
316 jvms->set_bci(_bci);
317 jvms->set_sp(_sp);
318 jvms->set_map(new SafePointNode(TypeFunc::Parms, jvms));
319 set_jvms(jvms);
320 for (uint i = 0; i < map()->req(); i++) map()->init_req(i, top());
321 set_all_memory(top());
322 while (map()->req() < jvms->endoff()) map()->add_req(top());
323 }
324 // (This is a kludge, in case you didn't notice.)
325 set_control(top());
326 }
327 JVMState* jvms = sync_jvms();
328 assert(!jvms->map()->has_exceptions(), "no exceptions on this map yet");
329 jvms->map()->set_next_exception(_exceptions);
330 _exceptions = nullptr; // done with this set of exceptions
331 return jvms;
332 }
333
334 static inline void add_n_reqs(Node* dstphi, Node* srcphi) {
335 assert(is_hidden_merge(dstphi), "must be a special merge node");
336 assert(is_hidden_merge(srcphi), "must be a special merge node");
337 uint limit = srcphi->req();
338 for (uint i = PhiNode::Input; i < limit; i++) {
339 dstphi->add_req(srcphi->in(i));
340 }
341 }
342 static inline void add_one_req(Node* dstphi, Node* src) {
343 assert(is_hidden_merge(dstphi), "must be a special merge node");
344 assert(!is_hidden_merge(src), "must not be a special merge node");
345 dstphi->add_req(src);
346 }
347
348 //-----------------------combine_exception_states------------------------------
349 // This helper function combines exception states by building phis on a
350 // specially marked state-merging region. These regions and phis are
351 // untransformed, and can build up gradually. The region is marked by
352 // having a control input of its exception map, rather than null. Such
353 // regions do not appear except in this function, and in use_exception_state.
354 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
355 if (failing_internal()) {
356 return; // dying anyway...
357 }
358 JVMState* ex_jvms = ex_map->_jvms;
359 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
360 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
361 // TODO 8325632 Re-enable
362 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
363 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
364 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
365 assert(ex_map->req() == phi_map->req(), "matching maps");
366 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
367 Node* hidden_merge_mark = root();
368 Node* region = phi_map->control();
369 MergeMemNode* phi_mem = phi_map->merged_memory();
370 MergeMemNode* ex_mem = ex_map->merged_memory();
371 if (region->in(0) != hidden_merge_mark) {
372 // The control input is not (yet) a specially-marked region in phi_map.
373 // Make it so, and build some phis.
374 region = new RegionNode(2);
375 _gvn.set_type(region, Type::CONTROL);
376 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
377 region->init_req(1, phi_map->control());
378 phi_map->set_control(region);
379 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
380 record_for_igvn(io_phi);
381 _gvn.set_type(io_phi, Type::ABIO);
382 phi_map->set_i_o(io_phi);
383 for (MergeMemStream mms(phi_mem); mms.next_non_empty(); ) {
384 Node* m = mms.memory();
385 Node* m_phi = PhiNode::make(region, m, Type::MEMORY, mms.adr_type(C));
386 record_for_igvn(m_phi);
387 _gvn.set_type(m_phi, Type::MEMORY);
388 mms.set_memory(m_phi);
389 }
390 }
391
392 // Either or both of phi_map and ex_map might already be converted into phis.
393 Node* ex_control = ex_map->control();
394 // if there is special marking on ex_map also, we add multiple edges from src
395 bool add_multiple = (ex_control->in(0) == hidden_merge_mark);
396 // how wide was the destination phi_map, originally?
397 uint orig_width = region->req();
398
399 if (add_multiple) {
400 add_n_reqs(region, ex_control);
401 add_n_reqs(phi_map->i_o(), ex_map->i_o());
402 } else {
403 // ex_map has no merges, so we just add single edges everywhere
404 add_one_req(region, ex_control);
405 add_one_req(phi_map->i_o(), ex_map->i_o());
406 }
407 for (MergeMemStream mms(phi_mem, ex_mem); mms.next_non_empty2(); ) {
408 if (mms.is_empty()) {
409 // get a copy of the base memory, and patch some inputs into it
410 const TypePtr* adr_type = mms.adr_type(C);
411 Node* phi = mms.force_memory()->as_Phi()->slice_memory(adr_type);
412 assert(phi->as_Phi()->region() == mms.base_memory()->in(0), "");
413 mms.set_memory(phi);
414 // Prepare to append interesting stuff onto the newly sliced phi:
415 while (phi->req() > orig_width) phi->del_req(phi->req()-1);
416 }
417 // Append stuff from ex_map:
418 if (add_multiple) {
419 add_n_reqs(mms.memory(), mms.memory2());
420 } else {
421 add_one_req(mms.memory(), mms.memory2());
422 }
423 }
424 uint limit = ex_map->req();
425 for (uint i = TypeFunc::Parms; i < limit; i++) {
426 // Skip everything in the JVMS after tos. (The ex_oop follows.)
427 if (i == tos) i = ex_jvms->monoff();
428 Node* src = ex_map->in(i);
429 Node* dst = phi_map->in(i);
430 if (src != dst) {
431 PhiNode* phi;
432 if (dst->in(0) != region) {
433 dst = phi = PhiNode::make(region, dst, _gvn.type(dst));
434 record_for_igvn(phi);
435 _gvn.set_type(phi, phi->type());
436 phi_map->set_req(i, dst);
437 // Prepare to append interesting stuff onto the new phi:
438 while (dst->req() > orig_width) dst->del_req(dst->req()-1);
439 } else {
440 assert(dst->is_Phi(), "nobody else uses a hidden region");
441 phi = dst->as_Phi();
442 }
443 if (add_multiple && src->in(0) == ex_control) {
444 // Both are phis.
445 add_n_reqs(dst, src);
446 } else {
447 while (dst->req() < region->req()) add_one_req(dst, src);
448 }
449 const Type* srctype = _gvn.type(src);
450 if (phi->type() != srctype) {
451 const Type* dsttype = phi->type()->meet_speculative(srctype);
452 if (phi->type() != dsttype) {
453 phi->set_type(dsttype);
454 _gvn.set_type(phi, dsttype);
455 }
456 }
457 }
458 }
459 phi_map->merge_replaced_nodes_with(ex_map);
460 }
461
462 //--------------------------use_exception_state--------------------------------
463 Node* GraphKit::use_exception_state(SafePointNode* phi_map) {
464 if (failing_internal()) { stop(); return top(); }
465 Node* region = phi_map->control();
466 Node* hidden_merge_mark = root();
467 assert(phi_map->jvms()->map() == phi_map, "sanity: 1-1 relation");
468 Node* ex_oop = clear_saved_ex_oop(phi_map);
469 if (region->in(0) == hidden_merge_mark) {
470 // Special marking for internal ex-states. Process the phis now.
471 region->set_req(0, region); // now it's an ordinary region
472 set_jvms(phi_map->jvms()); // ...so now we can use it as a map
473 // Note: Setting the jvms also sets the bci and sp.
474 set_control(_gvn.transform(region));
475 uint tos = jvms()->stkoff() + sp();
476 for (uint i = 1; i < tos; i++) {
477 Node* x = phi_map->in(i);
478 if (x->in(0) == region) {
479 assert(x->is_Phi(), "expected a special phi");
480 phi_map->set_req(i, _gvn.transform(x));
481 }
482 }
483 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
484 Node* x = mms.memory();
485 if (x->in(0) == region) {
486 assert(x->is_Phi(), "nobody else uses a hidden region");
487 mms.set_memory(_gvn.transform(x));
488 }
489 }
490 if (ex_oop->in(0) == region) {
491 assert(ex_oop->is_Phi(), "expected a special phi");
492 ex_oop = _gvn.transform(ex_oop);
493 }
494 } else {
495 set_jvms(phi_map->jvms());
496 }
497
498 assert(!is_hidden_merge(phi_map->control()), "hidden ex. states cleared");
499 assert(!is_hidden_merge(phi_map->i_o()), "hidden ex. states cleared");
500 return ex_oop;
501 }
502
503 //---------------------------------java_bc-------------------------------------
504 Bytecodes::Code GraphKit::java_bc() const {
505 ciMethod* method = this->method();
506 int bci = this->bci();
507 if (method != nullptr && bci != InvocationEntryBci)
508 return method->java_code_at_bci(bci);
509 else
510 return Bytecodes::_illegal;
511 }
512
513 void GraphKit::uncommon_trap_if_should_post_on_exceptions(Deoptimization::DeoptReason reason,
514 bool must_throw) {
515 // if the exception capability is set, then we will generate code
516 // to check the JavaThread.should_post_on_exceptions flag to see
517 // if we actually need to report exception events (for this
518 // thread). If we don't need to report exception events, we will
519 // take the normal fast path provided by add_exception_events. If
520 // exception event reporting is enabled for this thread, we will
521 // take the uncommon_trap in the BuildCutout below.
522
523 // first must access the should_post_on_exceptions_flag in this thread's JavaThread
524 Node* jthread = _gvn.transform(new ThreadLocalNode());
525 Node* adr = basic_plus_adr(top(), jthread, in_bytes(JavaThread::should_post_on_exceptions_flag_offset()));
526 Node* should_post_flag = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
527
528 // Test the should_post_on_exceptions_flag vs. 0
529 Node* chk = _gvn.transform( new CmpINode(should_post_flag, intcon(0)) );
530 Node* tst = _gvn.transform( new BoolNode(chk, BoolTest::eq) );
531
532 // Branch to slow_path if should_post_on_exceptions_flag was true
533 { BuildCutout unless(this, tst, PROB_MAX);
534 // Do not try anything fancy if we're notifying the VM on every throw.
535 // Cf. case Bytecodes::_athrow in parse2.cpp.
536 uncommon_trap(reason, Deoptimization::Action_none,
537 (ciKlass*)nullptr, (char*)nullptr, must_throw);
538 }
539
540 }
541
542 //------------------------------builtin_throw----------------------------------
543 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason) {
544 builtin_throw(reason, builtin_throw_exception(reason), /*allow_too_many_traps*/ true);
545 }
546
547 void GraphKit::builtin_throw(Deoptimization::DeoptReason reason,
548 ciInstance* ex_obj,
549 bool allow_too_many_traps) {
550 // If this throw happens frequently, an uncommon trap might cause
551 // a performance pothole. If there is a local exception handler,
552 // and if this particular bytecode appears to be deoptimizing often,
553 // let us handle the throw inline, with a preconstructed instance.
554 // Note: If the deopt count has blown up, the uncommon trap
555 // runtime is going to flush this nmethod, not matter what.
556 if (is_builtin_throw_hot(reason)) {
557 if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
558 // If the throw is local, we use a pre-existing instance and
559 // punt on the backtrace. This would lead to a missing backtrace
560 // (a repeat of 4292742) if the backtrace object is ever asked
561 // for its backtrace.
562 // Fixing this remaining case of 4292742 requires some flavor of
563 // escape analysis. Leave that for the future.
564 if (env()->jvmti_can_post_on_exceptions()) {
565 // check if we must post exception events, take uncommon trap if so
566 uncommon_trap_if_should_post_on_exceptions(reason, true /*must_throw*/);
567 // here if should_post_on_exceptions is false
568 // continue on with the normal codegen
569 }
570
571 // Cheat with a preallocated exception object.
572 if (C->log() != nullptr)
573 C->log()->elem("hot_throw preallocated='1' reason='%s'",
574 Deoptimization::trap_reason_name(reason));
575 const TypeInstPtr* ex_con = TypeInstPtr::make(ex_obj);
576 Node* ex_node = _gvn.transform(ConNode::make(ex_con));
577
578 // Clear the detail message of the preallocated exception object.
579 // Weblogic sometimes mutates the detail message of exceptions
580 // using reflection.
581 int offset = java_lang_Throwable::get_detailMessage_offset();
582 const TypePtr* adr_typ = ex_con->add_offset(offset);
583
584 Node *adr = basic_plus_adr(ex_node, ex_node, offset);
585 const TypeOopPtr* val_type = TypeOopPtr::make_from_klass(env()->String_klass());
586 Node *store = access_store_at(ex_node, adr, adr_typ, null(), val_type, T_OBJECT, IN_HEAP);
587
588 if (!method()->has_exception_handlers()) {
589 // We don't need to preserve the stack if there's no handler as the entire frame is going to be popped anyway.
590 // This prevents issues with exception handling and late inlining.
591 set_sp(0);
592 clean_stack(0);
593 }
594
595 add_exception_state(make_exception_state(ex_node));
596 return;
597 } else if (builtin_throw_too_many_traps(reason, ex_obj)) {
598 // We cannot afford to take too many traps here. Suffer in the interpreter instead.
599 assert(allow_too_many_traps, "not allowed");
600 if (C->log() != nullptr) {
601 C->log()->elem("hot_throw preallocated='0' reason='%s' mcount='%d'",
602 Deoptimization::trap_reason_name(reason),
603 C->trap_count(reason));
604 }
605 uncommon_trap(reason, Deoptimization::Action_none,
606 (ciKlass*) nullptr, (char*) nullptr,
607 true /*must_throw*/);
608 return;
609 }
610 }
611
612 // %%% Maybe add entry to OptoRuntime which directly throws the exc.?
613 // It won't be much cheaper than bailing to the interp., since we'll
614 // have to pass up all the debug-info, and the runtime will have to
615 // create the stack trace.
616
617 // Usual case: Bail to interpreter.
618 // Reserve the right to recompile if we haven't seen anything yet.
619
620 // "must_throw" prunes the JVM state to include only the stack, if there
621 // are no local exception handlers. This should cut down on register
622 // allocation time and code size, by drastically reducing the number
623 // of in-edges on the call to the uncommon trap.
624 uncommon_trap(reason, Deoptimization::Action_maybe_recompile,
625 (ciKlass*) nullptr, (char*) nullptr,
626 true /*must_throw*/);
627 }
628
629 bool GraphKit::is_builtin_throw_hot(Deoptimization::DeoptReason reason) {
630 // If this particular condition has not yet happened at this
631 // bytecode, then use the uncommon trap mechanism, and allow for
632 // a future recompilation if several traps occur here.
633 // If the throw is hot, try to use a more complicated inline mechanism
634 // which keeps execution inside the compiled code.
635 if (ProfileTraps) {
636 if (too_many_traps(reason)) {
637 return true;
638 }
639 // (If there is no MDO at all, assume it is early in
640 // execution, and that any deopts are part of the
641 // startup transient, and don't need to be remembered.)
642
643 // Also, if there is a local exception handler, treat all throws
644 // as hot if there has been at least one in this method.
645 if (C->trap_count(reason) != 0 &&
646 method()->method_data()->trap_count(reason) != 0 &&
647 has_exception_handler()) {
648 return true;
649 }
650 }
651 return false;
652 }
653
654 bool GraphKit::builtin_throw_too_many_traps(Deoptimization::DeoptReason reason,
655 ciInstance* ex_obj) {
656 if (is_builtin_throw_hot(reason)) {
657 if (method()->can_omit_stack_trace() && ex_obj != nullptr) {
658 return false; // no traps; throws preallocated exception instead
659 }
660 ciMethod* m = Deoptimization::reason_is_speculate(reason) ? C->method() : nullptr;
661 if (method()->method_data()->trap_recompiled_at(bci(), m) ||
662 C->too_many_traps(reason)) {
663 return true;
664 }
665 }
666 return false;
667 }
668
669 ciInstance* GraphKit::builtin_throw_exception(Deoptimization::DeoptReason reason) const {
670 // Preallocated exception objects to use when we don't need the backtrace.
671 switch (reason) {
672 case Deoptimization::Reason_null_check:
673 return env()->NullPointerException_instance();
674 case Deoptimization::Reason_div0_check:
675 return env()->ArithmeticException_instance();
676 case Deoptimization::Reason_range_check:
677 return env()->ArrayIndexOutOfBoundsException_instance();
678 case Deoptimization::Reason_class_check:
679 return env()->ClassCastException_instance();
680 case Deoptimization::Reason_array_check:
681 return env()->ArrayStoreException_instance();
682 default:
683 return nullptr;
684 }
685 }
686
687 //----------------------------PreserveJVMState---------------------------------
688 PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
689 DEBUG_ONLY(kit->verify_map());
690 _kit = kit;
691 _map = kit->map(); // preserve the map
692 _sp = kit->sp();
693 kit->set_map(clone_map ? kit->clone_map() : nullptr);
694 #ifdef ASSERT
695 _bci = kit->bci();
696 Parse* parser = kit->is_Parse();
697 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
698 _block = block;
699 #endif
700 }
701 PreserveJVMState::~PreserveJVMState() {
702 GraphKit* kit = _kit;
703 #ifdef ASSERT
704 assert(kit->bci() == _bci, "bci must not shift");
705 Parse* parser = kit->is_Parse();
706 int block = (parser == nullptr || parser->block() == nullptr) ? -1 : parser->block()->rpo();
707 assert(block == _block, "block must not shift");
708 #endif
709 kit->set_map(_map);
710 kit->set_sp(_sp);
711 }
712
713
714 //-----------------------------BuildCutout-------------------------------------
715 BuildCutout::BuildCutout(GraphKit* kit, Node* p, float prob, float cnt)
716 : PreserveJVMState(kit)
717 {
718 assert(p->is_Con() || p->is_Bool(), "test must be a bool");
719 SafePointNode* outer_map = _map; // preserved map is caller's
720 SafePointNode* inner_map = kit->map();
721 IfNode* iff = kit->create_and_map_if(outer_map->control(), p, prob, cnt);
722 outer_map->set_control(kit->gvn().transform( new IfTrueNode(iff) ));
723 inner_map->set_control(kit->gvn().transform( new IfFalseNode(iff) ));
724 }
725 BuildCutout::~BuildCutout() {
726 GraphKit* kit = _kit;
727 assert(kit->stopped(), "cutout code must stop, throw, return, etc.");
728 }
729
730 //---------------------------PreserveReexecuteState----------------------------
731 PreserveReexecuteState::PreserveReexecuteState(GraphKit* kit) {
732 assert(!kit->stopped(), "must call stopped() before");
733 _kit = kit;
734 _sp = kit->sp();
735 _reexecute = kit->jvms()->_reexecute;
736 }
737 PreserveReexecuteState::~PreserveReexecuteState() {
738 if (_kit->stopped()) return;
739 _kit->jvms()->_reexecute = _reexecute;
740 _kit->set_sp(_sp);
741 }
742
743 //------------------------------clone_map--------------------------------------
744 // Implementation of PreserveJVMState
745 //
746 // Only clone_map(...) here. If this function is only used in the
747 // PreserveJVMState class we may want to get rid of this extra
748 // function eventually and do it all there.
749
750 SafePointNode* GraphKit::clone_map() {
751 if (map() == nullptr) return nullptr;
752
753 // Clone the memory edge first
754 Node* mem = MergeMemNode::make(map()->memory());
755 gvn().set_type_bottom(mem);
756
757 SafePointNode *clonemap = (SafePointNode*)map()->clone();
758 JVMState* jvms = this->jvms();
759 JVMState* clonejvms = jvms->clone_shallow(C);
760 clonemap->set_memory(mem);
761 clonemap->set_jvms(clonejvms);
762 clonejvms->set_map(clonemap);
763 record_for_igvn(clonemap);
764 gvn().set_type_bottom(clonemap);
765 return clonemap;
766 }
767
768 //-----------------------------destruct_map_clone------------------------------
769 //
770 // Order of destruct is important to increase the likelyhood that memory can be re-used. We need
771 // to destruct/free/delete in the exact opposite order as clone_map().
772 void GraphKit::destruct_map_clone(SafePointNode* sfp) {
773 if (sfp == nullptr) return;
774
775 Node* mem = sfp->memory();
776 JVMState* jvms = sfp->jvms();
777
778 if (jvms != nullptr) {
779 delete jvms;
780 }
781
782 remove_for_igvn(sfp);
783 gvn().clear_type(sfp);
784 sfp->destruct(&_gvn);
785
786 if (mem != nullptr) {
787 gvn().clear_type(mem);
788 mem->destruct(&_gvn);
789 }
790 }
791
792 //-----------------------------set_map_clone-----------------------------------
793 void GraphKit::set_map_clone(SafePointNode* m) {
794 _map = m;
795 _map = clone_map();
796 _map->set_next_exception(nullptr);
797 DEBUG_ONLY(verify_map());
798 }
799
800
801 //----------------------------kill_dead_locals---------------------------------
802 // Detect any locals which are known to be dead, and force them to top.
803 void GraphKit::kill_dead_locals() {
804 // Consult the liveness information for the locals. If any
805 // of them are unused, then they can be replaced by top(). This
806 // should help register allocation time and cut down on the size
807 // of the deoptimization information.
808
809 // This call is made from many of the bytecode handling
810 // subroutines called from the Big Switch in do_one_bytecode.
811 // Every bytecode which might include a slow path is responsible
812 // for killing its dead locals. The more consistent we
813 // are about killing deads, the fewer useless phis will be
814 // constructed for them at various merge points.
815
816 // bci can be -1 (InvocationEntryBci). We return the entry
817 // liveness for the method.
818
819 if (method() == nullptr || method()->code_size() == 0) {
820 // We are building a graph for a call to a native method.
821 // All locals are live.
822 return;
823 }
824
825 ResourceMark rm;
826
827 // Consult the liveness information for the locals. If any
828 // of them are unused, then they can be replaced by top(). This
829 // should help register allocation time and cut down on the size
830 // of the deoptimization information.
831 MethodLivenessResult live_locals = method()->liveness_at_bci(bci());
832
833 int len = (int)live_locals.size();
834 assert(len <= jvms()->loc_size(), "too many live locals");
835 for (int local = 0; local < len; local++) {
836 if (!live_locals.at(local)) {
837 set_local(local, top());
838 }
839 }
840 }
841
842 #ifdef ASSERT
843 //-------------------------dead_locals_are_killed------------------------------
844 // Return true if all dead locals are set to top in the map.
845 // Used to assert "clean" debug info at various points.
846 bool GraphKit::dead_locals_are_killed() {
847 if (method() == nullptr || method()->code_size() == 0) {
848 // No locals need to be dead, so all is as it should be.
849 return true;
850 }
851
852 // Make sure somebody called kill_dead_locals upstream.
853 ResourceMark rm;
854 for (JVMState* jvms = this->jvms(); jvms != nullptr; jvms = jvms->caller()) {
855 if (jvms->loc_size() == 0) continue; // no locals to consult
856 SafePointNode* map = jvms->map();
857 ciMethod* method = jvms->method();
858 int bci = jvms->bci();
859 if (jvms == this->jvms()) {
860 bci = this->bci(); // it might not yet be synched
861 }
862 MethodLivenessResult live_locals = method->liveness_at_bci(bci);
863 int len = (int)live_locals.size();
864 if (!live_locals.is_valid() || len == 0)
865 // This method is trivial, or is poisoned by a breakpoint.
866 return true;
867 assert(len == jvms->loc_size(), "live map consistent with locals map");
868 for (int local = 0; local < len; local++) {
869 if (!live_locals.at(local) && map->local(jvms, local) != top()) {
870 if (PrintMiscellaneous && (Verbose || WizardMode)) {
871 tty->print_cr("Zombie local %d: ", local);
872 jvms->dump();
873 }
874 return false;
875 }
876 }
877 }
878 return true;
879 }
880
881 #endif //ASSERT
882
883 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
884 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
885 ciMethod* cur_method = jvms->method();
886 int cur_bci = jvms->bci();
887 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
888 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
889 return Interpreter::bytecode_should_reexecute(code) ||
890 (is_anewarray && (code == Bytecodes::_multianewarray));
891 // Reexecute _multianewarray bytecode which was replaced with
892 // sequence of [a]newarray. See Parse::do_multianewarray().
893 //
894 // Note: interpreter should not have it set since this optimization
895 // is limited by dimensions and guarded by flag so in some cases
896 // multianewarray() runtime calls will be generated and
897 // the bytecode should not be reexecutes (stack will not be reset).
898 } else {
899 return false;
900 }
901 }
902
903 // Helper function for adding JVMState and debug information to node
904 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
905 // Add the safepoint edges to the call (or other safepoint).
906
907 // Make sure dead locals are set to top. This
908 // should help register allocation time and cut down on the size
909 // of the deoptimization information.
910 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
911
912 // Walk the inline list to fill in the correct set of JVMState's
913 // Also fill in the associated edges for each JVMState.
914
915 // If the bytecode needs to be reexecuted we need to put
916 // the arguments back on the stack.
917 const bool should_reexecute = jvms()->should_reexecute();
918 JVMState* youngest_jvms = should_reexecute ? sync_jvms_for_reexecute() : sync_jvms();
919
920 // NOTE: set_bci (called from sync_jvms) might reset the reexecute bit to
921 // undefined if the bci is different. This is normal for Parse but it
922 // should not happen for LibraryCallKit because only one bci is processed.
923 assert(!is_LibraryCallKit() || (jvms()->should_reexecute() == should_reexecute),
924 "in LibraryCallKit the reexecute bit should not change");
925
926 // If we are guaranteed to throw, we can prune everything but the
927 // input to the current bytecode.
928 bool can_prune_locals = false;
929 uint stack_slots_not_pruned = 0;
930 int inputs = 0, depth = 0;
931 if (must_throw) {
932 assert(method() == youngest_jvms->method(), "sanity");
933 if (compute_stack_effects(inputs, depth)) {
934 can_prune_locals = true;
935 stack_slots_not_pruned = inputs;
936 }
937 }
938
939 if (env()->should_retain_local_variables()) {
940 // At any safepoint, this method can get breakpointed, which would
941 // then require an immediate deoptimization.
942 can_prune_locals = false; // do not prune locals
943 stack_slots_not_pruned = 0;
944 }
945
946 // do not scribble on the input jvms
947 JVMState* out_jvms = youngest_jvms->clone_deep(C);
948 call->set_jvms(out_jvms); // Start jvms list for call node
949
950 // For a known set of bytecodes, the interpreter should reexecute them if
951 // deoptimization happens. We set the reexecute state for them here
952 if (out_jvms->is_reexecute_undefined() && //don't change if already specified
953 should_reexecute_implied_by_bytecode(out_jvms, call->is_AllocateArray())) {
954 #ifdef ASSERT
955 int inputs = 0, not_used; // initialized by GraphKit::compute_stack_effects()
956 assert(method() == youngest_jvms->method(), "sanity");
957 assert(compute_stack_effects(inputs, not_used), "unknown bytecode: %s", Bytecodes::name(java_bc()));
958 assert(out_jvms->sp() >= (uint)inputs, "not enough operands for reexecution");
959 #endif // ASSERT
960 out_jvms->set_should_reexecute(true); //NOTE: youngest_jvms not changed
961 }
962
963 // Presize the call:
964 DEBUG_ONLY(uint non_debug_edges = call->req());
965 call->add_req_batch(top(), youngest_jvms->debug_depth());
966 assert(call->req() == non_debug_edges + youngest_jvms->debug_depth(), "");
967
968 // Set up edges so that the call looks like this:
969 // Call [state:] ctl io mem fptr retadr
970 // [parms:] parm0 ... parmN
971 // [root:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
972 // [...mid:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN [...]
973 // [young:] loc0 ... locN stk0 ... stkSP mon0 obj0 ... monN objN
974 // Note that caller debug info precedes callee debug info.
975
976 // Fill pointer walks backwards from "young:" to "root:" in the diagram above:
977 uint debug_ptr = call->req();
978
979 // Loop over the map input edges associated with jvms, add them
980 // to the call node, & reset all offsets to match call node array.
981
982 JVMState* callee_jvms = nullptr;
983 for (JVMState* in_jvms = youngest_jvms; in_jvms != nullptr; ) {
984 uint debug_end = debug_ptr;
985 uint debug_start = debug_ptr - in_jvms->debug_size();
986 debug_ptr = debug_start; // back up the ptr
987
988 uint p = debug_start; // walks forward in [debug_start, debug_end)
989 uint j, k, l;
990 SafePointNode* in_map = in_jvms->map();
991 out_jvms->set_map(call);
992
993 if (can_prune_locals) {
994 assert(in_jvms->method() == out_jvms->method(), "sanity");
995 // If the current throw can reach an exception handler in this JVMS,
996 // then we must keep everything live that can reach that handler.
997 // As a quick and dirty approximation, we look for any handlers at all.
998 if (in_jvms->method()->has_exception_handlers()) {
999 can_prune_locals = false;
1000 }
1001 }
1002
1003 // Add the Locals
1004 k = in_jvms->locoff();
1005 l = in_jvms->loc_size();
1006 out_jvms->set_locoff(p);
1007 if (!can_prune_locals) {
1008 for (j = 0; j < l; j++) {
1009 call->set_req(p++, in_map->in(k + j));
1010 }
1011 } else {
1012 p += l; // already set to top above by add_req_batch
1013 }
1014
1015 // Add the Expression Stack
1016 k = in_jvms->stkoff();
1017 l = in_jvms->sp();
1018 out_jvms->set_stkoff(p);
1019 if (!can_prune_locals) {
1020 for (j = 0; j < l; j++) {
1021 call->set_req(p++, in_map->in(k + j));
1022 }
1023 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1024 // Divide stack into {S0,...,S1}, where S0 is set to top.
1025 uint s1 = stack_slots_not_pruned;
1026 stack_slots_not_pruned = 0; // for next iteration
1027 if (s1 > l) s1 = l;
1028 uint s0 = l - s1;
1029 p += s0; // skip the tops preinstalled by add_req_batch
1030 for (j = s0; j < l; j++)
1031 call->set_req(p++, in_map->in(k+j));
1032 } else {
1033 p += l; // already set to top above by add_req_batch
1034 }
1035
1036 // Add the Monitors
1037 k = in_jvms->monoff();
1038 l = in_jvms->mon_size();
1039 out_jvms->set_monoff(p);
1040 for (j = 0; j < l; j++)
1041 call->set_req(p++, in_map->in(k+j));
1042
1043 // Copy any scalar object fields.
1044 k = in_jvms->scloff();
1045 l = in_jvms->scl_size();
1046 out_jvms->set_scloff(p);
1047 for (j = 0; j < l; j++)
1048 call->set_req(p++, in_map->in(k+j));
1049
1050 // Finish the new jvms.
1051 out_jvms->set_endoff(p);
1052
1053 assert(out_jvms->endoff() == debug_end, "fill ptr must match");
1054 assert(out_jvms->depth() == in_jvms->depth(), "depth must match");
1055 assert(out_jvms->loc_size() == in_jvms->loc_size(), "size must match");
1056 assert(out_jvms->mon_size() == in_jvms->mon_size(), "size must match");
1057 assert(out_jvms->scl_size() == in_jvms->scl_size(), "size must match");
1058 assert(out_jvms->debug_size() == in_jvms->debug_size(), "size must match");
1059
1060 // Update the two tail pointers in parallel.
1061 callee_jvms = out_jvms;
1062 out_jvms = out_jvms->caller();
1063 in_jvms = in_jvms->caller();
1064 }
1065
1066 assert(debug_ptr == non_debug_edges, "debug info must fit exactly");
1067
1068 // Test the correctness of JVMState::debug_xxx accessors:
1069 assert(call->jvms()->debug_start() == non_debug_edges, "");
1070 assert(call->jvms()->debug_end() == call->req(), "");
1071 assert(call->jvms()->debug_depth() == call->req() - non_debug_edges, "");
1072 }
1073
1074 bool GraphKit::compute_stack_effects(int& inputs, int& depth) {
1075 Bytecodes::Code code = java_bc();
1076 if (code == Bytecodes::_wide) {
1077 code = method()->java_code_at_bci(bci() + 1);
1078 }
1079
1080 if (code != Bytecodes::_illegal) {
1081 depth = Bytecodes::depth(code); // checkcast=0, athrow=-1
1082 }
1083
1084 auto rsize = [&]() {
1085 assert(code != Bytecodes::_illegal, "code is illegal!");
1086 BasicType rtype = Bytecodes::result_type(code); // checkcast=P, athrow=V
1087 return (rtype < T_CONFLICT) ? type2size[rtype] : 0;
1088 };
1089
1090 switch (code) {
1091 case Bytecodes::_illegal:
1092 return false;
1093
1094 case Bytecodes::_ldc:
1095 case Bytecodes::_ldc_w:
1096 case Bytecodes::_ldc2_w:
1097 inputs = 0;
1098 break;
1099
1100 case Bytecodes::_dup: inputs = 1; break;
1101 case Bytecodes::_dup_x1: inputs = 2; break;
1102 case Bytecodes::_dup_x2: inputs = 3; break;
1103 case Bytecodes::_dup2: inputs = 2; break;
1104 case Bytecodes::_dup2_x1: inputs = 3; break;
1105 case Bytecodes::_dup2_x2: inputs = 4; break;
1106 case Bytecodes::_swap: inputs = 2; break;
1107 case Bytecodes::_arraylength: inputs = 1; break;
1108
1109 case Bytecodes::_getstatic:
1110 case Bytecodes::_putstatic:
1111 case Bytecodes::_getfield:
1112 case Bytecodes::_putfield:
1113 {
1114 bool ignored_will_link;
1115 ciField* field = method()->get_field_at_bci(bci(), ignored_will_link);
1116 int size = field->type()->size();
1117 bool is_get = (depth >= 0), is_static = (depth & 1);
1118 inputs = (is_static ? 0 : 1);
1119 if (is_get) {
1120 depth = size - inputs;
1121 } else {
1122 inputs += size; // putxxx pops the value from the stack
1123 depth = - inputs;
1124 }
1125 }
1126 break;
1127
1128 case Bytecodes::_invokevirtual:
1129 case Bytecodes::_invokespecial:
1130 case Bytecodes::_invokestatic:
1131 case Bytecodes::_invokedynamic:
1132 case Bytecodes::_invokeinterface:
1133 {
1134 bool ignored_will_link;
1135 ciSignature* declared_signature = nullptr;
1136 ciMethod* ignored_callee = method()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
1137 assert(declared_signature != nullptr, "cannot be null");
1138 inputs = declared_signature->arg_size_for_bc(code);
1139 int size = declared_signature->return_type()->size();
1140 depth = size - inputs;
1141 }
1142 break;
1143
1144 case Bytecodes::_multianewarray:
1145 {
1146 ciBytecodeStream iter(method());
1147 iter.reset_to_bci(bci());
1148 iter.next();
1149 inputs = iter.get_dimensions();
1150 assert(rsize() == 1, "");
1151 depth = 1 - inputs;
1152 }
1153 break;
1154
1155 case Bytecodes::_ireturn:
1156 case Bytecodes::_lreturn:
1157 case Bytecodes::_freturn:
1158 case Bytecodes::_dreturn:
1159 case Bytecodes::_areturn:
1160 assert(rsize() == -depth, "");
1161 inputs = -depth;
1162 break;
1163
1164 case Bytecodes::_jsr:
1165 case Bytecodes::_jsr_w:
1166 inputs = 0;
1167 depth = 1; // S.B. depth=1, not zero
1168 break;
1169
1170 default:
1171 // bytecode produces a typed result
1172 inputs = rsize() - depth;
1173 assert(inputs >= 0, "");
1174 break;
1175 }
1176
1177 #ifdef ASSERT
1178 // spot check
1179 int outputs = depth + inputs;
1180 assert(outputs >= 0, "sanity");
1181 switch (code) {
1182 case Bytecodes::_checkcast: assert(inputs == 1 && outputs == 1, ""); break;
1183 case Bytecodes::_athrow: assert(inputs == 1 && outputs == 0, ""); break;
1184 case Bytecodes::_aload_0: assert(inputs == 0 && outputs == 1, ""); break;
1185 case Bytecodes::_return: assert(inputs == 0 && outputs == 0, ""); break;
1186 case Bytecodes::_drem: assert(inputs == 4 && outputs == 2, ""); break;
1187 default: break;
1188 }
1189 #endif //ASSERT
1190
1191 return true;
1192 }
1193
1194
1195
1196 //------------------------------basic_plus_adr---------------------------------
1197 Node* GraphKit::basic_plus_adr(Node* base, Node* ptr, Node* offset) {
1198 // short-circuit a common case
1199 if (offset == intcon(0)) return ptr;
1200 return _gvn.transform( new AddPNode(base, ptr, offset) );
1201 }
1202
1203 Node* GraphKit::ConvI2L(Node* offset) {
1204 // short-circuit a common case
1205 jint offset_con = find_int_con(offset, Type::OffsetBot);
1206 if (offset_con != Type::OffsetBot) {
1207 return longcon((jlong) offset_con);
1208 }
1209 return _gvn.transform( new ConvI2LNode(offset));
1210 }
1211
1212 Node* GraphKit::ConvI2UL(Node* offset) {
1213 juint offset_con = (juint) find_int_con(offset, Type::OffsetBot);
1214 if (offset_con != (juint) Type::OffsetBot) {
1215 return longcon((julong) offset_con);
1216 }
1217 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1218 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1219 return _gvn.transform( new AndLNode(conv, mask) );
1220 }
1221
1222 Node* GraphKit::ConvL2I(Node* offset) {
1223 // short-circuit a common case
1224 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1225 if (offset_con != (jlong)Type::OffsetBot) {
1226 return intcon((int) offset_con);
1227 }
1228 return _gvn.transform( new ConvL2INode(offset));
1229 }
1230
1231 //-------------------------load_object_klass-----------------------------------
1232 Node* GraphKit::load_object_klass(Node* obj, bool fold_for_arrays) {
1233 // Special-case a fresh allocation to avoid building nodes:
1234 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1235 if (akls != nullptr) return akls;
1236 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1237 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT, fold_for_arrays));
1238 }
1239
1240 //-------------------------load_array_length-----------------------------------
1241 Node* GraphKit::load_array_length(Node* array) {
1242 // Special-case a fresh allocation to avoid building nodes:
1243 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1244 Node *alen;
1245 if (alloc == nullptr) {
1246 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1247 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1248 } else {
1249 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1250 }
1251 return alen;
1252 }
1253
1254 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1255 const TypeOopPtr* oop_type,
1256 bool replace_length_in_map) {
1257 Node* length = alloc->Ideal_length();
1258 if (replace_length_in_map == false || map()->find_edge(length) >= 0) {
1259 Node* ccast = alloc->make_ideal_length(oop_type, &_gvn);
1260 if (ccast != length) {
1261 // do not transform ccast here, it might convert to top node for
1262 // negative array length and break assumptions in parsing stage.
1263 _gvn.set_type_bottom(ccast);
1264 record_for_igvn(ccast);
1265 if (replace_length_in_map) {
1266 replace_in_map(length, ccast);
1267 }
1268 return ccast;
1269 }
1270 }
1271 return length;
1272 }
1273
1274 //------------------------------do_null_check----------------------------------
1275 // Helper function to do a null pointer check. Returned value is
1276 // the incoming address with null casted away. You are allowed to use the
1277 // not-null value only if you are control dependent on the test.
1278 #ifndef PRODUCT
1279 extern uint explicit_null_checks_inserted,
1280 explicit_null_checks_elided;
1281 #endif
1282 Node* GraphKit::null_check_common(Node* value, BasicType type,
1283 // optional arguments for variations:
1284 bool assert_null,
1285 Node* *null_control,
1286 bool speculative,
1287 bool is_init_check) {
1288 assert(!assert_null || null_control == nullptr, "not both at once");
1289 if (stopped()) return top();
1290 NOT_PRODUCT(explicit_null_checks_inserted++);
1291
1292 if (value->is_InlineType()) {
1293 // Null checking a scalarized but nullable inline type. Check the null marker
1294 // input instead of the oop input to avoid keeping buffer allocations alive.
1295 InlineTypeNode* vtptr = value->as_InlineType();
1296 while (vtptr->get_oop()->is_InlineType()) {
1297 vtptr = vtptr->get_oop()->as_InlineType();
1298 }
1299 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1300 if (stopped()) {
1301 return top();
1302 }
1303 if (assert_null) {
1304 // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
1305 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1306 // replace_in_map(value, vtptr);
1307 // return vtptr;
1308 replace_in_map(value, null());
1309 return null();
1310 }
1311 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1312 return cast_not_null(value, do_replace_in_map);
1313 }
1314
1315 // Construct null check
1316 Node *chk = nullptr;
1317 switch(type) {
1318 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1319 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1320 case T_ARRAY : // fall through
1321 type = T_OBJECT; // simplify further tests
1322 case T_OBJECT : {
1323 const Type *t = _gvn.type( value );
1324
1325 const TypeOopPtr* tp = t->isa_oopptr();
1326 if (tp != nullptr && !tp->is_loaded()
1327 // Only for do_null_check, not any of its siblings:
1328 && !assert_null && null_control == nullptr) {
1329 // Usually, any field access or invocation on an unloaded oop type
1330 // will simply fail to link, since the statically linked class is
1331 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1332 // the static class is loaded but the sharper oop type is not.
1333 // Rather than checking for this obscure case in lots of places,
1334 // we simply observe that a null check on an unloaded class
1335 // will always be followed by a nonsense operation, so we
1336 // can just issue the uncommon trap here.
1337 // Our access to the unloaded class will only be correct
1338 // after it has been loaded and initialized, which requires
1339 // a trip through the interpreter.
1340 ciKlass* klass = tp->unloaded_klass();
1341 #ifndef PRODUCT
1342 if (WizardMode) { tty->print("Null check of unloaded "); klass->print(); tty->cr(); }
1343 #endif
1344 uncommon_trap(Deoptimization::Reason_unloaded,
1345 Deoptimization::Action_reinterpret,
1346 klass, "!loaded");
1347 return top();
1348 }
1349
1350 if (assert_null) {
1351 // See if the type is contained in NULL_PTR.
1352 // If so, then the value is already null.
1353 if (t->higher_equal(TypePtr::NULL_PTR)) {
1354 NOT_PRODUCT(explicit_null_checks_elided++);
1355 return value; // Elided null assert quickly!
1356 }
1357 } else {
1358 // See if mixing in the null pointer changes type.
1359 // If so, then the null pointer was not allowed in the original
1360 // type. In other words, "value" was not-null.
1361 if (t->meet(TypePtr::NULL_PTR) != t->remove_speculative()) {
1362 // same as: if (!TypePtr::NULL_PTR->higher_equal(t)) ...
1363 NOT_PRODUCT(explicit_null_checks_elided++);
1364 return value; // Elided null check quickly!
1365 }
1366 }
1367 chk = new CmpPNode( value, null() );
1368 break;
1369 }
1370
1371 default:
1372 fatal("unexpected type: %s", type2name(type));
1373 }
1374 assert(chk != nullptr, "sanity check");
1375 chk = _gvn.transform(chk);
1376
1377 BoolTest::mask btest = assert_null ? BoolTest::eq : BoolTest::ne;
1378 BoolNode *btst = new BoolNode( chk, btest);
1379 Node *tst = _gvn.transform( btst );
1380
1381 //-----------
1382 // if peephole optimizations occurred, a prior test existed.
1383 // If a prior test existed, maybe it dominates as we can avoid this test.
1384 if (tst != btst && type == T_OBJECT) {
1385 // At this point we want to scan up the CFG to see if we can
1386 // find an identical test (and so avoid this test altogether).
1387 Node *cfg = control();
1388 int depth = 0;
1389 while( depth < 16 ) { // Limit search depth for speed
1390 if( cfg->Opcode() == Op_IfTrue &&
1391 cfg->in(0)->in(1) == tst ) {
1392 // Found prior test. Use "cast_not_null" to construct an identical
1393 // CastPP (and hence hash to) as already exists for the prior test.
1394 // Return that casted value.
1395 if (assert_null) {
1396 replace_in_map(value, null());
1397 return null(); // do not issue the redundant test
1398 }
1399 Node *oldcontrol = control();
1400 set_control(cfg);
1401 Node *res = cast_not_null(value);
1402 set_control(oldcontrol);
1403 NOT_PRODUCT(explicit_null_checks_elided++);
1404 return res;
1405 }
1406 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1407 if (cfg == nullptr) break; // Quit at region nodes
1408 depth++;
1409 }
1410 }
1411
1412 //-----------
1413 // Branch to failure if null
1414 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1415 Deoptimization::DeoptReason reason;
1416 if (assert_null) {
1417 reason = Deoptimization::reason_null_assert(speculative);
1418 } else if (type == T_OBJECT || is_init_check) {
1419 reason = Deoptimization::reason_null_check(speculative);
1420 } else {
1421 reason = Deoptimization::Reason_div0_check;
1422 }
1423 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1424 // ciMethodData::has_trap_at will return a conservative -1 if any
1425 // must-be-null assertion has failed. This could cause performance
1426 // problems for a method after its first do_null_assert failure.
1427 // Consider using 'Reason_class_check' instead?
1428
1429 // To cause an implicit null check, we set the not-null probability
1430 // to the maximum (PROB_MAX). For an explicit check the probability
1431 // is set to a smaller value.
1432 if (null_control != nullptr || too_many_traps(reason)) {
1433 // probability is less likely
1434 ok_prob = PROB_LIKELY_MAG(3);
1435 } else if (!assert_null &&
1436 (ImplicitNullCheckThreshold > 0) &&
1437 method() != nullptr &&
1438 (method()->method_data()->trap_count(reason)
1439 >= (uint)ImplicitNullCheckThreshold)) {
1440 ok_prob = PROB_LIKELY_MAG(3);
1441 }
1442
1443 if (null_control != nullptr) {
1444 IfNode* iff = create_and_map_if(control(), tst, ok_prob, COUNT_UNKNOWN);
1445 Node* null_true = _gvn.transform( new IfFalseNode(iff));
1446 set_control( _gvn.transform( new IfTrueNode(iff)));
1447 #ifndef PRODUCT
1448 if (null_true == top()) {
1449 explicit_null_checks_elided++;
1450 }
1451 #endif
1452 (*null_control) = null_true;
1453 } else {
1454 BuildCutout unless(this, tst, ok_prob);
1455 // Check for optimizer eliding test at parse time
1456 if (stopped()) {
1457 // Failure not possible; do not bother making uncommon trap.
1458 NOT_PRODUCT(explicit_null_checks_elided++);
1459 } else if (assert_null) {
1460 uncommon_trap(reason,
1461 Deoptimization::Action_make_not_entrant,
1462 nullptr, "assert_null");
1463 } else {
1464 replace_in_map(value, zerocon(type));
1465 builtin_throw(reason);
1466 }
1467 }
1468
1469 // Must throw exception, fall-thru not possible?
1470 if (stopped()) {
1471 return top(); // No result
1472 }
1473
1474 if (assert_null) {
1475 // Cast obj to null on this path.
1476 replace_in_map(value, zerocon(type));
1477 return zerocon(type);
1478 }
1479
1480 // Cast obj to not-null on this path, if there is no null_control.
1481 // (If there is a null_control, a non-null value may come back to haunt us.)
1482 if (type == T_OBJECT) {
1483 Node* cast = cast_not_null(value, false);
1484 if (null_control == nullptr || (*null_control) == top())
1485 replace_in_map(value, cast);
1486 value = cast;
1487 }
1488
1489 return value;
1490 }
1491
1492 //------------------------------cast_not_null----------------------------------
1493 // Cast obj to not-null on this path
1494 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1495 if (obj->is_InlineType()) {
1496 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1497 vt->as_InlineType()->set_null_marker(_gvn);
1498 vt = _gvn.transform(vt);
1499 if (do_replace_in_map) {
1500 replace_in_map(obj, vt);
1501 }
1502 return vt;
1503 }
1504 const Type *t = _gvn.type(obj);
1505 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1506 // Object is already not-null?
1507 if( t == t_not_null ) return obj;
1508
1509 Node* cast = new CastPPNode(control(), obj,t_not_null);
1510 cast = _gvn.transform( cast );
1511
1512 // Scan for instances of 'obj' in the current JVM mapping.
1513 // These instances are known to be not-null after the test.
1514 if (do_replace_in_map)
1515 replace_in_map(obj, cast);
1516
1517 return cast; // Return casted value
1518 }
1519
1520 Node* GraphKit::cast_to_non_larval(Node* obj) {
1521 const Type* obj_type = gvn().type(obj);
1522 if (obj->is_InlineType() || !obj_type->is_inlinetypeptr()) {
1523 return obj;
1524 }
1525
1526 Node* new_obj = InlineTypeNode::make_from_oop(this, obj, obj_type->inline_klass());
1527 replace_in_map(obj, new_obj);
1528 return new_obj;
1529 }
1530
1531 // Sometimes in intrinsics, we implicitly know an object is not null
1532 // (there's no actual null check) so we can cast it to not null. In
1533 // the course of optimizations, the input to the cast can become null.
1534 // In that case that data path will die and we need the control path
1535 // to become dead as well to keep the graph consistent. So we have to
1536 // add a check for null for which one branch can't be taken. It uses
1537 // an OpaqueNotNull node that will cause the check to be removed after loop
1538 // opts so the test goes away and the compiled code doesn't execute a
1539 // useless check.
1540 Node* GraphKit::must_be_not_null(Node* value, bool do_replace_in_map) {
1541 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(value))) {
1542 return value;
1543 }
1544 Node* chk = _gvn.transform(new CmpPNode(value, null()));
1545 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::ne));
1546 Node* opaq = _gvn.transform(new OpaqueNotNullNode(C, tst));
1547 IfNode* iff = new IfNode(control(), opaq, PROB_MAX, COUNT_UNKNOWN);
1548 _gvn.set_type(iff, iff->Value(&_gvn));
1549 if (!tst->is_Con()) {
1550 record_for_igvn(iff);
1551 }
1552 Node *if_f = _gvn.transform(new IfFalseNode(iff));
1553 Node *frame = _gvn.transform(new ParmNode(C->start(), TypeFunc::FramePtr));
1554 Node* halt = _gvn.transform(new HaltNode(if_f, frame, "unexpected null in intrinsic"));
1555 C->root()->add_req(halt);
1556 Node *if_t = _gvn.transform(new IfTrueNode(iff));
1557 set_control(if_t);
1558 return cast_not_null(value, do_replace_in_map);
1559 }
1560
1561
1562 //--------------------------replace_in_map-------------------------------------
1563 void GraphKit::replace_in_map(Node* old, Node* neww) {
1564 if (old == neww) {
1565 return;
1566 }
1567
1568 map()->replace_edge(old, neww);
1569
1570 // Note: This operation potentially replaces any edge
1571 // on the map. This includes locals, stack, and monitors
1572 // of the current (innermost) JVM state.
1573
1574 // don't let inconsistent types from profiling escape this
1575 // method
1576
1577 const Type* told = _gvn.type(old);
1578 const Type* tnew = _gvn.type(neww);
1579
1580 if (!tnew->higher_equal(told)) {
1581 return;
1582 }
1583
1584 map()->record_replaced_node(old, neww);
1585 }
1586
1587
1588 //=============================================================================
1589 //--------------------------------memory---------------------------------------
1590 Node* GraphKit::memory(uint alias_idx) {
1591 MergeMemNode* mem = merged_memory();
1592 Node* p = mem->memory_at(alias_idx);
1593 assert(p != mem->empty_memory(), "empty");
1594 _gvn.set_type(p, Type::MEMORY); // must be mapped
1595 return p;
1596 }
1597
1598 //-----------------------------reset_memory------------------------------------
1599 Node* GraphKit::reset_memory() {
1600 Node* mem = map()->memory();
1601 // do not use this node for any more parsing!
1602 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1603 return _gvn.transform( mem );
1604 }
1605
1606 //------------------------------set_all_memory---------------------------------
1607 void GraphKit::set_all_memory(Node* newmem) {
1608 Node* mergemem = MergeMemNode::make(newmem);
1609 gvn().set_type_bottom(mergemem);
1610 map()->set_memory(mergemem);
1611 }
1612
1613 //------------------------------set_all_memory_call----------------------------
1614 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1615 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1616 set_all_memory(newmem);
1617 }
1618
1619 //=============================================================================
1620 //
1621 // parser factory methods for MemNodes
1622 //
1623 // These are layered on top of the factory methods in LoadNode and StoreNode,
1624 // and integrate with the parser's memory state and _gvn engine.
1625 //
1626
1627 // factory methods in "int adr_idx"
1628 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1629 MemNode::MemOrd mo,
1630 LoadNode::ControlDependency control_dependency,
1631 bool require_atomic_access,
1632 bool unaligned,
1633 bool mismatched,
1634 bool unsafe,
1635 uint8_t barrier_data) {
1636 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1637 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1638 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1639 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1640 Node* mem = memory(adr_idx);
1641 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1642 ld = _gvn.transform(ld);
1643
1644 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1645 // Improve graph before escape analysis and boxing elimination.
1646 record_for_igvn(ld);
1647 if (ld->is_DecodeN()) {
1648 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1649 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1650 // a Phi). Recording such cases is still perfectly sound, but may be
1651 // unnecessary and result in some minor IGVN overhead.
1652 record_for_igvn(ld->in(1));
1653 }
1654 }
1655 return ld;
1656 }
1657
1658 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1659 MemNode::MemOrd mo,
1660 bool require_atomic_access,
1661 bool unaligned,
1662 bool mismatched,
1663 bool unsafe,
1664 int barrier_data) {
1665 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1666 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
1667 const TypePtr* adr_type = nullptr;
1668 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1669 Node *mem = memory(adr_idx);
1670 Node* st = StoreNode::make(_gvn, ctl, mem, adr, adr_type, val, bt, mo, require_atomic_access);
1671 if (unaligned) {
1672 st->as_Store()->set_unaligned_access();
1673 }
1674 if (mismatched) {
1675 st->as_Store()->set_mismatched_access();
1676 }
1677 if (unsafe) {
1678 st->as_Store()->set_unsafe_access();
1679 }
1680 st->as_Store()->set_barrier_data(barrier_data);
1681 st = _gvn.transform(st);
1682 set_memory(st, adr_idx);
1683 // Back-to-back stores can only remove intermediate store with DU info
1684 // so push on worklist for optimizer.
1685 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1686 record_for_igvn(st);
1687
1688 return st;
1689 }
1690
1691 Node* GraphKit::access_store_at(Node* obj,
1692 Node* adr,
1693 const TypePtr* adr_type,
1694 Node* val,
1695 const Type* val_type,
1696 BasicType bt,
1697 DecoratorSet decorators,
1698 bool safe_for_replace,
1699 const InlineTypeNode* vt) {
1700 // Transformation of a value which could be null pointer (CastPP #null)
1701 // could be delayed during Parse (for example, in adjust_map_after_if()).
1702 // Execute transformation here to avoid barrier generation in such case.
1703 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1704 val = _gvn.makecon(TypePtr::NULL_PTR);
1705 }
1706
1707 if (stopped()) {
1708 return top(); // Dead path ?
1709 }
1710
1711 assert(val != nullptr, "not dead path");
1712 if (val->is_InlineType()) {
1713 // Store to non-flat field. Buffer the inline type and make sure
1714 // the store is re-executed if the allocation triggers deoptimization.
1715 PreserveReexecuteState preexecs(this);
1716 jvms()->set_should_reexecute(true);
1717 val = val->as_InlineType()->buffer(this, safe_for_replace);
1718 }
1719
1720 C2AccessValuePtr addr(adr, adr_type);
1721 C2AccessValue value(val, val_type);
1722 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1723 if (access.is_raw()) {
1724 return _barrier_set->BarrierSetC2::store_at(access, value);
1725 } else {
1726 return _barrier_set->store_at(access, value);
1727 }
1728 }
1729
1730 Node* GraphKit::access_load_at(Node* obj, // containing obj
1731 Node* adr, // actual address to store val at
1732 const TypePtr* adr_type,
1733 const Type* val_type,
1734 BasicType bt,
1735 DecoratorSet decorators,
1736 Node* ctl) {
1737 if (stopped()) {
1738 return top(); // Dead path ?
1739 }
1740
1741 C2AccessValuePtr addr(adr, adr_type);
1742 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1743 if (access.is_raw()) {
1744 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1745 } else {
1746 return _barrier_set->load_at(access, val_type);
1747 }
1748 }
1749
1750 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1751 const Type* val_type,
1752 BasicType bt,
1753 DecoratorSet decorators) {
1754 if (stopped()) {
1755 return top(); // Dead path ?
1756 }
1757
1758 C2AccessValuePtr addr(adr, adr->bottom_type()->is_ptr());
1759 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, nullptr, addr);
1760 if (access.is_raw()) {
1761 return _barrier_set->BarrierSetC2::load_at(access, val_type);
1762 } else {
1763 return _barrier_set->load_at(access, val_type);
1764 }
1765 }
1766
1767 Node* GraphKit::access_atomic_cmpxchg_val_at(Node* obj,
1768 Node* adr,
1769 const TypePtr* adr_type,
1770 int alias_idx,
1771 Node* expected_val,
1772 Node* new_val,
1773 const Type* value_type,
1774 BasicType bt,
1775 DecoratorSet decorators) {
1776 C2AccessValuePtr addr(adr, adr_type);
1777 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1778 bt, obj, addr, alias_idx);
1779 if (access.is_raw()) {
1780 return _barrier_set->BarrierSetC2::atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1781 } else {
1782 return _barrier_set->atomic_cmpxchg_val_at(access, expected_val, new_val, value_type);
1783 }
1784 }
1785
1786 Node* GraphKit::access_atomic_cmpxchg_bool_at(Node* obj,
1787 Node* adr,
1788 const TypePtr* adr_type,
1789 int alias_idx,
1790 Node* expected_val,
1791 Node* new_val,
1792 const Type* value_type,
1793 BasicType bt,
1794 DecoratorSet decorators) {
1795 C2AccessValuePtr addr(adr, adr_type);
1796 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1797 bt, obj, addr, alias_idx);
1798 if (access.is_raw()) {
1799 return _barrier_set->BarrierSetC2::atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1800 } else {
1801 return _barrier_set->atomic_cmpxchg_bool_at(access, expected_val, new_val, value_type);
1802 }
1803 }
1804
1805 Node* GraphKit::access_atomic_xchg_at(Node* obj,
1806 Node* adr,
1807 const TypePtr* adr_type,
1808 int alias_idx,
1809 Node* new_val,
1810 const Type* value_type,
1811 BasicType bt,
1812 DecoratorSet decorators) {
1813 C2AccessValuePtr addr(adr, adr_type);
1814 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS,
1815 bt, obj, addr, alias_idx);
1816 if (access.is_raw()) {
1817 return _barrier_set->BarrierSetC2::atomic_xchg_at(access, new_val, value_type);
1818 } else {
1819 return _barrier_set->atomic_xchg_at(access, new_val, value_type);
1820 }
1821 }
1822
1823 Node* GraphKit::access_atomic_add_at(Node* obj,
1824 Node* adr,
1825 const TypePtr* adr_type,
1826 int alias_idx,
1827 Node* new_val,
1828 const Type* value_type,
1829 BasicType bt,
1830 DecoratorSet decorators) {
1831 C2AccessValuePtr addr(adr, adr_type);
1832 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1833 if (access.is_raw()) {
1834 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1835 } else {
1836 return _barrier_set->atomic_add_at(access, new_val, value_type);
1837 }
1838 }
1839
1840 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1841 return _barrier_set->clone(this, src, dst, size, is_array);
1842 }
1843
1844 //-------------------------array_element_address-------------------------
1845 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1846 const TypeInt* sizetype, Node* ctrl) {
1847 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1848 uint shift;
1849 if (arytype->is_flat() && arytype->klass_is_exact()) {
1850 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1851 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1852 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1853 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1854 // though we don't need the address node in this case and throw it away again.
1855 shift = arytype->flat_log_elem_size();
1856 } else {
1857 shift = exact_log2(type2aelembytes(elembt));
1858 }
1859 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1860
1861 // short-circuit a common case (saves lots of confusing waste motion)
1862 jint idx_con = find_int_con(idx, -1);
1863 if (idx_con >= 0) {
1864 intptr_t offset = header + ((intptr_t)idx_con << shift);
1865 return basic_plus_adr(ary, offset);
1866 }
1867
1868 // must be correct type for alignment purposes
1869 Node* base = basic_plus_adr(ary, header);
1870 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1871 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1872 return basic_plus_adr(ary, base, scale);
1873 }
1874
1875 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* vk, bool is_null_free, bool is_not_null_free, bool is_atomic) {
1876 assert(vk->maybe_flat_in_array(), "element of type %s cannot be flat in array", vk->name()->as_utf8());
1877 if (!vk->has_nullable_atomic_layout()) {
1878 // Element does not have a nullable flat layout, cannot be nullable
1879 is_null_free = true;
1880 }
1881 if (!vk->has_atomic_layout() && !vk->has_non_atomic_layout()) {
1882 // Element does not have a null-free flat layout, cannot be null-free
1883 is_not_null_free = true;
1884 }
1885 if (is_null_free) {
1886 // TODO 8350865 Impossible type
1887 is_not_null_free = false;
1888 }
1889
1890 bool is_exact = is_null_free || is_not_null_free;
1891 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* flat */ true, is_null_free, is_atomic);
1892 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1893 arytype = arytype->cast_to_exactness(is_exact);
1894 arytype = arytype->cast_to_not_null_free(is_not_null_free);
1895 return _gvn.transform(new CastPPNode(control(), array, arytype, ConstraintCastNode::StrongDependency));
1896 }
1897
1898 //-------------------------load_array_element-------------------------
1899 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1900 const Type* elemtype = arytype->elem();
1901 BasicType elembt = elemtype->array_element_basic_type();
1902 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1903 if (elembt == T_NARROWOOP) {
1904 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1905 }
1906 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1907 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1908 return ld;
1909 }
1910
1911 //-------------------------set_arguments_for_java_call-------------------------
1912 // Arguments (pre-popped from the stack) are taken from the JVMS.
1913 void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
1914 PreserveReexecuteState preexecs(this);
1915 if (EnableValhalla) {
1916 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
1917 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
1918 jvms()->set_should_reexecute(true);
1919 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
1920 inc_sp(arg_size);
1921 }
1922 // Add the call arguments
1923 const TypeTuple* domain = call->tf()->domain_sig();
1924 uint nargs = domain->cnt();
1925 int arg_num = 0;
1926 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
1927 Node* arg = argument(i-TypeFunc::Parms);
1928 const Type* t = domain->field_at(i);
1929 // TODO 8284443 A static call to a mismatched method should still be scalarized
1930 if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
1931 // We don't pass inline type arguments by reference but instead pass each field of the inline type
1932 if (!arg->is_InlineType()) {
1933 assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
1934 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
1935 }
1936 InlineTypeNode* vt = arg->as_InlineType();
1937 vt->pass_fields(this, call, idx, true, !t->maybe_null());
1938 // If an inline type argument is passed as fields, attach the Method* to the call site
1939 // to be able to access the extended signature later via attached_method_before_pc().
1940 // For example, see CompiledMethod::preserve_callee_argument_oops().
1941 call->set_override_symbolic_info(true);
1942 // Register an evol dependency on the callee method to make sure that this method is deoptimized and
1943 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
1944 C->dependencies()->assert_evol_method(call->method());
1945 arg_num++;
1946 continue;
1947 } else if (arg->is_InlineType()) {
1948 // Pass inline type argument via oop to callee
1949 arg = arg->as_InlineType()->buffer(this, true);
1950 }
1951 if (t != Type::HALF) {
1952 arg_num++;
1953 }
1954 call->init_req(idx++, arg);
1955 }
1956 }
1957
1958 //---------------------------set_edges_for_java_call---------------------------
1959 // Connect a newly created call into the current JVMS.
1960 // A return value node (if any) is returned from set_edges_for_java_call.
1961 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1962
1963 // Add the predefined inputs:
1964 call->init_req( TypeFunc::Control, control() );
1965 call->init_req( TypeFunc::I_O , i_o() );
1966 call->init_req( TypeFunc::Memory , reset_memory() );
1967 call->init_req( TypeFunc::FramePtr, frameptr() );
1968 call->init_req( TypeFunc::ReturnAdr, top() );
1969
1970 add_safepoint_edges(call, must_throw);
1971
1972 Node* xcall = _gvn.transform(call);
1973
1974 if (xcall == top()) {
1975 set_control(top());
1976 return;
1977 }
1978 assert(xcall == call, "call identity is stable");
1979
1980 // Re-use the current map to produce the result.
1981
1982 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1983 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1984 set_all_memory_call(xcall, separate_io_proj);
1985
1986 //return xcall; // no need, caller already has it
1987 }
1988
1989 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1990 if (stopped()) return top(); // maybe the call folded up?
1991
1992 // Note: Since any out-of-line call can produce an exception,
1993 // we always insert an I_O projection from the call into the result.
1994
1995 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1996
1997 if (separate_io_proj) {
1998 // The caller requested separate projections be used by the fall
1999 // through and exceptional paths, so replace the projections for
2000 // the fall through path.
2001 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2002 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2003 }
2004
2005 // Capture the return value, if any.
2006 Node* ret;
2007 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2008 ret = top();
2009 } else if (call->tf()->returns_inline_type_as_fields()) {
2010 // Return of multiple values (inline type fields): we create a
2011 // InlineType node, each field is a projection from the call.
2012 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2013 uint base_input = TypeFunc::Parms;
2014 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2015 } else {
2016 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2017 ciType* t = call->method()->return_type();
2018 if (t->is_klass()) {
2019 const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
2020 if (type->is_inlinetypeptr()) {
2021 ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass());
2022 }
2023 }
2024 }
2025
2026 return ret;
2027 }
2028
2029 //--------------------set_predefined_input_for_runtime_call--------------------
2030 // Reading and setting the memory state is way conservative here.
2031 // The real problem is that I am not doing real Type analysis on memory,
2032 // so I cannot distinguish card mark stores from other stores. Across a GC
2033 // point the Store Barrier and the card mark memory has to agree. I cannot
2034 // have a card mark store and its barrier split across the GC point from
2035 // either above or below. Here I get that to happen by reading ALL of memory.
2036 // A better answer would be to separate out card marks from other memory.
2037 // For now, return the input memory state, so that it can be reused
2038 // after the call, if this call has restricted memory effects.
2039 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2040 // Set fixed predefined input arguments
2041 Node* memory = reset_memory();
2042 Node* m = narrow_mem == nullptr ? memory : narrow_mem;
2043 call->init_req( TypeFunc::Control, control() );
2044 call->init_req( TypeFunc::I_O, top() ); // does no i/o
2045 call->init_req( TypeFunc::Memory, m ); // may gc ptrs
2046 call->init_req( TypeFunc::FramePtr, frameptr() );
2047 call->init_req( TypeFunc::ReturnAdr, top() );
2048 return memory;
2049 }
2050
2051 //-------------------set_predefined_output_for_runtime_call--------------------
2052 // Set control and memory (not i_o) from the call.
2053 // If keep_mem is not null, use it for the output state,
2054 // except for the RawPtr output of the call, if hook_mem is TypeRawPtr::BOTTOM.
2055 // If hook_mem is null, this call produces no memory effects at all.
2056 // If hook_mem is a Java-visible memory slice (such as arraycopy operands),
2057 // then only that memory slice is taken from the call.
2058 // In the last case, we must put an appropriate memory barrier before
2059 // the call, so as to create the correct anti-dependencies on loads
2060 // preceding the call.
2061 void GraphKit::set_predefined_output_for_runtime_call(Node* call,
2062 Node* keep_mem,
2063 const TypePtr* hook_mem) {
2064 // no i/o
2065 set_control(_gvn.transform( new ProjNode(call,TypeFunc::Control) ));
2066 if (keep_mem) {
2067 // First clone the existing memory state
2068 set_all_memory(keep_mem);
2069 if (hook_mem != nullptr) {
2070 // Make memory for the call
2071 Node* mem = _gvn.transform( new ProjNode(call, TypeFunc::Memory) );
2072 // Set the RawPtr memory state only. This covers all the heap top/GC stuff
2073 // We also use hook_mem to extract specific effects from arraycopy stubs.
2074 set_memory(mem, hook_mem);
2075 }
2076 // ...else the call has NO memory effects.
2077
2078 // Make sure the call advertises its memory effects precisely.
2079 // This lets us build accurate anti-dependences in gcm.cpp.
2080 assert(C->alias_type(call->adr_type()) == C->alias_type(hook_mem),
2081 "call node must be constructed correctly");
2082 } else {
2083 assert(hook_mem == nullptr, "");
2084 // This is not a "slow path" call; all memory comes from the call.
2085 set_all_memory_call(call);
2086 }
2087 }
2088
2089 // Keep track of MergeMems feeding into other MergeMems
2090 static void add_mergemem_users_to_worklist(Unique_Node_List& wl, Node* mem) {
2091 if (!mem->is_MergeMem()) {
2092 return;
2093 }
2094 for (SimpleDUIterator i(mem); i.has_next(); i.next()) {
2095 Node* use = i.get();
2096 if (use->is_MergeMem()) {
2097 wl.push(use);
2098 }
2099 }
2100 }
2101
2102 // Replace the call with the current state of the kit.
2103 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2104 JVMState* ejvms = nullptr;
2105 if (has_exceptions()) {
2106 ejvms = transfer_exceptions_into_jvms();
2107 }
2108
2109 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2110 ReplacedNodes replaced_nodes_exception;
2111 Node* ex_ctl = top();
2112
2113 SafePointNode* final_state = stop();
2114
2115 // Find all the needed outputs of this call
2116 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2117
2118 Unique_Node_List wl;
2119 Node* init_mem = call->in(TypeFunc::Memory);
2120 Node* final_mem = final_state->in(TypeFunc::Memory);
2121 Node* final_ctl = final_state->in(TypeFunc::Control);
2122 Node* final_io = final_state->in(TypeFunc::I_O);
2123
2124 // Replace all the old call edges with the edges from the inlining result
2125 if (callprojs->fallthrough_catchproj != nullptr) {
2126 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2127 }
2128 if (callprojs->fallthrough_memproj != nullptr) {
2129 if (final_mem->is_MergeMem()) {
2130 // Parser's exits MergeMem was not transformed but may be optimized
2131 final_mem = _gvn.transform(final_mem);
2132 }
2133 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2134 add_mergemem_users_to_worklist(wl, final_mem);
2135 }
2136 if (callprojs->fallthrough_ioproj != nullptr) {
2137 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2138 }
2139
2140 // Replace the result with the new result if it exists and is used
2141 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2142 // If the inlined code is dead, the result projections for an inline type returned as
2143 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2144 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
2145 "unexpected number of results");
2146 C->gvn_replace_by(callprojs->resproj[0], result);
2147 }
2148
2149 if (ejvms == nullptr) {
2150 // No exception edges to simply kill off those paths
2151 if (callprojs->catchall_catchproj != nullptr) {
2152 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2153 }
2154 if (callprojs->catchall_memproj != nullptr) {
2155 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2156 }
2157 if (callprojs->catchall_ioproj != nullptr) {
2158 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2159 }
2160 // Replace the old exception object with top
2161 if (callprojs->exobj != nullptr) {
2162 C->gvn_replace_by(callprojs->exobj, C->top());
2163 }
2164 } else {
2165 GraphKit ekit(ejvms);
2166
2167 // Load my combined exception state into the kit, with all phis transformed:
2168 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2169 replaced_nodes_exception = ex_map->replaced_nodes();
2170
2171 Node* ex_oop = ekit.use_exception_state(ex_map);
2172
2173 if (callprojs->catchall_catchproj != nullptr) {
2174 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2175 ex_ctl = ekit.control();
2176 }
2177 if (callprojs->catchall_memproj != nullptr) {
2178 Node* ex_mem = ekit.reset_memory();
2179 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2180 add_mergemem_users_to_worklist(wl, ex_mem);
2181 }
2182 if (callprojs->catchall_ioproj != nullptr) {
2183 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2184 }
2185
2186 // Replace the old exception object with the newly created one
2187 if (callprojs->exobj != nullptr) {
2188 C->gvn_replace_by(callprojs->exobj, ex_oop);
2189 }
2190 }
2191
2192 // Disconnect the call from the graph
2193 call->disconnect_inputs(C);
2194 C->gvn_replace_by(call, C->top());
2195
2196 // Clean up any MergeMems that feed other MergeMems since the
2197 // optimizer doesn't like that.
2198 while (wl.size() > 0) {
2199 _gvn.transform(wl.pop());
2200 }
2201
2202 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2203 replaced_nodes.apply(C, final_ctl);
2204 }
2205 if (!ex_ctl->is_top() && do_replaced_nodes) {
2206 replaced_nodes_exception.apply(C, ex_ctl);
2207 }
2208 }
2209
2210
2211 //------------------------------increment_counter------------------------------
2212 // for statistics: increment a VM counter by 1
2213
2214 void GraphKit::increment_counter(address counter_addr) {
2215 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2216 increment_counter(adr1);
2217 }
2218
2219 void GraphKit::increment_counter(Node* counter_addr) {
2220 Node* ctrl = control();
2221 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2222 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2223 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2224 }
2225
2226
2227 //------------------------------uncommon_trap----------------------------------
2228 // Bail out to the interpreter in mid-method. Implemented by calling the
2229 // uncommon_trap blob. This helper function inserts a runtime call with the
2230 // right debug info.
2231 Node* GraphKit::uncommon_trap(int trap_request,
2232 ciKlass* klass, const char* comment,
2233 bool must_throw,
2234 bool keep_exact_action) {
2235 if (failing_internal()) {
2236 stop();
2237 }
2238 if (stopped()) return nullptr; // trap reachable?
2239
2240 // Note: If ProfileTraps is true, and if a deopt. actually
2241 // occurs here, the runtime will make sure an MDO exists. There is
2242 // no need to call method()->ensure_method_data() at this point.
2243
2244 // Set the stack pointer to the right value for reexecution:
2245 set_sp(reexecute_sp());
2246
2247 #ifdef ASSERT
2248 if (!must_throw) {
2249 // Make sure the stack has at least enough depth to execute
2250 // the current bytecode.
2251 int inputs, ignored_depth;
2252 if (compute_stack_effects(inputs, ignored_depth)) {
2253 assert(sp() >= inputs, "must have enough JVMS stack to execute %s: sp=%d, inputs=%d",
2254 Bytecodes::name(java_bc()), sp(), inputs);
2255 }
2256 }
2257 #endif
2258
2259 Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(trap_request);
2260 Deoptimization::DeoptAction action = Deoptimization::trap_request_action(trap_request);
2261
2262 switch (action) {
2263 case Deoptimization::Action_maybe_recompile:
2264 case Deoptimization::Action_reinterpret:
2265 // Temporary fix for 6529811 to allow virtual calls to be sure they
2266 // get the chance to go from mono->bi->mega
2267 if (!keep_exact_action &&
2268 Deoptimization::trap_request_index(trap_request) < 0 &&
2269 too_many_recompiles(reason)) {
2270 // This BCI is causing too many recompilations.
2271 if (C->log() != nullptr) {
2272 C->log()->elem("observe that='trap_action_change' reason='%s' from='%s' to='none'",
2273 Deoptimization::trap_reason_name(reason),
2274 Deoptimization::trap_action_name(action));
2275 }
2276 action = Deoptimization::Action_none;
2277 trap_request = Deoptimization::make_trap_request(reason, action);
2278 } else {
2279 C->set_trap_can_recompile(true);
2280 }
2281 break;
2282 case Deoptimization::Action_make_not_entrant:
2283 C->set_trap_can_recompile(true);
2284 break;
2285 case Deoptimization::Action_none:
2286 case Deoptimization::Action_make_not_compilable:
2287 break;
2288 default:
2289 #ifdef ASSERT
2290 fatal("unknown action %d: %s", action, Deoptimization::trap_action_name(action));
2291 #endif
2292 break;
2293 }
2294
2295 if (TraceOptoParse) {
2296 char buf[100];
2297 tty->print_cr("Uncommon trap %s at bci:%d",
2298 Deoptimization::format_trap_request(buf, sizeof(buf),
2299 trap_request), bci());
2300 }
2301
2302 CompileLog* log = C->log();
2303 if (log != nullptr) {
2304 int kid = (klass == nullptr)? -1: log->identify(klass);
2305 log->begin_elem("uncommon_trap bci='%d'", bci());
2306 char buf[100];
2307 log->print(" %s", Deoptimization::format_trap_request(buf, sizeof(buf),
2308 trap_request));
2309 if (kid >= 0) log->print(" klass='%d'", kid);
2310 if (comment != nullptr) log->print(" comment='%s'", comment);
2311 log->end_elem();
2312 }
2313
2314 // Make sure any guarding test views this path as very unlikely
2315 Node *i0 = control()->in(0);
2316 if (i0 != nullptr && i0->is_If()) { // Found a guarding if test?
2317 IfNode *iff = i0->as_If();
2318 float f = iff->_prob; // Get prob
2319 if (control()->Opcode() == Op_IfTrue) {
2320 if (f > PROB_UNLIKELY_MAG(4))
2321 iff->_prob = PROB_MIN;
2322 } else {
2323 if (f < PROB_LIKELY_MAG(4))
2324 iff->_prob = PROB_MAX;
2325 }
2326 }
2327
2328 // Clear out dead values from the debug info.
2329 kill_dead_locals();
2330
2331 // Now insert the uncommon trap subroutine call
2332 address call_addr = OptoRuntime::uncommon_trap_blob()->entry_point();
2333 const TypePtr* no_memory_effects = nullptr;
2334 // Pass the index of the class to be loaded
2335 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON |
2336 (must_throw ? RC_MUST_THROW : 0),
2337 OptoRuntime::uncommon_trap_Type(),
2338 call_addr, "uncommon_trap", no_memory_effects,
2339 intcon(trap_request));
2340 assert(call->as_CallStaticJava()->uncommon_trap_request() == trap_request,
2341 "must extract request correctly from the graph");
2342 assert(trap_request != 0, "zero value reserved by uncommon_trap_request");
2343
2344 call->set_req(TypeFunc::ReturnAdr, returnadr());
2345 // The debug info is the only real input to this call.
2346
2347 // Halt-and-catch fire here. The above call should never return!
2348 HaltNode* halt = new HaltNode(control(), frameptr(), "uncommon trap returned which should never happen"
2349 PRODUCT_ONLY(COMMA /*reachable*/false));
2350 _gvn.set_type_bottom(halt);
2351 root()->add_req(halt);
2352
2353 stop_and_kill_map();
2354 return call;
2355 }
2356
2357
2358 //--------------------------just_allocated_object------------------------------
2359 // Report the object that was just allocated.
2360 // It must be the case that there are no intervening safepoints.
2361 // We use this to determine if an object is so "fresh" that
2362 // it does not require card marks.
2363 Node* GraphKit::just_allocated_object(Node* current_control) {
2364 Node* ctrl = current_control;
2365 // Object::<init> is invoked after allocation, most of invoke nodes
2366 // will be reduced, but a region node is kept in parse time, we check
2367 // the pattern and skip the region node if it degraded to a copy.
2368 if (ctrl != nullptr && ctrl->is_Region() && ctrl->req() == 2 &&
2369 ctrl->as_Region()->is_copy()) {
2370 ctrl = ctrl->as_Region()->is_copy();
2371 }
2372 if (C->recent_alloc_ctl() == ctrl) {
2373 return C->recent_alloc_obj();
2374 }
2375 return nullptr;
2376 }
2377
2378
2379 /**
2380 * Record profiling data exact_kls for Node n with the type system so
2381 * that it can propagate it (speculation)
2382 *
2383 * @param n node that the type applies to
2384 * @param exact_kls type from profiling
2385 * @param maybe_null did profiling see null?
2386 *
2387 * @return node with improved type
2388 */
2389 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2390 const Type* current_type = _gvn.type(n);
2391 assert(UseTypeSpeculation, "type speculation must be on");
2392
2393 const TypePtr* speculative = current_type->speculative();
2394
2395 // Should the klass from the profile be recorded in the speculative type?
2396 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2397 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2398 const TypeOopPtr* xtype = tklass->as_instance_type();
2399 assert(xtype->klass_is_exact(), "Should be exact");
2400 // Any reason to believe n is not null (from this profiling or a previous one)?
2401 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2402 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2403 // record the new speculative type's depth
2404 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2405 speculative = speculative->with_inline_depth(jvms()->depth());
2406 } else if (current_type->would_improve_ptr(ptr_kind)) {
2407 // Profiling report that null was never seen so we can change the
2408 // speculative type to non null ptr.
2409 if (ptr_kind == ProfileAlwaysNull) {
2410 speculative = TypePtr::NULL_PTR;
2411 } else {
2412 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2413 const TypePtr* ptr = TypePtr::NOTNULL;
2414 if (speculative != nullptr) {
2415 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2416 } else {
2417 speculative = ptr;
2418 }
2419 }
2420 }
2421
2422 if (speculative != current_type->speculative()) {
2423 // Build a type with a speculative type (what we think we know
2424 // about the type but will need a guard when we use it)
2425 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2426 // We're changing the type, we need a new CheckCast node to carry
2427 // the new type. The new type depends on the control: what
2428 // profiling tells us is only valid from here as far as we can
2429 // tell.
2430 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2431 cast = _gvn.transform(cast);
2432 replace_in_map(n, cast);
2433 n = cast;
2434 }
2435
2436 return n;
2437 }
2438
2439 /**
2440 * Record profiling data from receiver profiling at an invoke with the
2441 * type system so that it can propagate it (speculation)
2442 *
2443 * @param n receiver node
2444 *
2445 * @return node with improved type
2446 */
2447 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2448 if (!UseTypeSpeculation) {
2449 return n;
2450 }
2451 ciKlass* exact_kls = profile_has_unique_klass();
2452 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2453 if ((java_bc() == Bytecodes::_checkcast ||
2454 java_bc() == Bytecodes::_instanceof ||
2455 java_bc() == Bytecodes::_aastore) &&
2456 method()->method_data()->is_mature()) {
2457 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2458 if (data != nullptr) {
2459 if (java_bc() == Bytecodes::_aastore) {
2460 ciKlass* array_type = nullptr;
2461 ciKlass* element_type = nullptr;
2462 ProfilePtrKind element_ptr = ProfileMaybeNull;
2463 bool flat_array = true;
2464 bool null_free_array = true;
2465 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2466 exact_kls = element_type;
2467 ptr_kind = element_ptr;
2468 } else {
2469 if (!data->as_BitData()->null_seen()) {
2470 ptr_kind = ProfileNeverNull;
2471 } else {
2472 assert(data->is_ReceiverTypeData(), "bad profile data type");
2473 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2474 uint i = 0;
2475 for (; i < call->row_limit(); i++) {
2476 ciKlass* receiver = call->receiver(i);
2477 if (receiver != nullptr) {
2478 break;
2479 }
2480 }
2481 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2482 }
2483 }
2484 }
2485 }
2486 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2487 }
2488
2489 /**
2490 * Record profiling data from argument profiling at an invoke with the
2491 * type system so that it can propagate it (speculation)
2492 *
2493 * @param dest_method target method for the call
2494 * @param bc what invoke bytecode is this?
2495 */
2496 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2497 if (!UseTypeSpeculation) {
2498 return;
2499 }
2500 const TypeFunc* tf = TypeFunc::make(dest_method);
2501 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2502 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2503 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2504 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2505 if (is_reference_type(targ->basic_type())) {
2506 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2507 ciKlass* better_type = nullptr;
2508 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2509 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2510 }
2511 i++;
2512 }
2513 }
2514 }
2515
2516 /**
2517 * Record profiling data from parameter profiling at an invoke with
2518 * the type system so that it can propagate it (speculation)
2519 */
2520 void GraphKit::record_profiled_parameters_for_speculation() {
2521 if (!UseTypeSpeculation) {
2522 return;
2523 }
2524 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2525 if (_gvn.type(local(i))->isa_oopptr()) {
2526 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2527 ciKlass* better_type = nullptr;
2528 if (method()->parameter_profiled_type(j, better_type, ptr_kind)) {
2529 record_profile_for_speculation(local(i), better_type, ptr_kind);
2530 }
2531 j++;
2532 }
2533 }
2534 }
2535
2536 /**
2537 * Record profiling data from return value profiling at an invoke with
2538 * the type system so that it can propagate it (speculation)
2539 */
2540 void GraphKit::record_profiled_return_for_speculation() {
2541 if (!UseTypeSpeculation) {
2542 return;
2543 }
2544 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2545 ciKlass* better_type = nullptr;
2546 if (method()->return_profiled_type(bci(), better_type, ptr_kind)) {
2547 // If profiling reports a single type for the return value,
2548 // feed it to the type system so it can propagate it as a
2549 // speculative type
2550 record_profile_for_speculation(stack(sp()-1), better_type, ptr_kind);
2551 }
2552 }
2553
2554
2555 //=============================================================================
2556 // Generate a fast path/slow path idiom. Graph looks like:
2557 // [foo] indicates that 'foo' is a parameter
2558 //
2559 // [in] null
2560 // \ /
2561 // CmpP
2562 // Bool ne
2563 // If
2564 // / \
2565 // True False-<2>
2566 // / |
2567 // / cast_not_null
2568 // Load | | ^
2569 // [fast_test] | |
2570 // gvn to opt_test | |
2571 // / \ | <1>
2572 // True False |
2573 // | \\ |
2574 // [slow_call] \[fast_result]
2575 // Ctl Val \ \
2576 // | \ \
2577 // Catch <1> \ \
2578 // / \ ^ \ \
2579 // Ex No_Ex | \ \
2580 // | \ \ | \ <2> \
2581 // ... \ [slow_res] | | \ [null_result]
2582 // \ \--+--+--- | |
2583 // \ | / \ | /
2584 // --------Region Phi
2585 //
2586 //=============================================================================
2587 // Code is structured as a series of driver functions all called 'do_XXX' that
2588 // call a set of helper functions. Helper functions first, then drivers.
2589
2590 //------------------------------null_check_oop---------------------------------
2591 // Null check oop. Set null-path control into Region in slot 3.
2592 // Make a cast-not-nullness use the other not-null control. Return cast.
2593 Node* GraphKit::null_check_oop(Node* value, Node* *null_control,
2594 bool never_see_null,
2595 bool safe_for_replace,
2596 bool speculative) {
2597 // Initial null check taken path
2598 (*null_control) = top();
2599 Node* cast = null_check_common(value, T_OBJECT, false, null_control, speculative);
2600
2601 // Generate uncommon_trap:
2602 if (never_see_null && (*null_control) != top()) {
2603 // If we see an unexpected null at a check-cast we record it and force a
2604 // recompile; the offending check-cast will be compiled to handle nulls.
2605 // If we see more than one offending BCI, then all checkcasts in the
2606 // method will be compiled to handle nulls.
2607 PreserveJVMState pjvms(this);
2608 set_control(*null_control);
2609 replace_in_map(value, null());
2610 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculative);
2611 uncommon_trap(reason,
2612 Deoptimization::Action_make_not_entrant);
2613 (*null_control) = top(); // null path is dead
2614 }
2615 if ((*null_control) == top() && safe_for_replace) {
2616 replace_in_map(value, cast);
2617 }
2618
2619 // Cast away null-ness on the result
2620 return cast;
2621 }
2622
2623 //------------------------------opt_iff----------------------------------------
2624 // Optimize the fast-check IfNode. Set the fast-path region slot 2.
2625 // Return slow-path control.
2626 Node* GraphKit::opt_iff(Node* region, Node* iff) {
2627 IfNode *opt_iff = _gvn.transform(iff)->as_If();
2628
2629 // Fast path taken; set region slot 2
2630 Node *fast_taken = _gvn.transform( new IfFalseNode(opt_iff) );
2631 region->init_req(2,fast_taken); // Capture fast-control
2632
2633 // Fast path not-taken, i.e. slow path
2634 Node *slow_taken = _gvn.transform( new IfTrueNode(opt_iff) );
2635 return slow_taken;
2636 }
2637
2638 //-----------------------------make_runtime_call-------------------------------
2639 Node* GraphKit::make_runtime_call(int flags,
2640 const TypeFunc* call_type, address call_addr,
2641 const char* call_name,
2642 const TypePtr* adr_type,
2643 // The following parms are all optional.
2644 // The first null ends the list.
2645 Node* parm0, Node* parm1,
2646 Node* parm2, Node* parm3,
2647 Node* parm4, Node* parm5,
2648 Node* parm6, Node* parm7) {
2649 assert(call_addr != nullptr, "must not call null targets");
2650
2651 // Slow-path call
2652 bool is_leaf = !(flags & RC_NO_LEAF);
2653 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2654 if (call_name == nullptr) {
2655 assert(!is_leaf, "must supply name for leaf");
2656 call_name = OptoRuntime::stub_name(call_addr);
2657 }
2658 CallNode* call;
2659 if (!is_leaf) {
2660 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2661 } else if (flags & RC_NO_FP) {
2662 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2663 } else if (flags & RC_VECTOR){
2664 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2665 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2666 } else {
2667 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2668 }
2669
2670 // The following is similar to set_edges_for_java_call,
2671 // except that the memory effects of the call are restricted to AliasIdxRaw.
2672
2673 // Slow path call has no side-effects, uses few values
2674 bool wide_in = !(flags & RC_NARROW_MEM);
2675 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2676
2677 Node* prev_mem = nullptr;
2678 if (wide_in) {
2679 prev_mem = set_predefined_input_for_runtime_call(call);
2680 } else {
2681 assert(!wide_out, "narrow in => narrow out");
2682 Node* narrow_mem = memory(adr_type);
2683 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2684 }
2685
2686 // Hook each parm in order. Stop looking at the first null.
2687 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2688 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2689 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2690 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2691 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2692 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2693 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2694 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2695 /* close each nested if ===> */ } } } } } } } }
2696 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2697
2698 if (!is_leaf) {
2699 // Non-leaves can block and take safepoints:
2700 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2701 }
2702 // Non-leaves can throw exceptions:
2703 if (has_io) {
2704 call->set_req(TypeFunc::I_O, i_o());
2705 }
2706
2707 if (flags & RC_UNCOMMON) {
2708 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2709 // (An "if" probability corresponds roughly to an unconditional count.
2710 // Sort of.)
2711 call->set_cnt(PROB_UNLIKELY_MAG(4));
2712 }
2713
2714 Node* c = _gvn.transform(call);
2715 assert(c == call, "cannot disappear");
2716
2717 if (wide_out) {
2718 // Slow path call has full side-effects.
2719 set_predefined_output_for_runtime_call(call);
2720 } else {
2721 // Slow path call has few side-effects, and/or sets few values.
2722 set_predefined_output_for_runtime_call(call, prev_mem, adr_type);
2723 }
2724
2725 if (has_io) {
2726 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2727 }
2728 return call;
2729
2730 }
2731
2732 // i2b
2733 Node* GraphKit::sign_extend_byte(Node* in) {
2734 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2735 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2736 }
2737
2738 // i2s
2739 Node* GraphKit::sign_extend_short(Node* in) {
2740 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2741 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2742 }
2743
2744
2745 //------------------------------merge_memory-----------------------------------
2746 // Merge memory from one path into the current memory state.
2747 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2748 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2749 Node* old_slice = mms.force_memory();
2750 Node* new_slice = mms.memory2();
2751 if (old_slice != new_slice) {
2752 PhiNode* phi;
2753 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2754 if (mms.is_empty()) {
2755 // clone base memory Phi's inputs for this memory slice
2756 assert(old_slice == mms.base_memory(), "sanity");
2757 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2758 _gvn.set_type(phi, Type::MEMORY);
2759 for (uint i = 1; i < phi->req(); i++) {
2760 phi->init_req(i, old_slice->in(i));
2761 }
2762 } else {
2763 phi = old_slice->as_Phi(); // Phi was generated already
2764 }
2765 } else {
2766 phi = PhiNode::make(region, old_slice, Type::MEMORY, mms.adr_type(C));
2767 _gvn.set_type(phi, Type::MEMORY);
2768 }
2769 phi->set_req(new_path, new_slice);
2770 mms.set_memory(phi);
2771 }
2772 }
2773 }
2774
2775 //------------------------------make_slow_call_ex------------------------------
2776 // Make the exception handler hookups for the slow call
2777 void GraphKit::make_slow_call_ex(Node* call, ciInstanceKlass* ex_klass, bool separate_io_proj, bool deoptimize) {
2778 if (stopped()) return;
2779
2780 // Make a catch node with just two handlers: fall-through and catch-all
2781 Node* i_o = _gvn.transform( new ProjNode(call, TypeFunc::I_O, separate_io_proj) );
2782 Node* catc = _gvn.transform( new CatchNode(control(), i_o, 2) );
2783 Node* norm = new CatchProjNode(catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci);
2784 _gvn.set_type_bottom(norm);
2785 C->record_for_igvn(norm);
2786 Node* excp = _gvn.transform( new CatchProjNode(catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci) );
2787
2788 { PreserveJVMState pjvms(this);
2789 set_control(excp);
2790 set_i_o(i_o);
2791
2792 if (excp != top()) {
2793 if (deoptimize) {
2794 // Deoptimize if an exception is caught. Don't construct exception state in this case.
2795 uncommon_trap(Deoptimization::Reason_unhandled,
2796 Deoptimization::Action_none);
2797 } else {
2798 // Create an exception state also.
2799 // Use an exact type if the caller has a specific exception.
2800 const Type* ex_type = TypeOopPtr::make_from_klass_unique(ex_klass)->cast_to_ptr_type(TypePtr::NotNull);
2801 Node* ex_oop = new CreateExNode(ex_type, control(), i_o);
2802 add_exception_state(make_exception_state(_gvn.transform(ex_oop)));
2803 }
2804 }
2805 }
2806
2807 // Get the no-exception control from the CatchNode.
2808 set_control(norm);
2809 }
2810
2811 static IfNode* gen_subtype_check_compare(Node* ctrl, Node* in1, Node* in2, BoolTest::mask test, float p, PhaseGVN& gvn, BasicType bt) {
2812 Node* cmp = nullptr;
2813 switch(bt) {
2814 case T_INT: cmp = new CmpINode(in1, in2); break;
2815 case T_ADDRESS: cmp = new CmpPNode(in1, in2); break;
2816 default: fatal("unexpected comparison type %s", type2name(bt));
2817 }
2818 cmp = gvn.transform(cmp);
2819 Node* bol = gvn.transform(new BoolNode(cmp, test));
2820 IfNode* iff = new IfNode(ctrl, bol, p, COUNT_UNKNOWN);
2821 gvn.transform(iff);
2822 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2823 return iff;
2824 }
2825
2826 //-------------------------------gen_subtype_check-----------------------------
2827 // Generate a subtyping check. Takes as input the subtype and supertype.
2828 // Returns 2 values: sets the default control() to the true path and returns
2829 // the false path. Only reads invariant memory; sets no (visible) memory.
2830 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2831 // but that's not exposed to the optimizer. This call also doesn't take in an
2832 // Object; if you wish to check an Object you need to load the Object's class
2833 // prior to coming here.
2834 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2835 ciMethod* method, int bci) {
2836 Compile* C = gvn.C;
2837 if ((*ctrl)->is_top()) {
2838 return C->top();
2839 }
2840
2841 // Fast check for identical types, perhaps identical constants.
2842 // The types can even be identical non-constants, in cases
2843 // involving Array.newInstance, Object.clone, etc.
2844 if (subklass == superklass)
2845 return C->top(); // false path is dead; no test needed.
2846
2847 if (gvn.type(superklass)->singleton()) {
2848 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2849 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2850
2851 // In the common case of an exact superklass, try to fold up the
2852 // test before generating code. You may ask, why not just generate
2853 // the code and then let it fold up? The answer is that the generated
2854 // code will necessarily include null checks, which do not always
2855 // completely fold away. If they are also needless, then they turn
2856 // into a performance loss. Example:
2857 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2858 // Here, the type of 'fa' is often exact, so the store check
2859 // of fa[1]=x will fold up, without testing the nullness of x.
2860 //
2861 // At macro expansion, we would have already folded the SubTypeCheckNode
2862 // being expanded here because we always perform the static sub type
2863 // check in SubTypeCheckNode::sub() regardless of whether
2864 // StressReflectiveCode is set or not. We can therefore skip this
2865 // static check when StressReflectiveCode is on.
2866 switch (C->static_subtype_check(superk, subk)) {
2867 case Compile::SSC_always_false:
2868 {
2869 Node* always_fail = *ctrl;
2870 *ctrl = gvn.C->top();
2871 return always_fail;
2872 }
2873 case Compile::SSC_always_true:
2874 return C->top();
2875 case Compile::SSC_easy_test:
2876 {
2877 // Just do a direct pointer compare and be done.
2878 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2879 *ctrl = gvn.transform(new IfTrueNode(iff));
2880 return gvn.transform(new IfFalseNode(iff));
2881 }
2882 case Compile::SSC_full_test:
2883 break;
2884 default:
2885 ShouldNotReachHere();
2886 }
2887 }
2888
2889 // %%% Possible further optimization: Even if the superklass is not exact,
2890 // if the subklass is the unique subtype of the superklass, the check
2891 // will always succeed. We could leave a dependency behind to ensure this.
2892
2893 // First load the super-klass's check-offset
2894 Node *p1 = gvn.transform(new AddPNode(superklass, superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2895 Node* m = C->immutable_memory();
2896 Node *chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2897 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2898 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2899 int chk_off_con = (chk_off_t != nullptr && chk_off_t->is_con()) ? chk_off_t->get_con() : cacheoff_con;
2900 bool might_be_cache = (chk_off_con == cacheoff_con);
2901
2902 // Load from the sub-klass's super-class display list, or a 1-word cache of
2903 // the secondary superclass list, or a failing value with a sentinel offset
2904 // if the super-klass is an interface or exceptionally deep in the Java
2905 // hierarchy and we have to scan the secondary superclass list the hard way.
2906 // Worst-case type is a little odd: null is allowed as a result (usually
2907 // klass loads can never produce a null).
2908 Node *chk_off_X = chk_off;
2909 #ifdef _LP64
2910 chk_off_X = gvn.transform(new ConvI2LNode(chk_off_X));
2911 #endif
2912 Node *p2 = gvn.transform(new AddPNode(subklass,subklass,chk_off_X));
2913 // For some types like interfaces the following loadKlass is from a 1-word
2914 // cache which is mutable so can't use immutable memory. Other
2915 // types load from the super-class display table which is immutable.
2916 Node *kmem = C->immutable_memory();
2917 // secondary_super_cache is not immutable but can be treated as such because:
2918 // - no ideal node writes to it in a way that could cause an
2919 // incorrect/missed optimization of the following Load.
2920 // - it's a cache so, worse case, not reading the latest value
2921 // wouldn't cause incorrect execution
2922 if (might_be_cache && mem != nullptr) {
2923 kmem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(C->get_alias_index(gvn.type(p2)->is_ptr())) : mem;
2924 }
2925 Node* nkls = gvn.transform(LoadKlassNode::make(gvn, kmem, p2, gvn.type(p2)->is_ptr(), TypeInstKlassPtr::OBJECT_OR_NULL));
2926
2927 // Compile speed common case: ARE a subtype and we canNOT fail
2928 if (superklass == nkls) {
2929 return C->top(); // false path is dead; no test needed.
2930 }
2931
2932 // Gather the various success & failures here
2933 RegionNode* r_not_subtype = new RegionNode(3);
2934 gvn.record_for_igvn(r_not_subtype);
2935 RegionNode* r_ok_subtype = new RegionNode(4);
2936 gvn.record_for_igvn(r_ok_subtype);
2937
2938 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2939 // SubTypeCheck node
2940 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2941 ciCallProfile profile = method->call_profile_at_bci(bci);
2942 float total_prob = 0;
2943 for (int i = 0; profile.has_receiver(i); ++i) {
2944 float prob = profile.receiver_prob(i);
2945 total_prob += prob;
2946 }
2947 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2948 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2949 for (int i = 0; profile.has_receiver(i); ++i) {
2950 ciKlass* klass = profile.receiver(i);
2951 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2952 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2953 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2954 continue;
2955 }
2956 float prob = profile.receiver_prob(i);
2957 ConNode* klass_node = gvn.makecon(klass_t);
2958 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2959 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2960
2961 if (result == Compile::SSC_always_true) {
2962 r_ok_subtype->add_req(iftrue);
2963 } else {
2964 assert(result == Compile::SSC_always_false, "");
2965 r_not_subtype->add_req(iftrue);
2966 }
2967 *ctrl = gvn.transform(new IfFalseNode(iff));
2968 }
2969 }
2970 }
2971
2972 // See if we get an immediate positive hit. Happens roughly 83% of the
2973 // time. Test to see if the value loaded just previously from the subklass
2974 // is exactly the superklass.
2975 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2976 Node *iftrue1 = gvn.transform( new IfTrueNode (iff1));
2977 *ctrl = gvn.transform(new IfFalseNode(iff1));
2978
2979 // Compile speed common case: Check for being deterministic right now. If
2980 // chk_off is a constant and not equal to cacheoff then we are NOT a
2981 // subklass. In this case we need exactly the 1 test above and we can
2982 // return those results immediately.
2983 if (!might_be_cache) {
2984 Node* not_subtype_ctrl = *ctrl;
2985 *ctrl = iftrue1; // We need exactly the 1 test above
2986 PhaseIterGVN* igvn = gvn.is_IterGVN();
2987 if (igvn != nullptr) {
2988 igvn->remove_globally_dead_node(r_ok_subtype);
2989 igvn->remove_globally_dead_node(r_not_subtype);
2990 }
2991 return not_subtype_ctrl;
2992 }
2993
2994 r_ok_subtype->init_req(1, iftrue1);
2995
2996 // Check for immediate negative hit. Happens roughly 11% of the time (which
2997 // is roughly 63% of the remaining cases). Test to see if the loaded
2998 // check-offset points into the subklass display list or the 1-element
2999 // cache. If it points to the display (and NOT the cache) and the display
3000 // missed then it's not a subtype.
3001 Node *cacheoff = gvn.intcon(cacheoff_con);
3002 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3003 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3004 *ctrl = gvn.transform(new IfFalseNode(iff2));
3005
3006 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3007 // No performance impact (too rare) but allows sharing of secondary arrays
3008 // which has some footprint reduction.
3009 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3010 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3011 *ctrl = gvn.transform(new IfFalseNode(iff3));
3012
3013 // -- Roads not taken here: --
3014 // We could also have chosen to perform the self-check at the beginning
3015 // of this code sequence, as the assembler does. This would not pay off
3016 // the same way, since the optimizer, unlike the assembler, can perform
3017 // static type analysis to fold away many successful self-checks.
3018 // Non-foldable self checks work better here in second position, because
3019 // the initial primary superclass check subsumes a self-check for most
3020 // types. An exception would be a secondary type like array-of-interface,
3021 // which does not appear in its own primary supertype display.
3022 // Finally, we could have chosen to move the self-check into the
3023 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3024 // dependent manner. But it is worthwhile to have the check here,
3025 // where it can be perhaps be optimized. The cost in code space is
3026 // small (register compare, branch).
3027
3028 // Now do a linear scan of the secondary super-klass array. Again, no real
3029 // performance impact (too rare) but it's gotta be done.
3030 // Since the code is rarely used, there is no penalty for moving it
3031 // out of line, and it can only improve I-cache density.
3032 // The decision to inline or out-of-line this final check is platform
3033 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3034 Node* psc = gvn.transform(
3035 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3036
3037 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3038 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3039 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3040
3041 // Return false path; set default control to true path.
3042 *ctrl = gvn.transform(r_ok_subtype);
3043 return gvn.transform(r_not_subtype);
3044 }
3045
3046 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3047 const Type* sub_t = _gvn.type(obj_or_subklass);
3048 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3049 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3050 obj_or_subklass = makecon(sub_t);
3051 }
3052 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3053 if (expand_subtype_check) {
3054 MergeMemNode* mem = merged_memory();
3055 Node* ctrl = control();
3056 Node* subklass = obj_or_subklass;
3057 if (!sub_t->isa_klassptr()) {
3058 subklass = load_object_klass(obj_or_subklass);
3059 }
3060
3061 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3062 set_control(ctrl);
3063 return n;
3064 }
3065
3066 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3067 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3068 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3069 set_control(_gvn.transform(new IfTrueNode(iff)));
3070 return _gvn.transform(new IfFalseNode(iff));
3071 }
3072
3073 // Profile-driven exact type check:
3074 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3075 float prob, Node* *casted_receiver) {
3076 assert(!klass->is_interface(), "no exact type check on interfaces");
3077 Node* fail = top();
3078 const Type* rec_t = _gvn.type(receiver);
3079 if (rec_t->is_inlinetypeptr()) {
3080 if (klass->equals(rec_t->inline_klass())) {
3081 (*casted_receiver) = receiver; // Always passes
3082 } else {
3083 (*casted_receiver) = top(); // Always fails
3084 fail = control();
3085 set_control(top());
3086 }
3087 return fail;
3088 }
3089 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3090 Node* recv_klass = load_object_klass(receiver);
3091 fail = type_check(recv_klass, tklass, prob);
3092
3093 if (!stopped()) {
3094 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3095 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3096 assert(recv_xtype->klass_is_exact(), "");
3097
3098 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3099 // Subsume downstream occurrences of receiver with a cast to
3100 // recv_xtype, since now we know what the type will be.
3101 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3102 Node* res = _gvn.transform(cast);
3103 if (recv_xtype->is_inlinetypeptr()) {
3104 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3105 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3106 }
3107 (*casted_receiver) = res;
3108 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3109 // (User must make the replace_in_map call.)
3110 }
3111 }
3112
3113 return fail;
3114 }
3115
3116 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3117 float prob) {
3118 Node* want_klass = makecon(tklass);
3119 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3120 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3121 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3122 set_control(_gvn.transform(new IfTrueNode (iff)));
3123 Node* fail = _gvn.transform(new IfFalseNode(iff));
3124 return fail;
3125 }
3126
3127 //------------------------------subtype_check_receiver-------------------------
3128 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3129 Node** casted_receiver) {
3130 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3131 Node* want_klass = makecon(tklass);
3132
3133 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3134
3135 // Ignore interface type information until interface types are properly tracked.
3136 if (!stopped() && !klass->is_interface()) {
3137 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3138 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3139 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3140 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3141 if (recv_type->is_inlinetypeptr()) {
3142 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3143 }
3144 (*casted_receiver) = cast;
3145 }
3146 }
3147
3148 return slow_ctl;
3149 }
3150
3151 //------------------------------seems_never_null-------------------------------
3152 // Use null_seen information if it is available from the profile.
3153 // If we see an unexpected null at a type check we record it and force a
3154 // recompile; the offending check will be recompiled to handle nulls.
3155 // If we see several offending BCIs, then all checks in the
3156 // method will be recompiled.
3157 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3158 speculating = !_gvn.type(obj)->speculative_maybe_null();
3159 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3160 if (UncommonNullCast // Cutout for this technique
3161 && obj != null() // And not the -Xcomp stupid case?
3162 && !too_many_traps(reason)
3163 ) {
3164 if (speculating) {
3165 return true;
3166 }
3167 if (data == nullptr)
3168 // Edge case: no mature data. Be optimistic here.
3169 return true;
3170 // If the profile has not seen a null, assume it won't happen.
3171 assert(java_bc() == Bytecodes::_checkcast ||
3172 java_bc() == Bytecodes::_instanceof ||
3173 java_bc() == Bytecodes::_aastore, "MDO must collect null_seen bit here");
3174 return !data->as_BitData()->null_seen();
3175 }
3176 speculating = false;
3177 return false;
3178 }
3179
3180 void GraphKit::guard_klass_being_initialized(Node* klass) {
3181 int init_state_off = in_bytes(InstanceKlass::init_state_offset());
3182 Node* adr = basic_plus_adr(top(), klass, init_state_off);
3183 Node* init_state = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3184 adr->bottom_type()->is_ptr(), TypeInt::BYTE,
3185 T_BYTE, MemNode::acquire);
3186 init_state = _gvn.transform(init_state);
3187
3188 Node* being_initialized_state = makecon(TypeInt::make(InstanceKlass::being_initialized));
3189
3190 Node* chk = _gvn.transform(new CmpINode(being_initialized_state, init_state));
3191 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3192
3193 { BuildCutout unless(this, tst, PROB_MAX);
3194 uncommon_trap(Deoptimization::Reason_initialized, Deoptimization::Action_reinterpret);
3195 }
3196 }
3197
3198 void GraphKit::guard_init_thread(Node* klass) {
3199 int init_thread_off = in_bytes(InstanceKlass::init_thread_offset());
3200 Node* adr = basic_plus_adr(top(), klass, init_thread_off);
3201
3202 Node* init_thread = LoadNode::make(_gvn, nullptr, immutable_memory(), adr,
3203 adr->bottom_type()->is_ptr(), TypePtr::NOTNULL,
3204 T_ADDRESS, MemNode::unordered);
3205 init_thread = _gvn.transform(init_thread);
3206
3207 Node* cur_thread = _gvn.transform(new ThreadLocalNode());
3208
3209 Node* chk = _gvn.transform(new CmpPNode(cur_thread, init_thread));
3210 Node* tst = _gvn.transform(new BoolNode(chk, BoolTest::eq));
3211
3212 { BuildCutout unless(this, tst, PROB_MAX);
3213 uncommon_trap(Deoptimization::Reason_uninitialized, Deoptimization::Action_none);
3214 }
3215 }
3216
3217 void GraphKit::clinit_barrier(ciInstanceKlass* ik, ciMethod* context) {
3218 if (ik->is_being_initialized()) {
3219 if (C->needs_clinit_barrier(ik, context)) {
3220 Node* klass = makecon(TypeKlassPtr::make(ik));
3221 guard_klass_being_initialized(klass);
3222 guard_init_thread(klass);
3223 insert_mem_bar(Op_MemBarCPUOrder);
3224 }
3225 } else if (ik->is_initialized()) {
3226 return; // no barrier needed
3227 } else {
3228 uncommon_trap(Deoptimization::Reason_uninitialized,
3229 Deoptimization::Action_reinterpret,
3230 nullptr);
3231 }
3232 }
3233
3234 //------------------------maybe_cast_profiled_receiver-------------------------
3235 // If the profile has seen exactly one type, narrow to exactly that type.
3236 // Subsequent type checks will always fold up.
3237 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3238 const TypeKlassPtr* require_klass,
3239 ciKlass* spec_klass,
3240 bool safe_for_replace) {
3241 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3242
3243 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3244
3245 // Make sure we haven't already deoptimized from this tactic.
3246 if (too_many_traps_or_recompiles(reason))
3247 return nullptr;
3248
3249 // (No, this isn't a call, but it's enough like a virtual call
3250 // to use the same ciMethod accessor to get the profile info...)
3251 // If we have a speculative type use it instead of profiling (which
3252 // may not help us)
3253 ciKlass* exact_kls = spec_klass;
3254 if (exact_kls == nullptr) {
3255 if (java_bc() == Bytecodes::_aastore) {
3256 ciKlass* array_type = nullptr;
3257 ciKlass* element_type = nullptr;
3258 ProfilePtrKind element_ptr = ProfileMaybeNull;
3259 bool flat_array = true;
3260 bool null_free_array = true;
3261 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3262 exact_kls = element_type;
3263 } else {
3264 exact_kls = profile_has_unique_klass();
3265 }
3266 }
3267 if (exact_kls != nullptr) {// no cast failures here
3268 if (require_klass == nullptr ||
3269 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3270 // If we narrow the type to match what the type profile sees or
3271 // the speculative type, we can then remove the rest of the
3272 // cast.
3273 // This is a win, even if the exact_kls is very specific,
3274 // because downstream operations, such as method calls,
3275 // will often benefit from the sharper type.
3276 Node* exact_obj = not_null_obj; // will get updated in place...
3277 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3278 &exact_obj);
3279 { PreserveJVMState pjvms(this);
3280 set_control(slow_ctl);
3281 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3282 }
3283 if (safe_for_replace) {
3284 replace_in_map(not_null_obj, exact_obj);
3285 }
3286 return exact_obj;
3287 }
3288 // assert(ssc == Compile::SSC_always_true)... except maybe the profile lied to us.
3289 }
3290
3291 return nullptr;
3292 }
3293
3294 /**
3295 * Cast obj to type and emit guard unless we had too many traps here
3296 * already
3297 *
3298 * @param obj node being casted
3299 * @param type type to cast the node to
3300 * @param not_null true if we know node cannot be null
3301 */
3302 Node* GraphKit::maybe_cast_profiled_obj(Node* obj,
3303 ciKlass* type,
3304 bool not_null) {
3305 if (stopped()) {
3306 return obj;
3307 }
3308
3309 // type is null if profiling tells us this object is always null
3310 if (type != nullptr) {
3311 Deoptimization::DeoptReason class_reason = Deoptimization::Reason_speculate_class_check;
3312 Deoptimization::DeoptReason null_reason = Deoptimization::Reason_speculate_null_check;
3313
3314 if (!too_many_traps_or_recompiles(null_reason) &&
3315 !too_many_traps_or_recompiles(class_reason)) {
3316 Node* not_null_obj = nullptr;
3317 // not_null is true if we know the object is not null and
3318 // there's no need for a null check
3319 if (!not_null) {
3320 Node* null_ctl = top();
3321 not_null_obj = null_check_oop(obj, &null_ctl, true, true, true);
3322 assert(null_ctl->is_top(), "no null control here");
3323 } else {
3324 not_null_obj = obj;
3325 }
3326
3327 Node* exact_obj = not_null_obj;
3328 ciKlass* exact_kls = type;
3329 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3330 &exact_obj);
3331 {
3332 PreserveJVMState pjvms(this);
3333 set_control(slow_ctl);
3334 uncommon_trap_exact(class_reason, Deoptimization::Action_maybe_recompile);
3335 }
3336 replace_in_map(not_null_obj, exact_obj);
3337 obj = exact_obj;
3338 }
3339 } else {
3340 if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3341 Node* exact_obj = null_assert(obj);
3342 replace_in_map(obj, exact_obj);
3343 obj = exact_obj;
3344 }
3345 }
3346 return obj;
3347 }
3348
3349 //-------------------------------gen_instanceof--------------------------------
3350 // Generate an instance-of idiom. Used by both the instance-of bytecode
3351 // and the reflective instance-of call.
3352 Node* GraphKit::gen_instanceof(Node* obj, Node* superklass, bool safe_for_replace) {
3353 kill_dead_locals(); // Benefit all the uncommon traps
3354 assert( !stopped(), "dead parse path should be checked in callers" );
3355 assert(!TypePtr::NULL_PTR->higher_equal(_gvn.type(superklass)->is_klassptr()),
3356 "must check for not-null not-dead klass in callers");
3357
3358 // Make the merge point
3359 enum { _obj_path = 1, _fail_path, _null_path, PATH_LIMIT };
3360 RegionNode* region = new RegionNode(PATH_LIMIT);
3361 Node* phi = new PhiNode(region, TypeInt::BOOL);
3362 C->set_has_split_ifs(true); // Has chance for split-if optimization
3363
3364 ciProfileData* data = nullptr;
3365 if (java_bc() == Bytecodes::_instanceof) { // Only for the bytecode
3366 data = method()->method_data()->bci_to_data(bci());
3367 }
3368 bool speculative_not_null = false;
3369 bool never_see_null = (ProfileDynamicTypes // aggressive use of profile
3370 && seems_never_null(obj, data, speculative_not_null));
3371
3372 // Null check; get casted pointer; set region slot 3
3373 Node* null_ctl = top();
3374 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3375
3376 // If not_null_obj is dead, only null-path is taken
3377 if (stopped()) { // Doing instance-of on a null?
3378 set_control(null_ctl);
3379 return intcon(0);
3380 }
3381 region->init_req(_null_path, null_ctl);
3382 phi ->init_req(_null_path, intcon(0)); // Set null path value
3383 if (null_ctl == top()) {
3384 // Do this eagerly, so that pattern matches like is_diamond_phi
3385 // will work even during parsing.
3386 assert(_null_path == PATH_LIMIT-1, "delete last");
3387 region->del_req(_null_path);
3388 phi ->del_req(_null_path);
3389 }
3390
3391 // Do we know the type check always succeed?
3392 bool known_statically = false;
3393 if (_gvn.type(superklass)->singleton()) {
3394 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3395 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3396 if (subk != nullptr && subk->is_loaded()) {
3397 int static_res = C->static_subtype_check(superk, subk);
3398 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3399 }
3400 }
3401
3402 if (!known_statically) {
3403 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3404 // We may not have profiling here or it may not help us. If we
3405 // have a speculative type use it to perform an exact cast.
3406 ciKlass* spec_obj_type = obj_type->speculative_type();
3407 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3408 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3409 if (stopped()) { // Profile disagrees with this path.
3410 set_control(null_ctl); // Null is the only remaining possibility.
3411 return intcon(0);
3412 }
3413 if (cast_obj != nullptr) {
3414 not_null_obj = cast_obj;
3415 }
3416 }
3417 }
3418
3419 // Generate the subtype check
3420 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
3421
3422 // Plug in the success path to the general merge in slot 1.
3423 region->init_req(_obj_path, control());
3424 phi ->init_req(_obj_path, intcon(1));
3425
3426 // Plug in the failing path to the general merge in slot 2.
3427 region->init_req(_fail_path, not_subtype_ctrl);
3428 phi ->init_req(_fail_path, intcon(0));
3429
3430 // Return final merged results
3431 set_control( _gvn.transform(region) );
3432 record_for_igvn(region);
3433
3434 // If we know the type check always succeeds then we don't use the
3435 // profiling data at this bytecode. Don't lose it, feed it to the
3436 // type system as a speculative type.
3437 if (safe_for_replace) {
3438 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3439 replace_in_map(obj, casted_obj);
3440 }
3441
3442 return _gvn.transform(phi);
3443 }
3444
3445 //-------------------------------gen_checkcast---------------------------------
3446 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3447 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3448 // uncommon-trap paths work. Adjust stack after this call.
3449 // If failure_control is supplied and not null, it is filled in with
3450 // the control edge for the cast failure. Otherwise, an appropriate
3451 // uncommon trap or exception is thrown.
3452 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node* *failure_control, bool null_free, bool maybe_larval) {
3453 kill_dead_locals(); // Benefit all the uncommon traps
3454 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3455 const Type* obj_type = _gvn.type(obj);
3456 if (obj_type->is_inlinetypeptr() && !obj_type->maybe_null() && klass_ptr_type->klass_is_exact() && obj_type->inline_klass() == klass_ptr_type->exact_klass(true)) {
3457 // Special case: larval inline objects must not be scalarized. They are also generally not
3458 // allowed to participate in most operations except as the first operand of putfield, or as an
3459 // argument to a constructor invocation with it being a receiver, Unsafe::putXXX with it being
3460 // the first argument, or Unsafe::finishPrivateBuffer. This allows us to aggressively scalarize
3461 // value objects in all other places. This special case comes from the limitation of the Java
3462 // language, Unsafe::makePrivateBuffer returns an Object that is checkcast-ed to the concrete
3463 // value type. We must do this first because C->static_subtype_check may do nothing when
3464 // StressReflectiveCode is set.
3465 return obj;
3466 }
3467
3468 // Else it must be a non-larval object
3469 obj = cast_to_non_larval(obj);
3470
3471 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3472 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3473 bool safe_for_replace = (failure_control == nullptr);
3474 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3475
3476 // Fast cutout: Check the case that the cast is vacuously true.
3477 // This detects the common cases where the test will short-circuit
3478 // away completely. We do this before we perform the null check,
3479 // because if the test is going to turn into zero code, we don't
3480 // want a residual null check left around. (Causes a slowdown,
3481 // for example, in some objArray manipulations, such as a[i]=a[j].)
3482 if (improved_klass_ptr_type->singleton()) {
3483 const TypeKlassPtr* kptr = nullptr;
3484 if (obj_type->isa_oop_ptr()) {
3485 kptr = obj_type->is_oopptr()->as_klass_type();
3486 } else if (obj->is_InlineType()) {
3487 ciInlineKlass* vk = obj_type->inline_klass();
3488 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3489 }
3490
3491 if (kptr != nullptr) {
3492 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3493 case Compile::SSC_always_true:
3494 // If we know the type check always succeed then we don't use
3495 // the profiling data at this bytecode. Don't lose it, feed it
3496 // to the type system as a speculative type.
3497 obj = record_profiled_receiver_for_speculation(obj);
3498 if (null_free) {
3499 assert(safe_for_replace, "must be");
3500 obj = null_check(obj);
3501 }
3502 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3503 return obj;
3504 case Compile::SSC_always_false:
3505 if (null_free) {
3506 assert(safe_for_replace, "must be");
3507 obj = null_check(obj);
3508 }
3509 // It needs a null check because a null will *pass* the cast check.
3510 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3511 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3512 Deoptimization::DeoptReason reason = is_aastore ?
3513 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3514 builtin_throw(reason);
3515 return top();
3516 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3517 return null_assert(obj);
3518 }
3519 break; // Fall through to full check
3520 default:
3521 break;
3522 }
3523 }
3524 }
3525
3526 ciProfileData* data = nullptr;
3527 if (failure_control == nullptr) { // use MDO in regular case only
3528 assert(java_bc() == Bytecodes::_aastore ||
3529 java_bc() == Bytecodes::_checkcast,
3530 "interpreter profiles type checks only for these BCs");
3531 if (method()->method_data()->is_mature()) {
3532 data = method()->method_data()->bci_to_data(bci());
3533 }
3534 }
3535
3536 // Make the merge point
3537 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3538 RegionNode* region = new RegionNode(PATH_LIMIT);
3539 Node* phi = new PhiNode(region, toop);
3540 _gvn.set_type(region, Type::CONTROL);
3541 _gvn.set_type(phi, toop);
3542
3543 C->set_has_split_ifs(true); // Has chance for split-if optimization
3544
3545 // Use null-cast information if it is available
3546 bool speculative_not_null = false;
3547 bool never_see_null = ((failure_control == nullptr) // regular case only
3548 && seems_never_null(obj, data, speculative_not_null));
3549
3550 if (obj->is_InlineType()) {
3551 // Re-execute if buffering during triggers deoptimization
3552 PreserveReexecuteState preexecs(this);
3553 jvms()->set_should_reexecute(true);
3554 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3555 }
3556
3557 // Null check; get casted pointer; set region slot 3
3558 Node* null_ctl = top();
3559 Node* not_null_obj = nullptr;
3560 if (null_free) {
3561 assert(safe_for_replace, "must be");
3562 not_null_obj = null_check(obj);
3563 } else {
3564 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3565 }
3566
3567 // If not_null_obj is dead, only null-path is taken
3568 if (stopped()) { // Doing instance-of on a null?
3569 set_control(null_ctl);
3570 if (toop->is_inlinetypeptr()) {
3571 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3572 }
3573 return null();
3574 }
3575 region->init_req(_null_path, null_ctl);
3576 phi ->init_req(_null_path, null()); // Set null path value
3577 if (null_ctl == top()) {
3578 // Do this eagerly, so that pattern matches like is_diamond_phi
3579 // will work even during parsing.
3580 assert(_null_path == PATH_LIMIT-1, "delete last");
3581 region->del_req(_null_path);
3582 phi ->del_req(_null_path);
3583 }
3584
3585 Node* cast_obj = nullptr;
3586 if (improved_klass_ptr_type->klass_is_exact()) {
3587 // The following optimization tries to statically cast the speculative type of the object
3588 // (for example obtained during profiling) to the type of the superklass and then do a
3589 // dynamic check that the type of the object is what we expect. To work correctly
3590 // for checkcast and aastore the type of superklass should be exact.
3591 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3592 // We may not have profiling here or it may not help us. If we have
3593 // a speculative type use it to perform an exact cast.
3594 ciKlass* spec_obj_type = obj_type->speculative_type();
3595 if (spec_obj_type != nullptr || data != nullptr) {
3596 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3597 if (cast_obj != nullptr) {
3598 if (failure_control != nullptr) // failure is now impossible
3599 (*failure_control) = top();
3600 // adjust the type of the phi to the exact klass:
3601 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3602 }
3603 }
3604 }
3605
3606 if (cast_obj == nullptr) {
3607 // Generate the subtype check
3608 Node* improved_superklass = superklass;
3609 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3610 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3611 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3612 // Additionally, the benefit would only be minor in non-constant cases.
3613 improved_superklass = makecon(improved_klass_ptr_type);
3614 }
3615 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3616 // Plug in success path into the merge
3617 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3618 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3619 if (failure_control == nullptr) {
3620 if (not_subtype_ctrl != top()) { // If failure is possible
3621 PreserveJVMState pjvms(this);
3622 set_control(not_subtype_ctrl);
3623 Node* obj_klass = nullptr;
3624 if (not_null_obj->is_InlineType()) {
3625 obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
3626 } else {
3627 obj_klass = load_object_klass(not_null_obj);
3628 }
3629 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3630 Deoptimization::DeoptReason reason = is_aastore ?
3631 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3632 builtin_throw(reason);
3633 }
3634 } else {
3635 (*failure_control) = not_subtype_ctrl;
3636 }
3637 }
3638
3639 region->init_req(_obj_path, control());
3640 phi ->init_req(_obj_path, cast_obj);
3641
3642 // A merge of null or Casted-NotNull obj
3643 Node* res = _gvn.transform(phi);
3644
3645 // Note I do NOT always 'replace_in_map(obj,result)' here.
3646 // if( tk->klass()->can_be_primary_super() )
3647 // This means that if I successfully store an Object into an array-of-String
3648 // I 'forget' that the Object is really now known to be a String. I have to
3649 // do this because we don't have true union types for interfaces - if I store
3650 // a Baz into an array-of-Interface and then tell the optimizer it's an
3651 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3652 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3653 // replace_in_map( obj, res );
3654
3655 // Return final merged results
3656 set_control( _gvn.transform(region) );
3657 record_for_igvn(region);
3658
3659 bool not_inline = !toop->can_be_inline_type();
3660 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3661 if (EnableValhalla && (not_inline || not_flat_in_array)) {
3662 // Check if obj has been loaded from an array
3663 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3664 Node* array = nullptr;
3665 if (obj->isa_Load()) {
3666 Node* address = obj->in(MemNode::Address);
3667 if (address->isa_AddP()) {
3668 array = address->as_AddP()->in(AddPNode::Base);
3669 }
3670 } else if (obj->is_Phi()) {
3671 Node* region = obj->in(0);
3672 // TODO make this more robust (see JDK-8231346)
3673 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3674 IfNode* iff = region->in(2)->in(0)->isa_If();
3675 if (iff != nullptr) {
3676 iff->is_flat_array_check(&_gvn, &array);
3677 }
3678 }
3679 }
3680 if (array != nullptr) {
3681 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3682 if (ary_t != nullptr) {
3683 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3684 // Casting array element to a non-inline-type, mark array as not null-free.
3685 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3686 replace_in_map(array, cast);
3687 array = cast;
3688 }
3689 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3690 // Casting array element to a non-flat-in-array type, mark array as not flat.
3691 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3692 replace_in_map(array, cast);
3693 array = cast;
3694 }
3695 }
3696 }
3697 }
3698
3699 if (!stopped() && !res->is_InlineType()) {
3700 res = record_profiled_receiver_for_speculation(res);
3701 if (toop->is_inlinetypeptr() && !maybe_larval) {
3702 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3703 res = vt;
3704 if (safe_for_replace) {
3705 replace_in_map(obj, vt);
3706 replace_in_map(not_null_obj, vt);
3707 replace_in_map(res, vt);
3708 }
3709 }
3710 }
3711 return res;
3712 }
3713
3714 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3715 // Load markword
3716 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3717 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3718 if (check_lock) {
3719 // Check if obj is locked
3720 Node* locked_bit = MakeConX(markWord::unlocked_value);
3721 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3722 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3723 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3724 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3725 _gvn.transform(iff);
3726 Node* locked_region = new RegionNode(3);
3727 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3728
3729 // Unlocked: Use bits from mark word
3730 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3731 mark_phi->init_req(1, mark);
3732
3733 // Locked: Load prototype header from klass
3734 set_control(_gvn.transform(new IfFalseNode(iff)));
3735 // Make loads control dependent to make sure they are only executed if array is locked
3736 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3737 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3738 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
3739 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3740
3741 locked_region->init_req(2, control());
3742 mark_phi->init_req(2, proto);
3743 set_control(_gvn.transform(locked_region));
3744 record_for_igvn(locked_region);
3745
3746 mark = mark_phi;
3747 }
3748
3749 // Now check if mark word bits are set
3750 Node* mask = MakeConX(mask_val);
3751 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3752 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3753 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3754 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3755 }
3756
3757 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3758 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3759 }
3760
3761 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3762 // We can't use immutable memory here because the mark word is mutable.
3763 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3764 // check is moved out of loops (mainly to enable loop unswitching).
3765 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3766 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3767 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3768 }
3769
3770 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3771 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3772 }
3773
3774 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3775 assert(vk->has_atomic_layout() || vk->has_non_atomic_layout(), "Can't be null-free and flat");
3776
3777 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
3778 if (!vk->has_non_atomic_layout()) {
3779 return intcon(1); // Always atomic
3780 } else if (!vk->has_atomic_layout()) {
3781 return intcon(0); // Never atomic
3782 }
3783
3784 // TODO 8350865 Don't fold this klass load because atomicity is currently not included in the typesystem
3785 Node* array_klass = load_object_klass(array, /* fold_for_arrays = */ false);
3786 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
3787 Node* layout_kind_addr = basic_plus_adr(array_klass, array_klass, layout_kind_offset);
3788 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
3789 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::ATOMIC_FLAT)));
3790 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3791 }
3792
3793 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
3794 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
3795 RegionNode* region = new RegionNode(3);
3796 Node* null_ctl = top();
3797 null_check_oop(val, &null_ctl);
3798 if (null_ctl != top()) {
3799 PreserveJVMState pjvms(this);
3800 set_control(null_ctl);
3801 {
3802 // Deoptimize if null-free array
3803 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
3804 inc_sp(nargs);
3805 uncommon_trap(Deoptimization::Reason_null_check,
3806 Deoptimization::Action_none);
3807 }
3808 region->init_req(1, control());
3809 }
3810 region->init_req(2, control());
3811 set_control(_gvn.transform(region));
3812 record_for_igvn(region);
3813 if (_gvn.type(val) == TypePtr::NULL_PTR) {
3814 // Since we were just successfully storing null, the array can't be null free.
3815 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
3816 ary_t = ary_t->cast_to_not_null_free();
3817 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
3818 if (safe_for_replace) {
3819 replace_in_map(ary, cast);
3820 }
3821 ary = cast;
3822 }
3823 return ary;
3824 }
3825
3826 //------------------------------next_monitor-----------------------------------
3827 // What number should be given to the next monitor?
3828 int GraphKit::next_monitor() {
3829 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3830 int next = current + C->sync_stack_slots();
3831 // Keep the toplevel high water mark current:
3832 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3833 return current;
3834 }
3835
3836 //------------------------------insert_mem_bar---------------------------------
3837 // Memory barrier to avoid floating things around
3838 // The membar serves as a pinch point between both control and all memory slices.
3839 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3840 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3841 mb->init_req(TypeFunc::Control, control());
3842 mb->init_req(TypeFunc::Memory, reset_memory());
3843 Node* membar = _gvn.transform(mb);
3844 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3845 set_all_memory_call(membar);
3846 return membar;
3847 }
3848
3849 //-------------------------insert_mem_bar_volatile----------------------------
3850 // Memory barrier to avoid floating things around
3851 // The membar serves as a pinch point between both control and memory(alias_idx).
3852 // If you want to make a pinch point on all memory slices, do not use this
3853 // function (even with AliasIdxBot); use insert_mem_bar() instead.
3854 Node* GraphKit::insert_mem_bar_volatile(int opcode, int alias_idx, Node* precedent) {
3855 // When Parse::do_put_xxx updates a volatile field, it appends a series
3856 // of MemBarVolatile nodes, one for *each* volatile field alias category.
3857 // The first membar is on the same memory slice as the field store opcode.
3858 // This forces the membar to follow the store. (Bug 6500685 broke this.)
3859 // All the other membars (for other volatile slices, including AliasIdxBot,
3860 // which stands for all unknown volatile slices) are control-dependent
3861 // on the first membar. This prevents later volatile loads or stores
3862 // from sliding up past the just-emitted store.
3863
3864 MemBarNode* mb = MemBarNode::make(C, opcode, alias_idx, precedent);
3865 mb->set_req(TypeFunc::Control,control());
3866 if (alias_idx == Compile::AliasIdxBot) {
3867 mb->set_req(TypeFunc::Memory, merged_memory()->base_memory());
3868 } else {
3869 assert(!(opcode == Op_Initialize && alias_idx != Compile::AliasIdxRaw), "fix caller");
3870 mb->set_req(TypeFunc::Memory, memory(alias_idx));
3871 }
3872 Node* membar = _gvn.transform(mb);
3873 set_control(_gvn.transform(new ProjNode(membar, TypeFunc::Control)));
3874 if (alias_idx == Compile::AliasIdxBot) {
3875 merged_memory()->set_base_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)));
3876 } else {
3877 set_memory(_gvn.transform(new ProjNode(membar, TypeFunc::Memory)),alias_idx);
3878 }
3879 return membar;
3880 }
3881
3882 //------------------------------shared_lock------------------------------------
3883 // Emit locking code.
3884 FastLockNode* GraphKit::shared_lock(Node* obj) {
3885 // bci is either a monitorenter bc or InvocationEntryBci
3886 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3887 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3888
3889 if( !GenerateSynchronizationCode )
3890 return nullptr; // Not locking things?
3891
3892 if (stopped()) // Dead monitor?
3893 return nullptr;
3894
3895 assert(dead_locals_are_killed(), "should kill locals before sync. point");
3896
3897 // Box the stack location
3898 Node* box = new BoxLockNode(next_monitor());
3899 // Check for bailout after new BoxLockNode
3900 if (failing()) { return nullptr; }
3901 box = _gvn.transform(box);
3902 Node* mem = reset_memory();
3903
3904 FastLockNode * flock = _gvn.transform(new FastLockNode(nullptr, obj, box) )->as_FastLock();
3905
3906 // Add monitor to debug info for the slow path. If we block inside the
3907 // slow path and de-opt, we need the monitor hanging around
3908 map()->push_monitor( flock );
3909
3910 const TypeFunc *tf = LockNode::lock_type();
3911 LockNode *lock = new LockNode(C, tf);
3912
3913 lock->init_req( TypeFunc::Control, control() );
3914 lock->init_req( TypeFunc::Memory , mem );
3915 lock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3916 lock->init_req( TypeFunc::FramePtr, frameptr() );
3917 lock->init_req( TypeFunc::ReturnAdr, top() );
3918
3919 lock->init_req(TypeFunc::Parms + 0, obj);
3920 lock->init_req(TypeFunc::Parms + 1, box);
3921 lock->init_req(TypeFunc::Parms + 2, flock);
3922 add_safepoint_edges(lock);
3923
3924 lock = _gvn.transform( lock )->as_Lock();
3925
3926 // lock has no side-effects, sets few values
3927 set_predefined_output_for_runtime_call(lock, mem, TypeRawPtr::BOTTOM);
3928
3929 insert_mem_bar(Op_MemBarAcquireLock);
3930
3931 // Add this to the worklist so that the lock can be eliminated
3932 record_for_igvn(lock);
3933
3934 #ifndef PRODUCT
3935 if (PrintLockStatistics) {
3936 // Update the counter for this lock. Don't bother using an atomic
3937 // operation since we don't require absolute accuracy.
3938 lock->create_lock_counter(map()->jvms());
3939 increment_counter(lock->counter()->addr());
3940 }
3941 #endif
3942
3943 return flock;
3944 }
3945
3946
3947 //------------------------------shared_unlock----------------------------------
3948 // Emit unlocking code.
3949 void GraphKit::shared_unlock(Node* box, Node* obj) {
3950 // bci is either a monitorenter bc or InvocationEntryBci
3951 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3952 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3953
3954 if( !GenerateSynchronizationCode )
3955 return;
3956 if (stopped()) { // Dead monitor?
3957 map()->pop_monitor(); // Kill monitor from debug info
3958 return;
3959 }
3960 assert(!obj->is_InlineType(), "should not unlock on inline type");
3961
3962 // Memory barrier to avoid floating things down past the locked region
3963 insert_mem_bar(Op_MemBarReleaseLock);
3964
3965 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3966 UnlockNode *unlock = new UnlockNode(C, tf);
3967 #ifdef ASSERT
3968 unlock->set_dbg_jvms(sync_jvms());
3969 #endif
3970 uint raw_idx = Compile::AliasIdxRaw;
3971 unlock->init_req( TypeFunc::Control, control() );
3972 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3973 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3974 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3975 unlock->init_req( TypeFunc::ReturnAdr, top() );
3976
3977 unlock->init_req(TypeFunc::Parms + 0, obj);
3978 unlock->init_req(TypeFunc::Parms + 1, box);
3979 unlock = _gvn.transform(unlock)->as_Unlock();
3980
3981 Node* mem = reset_memory();
3982
3983 // unlock has no side-effects, sets few values
3984 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3985
3986 // Kill monitor from debug info
3987 map()->pop_monitor( );
3988 }
3989
3990 //-------------------------------get_layout_helper-----------------------------
3991 // If the given klass is a constant or known to be an array,
3992 // fetch the constant layout helper value into constant_value
3993 // and return null. Otherwise, load the non-constant
3994 // layout helper value, and return the node which represents it.
3995 // This two-faced routine is useful because allocation sites
3996 // almost always feature constant types.
3997 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3998 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3999 if (!StressReflectiveCode && klass_t != nullptr) {
4000 bool xklass = klass_t->klass_is_exact();
4001 bool can_be_flat = false;
4002 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4003 if (UseArrayFlattening && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
4004 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4005 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4006 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4007 }
4008 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4009 jint lhelper;
4010 if (klass_t->is_flat()) {
4011 lhelper = ary_type->flat_layout_helper();
4012 } else if (klass_t->isa_aryklassptr()) {
4013 BasicType elem = ary_type->elem()->array_element_basic_type();
4014 if (is_reference_type(elem, true)) {
4015 elem = T_OBJECT;
4016 }
4017 lhelper = Klass::array_layout_helper(elem);
4018 } else {
4019 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4020 }
4021 if (lhelper != Klass::_lh_neutral_value) {
4022 constant_value = lhelper;
4023 return (Node*) nullptr;
4024 }
4025 }
4026 }
4027 constant_value = Klass::_lh_neutral_value; // put in a known value
4028 Node* lhp = basic_plus_adr(klass_node, klass_node, in_bytes(Klass::layout_helper_offset()));
4029 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4030 }
4031
4032 // We just put in an allocate/initialize with a big raw-memory effect.
4033 // Hook selected additional alias categories on the initialization.
4034 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4035 MergeMemNode* init_in_merge,
4036 Node* init_out_raw) {
4037 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4038 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4039
4040 Node* prevmem = kit.memory(alias_idx);
4041 init_in_merge->set_memory_at(alias_idx, prevmem);
4042 if (init_out_raw != nullptr) {
4043 kit.set_memory(init_out_raw, alias_idx);
4044 }
4045 }
4046
4047 //---------------------------set_output_for_allocation-------------------------
4048 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4049 const TypeOopPtr* oop_type,
4050 bool deoptimize_on_exception) {
4051 int rawidx = Compile::AliasIdxRaw;
4052 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4053 add_safepoint_edges(alloc);
4054 Node* allocx = _gvn.transform(alloc);
4055 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4056 // create memory projection for i_o
4057 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4058 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4059
4060 // create a memory projection as for the normal control path
4061 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4062 set_memory(malloc, rawidx);
4063
4064 // a normal slow-call doesn't change i_o, but an allocation does
4065 // we create a separate i_o projection for the normal control path
4066 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4067 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4068
4069 // put in an initialization barrier
4070 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4071 rawoop)->as_Initialize();
4072 assert(alloc->initialization() == init, "2-way macro link must work");
4073 assert(init ->allocation() == alloc, "2-way macro link must work");
4074 {
4075 // Extract memory strands which may participate in the new object's
4076 // initialization, and source them from the new InitializeNode.
4077 // This will allow us to observe initializations when they occur,
4078 // and link them properly (as a group) to the InitializeNode.
4079 assert(init->in(InitializeNode::Memory) == malloc, "");
4080 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4081 init->set_req(InitializeNode::Memory, minit_in);
4082 record_for_igvn(minit_in); // fold it up later, if possible
4083 _gvn.set_type(minit_in, Type::MEMORY);
4084 Node* minit_out = memory(rawidx);
4085 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4086 // Add an edge in the MergeMem for the header fields so an access
4087 // to one of those has correct memory state
4088 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
4089 set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
4090 if (oop_type->isa_aryptr()) {
4091 const TypeAryPtr* arytype = oop_type->is_aryptr();
4092 if (arytype->is_flat()) {
4093 // Initially all flat array accesses share a single slice
4094 // but that changes after parsing. Prepare the memory graph so
4095 // it can optimize flat array accesses properly once they
4096 // don't share a single slice.
4097 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4098 C->set_flat_accesses_share_alias(false);
4099 ciInlineKlass* vk = arytype->elem()->inline_klass();
4100 for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
4101 ciField* field = vk->nonstatic_field_at(i);
4102 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4103 continue; // do not bother to track really large numbers of fields
4104 int off_in_vt = field->offset_in_bytes() - vk->payload_offset();
4105 const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
4106 int fieldidx = C->get_alias_index(adr_type, true);
4107 // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node
4108 // can result in per flat array field Phis to be created which confuses the logic of
4109 // Compile::adjust_flat_array_access_aliases().
4110 hook_memory_on_init(*this, fieldidx, minit_in, nullptr);
4111 }
4112 C->set_flat_accesses_share_alias(true);
4113 hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
4114 } else {
4115 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4116 int elemidx = C->get_alias_index(telemref);
4117 hook_memory_on_init(*this, elemidx, minit_in, minit_out);
4118 }
4119 } else if (oop_type->isa_instptr()) {
4120 set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
4121 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4122 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4123 ciField* field = ik->nonstatic_field_at(i);
4124 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4125 continue; // do not bother to track really large numbers of fields
4126 // Find (or create) the alias category for this field:
4127 int fieldidx = C->alias_type(field)->index();
4128 hook_memory_on_init(*this, fieldidx, minit_in, minit_out);
4129 }
4130 }
4131 }
4132
4133 // Cast raw oop to the real thing...
4134 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4135 javaoop = _gvn.transform(javaoop);
4136 C->set_recent_alloc(control(), javaoop);
4137 assert(just_allocated_object(control()) == javaoop, "just allocated");
4138
4139 #ifdef ASSERT
4140 { // Verify that the AllocateNode::Ideal_allocation recognizers work:
4141 assert(AllocateNode::Ideal_allocation(rawoop) == alloc,
4142 "Ideal_allocation works");
4143 assert(AllocateNode::Ideal_allocation(javaoop) == alloc,
4144 "Ideal_allocation works");
4145 if (alloc->is_AllocateArray()) {
4146 assert(AllocateArrayNode::Ideal_array_allocation(rawoop) == alloc->as_AllocateArray(),
4147 "Ideal_allocation works");
4148 assert(AllocateArrayNode::Ideal_array_allocation(javaoop) == alloc->as_AllocateArray(),
4149 "Ideal_allocation works");
4150 } else {
4151 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4152 }
4153 }
4154 #endif //ASSERT
4155
4156 return javaoop;
4157 }
4158
4159 //---------------------------new_instance--------------------------------------
4160 // This routine takes a klass_node which may be constant (for a static type)
4161 // or may be non-constant (for reflective code). It will work equally well
4162 // for either, and the graph will fold nicely if the optimizer later reduces
4163 // the type to a constant.
4164 // The optional arguments are for specialized use by intrinsics:
4165 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4166 // - If 'return_size_val', report the total object size to the caller.
4167 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4168 Node* GraphKit::new_instance(Node* klass_node,
4169 Node* extra_slow_test,
4170 Node* *return_size_val,
4171 bool deoptimize_on_exception,
4172 InlineTypeNode* inline_type_node) {
4173 // Compute size in doublewords
4174 // The size is always an integral number of doublewords, represented
4175 // as a positive bytewise size stored in the klass's layout_helper.
4176 // The layout_helper also encodes (in a low bit) the need for a slow path.
4177 jint layout_con = Klass::_lh_neutral_value;
4178 Node* layout_val = get_layout_helper(klass_node, layout_con);
4179 bool layout_is_con = (layout_val == nullptr);
4180
4181 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4182 // Generate the initial go-slow test. It's either ALWAYS (return a
4183 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4184 // case) a computed value derived from the layout_helper.
4185 Node* initial_slow_test = nullptr;
4186 if (layout_is_con) {
4187 assert(!StressReflectiveCode, "stress mode does not use these paths");
4188 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4189 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4190 } else { // reflective case
4191 // This reflective path is used by Unsafe.allocateInstance.
4192 // (It may be stress-tested by specifying StressReflectiveCode.)
4193 // Basically, we want to get into the VM is there's an illegal argument.
4194 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4195 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4196 if (extra_slow_test != intcon(0)) {
4197 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4198 }
4199 // (Macro-expander will further convert this to a Bool, if necessary.)
4200 }
4201
4202 // Find the size in bytes. This is easy; it's the layout_helper.
4203 // The size value must be valid even if the slow path is taken.
4204 Node* size = nullptr;
4205 if (layout_is_con) {
4206 size = MakeConX(Klass::layout_helper_size_in_bytes(layout_con));
4207 } else { // reflective case
4208 // This reflective path is used by clone and Unsafe.allocateInstance.
4209 size = ConvI2X(layout_val);
4210
4211 // Clear the low bits to extract layout_helper_size_in_bytes:
4212 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4213 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4214 size = _gvn.transform( new AndXNode(size, mask) );
4215 }
4216 if (return_size_val != nullptr) {
4217 (*return_size_val) = size;
4218 }
4219
4220 // This is a precise notnull oop of the klass.
4221 // (Actually, it need not be precise if this is a reflective allocation.)
4222 // It's what we cast the result to.
4223 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4224 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4225 const TypeOopPtr* oop_type = tklass->as_instance_type();
4226
4227 // Now generate allocation code
4228
4229 // The entire memory state is needed for slow path of the allocation
4230 // since GC and deoptimization can happen.
4231 Node *mem = reset_memory();
4232 set_all_memory(mem); // Create new memory state
4233
4234 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4235 control(), mem, i_o(),
4236 size, klass_node,
4237 initial_slow_test, inline_type_node);
4238
4239 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4240 }
4241
4242 //-------------------------------new_array-------------------------------------
4243 // helper for newarray and anewarray
4244 // The 'length' parameter is (obviously) the length of the array.
4245 // The optional arguments are for specialized use by intrinsics:
4246 // - If 'return_size_val', report the non-padded array size (sum of header size
4247 // and array body) to the caller.
4248 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4249 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4250 Node* length, // number of array elements
4251 int nargs, // number of arguments to push back for uncommon trap
4252 Node* *return_size_val,
4253 bool deoptimize_on_exception,
4254 Node* init_val) {
4255 jint layout_con = Klass::_lh_neutral_value;
4256 Node* layout_val = get_layout_helper(klass_node, layout_con);
4257 bool layout_is_con = (layout_val == nullptr);
4258
4259 if (!layout_is_con && !StressReflectiveCode &&
4260 !too_many_traps(Deoptimization::Reason_class_check)) {
4261 // This is a reflective array creation site.
4262 // Optimistically assume that it is a subtype of Object[],
4263 // so that we can fold up all the address arithmetic.
4264 layout_con = Klass::array_layout_helper(T_OBJECT);
4265 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4266 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4267 { BuildCutout unless(this, bol_lh, PROB_MAX);
4268 inc_sp(nargs);
4269 uncommon_trap(Deoptimization::Reason_class_check,
4270 Deoptimization::Action_maybe_recompile);
4271 }
4272 layout_val = nullptr;
4273 layout_is_con = true;
4274 }
4275
4276 // Generate the initial go-slow test. Make sure we do not overflow
4277 // if length is huge (near 2Gig) or negative! We do not need
4278 // exact double-words here, just a close approximation of needed
4279 // double-words. We can't add any offset or rounding bits, lest we
4280 // take a size -1 of bytes and make it positive. Use an unsigned
4281 // compare, so negative sizes look hugely positive.
4282 int fast_size_limit = FastAllocateSizeLimit;
4283 if (layout_is_con) {
4284 assert(!StressReflectiveCode, "stress mode does not use these paths");
4285 // Increase the size limit if we have exact knowledge of array type.
4286 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4287 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4288 }
4289
4290 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4291 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4292
4293 // --- Size Computation ---
4294 // array_size = round_to_heap(array_header + (length << elem_shift));
4295 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4296 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4297 // The rounding mask is strength-reduced, if possible.
4298 int round_mask = MinObjAlignmentInBytes - 1;
4299 Node* header_size = nullptr;
4300 // (T_BYTE has the weakest alignment and size restrictions...)
4301 if (layout_is_con) {
4302 int hsize = Klass::layout_helper_header_size(layout_con);
4303 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4304 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4305 if ((round_mask & ~right_n_bits(eshift)) == 0)
4306 round_mask = 0; // strength-reduce it if it goes away completely
4307 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4308 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4309 assert(header_size_min <= hsize, "generic minimum is smallest");
4310 header_size = intcon(hsize);
4311 } else {
4312 Node* hss = intcon(Klass::_lh_header_size_shift);
4313 Node* hsm = intcon(Klass::_lh_header_size_mask);
4314 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4315 header_size = _gvn.transform(new AndINode(header_size, hsm));
4316 }
4317
4318 Node* elem_shift = nullptr;
4319 if (layout_is_con) {
4320 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4321 if (eshift != 0)
4322 elem_shift = intcon(eshift);
4323 } else {
4324 // There is no need to mask or shift this value.
4325 // The semantics of LShiftINode include an implicit mask to 0x1F.
4326 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4327 elem_shift = layout_val;
4328 }
4329
4330 // Transition to native address size for all offset calculations:
4331 Node* lengthx = ConvI2X(length);
4332 Node* headerx = ConvI2X(header_size);
4333 #ifdef _LP64
4334 { const TypeInt* tilen = _gvn.find_int_type(length);
4335 if (tilen != nullptr && tilen->_lo < 0) {
4336 // Add a manual constraint to a positive range. Cf. array_element_address.
4337 jint size_max = fast_size_limit;
4338 if (size_max > tilen->_hi) size_max = tilen->_hi;
4339 const TypeInt* tlcon = TypeInt::make(0, size_max, Type::WidenMin);
4340
4341 // Only do a narrow I2L conversion if the range check passed.
4342 IfNode* iff = new IfNode(control(), initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
4343 _gvn.transform(iff);
4344 RegionNode* region = new RegionNode(3);
4345 _gvn.set_type(region, Type::CONTROL);
4346 lengthx = new PhiNode(region, TypeLong::LONG);
4347 _gvn.set_type(lengthx, TypeLong::LONG);
4348
4349 // Range check passed. Use ConvI2L node with narrow type.
4350 Node* passed = IfFalse(iff);
4351 region->init_req(1, passed);
4352 // Make I2L conversion control dependent to prevent it from
4353 // floating above the range check during loop optimizations.
4354 lengthx->init_req(1, C->constrained_convI2L(&_gvn, length, tlcon, passed));
4355
4356 // Range check failed. Use ConvI2L with wide type because length may be invalid.
4357 region->init_req(2, IfTrue(iff));
4358 lengthx->init_req(2, ConvI2X(length));
4359
4360 set_control(region);
4361 record_for_igvn(region);
4362 record_for_igvn(lengthx);
4363 }
4364 }
4365 #endif
4366
4367 // Combine header size and body size for the array copy part, then align (if
4368 // necessary) for the allocation part. This computation cannot overflow,
4369 // because it is used only in two places, one where the length is sharply
4370 // limited, and the other after a successful allocation.
4371 Node* abody = lengthx;
4372 if (elem_shift != nullptr) {
4373 abody = _gvn.transform(new LShiftXNode(lengthx, elem_shift));
4374 }
4375 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4376
4377 if (return_size_val != nullptr) {
4378 // This is the size
4379 (*return_size_val) = non_rounded_size;
4380 }
4381
4382 Node* size = non_rounded_size;
4383 if (round_mask != 0) {
4384 Node* mask1 = MakeConX(round_mask);
4385 size = _gvn.transform(new AddXNode(size, mask1));
4386 Node* mask2 = MakeConX(~round_mask);
4387 size = _gvn.transform(new AndXNode(size, mask2));
4388 }
4389 // else if round_mask == 0, the size computation is self-rounding
4390
4391 // Now generate allocation code
4392
4393 // The entire memory state is needed for slow path of the allocation
4394 // since GC and deoptimization can happen.
4395 Node *mem = reset_memory();
4396 set_all_memory(mem); // Create new memory state
4397
4398 if (initial_slow_test->is_Bool()) {
4399 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4400 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4401 }
4402
4403 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4404 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4405
4406 Node* raw_init_value = nullptr;
4407 if (init_val != nullptr) {
4408 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4409 if (ary_type->is_flat()) {
4410 initial_slow_test = intcon(1);
4411 }
4412
4413 if (UseCompressedOops) {
4414 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4415 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4416 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4417 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4418 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4419 } else {
4420 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4421 }
4422 }
4423
4424 Node* valid_length_test = _gvn.intcon(1);
4425 if (ary_type->isa_aryptr()) {
4426 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4427 jint max = TypeAryPtr::max_array_length(bt);
4428 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4429 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4430 }
4431
4432 // Create the AllocateArrayNode and its result projections
4433 AllocateArrayNode* alloc
4434 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4435 control(), mem, i_o(),
4436 size, klass_node,
4437 initial_slow_test,
4438 length, valid_length_test,
4439 init_val, raw_init_value);
4440 // Cast to correct type. Note that the klass_node may be constant or not,
4441 // and in the latter case the actual array type will be inexact also.
4442 // (This happens via a non-constant argument to inline_native_newArray.)
4443 // In any case, the value of klass_node provides the desired array type.
4444 const TypeInt* length_type = _gvn.find_int_type(length);
4445 if (ary_type->isa_aryptr() && length_type != nullptr) {
4446 // Try to get a better type than POS for the size
4447 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4448 }
4449
4450 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4451
4452 array_ideal_length(alloc, ary_type, true);
4453 return javaoop;
4454 }
4455
4456 // The following "Ideal_foo" functions are placed here because they recognize
4457 // the graph shapes created by the functions immediately above.
4458
4459 //---------------------------Ideal_allocation----------------------------------
4460 // Given an oop pointer or raw pointer, see if it feeds from an AllocateNode.
4461 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr) {
4462 if (ptr == nullptr) { // reduce dumb test in callers
4463 return nullptr;
4464 }
4465
4466 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4467 ptr = bs->step_over_gc_barrier(ptr);
4468
4469 if (ptr->is_CheckCastPP()) { // strip only one raw-to-oop cast
4470 ptr = ptr->in(1);
4471 if (ptr == nullptr) return nullptr;
4472 }
4473 // Return null for allocations with several casts:
4474 // j.l.reflect.Array.newInstance(jobject, jint)
4475 // Object.clone()
4476 // to keep more precise type from last cast.
4477 if (ptr->is_Proj()) {
4478 Node* allo = ptr->in(0);
4479 if (allo != nullptr && allo->is_Allocate()) {
4480 return allo->as_Allocate();
4481 }
4482 }
4483 // Report failure to match.
4484 return nullptr;
4485 }
4486
4487 // Fancy version which also strips off an offset (and reports it to caller).
4488 AllocateNode* AllocateNode::Ideal_allocation(Node* ptr, PhaseValues* phase,
4489 intptr_t& offset) {
4490 Node* base = AddPNode::Ideal_base_and_offset(ptr, phase, offset);
4491 if (base == nullptr) return nullptr;
4492 return Ideal_allocation(base);
4493 }
4494
4495 // Trace Initialize <- Proj[Parm] <- Allocate
4496 AllocateNode* InitializeNode::allocation() {
4497 Node* rawoop = in(InitializeNode::RawAddress);
4498 if (rawoop->is_Proj()) {
4499 Node* alloc = rawoop->in(0);
4500 if (alloc->is_Allocate()) {
4501 return alloc->as_Allocate();
4502 }
4503 }
4504 return nullptr;
4505 }
4506
4507 // Trace Allocate -> Proj[Parm] -> Initialize
4508 InitializeNode* AllocateNode::initialization() {
4509 ProjNode* rawoop = proj_out_or_null(AllocateNode::RawAddress);
4510 if (rawoop == nullptr) return nullptr;
4511 for (DUIterator_Fast imax, i = rawoop->fast_outs(imax); i < imax; i++) {
4512 Node* init = rawoop->fast_out(i);
4513 if (init->is_Initialize()) {
4514 assert(init->as_Initialize()->allocation() == this, "2-way link");
4515 return init->as_Initialize();
4516 }
4517 }
4518 return nullptr;
4519 }
4520
4521 // Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
4522 void GraphKit::add_parse_predicate(Deoptimization::DeoptReason reason, const int nargs) {
4523 // Too many traps seen?
4524 if (too_many_traps(reason)) {
4525 #ifdef ASSERT
4526 if (TraceLoopPredicate) {
4527 int tc = C->trap_count(reason);
4528 tty->print("too many traps=%s tcount=%d in ",
4529 Deoptimization::trap_reason_name(reason), tc);
4530 method()->print(); // which method has too many predicate traps
4531 tty->cr();
4532 }
4533 #endif
4534 // We cannot afford to take more traps here,
4535 // do not generate Parse Predicate.
4536 return;
4537 }
4538
4539 ParsePredicateNode* parse_predicate = new ParsePredicateNode(control(), reason, &_gvn);
4540 _gvn.set_type(parse_predicate, parse_predicate->Value(&_gvn));
4541 Node* if_false = _gvn.transform(new IfFalseNode(parse_predicate));
4542 {
4543 PreserveJVMState pjvms(this);
4544 set_control(if_false);
4545 inc_sp(nargs);
4546 uncommon_trap(reason, Deoptimization::Action_maybe_recompile);
4547 }
4548 Node* if_true = _gvn.transform(new IfTrueNode(parse_predicate));
4549 set_control(if_true);
4550 }
4551
4552 // Add Parse Predicates which serve as placeholders to create new Runtime Predicates above them. All
4553 // Runtime Predicates inside a Runtime Predicate block share the same uncommon trap as the Parse Predicate.
4554 void GraphKit::add_parse_predicates(int nargs) {
4555 if (UseLoopPredicate) {
4556 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4557 if (UseProfiledLoopPredicate) {
4558 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4559 }
4560 }
4561 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4562 // Loop Limit Check Predicate should be near the loop.
4563 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4564 }
4565
4566 void GraphKit::sync_kit(IdealKit& ideal) {
4567 set_all_memory(ideal.merged_memory());
4568 set_i_o(ideal.i_o());
4569 set_control(ideal.ctrl());
4570 }
4571
4572 void GraphKit::final_sync(IdealKit& ideal) {
4573 // Final sync IdealKit and graphKit.
4574 sync_kit(ideal);
4575 }
4576
4577 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4578 Node* len = load_array_length(load_String_value(str, set_ctrl));
4579 Node* coder = load_String_coder(str, set_ctrl);
4580 // Divide length by 2 if coder is UTF16
4581 return _gvn.transform(new RShiftINode(len, coder));
4582 }
4583
4584 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4585 int value_offset = java_lang_String::value_offset();
4586 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4587 false, nullptr, Type::Offset(0));
4588 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4589 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4590 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true),
4591 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4592 Node* p = basic_plus_adr(str, str, value_offset);
4593 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4594 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4595 return load;
4596 }
4597
4598 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4599 if (!CompactStrings) {
4600 return intcon(java_lang_String::CODER_UTF16);
4601 }
4602 int coder_offset = java_lang_String::coder_offset();
4603 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4604 false, nullptr, Type::Offset(0));
4605 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4606
4607 Node* p = basic_plus_adr(str, str, coder_offset);
4608 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4609 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4610 return load;
4611 }
4612
4613 void GraphKit::store_String_value(Node* str, Node* value) {
4614 int value_offset = java_lang_String::value_offset();
4615 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4616 false, nullptr, Type::Offset(0));
4617 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4618
4619 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4620 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4621 }
4622
4623 void GraphKit::store_String_coder(Node* str, Node* value) {
4624 int coder_offset = java_lang_String::coder_offset();
4625 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4626 false, nullptr, Type::Offset(0));
4627 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4628
4629 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4630 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4631 }
4632
4633 // Capture src and dst memory state with a MergeMemNode
4634 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4635 if (src_type == dst_type) {
4636 // Types are equal, we don't need a MergeMemNode
4637 return memory(src_type);
4638 }
4639 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4640 record_for_igvn(merge); // fold it up later, if possible
4641 int src_idx = C->get_alias_index(src_type);
4642 int dst_idx = C->get_alias_index(dst_type);
4643 merge->set_memory_at(src_idx, memory(src_idx));
4644 merge->set_memory_at(dst_idx, memory(dst_idx));
4645 return merge;
4646 }
4647
4648 Node* GraphKit::compress_string(Node* src, const TypeAryPtr* src_type, Node* dst, Node* count) {
4649 assert(Matcher::match_rule_supported(Op_StrCompressedCopy), "Intrinsic not supported");
4650 assert(src_type == TypeAryPtr::BYTES || src_type == TypeAryPtr::CHARS, "invalid source type");
4651 // If input and output memory types differ, capture both states to preserve
4652 // the dependency between preceding and subsequent loads/stores.
4653 // For example, the following program:
4654 // StoreB
4655 // compress_string
4656 // LoadB
4657 // has this memory graph (use->def):
4658 // LoadB -> compress_string -> CharMem
4659 // ... -> StoreB -> ByteMem
4660 // The intrinsic hides the dependency between LoadB and StoreB, causing
4661 // the load to read from memory not containing the result of the StoreB.
4662 // The correct memory graph should look like this:
4663 // LoadB -> compress_string -> MergeMem(CharMem, StoreB(ByteMem))
4664 Node* mem = capture_memory(src_type, TypeAryPtr::BYTES);
4665 StrCompressedCopyNode* str = new StrCompressedCopyNode(control(), mem, src, dst, count);
4666 Node* res_mem = _gvn.transform(new SCMemProjNode(_gvn.transform(str)));
4667 set_memory(res_mem, TypeAryPtr::BYTES);
4668 return str;
4669 }
4670
4671 void GraphKit::inflate_string(Node* src, Node* dst, const TypeAryPtr* dst_type, Node* count) {
4672 assert(Matcher::match_rule_supported(Op_StrInflatedCopy), "Intrinsic not supported");
4673 assert(dst_type == TypeAryPtr::BYTES || dst_type == TypeAryPtr::CHARS, "invalid dest type");
4674 // Capture src and dst memory (see comment in 'compress_string').
4675 Node* mem = capture_memory(TypeAryPtr::BYTES, dst_type);
4676 StrInflatedCopyNode* str = new StrInflatedCopyNode(control(), mem, src, dst, count);
4677 set_memory(_gvn.transform(str), dst_type);
4678 }
4679
4680 void GraphKit::inflate_string_slow(Node* src, Node* dst, Node* start, Node* count) {
4681 /**
4682 * int i_char = start;
4683 * for (int i_byte = 0; i_byte < count; i_byte++) {
4684 * dst[i_char++] = (char)(src[i_byte] & 0xff);
4685 * }
4686 */
4687 add_parse_predicates();
4688 C->set_has_loops(true);
4689
4690 RegionNode* head = new RegionNode(3);
4691 head->init_req(1, control());
4692 gvn().set_type(head, Type::CONTROL);
4693 record_for_igvn(head);
4694
4695 Node* i_byte = new PhiNode(head, TypeInt::INT);
4696 i_byte->init_req(1, intcon(0));
4697 gvn().set_type(i_byte, TypeInt::INT);
4698 record_for_igvn(i_byte);
4699
4700 Node* i_char = new PhiNode(head, TypeInt::INT);
4701 i_char->init_req(1, start);
4702 gvn().set_type(i_char, TypeInt::INT);
4703 record_for_igvn(i_char);
4704
4705 Node* mem = PhiNode::make(head, memory(TypeAryPtr::BYTES), Type::MEMORY, TypeAryPtr::BYTES);
4706 gvn().set_type(mem, Type::MEMORY);
4707 record_for_igvn(mem);
4708 set_control(head);
4709 set_memory(mem, TypeAryPtr::BYTES);
4710 Node* ch = load_array_element(src, i_byte, TypeAryPtr::BYTES, /* set_ctrl */ true);
4711 Node* st = store_to_memory(control(), array_element_address(dst, i_char, T_BYTE),
4712 AndI(ch, intcon(0xff)), T_CHAR, MemNode::unordered, false,
4713 false, true /* mismatched */);
4714
4715 IfNode* iff = create_and_map_if(head, Bool(CmpI(i_byte, count), BoolTest::lt), PROB_FAIR, COUNT_UNKNOWN);
4716 head->init_req(2, IfTrue(iff));
4717 mem->init_req(2, st);
4718 i_byte->init_req(2, AddI(i_byte, intcon(1)));
4719 i_char->init_req(2, AddI(i_char, intcon(2)));
4720
4721 set_control(IfFalse(iff));
4722 set_memory(st, TypeAryPtr::BYTES);
4723 }
4724
4725 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4726 if (!field->is_constant()) {
4727 return nullptr; // Field not marked as constant.
4728 }
4729 ciInstance* holder = nullptr;
4730 if (!field->is_static()) {
4731 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4732 if (const_oop != nullptr && const_oop->is_instance()) {
4733 holder = const_oop->as_instance();
4734 }
4735 }
4736 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4737 /*is_unsigned_load=*/false);
4738 if (con_type != nullptr) {
4739 Node* con = makecon(con_type);
4740 if (field->type()->is_inlinetype()) {
4741 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4742 } else if (con_type->is_inlinetypeptr()) {
4743 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4744 }
4745 return con;
4746 }
4747 return nullptr;
4748 }
4749
4750 //---------------------------load_mirror_from_klass----------------------------
4751 // Given a klass oop, load its java mirror (a java.lang.Class oop).
4752 Node* GraphKit::load_mirror_from_klass(Node* klass) {
4753 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
4754 Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4755 // mirror = ((OopHandle)mirror)->resolve();
4756 return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
4757 }
4758
4759 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4760 const Type* obj_type = obj->bottom_type();
4761 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4762 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4763 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4764 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4765 obj = casted_obj;
4766 }
4767 if (sig_type->is_inlinetypeptr()) {
4768 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4769 }
4770 return obj;
4771 }